The 8th International Conference on Ambulatory Monitoring of Physical Activity and Movement

in Journal for the Measurement of Physical Behaviour

Click name to view affiliation

Open access

We are pleased to publish this special issue of the Journal for the Measurement of Physical Behaviour to feature abstracts that were presented at the 8th International Conference on Ambulatory Monitoring of Physical Activity and Movement (ICAMPAM).

The ICAMPAM conferences are hosted biannually by the International Society for the Measurement of Physical Behaviour (ISMPB). Our society is a non-profit scientific society that focuses on issues related to ambulatory monitoring, wearable monitors, movement sensors, physical activity, sedentary behaviour, movement behaviour, body postures, sleep, and constructs related to physical behaviours. Therefore, the published abstracts of this special issue focus on device-based measurement and quantification of physical behaviours such as:

  1. Physical behaviours (including sleep) such as volumes and patterns and how they associate with health outcomes
  2. Use of wearable devices in clinical settings and populations to assess gait, disease risk and as digital biomarkers
  3. Integration of signals including geospatial data, heart rate, and acceleration into sensors

The ICAMPAM conferences are designed to provide a forum for researchers and students to discuss the latest developments in physical behaviour monitoring using wearable devices. The 2022 conference contained both in-person posters and oral presentations and virtual poster presentations from an international group of scientists in the field of health sciences, engineering, medical sciences, physiology, psychology, sports sciences and more. The organizing committee paid special attention to create a conference program where many young scientists had the opportunity to present their work and awarded prizes to students who presented the best in-person poster, oral presentation and virtual posters. The conference featured seven keynote presentations, six workshops and six symposium sessions on a range of topics. The breadth and depth of the science that was presented at ICAMPAM are summarized in this special issue.

We hope that the contents of this special issue will be informative to our readers and to new readers interested in the objective measurement of physical behaviors.

On behalf of the organizing committee,

Jeff Hausdorff & Sarah Keadle, ICAMPAM 2022 Co-chairs

Bronwyn Clark, ISMPB President

ICAMPAM 2022 Conference Abstracts

Oral Presentations

Clinical Applications 1

Are Physical Behavior and Momentary Fatigue Bidirectionally Associated After Subarachnoid Hemorrhage?; Merging Accelerometry and Electronic Diary Data

Lianne de Vries1, Elisabeth de Vries1, Marco Giurgiu2, Fop van Kooten1, Gerard Ribbers1, Majanka Heijenbrok-Kal1, Rita van den Berg-Emons1, Hans Bussmann1

1Erasmus University, 2Karlsruhe Institute of Technology

Objective: To examine bidirectional associations between physical behavior and momentary fatigue after subarachnoid hemorrhage (SAH). Methods: Observational study with sensor-based measures of physical behavior using an accelerometer (Activ8) and repeated-measures of fatigue during 7 consecutive days according to Ecological Momentary Assessment (EMA) using an e-diary application. Patients with SAH who suffer from chronic fatigue were included. Physical behavior was expressed as total minutes of physical activity (PA) (walking, running, cycling) and sedentary behavior (SB) (sitting, lying) in a period of 45 minutes before and after a momentary fatigue assessment. Momentary fatigue was assessed 10-12 times per day on a 1-7 scale (no-extreme fatigue), for 7 consecutive days using EMA. Multilevel model analyses were conducted. Results: In total 38 patients participated, with a mean age of 53.2 years (SD=13.4), 58% female, and mean time post SAH onset of 9.5 months (SD=2.1). More PA significantly (B=0.01, p< 0.05) predicted higher subsequent fatigue, but SB did not predict subsequent fatigue. In addition, higher fatigue significantly predicted more subsequent SB (B=0.67, p< 0.05), but fatigue did not predict subsequent PA. The associations between PA and subsequent fatigue and between fatigue and subsequent SB significantly (p< 0.001) differed between participants. All models were adjusted for sex, age and day type (week vs weekend). Conclusions: By combining EMA measures of fatigue with accelerometry outcomes we found that longer engagement in physical activity was followed by higher fatigue and that higher fatigue was followed by longer engagement in sedentary behavior. Strikingly, these associations differed between participants. Therefore, these findings should be taken into account in developing personalized rehabilitation programs aimed at reducing fatigue and enhancing physical behavior after subarachnoid hemorrhage.

Frequency of Inpatient Out-of-Bed Activities by ActivPAL vs Johns Hopkins Highest Level of Mobility Scale After Major Abdominal Surgery

Mikita Fuchita1, Kyle Ridgeway1, Edward Melanson1, Ana Fernandez-Bustamante1

1University of Colorado

Objective: Postoperative mobilization may reduce complications following major abdominal surgery, but the optimal timing, frequency, duration, or intensity of mobilization remains unknown. Our institution uses the Johns Hopkins Highest Level of Mobility (JH-HLM) scale to document mobility of hospitalized patients. Although inpatient staff is instructed to record every mobility event using JH-HLM, it is unclear if JH-HLM documentation reflects the frequency of mobility events. We compared the number of out-of-bed activities reported by JH-HLM to those detected by a research-grade accelerometer, activPAL, from patients recovering after major abdominal surgery in the hospital. Methods: This is a single-center, prospective cohort study of ambulatory adult patients who underwent elective open abdominal surgery at a tertiary care academic hospital. Consented participants continuously wore an activPAL on the anterior thigh, starting immediately postoperatively until discharge or up to 7 days. We defined out-of-bed events as JH-HLM ≥ 4 and activPAL-recorded standing bouts. Results: Fifty-five patients wore an activPAL for a median [IQR] of 4 [3-6] days, cumulating to a total of 1,582 out-of-bed events recorded by activPAL. Daily numbers of out-of-bed mobility events as reported by JH-HLM vs ActivPAL are shown in Figure 1. Forty-five percent of out-of-bed events were bedside, non-ambulation events (standing or transferring to a chair) while the remaining 55% were ambulation events. Only 26% of all out-of-bed events were documented by JH-HLM, and documentation was less frequent for non-ambulation events compared to ambulation events (6% vs 43%, p< 0.001). Conclusions: Overall, only 26% of postoperative out-of-bed activities were documented via JH-HLM. Non-ambulation events were less likely than ambulation events to be captured by JH-HLM. The clinical significance of non-ambulatory, out-of-bed mobility vs ambulation is unknown and requires further investigation.

Gait During Daily Life in Men Treated With Androgen Deprivation Therapy for Prostate Cancer: Evidence for Accelerated Aging?

Deanne Tibbitts1, Martina Mancini1, Sydnee Stoyles1, Ramyar Eslami1, Christopher Palmer1, Mahmoud El-Gohary2, Fay Horak1, Kerri Winters-Stone1

1Oregon Health and Science University, 2APDM Wearable Technologies

Objective: Men with prostate cancer receiving androgen deprivation therapy (ADT) report declines in physical functioning that may be due to side effects from ADT. We had a unique opportunity to characterize gait during daily life in men treated with ADT and compare them to a sample of healthy older men. Our objective was to understand if men on ADT show evidence of accelerated aging using an ecologically valid measure of gait. Methods: Cross-sectional analysis of healthy older men and baseline data from participants in a fall and frailty prevention exercise trial for men treated with ADT. Men with prostate cancer self-reported current vs past use of ADT. Gait was collected using instrumented socks (APDM Wearable Technologies) that housed inertial sensors on the feet, and were worn at home during waking hours for up to 7 days. We derived measures of quantity and quality of gait by averaging across all gait bouts from at least 5 days of passive monitoring. Linear models were used to investigate differences among groups. Results: Our sample was 94 older men, including healthy older men (n=28; mean age, 70y) and men with prostate cancer (n=66; mean age, 72y). Of the men with prostate cancer, 36% were current and 64% were past users of ADT. Gait was similar among current and past users of ADT, so these groups were combined. Gait analysis showed that mean cadence (steps/min; beta=-3.3, p< 0.01) and foot angle at toe-off (beta=-2.5, p< 0.01) were significantly lower, while stride duration (beta=0.1, p< 0.01) and double support time (beta=2.4, p< 0.01) were significantly higher in men treated with ADT compared to healthy older men. Conclusions: Men treated with ADT show specific gait impairments during daily life that reflect weakness, fatigue, and altered dynamic balance. Compared to gait patterns in healthy men of similar age, men treated with ADT have a worse gait pattern, suggesting that ADT treatment may be associated with an accelerated aging phenotype.

Validation and Ranking of Algorithms for Gait Sequence Detection in Healthy Controls and People With Parkinson’s Disease

María Encarnación Micó Amigo1, Martin Ulrich2, Anisoara Paraschiv-Ionescu3, Eran Gazit4, Tecla Bonci5, Francesca Salis6, Kristy Scott5, Stefano Bertuletti6, Andrea Cereatti7, Lynn Rochester1, Claudia Mazzà5, Silvia Del Din1

1Newcastle University, 2Friedrich-Alexander-Universität Erlangen-Nürnberg, 3École Polytechnique Fédérale de Lausanne, 4Tel Aviv Sourasky Medical Center 5The University of Sheffield, 6University of Sassari, 7Politecnico di Torino

Objective: Free-living monitoring of gait enables the quantification of a large spectrum of ecologically valid digital mobility outcomes in diverse clinical cohorts. However, this is contingent on identification of valid gait sequences (walking activities). Here, we validated and ranked nine algorithms for gait sequence (GS) detection from a wearable device (WD) worn on the lower-back. Methods: As part of the Mobilise-D1 consortium, 19 people with Parkinson’s disease (pwPD, 65.8± 12.3 y) and 20 healthy controls (HC, 65.9± 13.1 y) were assessed in the laboratory (lab) and during 2.5 hours of free-living conditions with a single WD worn on the lower-back. A stereophotogrammetric system was the reference system in the lab, and a combination of inertial module with distance sensors and pressure insoles (INDIP) was the reference system in free-living. Nine algorithms were implemented to detect GS from inertial data. Validity measures included: accuracy, sensitivity, specificity and positive predictive value for GS detection. Mean absolute errors and interclass-correlation-coefficients (ICC) of GS duration were also assessed. Based on this analysis we developed and implemented a methodology to rank all algorithms, and the top performer was independently selected for each cohort. Results: Combining lab and free-living results, the top algorithm’s performances for the HC were: accuracy 0.84±0.12; sensitivity 0.94±0.15; specificity 0.72±0.22 and positive predictive value 0.71±0.19. The mean absolute error in gait sequence duration was 2.62±1.8 s and the ICC 0.95. For the pwPD group, a different algorithm gave the best performance with accuracy 0.86±0.09; sensitivity 0.94±0.14; specificity 0.76±0.19 and positive predictive value 0.77±0.16. Mean absolute error was 3.93±3.42 s and the ICC 0.87. Discussion: The results highlight good performance of the top algorithms for GS detection. The algorithms were tailored for each cohort indicating that one size does not fit all.

Validation of the Apple Watch and Fitbit for Assessing Heart Rate During Rest and Wheelchair Propulsion in Able-Bodied Participants and Wheelchair Users

Julia Baumgart1, Melanie Vergeer2, Guy Plasqui2, Marius Lyng Danielsson1

1Norwegian University of Science and Technology, 2Maastricht University

Objective: To investigate the criterion validity of 2 commercial wrist-worn devices (Apple Watch W4 and Fitbit Versa) for assessing heart rate (HR) during rest and wheelchair propulsion in able-bodied participants (AB) and manual wheelchair users (MWU). Methods: Currently included are 20 healthy AB (age:33±11years, height:176±10cm, body mass:75±11kg). Data collection is ongoing for inclusion of 20 MWU. Average HR was collected during two 10-min resting periods (one sitting, one lying), and three 4-min stages of treadmill wheelchair propulsion at 2.5% incline (women:2,3,4, men:3,4,5km/h). Participants simultaneously wore the Apple Watch and Fitbit on the non-dominant arm, in addition to the Polar H10 HR belt (criterion measure). Placement of the devices (i.e., closeness to the wrist) was counterbalanced. Bland-Altman plots with mean differences (MD) and lower and upper limits of agreement (LoA) were used to establish absolute agreement, and intraclass correlation coefficients (ICCA,1) for relative agreement. Results: During sitting and lying rest, MDs were close to zero with accompanying acceptable lower and upper LoAs [in square brackets] both for the Apple Watch (-0.4[-5,4] and -0.7[6,7]) and Fitbit (-2[-5,2] and -3[-7,1]), in addition to high ICCs for both the Apple Watch (0.95 and 0.93) and Fitbit (0.89 and 0.82). During the 3 stages, MDs were more variable and LoAs high for the Apple Watch (Stage1: 0[-20,21], Stage2: -3[-21,14], Stage3: -4[-31,24]) and Fitbit (Stage1: 6[-26,36], Stage2: 0[-31,31] and Stage3: -12[-46,22]). Furthermore, ICCs were low to moderate across all stages for the Apple Watch (0.57-0.72) and Fitbit (0.02-0.26). Conclusion: The Apple Watch and Fitbit display acceptable agreement for tracking HR at rest. Preliminary findings based on AB for now indicate that although agreement is slightly better for the Apple Watch than Fitbit during wheelchair propulsion, it is too low for both devices to recommend their use during wheelchair exercise.

Clinical Applications 2

Activity and Rest Fragmentation Analysis of Daily-Living Physical Activity Fluctuations Among People With MS

Amit Salomon1, David Buzaglo1, Irina Galperin1, Anat Mirelman1, Keren Regev1, Arnon Karni2, Tanja Schmitz-Hübsch3, Friedemann Paul3, Hannes Devos4, Jacob Sosnoff4, Raz Tamir5, Nathaniel Shimoni5, Yarden Rotem5, Yehudit Michaelis5, Jeffrey Hausdorff1

1Tel Aviv Sourasky Medical Center, 2Tel Aviv University, 3Universitaetsmedizin Berlin, 4University of Kansas, 5Owlytics Healthcare Ltd.

Objective: Physical activity declines in people with multiple sclerosis (pwMS). Previous work focused on studying activity levels, but the work exploring daily-living activity/rest fragmentation (FRG) patterns in pwMS is limited. Here, we aimed to evaluate FRG metrics among pwMS. Our questions were: (1) Can we identify FRG metrics that differ between pwMS and healthy controls (HCs)? (2) Are FRG metrics correlated with disease severity? (3) Are they correlated with more conventional physical activity metrics, e.g., step count and mean signal vector magnitude (mSVM)? Methods: 146 people with relapsing-remitting MS (47±11y, 69%f, EDSS: 3.2±1.6 IQR=2-4) and 93 HCs (46±11y, 49%f) were asked to wear a 3D accelerometer on their lower back for 7 days. Gait-based FRG metrics included: binary index (BI, a measure of activity FRG), average walking bout duration (AWBD), average rest bout duration (ARBD), gait and rest gini index and average hazard (AH), K_AR and K_RA (gait to rest and rest to gait transition frequency, respectively). Mann-Whitney tests compared the pwMS and HCs. Nonparametric correlations evaluated the associations between FRG metrics and other measures. Results: BI (pwMS: 0.052±0.0067, HCs: 0.050±0.0072, p=0.02), gait AH (pwMS: 0.031±0.012, HCs: 0.026±0.0067; p=0.001), rest AH (pwMS: 0.028±0.0086, HCs: 0.024±0.0047; p=0.0001), and K_AR (pwMS: 0.17±0.043, HCs: 0.15±0.038, p=0.02) were higher in pwMS compared to HCs. K_RA was lower (pwMS: 0.046±0.015, HCs: 0.052±0.013, p=0.0009). FRG metrics were mild-to-moderately associated with disease severity and step count (Figure 2). Conclusions: These findings suggest that selected free-living activity-rest FRG patterns differ between pwMS and HCs. A subset of FRG metrics are also mildly related to disease severity and to conventional activity metrics. Among those, two of the metrics with the strongest association to disability levels were related to rest patterns, perhaps indirectly capturing an aspect of fatigue.

Impact of Frailty on Free-Living Walking Performance in People Living With MS

Tobia Zanotto1, Irina Galperin2, Anat Mirelman2, Lingjun Chen1, Keren Regev2, Arnon Karni2, Tanja Schmitz-Hubsch3, Friedemann Paul3, Sharon Lynch1, Abiodun Akinwuntan1, Hannes Devos1, Jeffrey Hausdorff2, Jacob Sosnoff1

1University of Kansas, 2Tel Aviv Sourasky Medical Center, 3Universitaetsmedizin Berlin

Objectives: To explore: 1) the impact of frailty on free-living walking performance (FLWP) in people with multiple sclerosis (pwMS) and 2) the mediating effect of frailty on the relationship between disability and FLWP. Methods: Ninety-nine people with relapsing-remitting MS [age=49.2 years (SD=9.9); 73.5% female; expanded disability status scale (EDSS) range=2.0-6.0] were studied in this cross-sectional analysis. A frailty index was calculated by following standard procedures. Participants wore a tri-axial accelerometer for 7 days. Derived measures reflected the quantity (daily steps, number of 30-seconds walking bouts, signal vector magnitude) and the quality (gait speed, step cadence, step and stride regularity, and sample entropy) of FLWP. For each walking quality measure, the typical (median), best (90%), and worst (10%) values were extracted. Results: Participants were classified as non-frail (n=37), moderately (n=28), and severely frail (n=34). Severely frail participants had worse performance in all walking quantity measures compared to non-frail individuals. Mixed-ANOVA simple main effects revealed both moderately (p=0.026) and severely (p=0.005) frail participants had lower best values of gait speed compared to the non-frail. Severely frail participants also had lower best values of step cadence compared to the non-frail (p=0.005). Frailty did not mediate the relationship between EDSS and gait quality (e.g., gait speed, step cadence, step and stride regularity). Conversely, frailty partially mediated the relationship between EDSS and daily steps and number of 30-seconds walking bouts, and fully mediated the relationship between EDSS and signal vector magnitude. Conclusions: The current study indicates that higher frailty negatively affects both the quantity and quality of walking in pwMS. While disability exhibited a greater impact on pace, rhythm, variability, and symmetry of walking, frailty may have a larger impact on activity curtailment in pwMS.

Objective Estimation of Disability Levels and Physical Fatigue Among People With Multiple Sclerosis Using a Single Sensor Worn During Daily-Living

Amit Salomon1, Irina Galperin1, David Buzaglo1, Anat Mirelman1, Keren Regev1, Arnon Karni2, Tanja Schmitz-Hübsch3, Friedemann Paul3, Hannes Devos4, Jacob Sosnoff4, Raz Tamir5, Nathaniel Shimoni5, Yarden Rotem5, Yehudit Michaelis5, Jeffrey Hausdorff1

1Tel Aviv Sourasky Medical Center, 2Tel Aviv University, 3Universitaetsmedizin Berlin, 4University of Kansas, 5Owlytics Healthcare Ltd.

Objective: The most widely used measure of disease severity in people with multiple sclerosis (pwMS) is the Expanded Disability Status Scale (EDSS). This requires an expert clinician to examine the patient. Such assessment is typically done about once every 6 months. Physical fatigue (PF) is a common problem that affects many pwMS. Typically, self-report is used to quantify PF. Here we examine if daily-living gait measures obtained via a single wearable sensor can estimate the EDSS and PF. Methods: 160 pwMS (age: 47.5±11.2; 69.4% women; EDSS: 3.2±1.6) were asked to wear a 3D accelerometer on their lower back for 7 days. EDSS was evaluated along with PF, measured via the Modified Fatigue Impact Scale. Daily-living domains included: gait quality (e.g., gait speed, variability), across bout gait variability (e.g., SD of gait quality bouts), physical activity (e.g., sedentary time, step count), movement at night (e.g., turns in bed), activity fragmentation (e.g., gini index) and activity per time of day. A series of stepwise multiple regressions examined the % of variance explained by each domain, with the EDSS and PF as the dependent variables. To minimize overfitting, the model was developed on 70% of the data and re-tested on a random 30%. Results: Gait quality and across bout gait quality were most strongly associated with EDSS scores, followed by other domains. A combination of daily-living measures accounted for 62% (58% when retested) of variance of the EDSS (Figure 3-top). In contrast, measures of activity and gait quality explained 25% of the PF variance (Figure 3). Interestingly, the final model predictors were unique for EDSS and PF. Conclusions: A linear combination of multiple domains of daily-living gait and physical activity capture much of the variance in EDSS scores. This approach could be used to aid in the monitoring of disease severity, remotely, in-between routine clinical visits. A different modeling approach is needed to better track PF.

Setting the Building Blocks for Long Term Remote and Continuous Real-Time Monitoring of MS Patients in Their Daily Living Environment Using a Wrist-Worn Smart Watch

Nathaniel Shimoni1, Raz Tamir1, Yarden Rotem1, Efrat Yatziv1, Yehudit Michaelis1, Eran Gazit2, David Buzaglo2, Irina Galperin2, Jeffrey Hausdorff2, Keren Regev2, Arnon Karni2

1Owlytics Healthcare Ltd., 2Tel Aviv Sourasky Medical Center

Background: Severity and progression assessment of Multiple Sclerosis (MS) is done by a periodic visit at a specialized clinic. Periodic In-lab protocols for quantifying disease progression (EDSS) may be biased due to patients’ state at time of visit. Passive monitoring in home environments, while performing activities of daily living (ADL), reduces the recency and point-in-time effects of in-clinic assessment. We present validated ADL metrics that show differences among groups of EDSS, enabling real-time and continuous collection, analysis and monitoring of MS patients. Methods: 60 MS patients (age(42.7±11.36), EDSS(2.29±1.58)) split into Mild (N=41, age(40.1±10.86), EDSS(1.41±0.99)) and Severe (N=19, age(48.32±10.59), EDSS(4.18±0.69)) and 41 healthy controls (HC, age(38.02±10.29)) participated in a protocol consisting of in-lab evaluation and daily living environment monitoring wearing a wristwatch. Subjects were remotely and continuously monitored between 7 days(HC) and up to 344 days(MS). Deep neural networks were trained to estimate gait speed and duration, validated using insoles, back worn and mat sensors, used to identify group differences by gait speed, activity intensity and walking bouts duration in home environments. Results: Gait-speed (calibrated RMSE(8.27±2.19cm/s), RMSE(12.68±7.56cm/s)), Activity Intensity (HC:91.75±39.08, MS:71.12±32.49, p=.005; Mild:83.23±30.74, Severe:45.0±17.51, p<.001) Gait-Speed (HC:115.02±9.78, MS:106.78±11.01, p< .001;Mild: 109.94±8.58, Severe:99.97±12.74, p<.001) Long-Bouts Ratio (HC:0.025±0.017, MS:0.009±0.009, p<.001; Mild:0.011 0.01, Severe:0.004±0.003, p<.001). Conclusions: We show that EDSS can be estimated by measuring daily activities of MS patients at their homes. These results lay the foundations to an ongoing assessment of disease progression allowing a comprehensive data-driven approach during clinic visits, by providing patients and physicians tools for better managing and evaluating treatment effectiveness. (Figure 4)

Using a Wrist-Worn Sensor to Objectively Monitor Gait Quality in People With Multiple Sclerosis: Initial Findings

Eran Gazit1, Arnon Karni1, Keren Regev1, Irina Galperin1, David Buzaglo1, Nathaniel Shimoni2, Yarden Rotem2, Yehudit Michaelis2, Raz Tamir2, Jeffrey Hausdorff1

1Tel Aviv Sourasky Medical Center, 2Owlytics Healthcare Ltd.

Objectives: To show that measures derived from a wrist-worn IMU differ in people with multiple relapsing-remitting sclerosis (pwMS) and health controls (HC) and that these measures are related to disease severity. Method: 56 pwMS (age 42.7.0±11.4, 66.6% women, Expanded Disability Scale Score(EDSS) 2.4±1.6) and 34 HC(age 37.3±10.4, 50% women) performed a 6-minute walk test with an “Opal” sensor on the right wrist (sample rate 128Hz). The subjects walked back and forth along a 25 meters corridor for 6 minutes with the instruction to achieve the longest distance they could. PCA analysis was used to decompose the 3D acceleration signal and the most informative axis was used for the gait quality assessment. Derived measures included stride regularity, RMS, jerk, sample entropy, and in the frequency domain, the frequency (reflection of cadence) and amplitude of the strongest peak. T-test was used to test for differences between the groups, linear regression model was used for predicting EDSS scores and binary logistic to differentiate between severity of the pwMS disease. Results: RMS, jerk, stride regularity, and sample entropy differed significantly between the pwMS and the HC. A forward linear regression model using RMS and sample entropy measures predicted EDSS scores (R=0.612, R^2=0.375, p< 0.01). A Binary logic regression model, with the same measures, separated the pwMS into those with relatively high and low levels of disease severity (Accuracy 81%, EDSS above or below 3). Conclusions: The results demonstrate that during the 6 min test, arm movement of pwMS is less consistent, smaller and with lower entropy (suggesting more impaired system) compared to healthy controls. More generally, the results suggest that it is possible to extract meaningful measures of gait quality from a wrist-worn sensor in patients with MS. The ability to differentiate between pwMS and HC suggests that it could be applied for long-term monitoring of MS patients using a wrist-worn sensor. (Figure 5)

Clinical Applications: Knee and Back Pain and Fatigue

Applying the Pittsburgh Performance Fatigability Index to a 6-Minute Walk in Older Adults

Yujia (Susanna) Qiao1, Jennifer Brach1, Andrea Rosso1, Kyle Moored1, Robert Boudreau1, Jaroslaw Harezlak2, Nancy Glynn1

1University of Pittsburgh, 2Indiana University

Objective: The Pittsburgh Performance Fatigability Index (PPFI) is a novel, sensitive and objective performance fatigability measure for older adults. Given the importance and common use of 6-minute walk (6MWT) in clinical and research settings, we adapted our original PPFI equation developed for 400m walks for use with the 6MWT. Method: Participants, from the randomized exercise trial “Program to Improve Mobility in Aging” (N=227, mean age=79 years, 67% women), wore ActiGraph GT3X+ (sampling=100 or 30Hz) during a 6MWT on non-dominant wrist at baseline. Raw triaxial accelerations were processed via Adaptive Empirical Pattern Transformation R package to estimate stride-to-stride cadence (steps/sec). Then, individual-level smoothed cadence-versus-time trajectories were estimated with penalized regression splines via “mgcv” R package. PPFI quantifies percent of performance decrement during 6MWT by comparing area under the observed cadence trajectories to a hypothetical area observed in the absence of fatigue (i.e., sustained maximal cadence) (Figure 6). To maintain comparability with our 400m equation, those covering >480m during 6MWT were assigned PPFI=0 as they walked very fast with no performance deterioration. To minimize the influence of increased cadence at very end of task, individualized weights, proportional to the time used to cover the distance, were applied to two parts (1st: 370m; 2nd: remaining distance) of each cadence trajectory. Results: PPFI scores from 6MWT ranged from 0-13.1% (mean=2.8%±2.6%). Higher PPFI scores were correlated with worse physical function and leg peak power, less time spent in moderate-to-vigorous physical activity, and fewer daily step counts (corr=0.19-0.37, P<.05), similar as correlations seen in previous study of PPFI from 400m walk. Conclusion: Our work advanced comparability of fatigability measure across walking tasks. PPFI is a promising objective risk assessment tool for measuring performance fatigability in older adults.

Associations of Digital Measures of Gait With Sleep and Fatigue: A Real World Feasibility Study

Rana Zia Ur Rehman1, Diogo Branco2, Dan Jackson1, Meenakshi Chatterjee3, Ahmaniemi Teemu4, Tiago Guerreiro2, Yannis Pandis3, Kristen Davies1, Victoria Macrae5, Svenja Aufenberg6, Emma Paulides7, Hanna Hildesheim8, Jennifer Kudelka8, Kirsten Emmert8, Lynn Rochester1, C. Janneke van der Woude7, Ralf Reilmann6, Walter Maetzler8, Wan-Fai Ng1, Silvia Del Din1

1Newcastle University, 2University of Lisbon, 3Janssen Research & Development, 4VTT, 5NIHR Newcastle, 6University of Muenster, 7Erasmus University, 8University Medical Center Schleswig- Holstein

Background and Aims: Sleep disturbances and fatigue are commonly reported symptoms in individuals with neurological and immune disorders. Current assessment of sleep and fatigue rely on patient reported outcomes (PROs), which are subjective, prone to recall biases and do not capture variability over time. Wearable technologies provide objective and reliable estimates of gait that are sensitive to change. Here, we evaluate the performance of a McRoberts device worn at the lower back to assess gait and its association with fatigue and sleep in IDEA-FAST feasibility study. Methods: 159 participants from six different disease groups (Parkinson’s disease (PD=25), Huntington’s disease (HD=14), Rheumatoid arthritis (RA=24), Systemic Lupus Erythematosus (SLE=18), Primary Sjogren’s syndrome (PSS=18), Inflammatory Bowel Disease (IBD=18)) and healthy controls (HC=42) were enrolled for 60 days. Micro gait characteristics (pace, rhythm, variability, asymmetry and postural control) were estimated from the device worn at lower back. Participants completed sleep, and fatigue related PROs up to 4 times a day, using a mobile phone application. The performance of device was assessed by evaluating coverage, data quality, and outcomes association with PROs. Results: Participants wore the device for over 10 days, resulting in a coverage >80%. Overall, step velocity and length variability were significantly associated with physical fatigue -0.4≤r ≤-0.3 (p<0.05). Step length and velocity had significant positive relationship with mental fatigue PROs 0.3≤r≤0.38 (p<0.05). Statistics of swing time asymmetry had highest association with the sleepiness index (r=-0.452, p<0.05). Conclusion: Selective gait characteristics such as step length and step velocity representing gait pace had moderate correlation with physical and mental fatigue PROs. In future studies, other aspects of mobility (e.g. turning) and data aggregation could be explored to assess their association with sleep and fatigue.

Continuous Longitudinal Monitoring of Early Physical Activity Recovery Following Knee Arthroplasty

Scott Small1, Aiden Doherty1, Sara Khalid1, Andrew Price1

1University of Oxford

Physical activity monitoring in clinical orthopaedics is frequently limited to comparisons of preoperative and 6-12-month postoperative activity. The current study was designed to assess the trajectory of early recovery in knee arthroplasty patients through continuous wrist-based activity monitoring during the first six postoperative weeks. Patients with end-stage knee osteoarthritis, listed for primary total (TKA) or unicompartmental (UKA) knee arthroplasty, were recruited at a single site. Participants wore an Axivity AX3 accelerometer on the dominant wrist in a 24-hour protocol for 7 days, prior to surgery, as a baseline measurement. Following surgery, participants wore the same accelerometer continuously for 42 days. A total of 141 participants, median [IQR] 70.4 [61.6-76.4] years, completed baseline monitoring, with 105 participants proceeding to surgery and contributing postoperative data. 17 participants elected not to continue following baseline collection and 19 participants did not proceed to surgery or were lost to follow-up. Wear compliance was high, with participants contributing a median 37.3 days of postoperative wear. Preoperative activity of 6,855 [3,201-9,003] daily steps and overall acceleration of 21.5 [17.4-25.9] mg was observed. Postoperatively, activity increased from 653 [258-1,141] daily steps and 14.6 [12.3-18.2] mg overall acceleration during the first postoperative week to 3,983 [2,026-6,390] steps and 19.5 [16.2-24.0] mg of acceleration in Week 6. Compared to TKA, less invasive UKA was associated with higher postoperative activity beginning at Week 2 (p=0.015). In this study, continuous monitoring over 6 weeks was well tolerated by an orthopaedic patient population. Perioperative physical activity monitoring provides the opportunity for objective assessment of individual-level patient rehabilitation in the at-home, free-living setting, and can inform comparisons of early recovery between surgical techniques or clinical pathways. (Figure 7)

Patterns of Physical Activity Accumulation as a Potential Biomarker for Low Back Pain Phenotyping

Ruopeng Sun1, Dokyoung You1, Anuradha Roy1, Beth Darnall1, Sean Mackey1, Matthew Smuck1

1Stanford University

Introduction: Physical inactivity is a known risk factor for low back pain (LBP) disability, yet objective quantification of physical activity (PA) and its relationship with pain-related limitations are still sparse. In this study, we aim to investigate the connection between accelerometer-based PA patterns and pain-related limitations. Methods: As part of a clinical trial, 82 LBP participants (51% female, mean 51.8 yrs old) wore an Actigraph sensor (GT3X+, non-dominant wrist) for 7+ days before the intervention. Waking hours (5am-0am) signals were extracted to derive the following PA data: step count, sedentary time (SED), moderate-vigorous physical activity (MVPA) time, and hourly steps accumulation. Associations between PA and patient-reported outcomes (PROMIS-29) were examined. PA data were further divided into three terciles based on the pain interference (PI) index to identify potentially sensitive activity markers for stratifying LBP cohorts. Results: Daily PA measures (steps, SED, and MVPA time) were not associated with any patient-reported pain outcome, and were not able to stratify the LBP cohort based on the PI index. However, the hourly steps accumulation pattern identified a distinguishable morning/evening pattern (Figure 8) that can stratify the LBP cohort. More specifically, individuals in the high PI tercile accumulated fewer steps in the morning (9-11am) than those in the low PI tercile (600 vs 800 steps/hr, p<.05), and more steps (350 vs 200 steps/ hr, p<.05) during late-night(10pm-0am). Discussions: Even though the total amount of daily PA was not associated with any of the pain-related measures, we observed a unique morning/evening activity accumulation pattern that was associated with PI. Future research is needed to investigate the underlying mechanism for the differential activity patterns between individuals with low/high PI, and to investigate whether such activity patterns can be used to optimize individualized LBP rehabilitation programs.

Epidemiologic Studies With Health Outcomes

Association of Profiles of Objectively-Measured Physical Activity and Sedentary Behavior With All-Cause Mortality Risk in Older Adults

Manasa Shanta Yerramalla1, Mathilde Chen1, Vincent van Hees2, Quentin Le Cornu1, Aline Dugravot1, Séverine Sabia1

1Université de Paris, 2Accelting

Objective: To identify profiles of physical activity and sedentary behavior based on accelerometer-derived characteristics, and examine the association of the derived profiles with all-cause mortality among older adults. Method: Whitehall II accelerometer substudy participants (N=3991) aged 60-83 years in 2012-2013 wore a wrist-accelerometer over 7 days. K-means cluster analysis using 13 variables characterizing the total duration, frequency, bout duration, timing and activity intensity distribution of daily movement behaviors was performed to identify profiles. Cox regression models were used to assess the association between derived profiles and mortality risk. Results: A total of 410 deaths were recorded over a mean follow-up of 8.1 (standard deviation=1.3) years. Five distinct profiles were identified and labelled as: “Active breakers”, most active and least sedentary; “Active sitters”, high intensity activity in prolonged bout and less interrupted sedentary behavior; “Light movers”, less high intensity and more light activity; “Prolonged sitters”, highly sedentary and not active; “Couch potatoes”, highest level of sedentariness and least interruptions in sedentary behavior. In model adjusted for sociodemographic, lifestyle and health-related risk factors, as compared to active breakers, all profiles were associated with higher risk of mortality. The hazard ratios were: 1.57 (95% confidence interval (CI): 1.01-2.44) for active sitters, 1.75 (1.17-2.63) for light movers, 1.67 (1.11-2.51) for prolonged sitters, and 3.25 (2.10-5.02) for couch potatoes. Conclusions: Older adults may be encouraged to model their behavior as “active breakers”. Given the 3-fold increase in risk of death among those with a “couch potato” profile, public health interventions shall specifically target this group where any improvement in physical activity and sedentary behavior might be useful.

Impact of Patterns of Physical Activity at Pre- and Post-Diagnosis With Mortality of Asian Cancer Patients: Results From Health Examinees-G Study in Korea

Jaesung Choi1, Joo-Yong Park1, Ji-Eun Kim1, Miyoung Lee1, Kyuwan Lee1, Daehee Kang1, Aesun Shin1, Ji-Yeob Choi1

1Seoul National University

Background: Physical activity is recommended to improve the survival of cancer patients. However, the effect of a specific pattern of physical activity remains poorly understood. We investigated the associations of the duration, type, intensity, diversity of physical activity at pre and post-diagnosis with mortality in Korean cancer patients utilizing data from the Health Examinees-G (HEXA-G) study. Method: Among the 112,108 participants aged 40-69 recruited from the HEXA-G study, patients diagnosed with cancer after (n=4,207) and before (n=2,047) baseline were included in the analyses for leisure-time physical activity (LTPA). Cancer diagnosis and mortality were ascertained by linking with the Korea Central Cancer Registry and Statistics Korea. Patterns of LTPA including self-reported duration per week, intensity, types, and diversity were assessed. Cox-proportional hazard models adjusting for demographic, behaviors, comorbidity, and SEER stage were used to identify the association between patterns of LTPA and mortality. Results: At pre-diagnosis, patients participating in LTPA more than 150 minutes/week, vigorous-intensity, and more than 2 activities had significantly lower all-cause and cancer mortality. Importantly, reduced risk of all-cause mortality in those participating in more than 2 different LTPA activities was strengthened when they participated in the vigorous-intensity activity (HR: 0.54, 95% CI: 0.41-0.70) or mountain climbing (HR: 0.50, 95% CI: 0.37-0.68). At post-diagnosis, cancer mortality was lower (HR: 0.68, 95% CI: 0.46-0.99) in those participating in LTPA more than 150 min/week when compared to the inactive group. All-cause mortality was lower in men participating in mountain climbing (HR: 0.53, 95% CI: 0.29-0.99). Conclusion: Specific patterns of physical activity may influence the survival of cancer patients. Future randomized controlled trials are warranted to demonstrate the effects of physical activity in cancer patients.

Implementation of Wrist Accelerometry Into the National Health and Aging Trends Study (NHATS) to Expand Physical Activity Assessment in Older Adults

Jennifer Schrack1, Vadim Zipunnikov1, Vicki Freedman2

1Johns Hopkins University, 2University of Michigan

Background: Wearable accelerometers to monitor daily physical activity have been implemented in multiple large-scale studies, yet representative longitudinal data in older adults is lacking. To address this gap, we implemented wrist worn accelerometry into round 11 (2021) of NHATS, an ongoing panel study of US Medicare beneficiaries aged ≥ 65 years. Longitudinal data will be collected in rounds 12 & 13. Methods: Data collection in Round 11 was completed in-person and via phone by trained interviewers in 3184 participants, 1000 of whom were selected to wear an Actigraph Centrepoint Insight Watch on their non-dominant wrist for 7-days. Watches were placed on participants during their in-person visit and mailed to the study center upon completion. Data were collected at 64 Hz and uploaded to the Centrepoint system. Data were processed into 1-minute intervals of activity counts. Raw data will be available upon request. Results: 872 (87.2%) of the 1000 participants were eligible to participate. Of the 128 not eligible, 56 (5.6%) chose not to complete the NHATS interview and 72 (7.2%) were either deceased or unable to complete the participant interview. Of the 872 eligible participants, 747 (85.7%) completed at least one day of data collection, 58 (6.7%) refused, 25 (2.9%) returned the watch with no data, 36 (4.1%) had no in-person interview, and 6 (.7%) did not return the watch. Participants wore the watch a mean 7.0(1.5) days, 1431(17.3) minutes/day. Participant age averaged 78.0(5.9) years (range 70-99) and 54.4% were women. Race and ethnicity included: 80.5% White, 7.9% Black non-Hispanic, 5.9% Hispanic, and 5.7% other. Total daily activity averaged 1,665,045(639,774) counts/day and 5.6(2.1) hours/day spent active. Conclusion: Implementation of accelerometry into NHATS was largely successful with strong wear-time compliance. Weighting and longitudinal follow-up will increase generalizability to the larger US Medicare beneficiary population.

Multidimensional Movement Behavior and Mortality in Older Adults From the Whitehall II Accelerometer Sub-Study: A Machine Learning Approach

Mathilde Chen1, Vincent van Hees2, Manasa Shanta Yerramalla1, Mohamed Amine Benadjaoud3, Séverine Sabia1

1Université de Paris, 2Accelting, 3IRSN

Objective: Current guidelines mainly rely on total physical activity (PA) or sedentary behavior (SB) duration, ignoring other characteristics related to frequency, intensity distribution, timing, or typical duration of movement behavior. We aim to identify key movement behavior dimensions predicting mortality in old age and compare their predictive ability with other established predictors. Methods: Twenty-one movement behavior variables were extracted from accelerometer data of 3991 Whitehall II study participants aged 60-83 years in 2012-2013. Relevant accelerometer variables for 8-year mortality prediction were identified using a sparse Partial Least Square Cox procedure to jointly perform variable selection and dimension reduction. Selected variables’ predictive ability was assessed and compared to sociodemographic, lifestyle, and health-related predictors using Akaike’s information criterion (AIC) and C-index. Results: In total 410 participants died. In a model including all predictors, 1 standard deviation increase in a score composed of mean sedentary bout duration, time in 10-30 and ≥30 min SB bouts, total moderate-to-vigorous physical activity (MVPA) duration, time in <10min MVPA bouts, number of MVPA bouts, number of days with ≥30min MVPA, average acceleration, and intensity gradient was found to be associated with a 13% increase in mortality risk (95% confidence interval: 9-18%). This score was the strongest predictor of mortality after age. Gain in predictive value was higher for this score (ΔC-index=0.011; ΔAIC=-30.5) compared to lifestyle (ΔC-index=0.002; ΔAIC=-0.3) and health-related (ΔC-index=0.006; ΔAIC=-10.9) predictors, but lower compared to sociodemographic predictors (ΔC-index=0.056; ΔAIC=-112.4). Conclusions: More frequent and longer time in MVPA, more fragmented SB, and better activity intensity distribution were identified as core predictors of mortality in old age, above other modifiable risk factors.

The Association Between Moderate- to- Vigorous Physical Activity During Commuting and Metabolic Markers

Abolanle Gbadamosi1, Alexandra Clarke-Cornwell1, Malcolm Granat1

1University of Salford

Background: Commuting to and from work can increase moderate-to-vigorous physical activity (MVPA) and objective measures of quantifying MVPA during commuting vary depending on the objective device used. Cadence has been suggested as a practical measure of estimating the intensity of activities. The aim of this study was to investigate the association between objectively-measured MVPA during commuting and metabolic markers. Methods: Forty participants were recruited from a sample of staff and postgraduate research students at the University of Salford to wear the activPAL? and filled out an activity diary to collect data on commute duration, mode, and demographic and diet questions. The mode of commute was categorised into car, walk, cycling, and mixed-mode (those using more than one mode of commute). Metabolic markers were measured to investigate the associations between commute MVPA outcomes and metabolic markers. The definition of MVPA used was a minimum walking cadence of either 76, 100, or 109 steps/minute. Results: Car (25.7 (23.7-27.4) kg/m2) and walking commuters (25.2 (21.5-29.2) kg/m2) were slightly overweight, compared to the mixed-mode commuters, who had an average BMI in the 'normal’ range (23.8 (21.7-25.7) kg/m2). The participants were in the normal/healthy ranges for fasting blood glucose, high HDL-cholesterol, low triglycerides, systolic blood pressure, and diastolic blood pressure as defined by National Cholesterol Education Program Adult Treatment Panel III criteria. Commute time in MVPA, before and after grouping (where short interruptions of time between walking events were combined based on an average walking cadence), was significantly negatively associated with BMI (Figure 9). Non-commute and total time in MVPA did not give any significant results. Conclusion: Commuting can be a major contributor to total daily MVPA, with the mode of commute playing a significant role: active commuting may provide a protective effect against metabolic syndrome.

Integrated Systems to Assess Physical Behavior

Assessment of Activities of Daily Living Using Markerless Motion Capture in a Virtual Reality Setting

Kevin Abbruzzese1, Andre Freligh1, Vincent Alipit1, Sally LiArno1

1Stryker Orthopaedics

Unity3D gaming software can be utilized to create virtual applications to assess Activities of Daily Living(ADL). Virtual Reality (VR) platforms provide a systematic method to assess accuracy and repeatability of measurements. Chair rise activities can provide insight into functional movement and rely on clinical assessments to monitor performance. Markerless motion capture (MMC) represents a potential modality to quantify ADLs. The objective of this study was to validate virtual object scale and determine accuracy of virtual object height during a seated task in the sagittal and frontal plane. A VR task was performed to assess virtual object height measurements in the frontal (FP) and sagittal plane (SP) using MMC to map avatar movement. Various chair heights were measured in a VR interface in Unity3D and confirmed with a tape measure. Ten measurements were recorded for five different object heights for both SP and FP measurements for a single participant. The user was required to sit in a physical chair where the virtual object height was adjusted until the virtual object contacted the avatar in the seated position. A paired 2 sample T-test was performed to assess significance. No significant differences in SP measurements were detected with average virtual object height of 17.63in±3.47in and average physical object height of 15.1in±3.54in(p=0.144). Average observed difference in object height was 2.53in for SP measurements. Significant differences were detected between measurements in the FP (p=0.003). Significant differences were observed between planar measurements (p=0.00034) with FP measurements resulting in greater errors (6.22in). A VR interface with Azure Kinect camera demonstrated comparable results to analog measurements in the SP with low error. This work demonstrates the potential to monitor patient progress with VR during a seated activity with MMC. This modality may provide a systematic method to assess patient kinematics using MMC during an ADL. (Figure 10)

Effects on Heart Rate, Physical Activity and Ambulatory Blood Pressure From Occupational Physical Activity With and Without Lifting Among Farmers in Denmark

Mette Korshøj1, Mathilde Baumann1, Michael Olsen1, Ole Mortensen1

1Holbæk Hospital

Objective: High levels of occupational physical activity associate to increased risk of cardiovascular disease. However, knowledge regarding the acute effects of different components of the occupational physical activity, such as lifting, on risk factors for cardiovascular disease remains uninvestigated during every day work. Thus, the aim was to investigate the acute effects from exposure to occupational physical activity with and without lifting on heart rate, physical activity and ambulatory blood pressure. Methods: A randomized cross-over study among 18 farming workers in Denmark, all working in the stables of pig-producing farms. Workday measurements of heart rate (Actiheart), physical activity (Axivity) and ambulatory blood pressure (Spacelabs 90217, measuring every 20th minute) were collected at a workday with and a workday without occupational lifting. The wash out period between the measurements was 48 hours. Data were processed in the Acti4 software. Results: During workdays with lifting compared to workdays without lifting we observed non-significantly higher intensity of occupational physical activity (Δ 3.75 % heart rate reserve), number of steps/workday (Δ 4,346 steps), standing/walking activities (Δ 78 min/workday), as well as higher heart rate (Δ 6.57 bpm) and higher ambulatory blood pressures, both systolic (Δ 3.77 mmHg) and diastolic (Δ 1.37 mmHg). The average burden of the occupational lifting were 2,425 kg/workday and amount of lifts/workday were 239 lifts. Conclusions: This pilot project indicated that occupational lifting are adding strenuousness on top of the general occupational physical activity, and influence blood pressure and heart rate at clinically relevant magnitudes. Disentangling the potential relations between one component of occupational physical activity, such as lifting, and risk for cardiovascular disease is key in the development of initiatives for specific prevention, exposure recommendations and vocational rehabilitation.

Estimation of Metabolic Rate During Submaximal Exercise Using Heart Rate, Sex, Age, Training Status and Exercise Mode in Participants With and Without a Disability

Julia Baumgart1, Emma den Hartog1

1Princess Máxima Center for Pediatric Oncology

Objective: The primary objective of this study was to investigate how well metabolic rate (MR) can be estimated from heart rate (HR) when accounting for body mass, sex, age, having a disability or not, training status (i.e., recreational, sub-elite, and elite), and exercise mode (i.e., upper-body, lower-body, and whole-body exercise) during submaximal exercise. In addition, it was investigated whether the use of percentage of peak HR (%HRpeak) instead of HR provides a better estimation of MR. Methods: Submaximal tests with increasing intensity were performed by 203 individuals who were at least recreationally trained. Linear mixed models were used to estimate MR from HR or %HRpeak while adjusting for body mass, sex, age, having a disability or not, training status, and exercise mode. The amount of variation explained by the model was assessed by the R-squared (r2). Agreement between the measured and estimated MR was assessed with Bland-Altman plots. Results: 71% of the variation in measured MR was explained by an estimation model based on HR, age, sex, having a disability or not, training status, and exercise mode. When using %HRpeak instead of absolute HR values, 74% of the variation in individual actual MR was explained by the model. Despite relatively high r2 values and good agreement on group level (i.e., the mean difference between actual and estimated MR is 0), variation between the individually measured and estimated MR ranged between 300 and 450 W for HR (larger variation for higher HR) and was around 380 W for %HRpeak. Conclusions: Models for estimating MR from either HR or %HRpeak (as well as body mass, sex, age, having a disability or not, training status, and exercise mode) are in good agreement with direct measurements of MR during submaximal exercise. However, caution is required due to the large variation between individually measured and estimated MR in both models, indicating the need for more advanced, non-linear modelling procedures.

Exploring Effects of Central Sensitization on Gait in Chronic Low Back Pain By Using Machine Learning Approach

Xiaoping Zheng1, Michiel Reneman1, Jone Echeita1, Schiphorst Preuper1, Herbert Kruitbosch1, Egbert Otten1, Claudine Lamoth1

1University of Groningen

Background: A major issue for interventions for patients with chronic low back pain (CLBP) is the heterogeneity of the patient population. Many studies have shown that gait patterns among patients with CLBP are inconsistent. One of the explanations for the inconsistent finding could be the presence of central sensitization (CS). This study aimed to explore whether patients with CLBP with low or high CS levels (a CS Inventory score lower than 40 (CLBP-), or 40-100 (CLBP+)) have different gait performance. Methods: Forty-two patients with CLBP were included (23 CLBP- and 19 CLBP+). Patients wore an accelerometer for about one week and 4 days of accelerometer-data was selected randomly. For each day data, continuous gait cycles were extracted to compute 36 gait outcomes which representing the pace, regularity, synchronization, smoothness, stability, and predictability. A Random Forest classifier was trained to classify CLBP- and CLBP+ groups based on gait outcomes and SHapley Additive exPlanations (SHAP) method was used to explain the differences between groups in gait outcomes. Results: The Random Forest method had high classification accuracy (84.4%). The top ten gait outcomes indicated by SHAP were: index of harmonicity-vertical and harmonic ratio-mediolateral (smoothness), stride frequency variability- mediolateral/anteroposterior, stride length variability (variability), stride regularity-mediolateral (regularity), maximal Lyapunov exponent-vertical/mediolateral and maximal Lyapunov exponent per stride-vertical (stability), and sample entropy-anteroposterior (predictability). Conclusion: Differences in gait performance could be classified with high accuracy. The results suggest that CLBP- and CLBP+ presented different motor control strategies. CLBP- presented a more “loose” control, including higher gait smoothness and stability. CLBP+ presented a more “tight” control, including a more regular, less variable and more predictable gait pattern.

Towards Eco-Design of Self-Powered Wearable Devices: Analysis of Available Energy on the Human Body for Lead-Free Piezoelectric Energy Harvester Positioning

Damien Hoareau1, Gurvan Jodin1, Jacques Prioux1, Abdo-Rahmane Anas Laaraibi1, Ausrine Bartasyte2, Samuel Margeuron2, Guylaine Poulin-Vittrant3, Maxime Bavencoffe3, Alexis Brenes4, Elie Lefeuvre4, Florence Razan1

1ENS Rennes, 2Université de Franche-Comté, 3Université de Tours, 4Université Paris Saclay

Introduction: Today, wearable sensors are used in various applications, such as patient health monitoring to sports activity monitoring. Their growing number raises the question of their environmental impact in terms of materials and energy consumption. In this sense, harvesting biomechanical energy using piezoelectric generators made of lead-free alloys is an appealing approach. Method: In this study, one subject performing sports actions was equipped with 17 inertial measurement units (IMU) including 3-axis acceleration sensors. The 17 sensors (MNV Link, Xsens) were located (Figure 11) on the different segments of the human body. The recorded acceleration signals were used to estimate the average energy which could be harvested using a 70 x 10 x 0.2 mm3 LiNbO3 piezoelectric device. For this purpose, an electromechanical model of the harvester was implemented using Matlab software. Results: Temporal and frequency analysis of the recorded data show that the main frequencies are below 10Hz, except in the hands and feet location where impacts exhibit broader spectrum. For the right hand, the model results show a maximum energy production of 50mJ for a 13min acquisition on the z-axis data of the accelerometer. Actions like 'Run’ or 'Dribble’ allow an optimal energy generation because they have repetitive impacts. The best positions for energy harvesting in our study are hands and feet. In the general case, it depends on the type of action performed, the period and the orientation of the generator on the human body. Conclusion: The energy harvested could power some conventional accelerometers and avoid the need for bulky energy storage. Design of low frequency energy harvester in embedded sustainable system is a real challenge. Harvester parameter optimization will then improve energy production.

Measuring Steps

A Step Towards More Intuitive Physical Activity Prescription: Validity of Stepping-Based Metrics Derived From Wrist-Worn Accelerometry

Ben Maylor1, Charlotte Edwardson1, Paddy Dempsey1, Matthew Patterson2, Tom Yates1, Alex Rowlands1

1University of Leicester, 2Shimmer Sensing

Objective: Accelerometer-assessed physical activity (PA) is commonly reported as time spent above acceleration thresholds as a measure of moderate to vigorous PA. However, reporting PA as number of steps per day is intuitive and potentially more understandable. This study sought to assess the precision and validity of an algorithm which estimates steps from wrist-worn accelerometry. Methods: Baseline data from the SMART Work and Life study of office-based employees were used. Participants (N = 656, mean age 44.7 years, 72% women) wore an activPAL on the thigh and an Axivity AX3 on the non-dominant wrist concurrently for 8 days. Steps derived from the thigh-worn activPAL were the criterion. Steps were estimated from the wrist-worn Axivity using the step-count external function algorithm within GGIR. Steps per day were reported for all days, working days and non-working days. Bland-Altman analysis was used to assess agreement between steps/day estimates from the two monitors (matched for wear-time). Results: Axivity AX3 reported an average (SD) of 10250 (2758) steps/day, with a mean bias (95% Confidence Interval) bias of +851 (709 to 995) steps/day relative to ActivPAL. The 95% limits of agreement (LoA) were large (-2807 to 4510 steps/day). Mean bias was greater on working days (mean bias = +902 steps/day, 95% LoA = -2568 to 4373) than non-working days (mean bias = +809 steps/day, 95% LoA = -3955 to 5573), but LoA were wider on non-working days. Consistent with this, hour-by-hour comparisons showed greater bias during desk working hours. Conclusions: Step-count estimation from the wrist overestimated relative to the thigh-worn activPAL by 8.7% in this sample of office-based employees, with greater overestimation on working days. Further investigation is required to assess performance of the algorithm in other populations and whether the algorithm can be refined to improve performance.

Changes in Brisk Stepping Cadence Are Associated With Improvements in Adiposity, HDL-C, and HbA1c in People With Non-Diabetic Hyperglycaemia

Phil McBride1, Joseph Henson1, Charlotte Edwardson1, Melanie Davies1, Kamlesh Khunti1, Benjamin Maylor1, Thomas Yates1

1University of Leicester

Objective: To investigate the associations between change in step cadence over 48 months and markers of cardiometabolic health in people with non-diabetic hyperglycaemia and to explore whether these associations are modified by demographic factors. Methods: 794 adults, with non-diabetic hyperglycaemia, were assessed for various markers of cardiometabolic health and stepping activity at three time-points (baseline, 12 months, and 48 months). Free-living stepping was measured using activPAL3 monitors. Brisk stepping was defined as ≥100 steps/minute; the mean stepping cadence during the most active 5, 10 and 30 minutes of the day were also derived. Generalised estimating equations examined associations between 48-month change in step cadence and change in cardiometabolic health markers, with interactions by ethnicity, sex, and randomisation group assessed. Results: 794 participants were included (age = 59.8 ±8.9, 48.7% women, 21.9% South Asian (SA), total average steps = 8445 ± 3364 steps/day, brisk steps/day = 4794 ± 2865, peak 5-minute cadence = 136 ± 10 steps/minute, peak 10-minute cadence = 128 ± 10 steps/minute, peak 30-minute cadence = 114 ± 11 steps/minute). Beneficial associations were observed between change in brisk stepping and change in body mass index (BMI), waist circumference (WC), HDL-C, and HbA1c; peak 30-minute step cadence and BMI; all peak step cadence metrics and HDL-C and WC; and slow stepping and BMI. Interactions by ethnicity revealed change in peak 30-minute step cadence had a stronger association with HbA1c in White Europeans (WEs), whereas associations between 5-minute and 10-minute peak step cadence with measures of adiposity were stronger in SAs. Conclusion: In this study, change in brisk walking was associated with beneficial change in adiposity, HDL-C, and HbA1c; however, potential benefits may be dependent on ethnicity for outcomes related to HbA1c and adiposity.

Comparison of the Performances of Step Counting Algorithms in Different Physical Activities

Dawid Gerstel1, Joe Nguyen1, Rakesh Pilkar1, Tyler Guthrie1, Matt Biggs1, Ali Neishabouri1, Christine Guo1

1ActiGraph

Research Objectives: Accurate and objective measures of step counts in the real world are of great clinical importance as walking is strongly associated with functional health, activities of daily living and overall quality of life. Our objective is to measure and compare the accuracies of three step counting algorithms (Lee, Choi, and Lee 2015 (A1); Femiano et al. 2022 Autocorrelation (A2) and Windowed Peak Detection (A3)) and two ActiGraph algorithms (Moving Average Vector Magnitude Algorithm (MAVM) and ActiLife) applied on raw accelerometer data. Methods: ActiGraph internal dataset collected at 30 Hz using GT9X-Link devices during various activities such as walking (outdoors and treadmill), running and free-living from healthy adults (age: above 18) or adolescents (age: 8-17). Three algorithms (A1, A2, A3) are implemented in Python to determine step counts for each activity and the outputs are compared to the ground truth. Results: A1 showed the best results when running, with a mean relative error of approx. 3% and was relatively accurate for walking (approx. 15% relative error). A2 managed relative error rates of 20% and 5%, for running and walking respectively. A3 had a mean of -41% for running and -10% for walking. MAVM resulted in errors of approx. -10% for running and -15% for walking. ActiLife showed a much larger error of approx. -48% for running and -30% for walking. We also tested our algorithms on a dataset of short walking bouts captured in more realistic conditions. A1, A2, A3 and ActiLife had respective errors of -42%, -18%, -30% and -44%. Because MAVM requires walking bouts of at least 10 seconds, its performance on this dataset was very poor (-92%). Conclusions: The choice of the optimal algorithm depends largely on the activity type. Our work aims to provide researchers an evidence-based approach to guide the choice of algorithms in their context of use.

Development of an Externally Validated Free-Living Step Counting Algorithm With Deployment in the UK Biobank

Scott Small1, Lennart von Fritsch1, Shing Chan1, Andrew Creagh1, Andrew Price1, Sara Khalid1, Aiden Doherty1

1University of Oxford

Daily step count is an intuitive measure of physical activity, readily accessible to the public and frequently utilised by researchers in a range of clinical research areas. However, a major limitation to step counting is the difficulty of accurate step detection from wrist-worn accelerometry, particularly in the free-living environment. Ground truth video was captured from heathy volunteers during 1 hour of unscripted, free living while wearing Axivity AX3 accelerometers on the dominant wrist. Individual steps were then manually annotated by two annotators. We trained a hybrid step detection model using a binary random forest walking classification model followed by tuned peak detection. External validation was performed using the Clemson University open-source dataset of 30 participants wearing a wrist based Shimmer3 IMU. To access face validity, the model was then deployed in the UK Biobank physical activity dataset, with median daily step count and peak cadence calculated for all 100,000 participants. Thirty-nine video-recorded participants (aged 19-81) took a mean 1,604 steps during the 1 hour window. The final hybrid model estimated step count with a root mean square error of 11.6% and population mean bias of 0.8% in the external Clemson validation dataset. In the UK Biobank, high correlation (0.67-0.79) was observed between overall acceleration and median daily steps across age and sex. Aged 45-54 females and males recorded a median [IQR] 10,399 [7,891-13,316] and 9,730 [7,446-12,493] daily steps, decreasing to 7,589 [5,087-10,037] and 7,347 [5,157-10,117] steps in aged 75-79 females and males, respectively. Peak cadence was a mean 10 steps/min higher in self-reported brisk walkers compared to slow walkers (p<0.01) and decreased by 2 steps/min per decade of aging (p<0.01). This study offers a free living, wrist-based step detection algorithm with apparently low error and good association, in an expected direction, with other metrics in the UK Biobank. (Figure 12)

Device Comparison of Free-Living Steps per Day: A Systematic Review and Meta-Analysis

Amanda Paluch1, Eric Eberl1, Kelly Evenson2, Erika Rees-Punia3, Susan Park1, Lindsay Toth4, David Bassett5

1University of Massachusetts Amherst, 2University of North Carolina Chapel Hill, 3American Cancer Society, 4University of North Florida, 5University of Tennessee Knoxville

Objective: There has been a rapid rise in use and variety of activity trackers to measure steps for consumer and research purposes. It is uncertain whether these devices are comparable in estimating steps per day (steps/d). We conducted a systematic review and meta-analysis to summarize device comparison studies of free-living steps/d. Methods: The search was performed in PubMed, PsychInfo, and Scopus through November 2020. Device brands included ActiGraph, StepWatch, activPAL, Actical, Omron, Yamax, Fitbit, Garmin, New Lifestyles, and Apple Watch. Studies comparing steps/d with at least one full day (≥10 hrs) of simultaneous wear in free-living conditions, among generally healthy community-dwelling adults were included. Study-level steps/d means (s.d.) were extracted for each study. Mean differences (device 2-device 1), ratio of means (device 2/device 1), and 95% confidence intervals (CI) of study-level steps/d were pooled using random-effects models. Heterogeneity was measured using I2. Results: In total, 285 device comparisons from 55 studies were included. Most commonly, devices were compared against ActiGraph (205 comparisons). For example, there were 44 comparisons between the ActiGraph GT model devices worn at the hip (normal filter) with eight other devices/wear locations (Figure 13). Compared to ActiGraph GT, seven devices did not significantly differ in steps/d, with ratio of means ranging from 0.99 [95%CI, 0.91-1.07, n=3 studies] for Omron (hip worn) to 1.11 [95%CI, 0.99-1.22, n=10 studies] for Fitbit (wrist worn). StepWatch (ankle worn) had significantly higher steps/d (ratio of means: 1.39 [95%CI, 1.23-1.56, n=4 studies]) compared to the ActiGraph GT. Heterogeneity was low to moderate (I2, 0-55%). Conclusion: This study can inform calibration and harmonization of steps/d estimates across common devices. Manufacturer, versions, processing, and wear locations are important considerations for the comparability of free-living step estimates between devices.

Novel Statistical Approaches and Applications

A Fully Bayesian Semi-Parametric Scalar-on-Function Regression (SoFR) With Measurement Error Using Instrumental Variables

Roger Zoh1

1Indiana University

Wearable devices such as the ActiGraph are now commonly used in research to monitor or track physical activity. This trend corresponds with the growing need to assess the relationships between physical activity and health outcomes, such as obesity, accurately. The device-based physical activity measures are best treated as functions when assessing their associations with scalar-valued outcomes such as body mass index. Scalar-on-function regression (SoFR) is a suitable regression model in this setting. Most estimation approaches in SoFR involve an assumption that the measurement error in functional covariates is white noise. Violating this assumption can lead to under-estimating model parameters. There are limited approaches to correcting measurement error for frequentist methods and none for Bayesian methods in this area. We present a fully non-parametric Bayesian measurement error-corrected SoFR model that relaxes all the constraining assumptions often involved with these models. Our estimation relies on an instrumental variable which is allowed to have a time-varying biasing factor, a significant departure from the current approach. Our method is easy to implement, and we demonstrate its finite sample properties in extensive simulations. Finally, we applied our method to data from the National Health and Examination Survey to assess the relationship between wearable device-based measures of physical activity and body mass index in adults in the United States.

Association of Gait Quality With Daily Life Mobility: An Actigraphy and Global Positioning System Based Analysis in Older Adults

Anisha Suri1, Jessie VanSwearingen1, Emma Baillargeon1, Breanna Crane2, Michelle Carlson2, Kyle Moored1, Pamela Dunlap1, Patrick Donahue2, Mark Redfern1, Jennifer Brach1, Ervin Sejdic3, Andrea Rosso1

1University of Pittsburgh, 2Johns Hopkins University, 3University of Toronto

Objective: Walking is a key component of daily-life mobility. We aimed to test associations between laboratory gait quality and daily-life mobility from actigraphy and Global Positioning System (GPS). We hypothesized that better gait quality would be associated with greater physical activity and more time outside home (TOH). Methods: Cross-sectional, baseline data of older adults from an intervention study were used. Activity was measured for a seven-day wear period by step-count from actigraphy and TOH (%) from GPS. Gait quality was derived from six passes over a 4-m instrumented walkway [pace (gait speed) and variability (stride time coefficient of variation)] and from accelerometry signals at the lower back during a 6-minute walk test [adaptability (standard deviation signal amplitude, Anterior-Posterior [AP]), smoothness (harmonic ratio AP), power (peak frequency, Vertical [V]), and regularity (entropy rate V)]. Individuals were divided based on tertile thresholds for step-count and TOH. Separate analyses for each metric compared groups using ANCOVA and Tukey (age, body mass index, sex as covariates). Results: Participants (n=121, age=77±5 years, 70% females, 90% White) were placed into high (>4800 steps, n=41), medium (3100-4800 steps, n=38), and low (steps<3100, n=42) activity groups. The high activity group had faster gait speed (1.16 ± 0.13 vs 1.13 ± 0.16 vs 1.00 ± 0.17 m/s) and greater power (1.94 ± 0.83 vs 1.68 ± 0.54 vs 1.34 ± 0.87 Hz) than low activity group, p<.05. A decreasing trend in adaptability and an increasing trend in regularity was noted (p<0.1). Based on TOH [high (>21%TOH, n=39), medium (12-21%TOH, n=42), and low (<12%TOH, n=40)], medium TOH group had greater smoothness than low TOH group (2.84 ± 0.67 vs 3.03 ± 0.88 vs 2.57 ± 0.62), p<.05. Increasing trend in regularity was noted (p<0.1). Conclusions: Step-count and TOH were associated with unique gait measures. Linking clinical gait measures to daily-life mobility can inform interventions.

Combining Compositional Data Analyses and Ecological Momentary Assessment: Insights on the Association Between Physical Behavior on Mood in Daily Life

Marco Giurgiu1, Ulrich Ebner-Priemer1, Dorothea Dumuid2

1Karlsruhe Institute of Technology, 2University of South Australia

Objective: Given the increasing number of mental disorders, a growing body of studies now focuses on the relationship between physical behavior (PB) (i.e., physical activity (PA), sedentary behavior (SB)) and mental health. However, momentary mechanisms and interrelatedness between PA, SB, and mood in daily life are largely unknown. Methods: To examine whether the composition of light physical activity (LPA), moderate-to-vigorous physical activity (MVPA), and SB influences momentary mood, we conducted an Ecological Momentary Assessment (EMA) study in the everyday life of 103 university students over five days. We measured PB continuously via accelerometers and assessed mood up to six times each day on smartphone diaries. We integrated compositional data analyses (CoDA) with multilevel modeling to analyze within-person associations of the behavior composition with mood. Results: Higher ratio of LPA to SB and MVPA positively influenced energetic arousal (p < 0.001) and a higher ratio of MVPA to SB and LPA positively influenced valence (p = 0.004) and energetic arousal (p = 0.022). Furthermore, a higher ratio of SB to LPA and MVPA within the 60 minutes prior to a diary rating negatively influenced valence (p = 0.006) and energetic arousal (p < 0.001). Simulation analyses revealed that replacing 20 minutes of SB with PA can influence mood rating up to 3.39 units on a scale ranging from 0-100. Conclusions: Findings suggest that replacements of SB with PA may lead to mood enhancements. Given the high prevalence of mental disorders, more studies are warranted to deepen the understanding of momentary compositional mechanisms between PB and mood. Applying CoDA to intensive longitudinal data can serve as a starting point to identify the optimal composition of SB, LPA, and MVPA for mood enhancements in everyday life.

Methods to Determine Common Periods of Wear in Concurrently Worn Activity Monitors

Craig Speirs1, Malcolm Granat2, David Loudon3

1University of Strathclyde, 2University of Salford, 3PAL Technologies Ltd

Objective: The use of device-specific wear protocols and criteria to identify periods of non-wear make it challenging to identify periods of common wear. A wear time validation based on shared characteristics could identify common wear periods, allowing robust data analysis. We investigated an approach to use acceleration and stepping data to identify common periods of valid wear. Methods: We identified 922 older adults from the iData study with concurrently worn thigh-worn activPAL (aP) and hip-worn ActiGraph (Ag) data. Hourly step count and acceleration volume (aP - sum of absolute difference, Ag - sum of vector magnitude) were calculated. The ratios of aP acceleration to Ag acceleration (acceleration comparison ratio - ACR) and aP steps to Ag steps (step comparison ratio - SCR), were then calculated. Results: 11,342 days of data were analysed. There were 110,474 hours (40.8%) with no Ag acceleration. The median ACR was 7.6 (IQR - 5.6). The median SCR was 0.54 (IQR - 0.33). Ag step count was 51% greater than aP step count in low stepping (< 500 steps) periods. As hourly step count increased above 2,500 steps, agreement between the measured step counts increased. Conclusions: There were periods with large volumes of acceleration across both devices with a similar ACR. These are likely to be common wear with continuous stepping. The periods with zero Ag and low aP acceleration may be valid wear with minimal movement or periods of non-wear. These could be classified based on surrounding periods, with prolonged zero/low acceleration in one or more devices indicating non-wear and periods of previously identified common wear indicating a period of common wear. During low stepping periods, Ag step count was substantially greater than aP step count. For many periods, sizeable Ag step counts were accumulated in aP defined sedentary periods. Some of this difference may be Ag classifying periods of noisy sedentary activity, such as seated transportation, as LPA. (Figure 14)

Unknown Distributions; Modelling Distributions of Real-World Walking Speed in People With Parkinsons

Cameron Kirk1, Rana Zia Ur Rehman1, Brook Galna2, Saverio Ranciati3, Encarna Mico-Amigo1, Lynn Rochester1, Alison Yarnall1, Silvia Del-Din1

1Newcastle University, 2Murdoch University, 3University of Bologna

Introduction: Traditionally, researchers have condensed assessment of real-world walking speed (RWS) into a single mean or median value, which may hide clinically valuable information. This is relevant for Parkinson’s disease (PD), where fluctuations of symptoms and medication state could play an influence upon RWS and result in a pattern of several modes in the distribution. While previous research has assumed a bimodal distribution in the RWS of PD, no study has estimated a data-driven number of distributions (modes). Here we present an application of gaussian mixture modelling (GMM) to establish the number of modes within distributions of RWS in people with PD. Methods: Sixty-one PD participants (Age = 69.3± 10 years, MDS-UPDRS III = 38±12.4 points) were recruited from the Incidence of Cognitive Impairment with Longitudinal Evaluation - GAIT (ICICLE-GAIT) study. Participants were assessed for RWS using an accelerometer (Axivity AX3) for seven days. RWS was the weekly average aggregated across 'all’ WBs. To model the number of modes for each participant, we employed Gaussian Mixture Modelling (GMMs), using the “mclust” package in R. Results: The largest proportion of participants had a four-mode distribution (n = 27, 44%) followed by a trimodal distribution (n = 23, 37%). The remaining participants were characterised as having five modes, (7, 11%), six modes (4, 6%) or seven modes (1, 1%). Discussion: The GMMs identified distributions between individuals ranging from three to seven modes, with most participants having three or four mode distribution. These findings challenge the notion of a consistent bimodal distribution of RWS and demonstrates the importance of modelling, rather than assuming a uniform number of modes exist. Further work is required to understand if the value of RWS within modes of gaussian distributions are clinically useful (e.g. sensitive to disease severity or medication state), and establish their validity and reliability. (Figure 15)

Physical Activity Determinants and COVID-19

Does Context Matter? The Association Between Affective States and Physical Behavior and Its Moderation By Weather Factors Measured With Ambulatory Assessment

Irina Timm1, Markus Reichert2, Ulrich Ebner-Priemer1, Marco Giurgiu1

1Karlsruhe Institute of Technology, 2Ruhr-University Bochum

Introduction: Research findings indicate that there are associations between affective states and daily physical behavior (PB) (i.e., physical activity and sedentary behavior). Affective states as drivers of physical activity may influence human health behavior. Whether the relationship is influenced and moderated by other contextual factors is largely unknown. This study examined the within-subject relationship between affective states and subsequent PB. Additionally, the effects of contextual weather factors on PB were analyzed and interactions were exploratively considered. Methods: Utilizing ambulatory assessment, 79 participants completed electronic diaries about their affective states up to six times a day over five days, and simultaneously their PB was recorded via accelerometers. The values from the weather station of the Climate Data Center of the German Weather Service were included in the multilevel analysis. Results: Increased valence and energetic arousal were positively associated with physical activity (beta = 0.007; p < .001), whereas calmness predicted lower levels of physical activity (beta = -0.006; p < .001). Higher levels of calmness showed a positive association with sedentary behavior (beta = 0.054; p = .003). Temperature was positively associated with physical activity (beta = 0.025; p = .015). Exploratory analyses showed that temperature significantly moderated the relationship between calmness, valence, and PB. Discussion: The findings suggested that affective states in everyday life were related to PB in adults. Consideration of contextual weather factors referred to the complexity of the relationship. Health care practitioners could benefit from knowing how to incorporate contextual factors to promote a sustainable physical active lifestyle.

Multiple Accelerometry Assessed Physical Behavior Across 24-Hour Period in Older Adults With Different Level of Physical Fitness: A Pilot Study During COVID-19 Pandemic

Jan Vindis1, Denisa Nohelova1, Jana Pelclova1

1Palacký University Olomouc

Background: The daily physical behavior (PB) profiles of older adults were affected during the pandemic restrictions (during COVID-19). The difference in PB profiles of older adults with different physical fitness (PF) levels during the pandemic is unknown. Objectives: The purpose of this study was to describe the differences in the profiles of 24-hour PB in older adults with different levels of PF. Methods: Data were collected in 192 older adults (70.9 ± 5.3 years, 128 women) by three accelerometers worn on different body segments (thigh, hip, wrist) during consecutive seven days. Accelerometers were set at 25 Hz (Axivity AX3) and 30 Hz (ActiGraph wGT3X-BT) frequency with a dynamic range of ± 8 g. The raw data processing was performed using two software R package GGIR and Acti4. The level of PF was assessed according to the Short Physical Performance Battery. Subsequently, differences between PB profiles in older adults with different levels of PF were compared by t-test and Hotelling test. Results: Relative PB profiles significantly differed between older adults with different levels of PF (p < 0.001). Based on compositional geometric means, older adults with lower PF were characterized by a relatively lower amount of moderate to vigorous physical activity (by 26.1%; 22.2 min/day) and a higher amount of sedentary behavior (by 15.0%; 57.4 min/day) compared to older adults with good PF. Older adults with lower PF spent 19.1 min/day more time walking (p=0.002) and 16.1 min/day more time fast walking (p=0.002) than older adults with higher PF. Conclusion: Differences in daily PB profiles between older adults with different levels of PF were apparent also during pandemic restrictions. Therefore, public health interventions are needed to prevent unhealthy changes in PB in the older adults with lower PF.

Temporal Patterns of Sitting and Non-Sitting in Normal-Weight and Overweight Brazilian Office Workers Working From Home During the COVID-19 Pandemic

Luiz Augusto Brusaca1, Svend Erik Mathiassen2, David M. Hallman2, Nidhi Gupta3, Dechristian França Barbieri4, Ana Beatriz Oliveira1

1Federal University of São Carlos, 2University of Gävle, 3National Research Centre for the Working Environment, 4Clemson University

Objective: This study documented the temporal patterns of sitting, non-sitting and time-in-bed (TIB) of Brazilian office workers working from home during the pandemic; and also determined the extent to which these patterns differed between normal-weight (NWW) and overweight workers (OWW). Methods: Behaviors were monitored over 5 days using accelerometers in 33 NWW (BMI <25 kg/m2) and 40 OWW (BMI ≥25 kg/m2). Time-use compositions were described in terms of sitting, non-sitting and TIB. Temporal patterns of sitting/non-sitting were quantified according to Exposure Variation Analysis into short (≤5 min), moderate (>5 and ≤20) and long uninterrupted bouts (>20). Following compositional data analysis, isometric log-ratios (ilr) were calculated; ilr1-TIB relative to time spent awake, ilr2-sitting (all bouts) relative to non-sitting (all bouts), ilr3-sitting in short relative to moderate and long bouts, ilr4-sitting in moderate relative to long bouts; ilr5 and ilr6 represent the same behavior contrasts as ilr3 and ilr4, but for non-sitting. We examined differences between groups using MANOVA, followed by univariate post-hoc tests of pairwise differences. Results: NWW spent more time sitting in short bouts (50 min) and less time in moderate and long bouts (154 and 552) than OWW (42, 155 and 585). For non-sitting, NWW spent more time in short and moderate bouts (71 and 96) and less time in long bouts (45) than OWW (60, 83 and 54). NWW had longer TIB (473) than OWW (461). NWW and OWW differed in the set of ilrs as a whole (p=0.05). The post-hoc tests showed that time spent sitting in short relative to longer bouts (ilr3) was smaller for OWW than for NWW (p=0.05). This indicates that OWW had less variation in sitting behaviors. Conclusions: OWW spent less time at work sitting in short uninterrupted bouts, relative to sitting for longer bouts, than NWW, while the relative time-use did not differ for other behaviors.

The Impact of UK COVID-19 Restrictions on Objectively Measured Physical Behaviour

Alexandra Clarke-Cornwell1, Benjamin Griffiths1, Benjamin Maylor2, Malcolm Granat1, Charlotte Edwardson2

1University of Salford, 2University of Leicester

Background: The COVID-19 pandemic led to a national lockdown in the UK in March 2020: employees were encouraged to work at home where possible. Although studies reported a decrease in physical activity and an increase in sitting time as a result of the lockdown, many used self-reported surveys and relied on participants to accurately recall their behaviours prior to the restrictions. Too much sitting time is associated with a number of health-related outcomes in office workers, which can have negative consequences in terms of economic impact and reduced productivity. With many office workers expecting to regularly work from home now that all restrictions have been lifted, there is a need to assess the impact of these changes on objectively measured physical behaviour. Methods: Pre-COVID-19 data from 756 office workers from local government councils were available as part of a workplace intervention to reduce sitting time. These employees were invited to complete a short online survey in May 2020 to identify changes in daily lifestyle behaviours as a consequence of COVID-19 restrictions; they were invited to complete a second survey one year after the first (May 2021). Sitting time and physical activity were assessed using the activPAL? in early 2020 (pre-COVID-19) and a sub-sample of participants agreed to wear the activPAL? in May 2020 and May 2021. Results: During the first lockdown, there was an increase in % of daily sitting time compared to pre-COVID 19, and a decrease in % of standing and stepping; however, none of these changes were significant. A year later, % of sitting time remained higher than pre-COVID-19, % standing had further decreased, and % stepping increased. Conclusion: The COVID-19 restrictions changed the distribution of physical behaviours in office workers. The workplace should ideally enhance both health and productivity and therefore employers should look at ways of increasing standing and walking in employees who continue to work at home. (Figure 16)

Typical Day and Influence of Weekend on Accelerometer Measured Physical Activity

Alexander Burchartz1, Simon Kolb1, Steffen Schmidt1, Birte von Haaren-Mack, Claudia Niessner1, Alexander Woll1

1Karlsruhe Institute of Technology

Background: Structured activities in which children participate for example during school are consistent and limited in scope. In contrast, after-school or weekend activities involve a wider range of behaviors. Studies have shown that physical activity (PA), as measured by accelerometers, is lower on weekends compared to weekdays or school days, whereas PA does not differ between weekdays. In the present study we examined accelerometer data in the spotlight of the different week- and weekend days among children and adolescents living in Germany. Sedentary behavior, light, moderate, and vigorous PA were analyzed. Methods: Physical activity of n = 2,278 children 6-17 years (52.3% female) was measured using ActiGraph accelerometers (GT3X+/wGT3X-BT). Absolute and percentage intensity distributions were evaluated daily. Results: The wear-time was on average daily 807 min on Monday to Thursday with significant deviations from the mean on Friday + 38 min, Saturday -76 min, and Sunday -141min. Absolute moderate to vigorous PA times were lower on weekends than during the week, however, the percentage intensity distribution remains uniformly consistent across all days. Girls are less physically active than boys and adolescents are significantly less active than younger children. Awake times increase with age. Conclusion: Shorter awake periods limit the possible active times on weekends, resulting in lower PA and sedentary behavior times compared to weekdays. The percentage distribution of the different activity intensities is the same across all weekdays and weekend days. We could not find a justification for specific weekend interventions; instead, interventions should generally try to shift activity behavior away from sedentary behavior toward a more active lifestyle. (Figure 17)

Physical Activity Interventions

A Physical Activity Intervention Results in Higher Randomness of Postural Control Accelerations During Dual-Task Conditions

Kayla Bohlke1, Patrick Sparto1, Mark Redfern1, Ervin Sejdic2, Andrea Rosso1

1University of Pittsburgh, 2University of Toronto

Automaticity of balance decreases with age and increasing automaticity may improve function and reduce falls in older adults. This study examined the intervention-related changes in standing postural control with and without a concurrent dual-task to examine automaticity. We hypothesized that training would improve postural control, particularly during dual-task condition. Sway measures were recorded by an accelerometer (Actigraph wGT3X) placed on the low back before and after a 12-week physical activity intervention to improve gait speed. Twenty-nine older adults (76±7 years old, 17 female) participated. Quiet standing followed by standing plus reciting every other letter of the alphabet were the single- and dual-task conditions. Time and frequency domain features were extracted from the accelerometer signals to quantify postural control (e.g. root-mean-square, centroid frequency, entropy rate, and wavelet entropy). Wilcoxon matched-pair signed rank tests (p<.05) compared the following: (post dual - post single) - (pre dual - pre single). Of the six accelerometry variables examined, only medial-lateral wavelet entropy (ML WE) single- to dual-task changes differed from pre- to post-intervention (p=.034). Post-hoc analyses by visit (post single vs. pre single) and task type (pre single vs. pre dual) found the difference was driven by an increase from single- to dual-task ML WE during post-intervention (p=.011). Additionally, seven within-task comparisons from pre- to post-intervention showed significant changes, all of which indicate improvements in balance (e.g. lower single- (p=.046) and dual-task (p=.024) AP root-mean-square). Higher WE indicates more randomness and could imply increased adaptability and thus improved automatic postural control responses during dual-task. The changes from pre- to post-intervention indicate that the older adults improved their balance from an intervention focused on gait, possibly through increased automaticity of motor control.

An Empirical Approach to Understand Mhealth Application Engagement and Its Associations With Daily Changes in Physical Activity in a Lifestyle Intervention Among US Veterans With Prediabetes

Krista Leonard1, Abdullah Mamun1, Hassan Ghasemzadeh1, Matthew Buman1

1Arizona State University

Mobile health (mHealth) interventions have potential to promote physical activity (PA). Yet, mHealth interventions suffer from poor engagement. Understanding mHealth engagement and its associations with PA may improve efforts to maximize the effectiveness of mHealth interventions. This study examined if patterns of mHealth engagement predicted changes in PA during a 9-month mHealth intervention. Data were drawn from US Veterans with prediabetes (N = 53; 74% male, 34.9±7.0 BMI, 55.9±10.5 yr) participating in BeWell24, a multicomponent lifestyle mHealth intervention targeting PA and related behaviors (sleep, sedentary time, diet). The BeWell24 app provided health education, adaptive behavior change content, and ongoing objective feedback. The app passively recorded the number of times the app was used per day. Participants wore a Fitbit Charge 2 continuously. Minute-level PA was summarized into light (1-5-4 METs; LPA) or moderate-vigorous PA (MVPA; ≥4 METs). Similarity scores were calculated to represent the fraction of days where PA increased as frequency of app use increased and days where PA decreased/remained stable as app usage decreased/remained stable. Scores ≥70% were considered a strong correlation. On average, participants used the app 4.1±6.3 times/day and engaged in 193.0±92.4 min/day of LPA and 90.4±80.0 min/day of MVPA. Average similarity scores ranged from 43-49% with a similarity score ≥70% for 8% of participants. On days when the frequency of app use increased from the previous day, 2% of participants increased their LPA/MVPA the next day at least 70% of the time. On days when the frequency of app use decreased/remained stable from the previous day, 6% of participants reduced their LPA/MVPA the next day at least 70% of the time. The frequency of app use on a given day was not a strong indicator of PA the following day. More research is needed to identify if alternative metrics of engagement (e.g. duration) may be stronger indicators of PA promotion.

Detecting and Modifying Daily Inactivity Among Adults Over 60 Years Using an Integrated Two-Way Communication-Based Near-Real-Time Sensing System: A Randomized Clinical Trial

Diego J Arguello1, Ethan Rogers1, Grant Denmark1, Gregory Cloutier1, Carmen Castaneda-Sceppa1, Charles Hillman1, Arthur Kramer1, Dinesh John1

1Northeastern University

Supervised personal training is most effective in improving the health effects of exercise in older adults (OA) >60y. Yet, low frequency (60 min, 1-3x/wk) of trainer contact limits influence on behavior change outside weekly sessions. Strategies to extend the effect of trainer contact outside sessions may motivate OA to 'sit less and move more’ and sustain positive behaviors to further improve health. Purpose: To test if supplementing a traditional supervised exercise program with daily near-real-time adaptive coaching based on participant needs improves daily physical activity (PA) and sedentary behaviors (SB) in overweight/obese OA. Methods: 21 OA (age: 67.8±5.8y; BMI: 30.1±4.8 kg^m-2) who completed an ongoing trial (n=46) were randomized to a control (C, n=9) or treatment arm (T, n=12). Both received 4-months (4M) of supervised aerobic/resistance training 2x/wk. T also received a near-real-time adaptive intervention to reduce SB that was grounded in Self-Determination Theory and Motivational Interviewing strategies using knowledge of recent/ongoing behaviors and its context gathered via continuously worn wrist-sensors and text-conversations. Intent-to-treat intervention effects of 24-h PA and SB measured with thigh worn activPAL3CTM for 7-continuous days at baseline and after 4M were analyzed using random-intercept mixed linear models accounting for repeated measures after adjustment for baseline. Results: T significantly reduced sitting (mean Δ: -1.1 h/day [95% CI: -2.0, -0.2], P=.02) and increased standing (mean Δ: 0.7 h/day [95% CI: 0.1, 1.3], P=.04) after 4M, relative to C. Contrarily, C spent less daytime lying than T after 4M (mean Δ: -0.6 h/day [95% CI: -1.1, -0.1], P=.02). Daily stepping and sleep did not differ significantly between groups. Conclusions: Supplementing a traditional supervised exercise program with socially engaging and contextually salient coaching to influence behavior in near-real-time motivates OA to sit less in favor of light PA. (Figure 18)

Development and Pilot Testing of the ActiveGOALS Online Physical Activity Intervention for Primary Care Patients

Bonny Rockette-Wagner1, Gary Fischer1, Andrea Kriska1, Molly Conroy2, David Dunstan3, Sarah Deperrior1, Reagan Moffit1, Neel Rao1, Kathleen McTigue1

1University of Pittsburgh, 2University of Utah, 3Baker Heart and Diabetes Institute

Clinical recommendations endorse physical activity (PA) in the treatment of common health conditions. PA referral is not widely implemented in primary care, in part due to the lack of coordination between PA interventions and routine care. With input from clinical and patient partners, the ActiveGOALS social-cognitive theory-based intervention is being developed for primary care patients not meeting PA guidelines (≥150 minutes/week of moderate-vigorous (MV) PA. Initial pilot testing was conducted with primary care patients (n=79; aged 21-70 yrs) randomized (1:1) to a waitlisted control or the ActiveGOALS program (13 weekly online sessions, remote coaching, Omron Alvita Step counter). At 3 months participants randomized to ActiveGOALS (vs waitlist) recorded higher mean [95% CI] MVPA minutes/day occurring in 10 minute bouts (19.4 [11.5,27.3] vs 7.4 [4.6, 10.2], p=0.01) from ActiGraph GT3X monitors; while differences in total MVPA minutes/day (46.6 [35.2, 58.1] vs 29.8 [23.8, 35.8]; p=0.07) and steps/day (6563 [5597, 7528] vs 5406 [4741, 6072]; p=0.16) were clinically important, but not statistically significant. After 3 months waitlisted participants were provided with the intervention, with the substitution of a Fitbit Alta HR tracker. Pre-to-post-intervention improvements in PA were similar for immediate and waitlisted intervention participants. Satisfaction was high with both the Omron and Fitbit, with the Fitbit scoring slightly higher in most categories (appearance, comfort, convenience, ease of use, and durability, but not ease of setup). The results of this study have guided the development of a 1-year PA intervention with increased integration of the tracker; as was suggested by participants. In this next stage of program development we are currently testing the 1-year ActiveGOALS program with Fitbit against an active-control (Fitbit only) group towards providing better options for PA interventions designed to be coordinated with patient primary care.

Wear Fatigue: Does Device Wear Compliance Wane Over a Free-Living Assessment Period?

Samuel LaMunion1, Robert Brychta1, Kong Chen1

1National Institute of Diabetes and Digestive and Kidney Diseases

Background: Accelerometers are often used to objectively measure physical behaviors in free-living environments, typically for ≥7 consecutive days. We hypothesized that over a typical assessment period some participants may be prone to “wear fatigue”, defined as the rate of reduction in wear time from the beginning to the end of the study period. To test this hypothesis, we examined accelerometer data from the 2011-2014 National Health and Nutrition Examination Survey (NHANES) cycles and the 2012 NHANES National Youth Fitness Survey Data. Methods: Participants were instructed to wear an ActiGraph GT3X+ on their non-dominant wrist for 24 h/day for 7 consecutive days. Publicly available accelerometer wear-time data were scored by the National Center for Health Statistics used in the presented analyses. All participants with 7 complete days of recorded data, regardless of wear status, were included in the analyses (N = 15,585). The first and last partial days were removed. Results: Participants recorded 1247.4 ± 288.2 (Mean ± SD) minutes per day of wear time (86.6% ± 20.0%). For all participants, wear fatigue was 21.2 ± 59.7 minutes/day resulting in a significant decrease of 127.0 ± 358.2 minutes (p < 0.001) between day 1 (1298.0 ± 257.8 minutes or 90.1% ± 17.9%) and day 7 (1171.0 ± 414.9 minutes or 81.3% ± 28.8%, Figure 19). Wear fatigue varied temporally, with greater rates occurring at night, and by age group, with adolescents having the highest (26.7 ± 65.8 mins/day) and older adults having the lowest rates (12.3 ± 45.7 mins/day, Figure 1). Conclusion: Further research is needed to determine if the wear fatigue would be different in longer studies, interventional studies, or in other countries/cultures. To minimize wear fatigue, periodic reminders may be needed to maintain high wear compliance, particularly when sleep monitoring is important or for younger age groups and when the 24h activity cycle is a primary endpoint.

Technical Challenges and Considerations

Comparing ActiGraph CentrePoint Insight Watch, GT9X Link, and wGT3X-BT Accelerometers to NHANES 2011-2014 GT3X+ Devices Using an Orbital Shaker

Samuel LaMunion1, Joe Nguyen2, Robert Brychta1, Richard Troiano3, Karl Friedl4, Kong Chen1

1National Institute of Diabetes and Digestive and Kidney Diseases, 2ActiGraph, 3National Cancer Institute, 4United States Army Research Institute of Environmental Medicine

Background: With many researchers seeking to compare their accelerometry data to the NHANES dataset, it is important to establish equivalence between current ActiGraph device generations and the GT3X+. Doing so may help to characterize potential inter-generational differences in data stemming from hardware or firmware changes over time. Methods: ActiGraph devices (15 of each) wGT3X-BT, GT9X CentrePoint Insight Watch (CPIW), and original GT3X+ devices used in the 2011-2014 NHANES data collection were tested on a modified VWR benchtop orbital shaker between 0-250 RPM [∼0-3500 milli-g (mg)] in increments of 25 RPM. All devices were tested at 80 Hz except CPIW (64 Hz). Raw vector magnitude (VM) was evaluated as the primary outcome. Equivalence was evaluated two ways: 1) using a ±50 mg threshold and 2) using a ±5% equivalence zone based on the mean GT3X+ VM at each frequency suggested by ActiGraph. Devices are considered statistically equivalent if the 90% confidence intervals fall completely within the equivalence zone. Results: Using either the ±50 mg (Figure 1A) or ±5% prespecified equivalence zone (Figure 20), all devices were statistically equivalent throughout the range of accelerations except the GT9X and wGT3X-BT devices at the highest accelerations. However, the VMs of the newer devices were consistently below the GT3X+, except the CPIW at the highest accelerations (225 and 250 RPM or 3000 and 3600 mg). Conclusion: All ActiGraph generations were found to be statistically equivalent across a range of simulated accelerations to those observed in human studies. However, it is unknown if the consistent small disparities in raw accelerations between device generations detected by mechanical oscillator will lead to practical differences in week-long remote monitoring assessments of physical behavior.

Comparison of a Head-Worn Accelerometer to a Hip-Worn ActiGraph GT9X for Classifying Activity Type and Estimating Energy Expenditure

Edward Sazonov1, Samuel LaMunion2, Billal Hossain1, Scott Crouter3

1University of Alabama, 2National Institutes of Health, 3University of Tennessee

Introduction: The purpose of this study was to compare the head-mounted automated ingestion monitor (AIM) to a hip-worn ActiGraph GT9X (GT9X) for classifying activity type and estimating energy expenditure (EE). Methods: Adult participants (N=9, 8 males; mean±SD; 25.1±3.5 y, 179.3±7.4 cm, 81.7±18.1 kg) completed eight structured activities ranging from sedentary (e.g., seated computer work) to vigorous intensity (e.g., treadmill running at 6 MPH, 0% grade). Participants wore an AIM affixed to the right arm of a generic pair of eyeglasses, a GT9X on the right hip, and a Cosmed K5 was used as the criterion measure of EE. Both the AIM and GT9X collected accelerometer (128 Hz and 90 Hz, respectively) and gyroscope (128 Hz and 100 Hz, respectively) data. GT9X gyroscope data were down-sampled to 90 Hz and merged with the GT9X accelerometer data. Vector magnitude was calculated for the AIM and GT9X prior to being collapsed to 1-s means of each sensor signal and computing basic time domain features including: mean, standard deviation (SD), coefficient of variation (CV), variance, mean amplitude deviation (MAD), min, and max in rolling 10-s windows. Two series of random forest models were developed and cross-validated using a leave-one-participant-out-cross-validation procedure to A) classify activity type (7 participants) and B) estimate EE (6 participants) using 1) only accelerometer data, 2) only gyroscope data, and 3) accelerometer + gyroscope data combined. Model performance was reported as mean F1 (harmonic mean of precision and recall) for activity classification and root mean square error (RMSE) for the EE estimation. Results: For all three approaches, the GT9X achieved higher F1 scores by an average of 8.6%, while the AIM achieved a lower RMSE by an average of 0.43 METs. Conclusion: These results show proof-of-concept that a head-mounted device can be used to characterize physical activity parameters typically collected at other attachment sites such as the hip. (Figure 21)

Impact of Using a 60, 80, 90, or 100 Hz Versus 30 Hz ActiGraph Sampling Rate on Free-Living Physical Activity Assessment in Youth

Kimberly Clevenger1, Jan Brønd2, Kelly Mackintosh3, Karin Pfeiffer4, Alexander Montoye5, Melitta McNarry3

1National Cancer Institute, 2University of Southern Denmark, 3Swansea University, 4Michigan State University, 5Alma College

Using an ActiGraph sampling rate of more than 30 Hz has been shown to result in overestimation of activity counts in both children and adults but research on free-living individuals has not included the full range of sampling frequencies used by researchers. Objective: The present study compared count- and raw-acceleration-based metrics from free-living children and adolescents across a range of ActiGraph sampling frequencies. Methods: Children and adolescents (n=457; 10-15 y) wore an ActiGraph accelerometer over the right hip for at least one 8-h day. Vector Magnitude counts, Mean Amplitude Deviation, Monitor-Independent Movement Summary (MIMS) units, and activity intensity classified using six different methods (four cut-point based approaches, a two-regression model, and an artificial neural network) were compared between 30 Hz and 60, 80, 90, and 100 Hz sampling frequencies using mean absolute differences, correlations, and equivalence testing. Results: All outcomes were considered statistically equivalent, and correlation coefficients were ≥0.984. Absolute differences were largest for the 30 vs. 80 and 30 vs. 100 Hz count comparisons. For comparisons of 30 with 60, 80, 90, or 100 Hz, mean (and maximum) absolute differences in minutes of moderate-to-vigorous physical activity per day ranged from 0.04 to 0.13 (0.23 to 0.96), 0.13 to 1.00 (1.01 to 4.30), 0.09 to 0.18 (0.68 to 1.44), and 0.13 to 2.00 (0.99 to 9.90). At the epoch-level (per 5-sec), mean absolute percent differences between 30 and 60, 80, 90, or 100 Hz were highest for counts (2.9, 7.5, 4.0, 9.7%) and lowest for MIMS (0.9, 0.9, 0.9, 0.8%). Discussion: Our findings indicate that 60, 80, or 90 Hz sampling frequencies resulted in minimal differences in outcomes compared to the standard 30 Hz. Maximum individual differences and epoch-level differences in counts preclude us from recommending use of a 100 Hz sampling rate. Acceleration-based outcomes are comparable across the full range of sampling rates.

Interrelationships Between Open-Source, Proprietary, and Machine Learning-Derived Accelerometry Metrics

Christopher Moore1, Kelly Evenson1, Eric Shiroma2, Carmen Cuthbertson1, Julie Buring3, I-Min Lee3

1University of North Carolina, 2National Institute on Aging, 3Brigham and Women’s Hospital

Purpose: To evaluate correlations among various accelerometer metrics evaluated in the same data. Methods: In 2011-2015, the Women’s Health Study asked women to wear an ActiGraph GT3X+ on their hip for 7 days. After selecting 100 of these women (mean age, 71 years) and removing nonwear time and days with <10 wear hours, we computed physical behavior metrics at the 5-sec-, minute-, and day-level. Five-sec-level metrics included mean amplitude deviation (MAD), Euclidean norm minus one (ENMO), accelerometer activity index (AAI), and monitor-independent movement summary (MIMS). Minute-level metrics included 5-sec metrics averaged over 1 min and activity counts. Day-level metrics included daily averages of the 5-sec metrics, total daily counts, time in count-classified intensities (sedentary, light low, light high, and moderate-to-vigorous), and time in behaviors classified with the two-level behavioral classification (TLBC) algorithm (sitting, riding in a vehicle, standing still, standing/moving, and walking/running). For each combination of metrics within each data level, we estimated person-specific Spearman correlations and averaged them across the sample. Results: At all three data levels, AAI-MIMS correlations were ≥0.94 while MAD-ENMO correlations were ≥0.65 (Figure 22). Both MAD and ENMO had weaker correlations with AAI and MIMS at the 5-sec- and minute-level (0.21 to 0.46) than the day-level (0.59 to 0.88). Counts per minute and per day were more strongly correlated with AAI/MIMS (0.38 to 0.42) than MAD/ENMO (0.11 to 0.32). The strongest correlations among TLBC-classified behaviors were for sitting (-0.64 to -0.59 with MAD, AAI, and MIMS) and standing/moving (0.58 to 0.63 with AAI and MIMS). Conclusions: The stronger MAD-ENMO and AAI-MIMS correlations suggest that these pairs similarly characterize movement in older women. However, the divergence between MAD/ENMO, AAI/MIMS, and counts may be an important consideration when comparing or pooling data across studies.

Let the Epoch Length Float for More Reliable Measurements

Henri Vähä-Ypyä1, Ari Mänttäri1, Pauliina Husu1, Kari Tokola1, Harri Sievänen1, Tommi Vasanakari1

1The UKK Institute for Health Promotion Research

Introduction: Accelerometer measurements are typically based on equations developed using steadily-paced activities. However, habitual physical activities contain turning, braking, and acceleration, which increase the energy expenditure (EE) compared to constant speed locomotion. The purpose of the present study was to compare EE estimation models for different types of activities. Methods: The study comprised a constant speed test (CS) and a test containing acceleration and deceleration phases (AC/DC), during which VO2 (in MET) was measured with a portable gas analyzer and participants had a hip-worn tri-axial accelerometer (Hookie AM20). In CS, 29 adults performed a pace-conducted non-stop test on a 200m track. The initial speed was 0.6m/s and it was increased by 0.4m/s at every 2.5min. In AC/DC, 61 adults walked back and forth on a 15m track as fast as possible for 6min. The acceleration data was analyzed with 6s and floating epochs using the MAD algorithm. The floating epoch length was set by the stride rate containing a pair of steps. The EE equations were determined for CS with 6s epochs and for AC/DC with floating epochs. Results: The EE equation based on 6s epoch and MAD yielded R2=88%, bias=0.0MET, and SEE=1.1MET for CS, but R2=31%, bias -0.7MET, and SEE=1.4MET for AC/DC. The best predictors with floating epochs were MAD, absolute change in MAD between adjacent epochs, and inverse epoch length. The EE equation yielded R2=75%, bias=0.0MET and SEE=0.8MET for AC/DC, and R2=87%, bias=-0.4MET and SEE=1.3MET for CS. Discussion: The EE estimation equation developed with the sporadic activity performed adequately with steady activity but not vice versa. The floating epoch improved the EE estimation by detecting the variations between steps due to braking and acceleration. Habitual activities have more likely sporadic than steadily paced patterns and their assessment could be more reliable with methods based on sporadic activities.

Use of Devices in Children and Adolescents

Active and Sitting Time Precursors to Mood in Young Adults

Bronwyn Clark1, Elisabeth Winkler1, Marco Giurgiu2, Markus Reichert3, Eric Vanman1, Fiona Maccallum1

1University of Queensland, 2Karlsruhe Institute of Technology, 3Ruhr University Bochum

Objective: Physical behaviour and mood fluctuate through the day. We assessed recent and average sitting and stepping in relation to mood, collected via ecological momentary assessment. Methods: University students (n=57, age [mean±SD] 20±2 years, 83% female, BMI 21.7± 3.4 kg/m2) wore the Movisens Move 4 monitor on the thigh and reported (scale: 0-100) six mood items in response to mobile phone prompts (4.8±0.6 days; total 1472 responses). Associations of mood with each behaviour (% sedentary, % prolonged sedentary [≥30 min bout], and step count) on average, and over a prior 2 hour window (selected for lowest AIC), were tested in linear mixed models. Models included both recent and average behaviour and adjusted for time-varying (day of week; time of day) and fixed (sex; age; work hours; study hours; language; education; self-report BMI) covariates. Results: Above and beyond average behaviour, each additional 10% in the prior 2 hours (12 min) of prolonged sedentary/sedentary time were respectively associated with significantly (p<0.05) lower energy, alertness and wellness (≈-0.6 to -0.9 points/-0.9 to -1.3 points) and lower contentment, calmness and relaxation (≈-0.4 to -0.5, p<0.05/≈-0.3 to -0.4 points; p≥0.05). Conversely, each additional 1000 steps in the prior 2 hours was associated with significantly higher energy, alertness, and wellness (≈1.2-1.5 points; p<0.05) and non-significantly higher calmness (≈0.5 points). Average behaviour (accounting for recent behaviour) showed no significant associations, with effect sizes being sometimes similar, smaller, or in the reverse direction. Conclusions: Being more active and sitting less were significantly associated with better mood, with recent behaviour having an importance that is not explained by average levels. (Figure 23)

An Objective Assessment of Toddler Physical Activity Type and Context at the Childcare Center and Home

Cailyn Van Camp1, Darcy Thompson2, Karin Pfeiffer1

1Michigan State University, 2University of Colorado

Evidence suggests that certain opportunities1 within childcare centers or outdoors are associated with increases in PA in preschoolers. Evidence in toddlers, however, is lacking. One known study has evaluated the types of PA in toddlers; however, the context surrounding the PA was not described. Objective: To describe the PA in which toddlers participate by activity type and context in indoor and outdoor environments, in both a childcare and in-home setting. Methods: Toddlers (12-36 months) were recruited from a childcare center and from a registry of families with toddlers. For at least 30-minutes of free-play, cameras were used to video record the toddlers’ activity. Activity type and context were coded using Behavioral Observation Research Interactive Software (BORIS). This software allows for the use of continuous direct observation. The coding rules stated in the Observational System for Recording Physical Activity-Preschool (OSRAC-P) were followed. Results: Eleven toddlers participated in the study (7 males). In all settings, sit/squat, stand, and walk accounted for most activity (76.5-93.7%). Other activities in the childcare center, including push/pull outdoors (6.7%) and crawl in the gym (10.7%), and activities at home including lie down (5.1%), and ride outdoors (5.5%), accounted for >5% of activity type. Across each childcare center setting, the most observed activity contexts were sociodramatic props (29%), other (45.2%), and open space (41.3%). At home, the most observed were educational (33.4%) and wheel toys (24.6%). Conclusion: Toddlers spend most of their time in sit/squat, stand, and walk. This is comparable to previous findings reporting 81.7% of time spent in these activities.2 Compared to preschoolers', similarities are seen in outdoor activity contexts, but not indoors.3 To encourage PA in toddlers, a different approach is needed. PA should be observed further to provide insight into engaging toddlers in optimal levels of activity.

Comparison of Youth-Specific Cut-Point and Machine Learning Methods for Classifying Physical Activity Intensity From Wrist Accelerometer Data

Matthew Ahmadi1, Stewart Trost2

1University of Sydney, 2The University of Queensland

Background: No study has compared the performance of machine learning (ML) models and cut-points in an independent sample children. The absence of such comparisons has hindered the progress of processing methods and uptake of ML approaches. Further, the generalizability of ML models to different age groups has not been investigated. We compared the PA intensity classification of two youth cut-points (Hildrebrand, Phillips), a random forest (RF) classifier for youth, and an adult RF classifier in a sample of high school students. Methods: 22 children (age=15±0.7y) wore an ActiGraph accelerometer on their non-dominant wrist and an indirect calorimetry unit while performing the following activities: Sitting, standing, walking, active sports, and running. VO2 was converted to METs using the Schofield equation and classified as sedentary (≤1.5METs and sitting), light (>1.5 and <3.0METs), and MVPA (≥3METs). Performance was evaluated using overall accuracy and weighted Kappa statistics. Activity confusion matrices were generated to determine the proportion and direction of intensity misclassification. Results: The youth RF had the highest accuracy (89.0%[87.6%,90.3%]) followed by the adult RF (78.4%[77.0%,79.8%]) and Hildebrand cut-point (75.4%[73.9%,6.8%]). Kappa statistics indicated substantial agreement for the school-age RF classifier (0.75) and moderate agreement for the adult RF classifier (0.54), Hildebrand cut-point (0.57) and Phillips cut-points (0.46). The Hildebrand and Phillips cut-points consistently misclassified the intensity of walking (>80%). The youth RF misclassified the intensity of walking by <10%. Conclusion: A RF classifier trained in children outperformed cut-points for prediction of PA intensity from raw accelerometer data from the wrist. ML models trained in adults did not show good generalizability to a younger age-group.

Validating Youth Accelerometer Methods Using Direct Observation in Free-Living Settings

John Sirard1, Robert Marcotte1, Marcos Amalbert-Birriel1, John Chase1, Melanna Cox1, Nicholas Remillard1, Patty Freedson1, John Staudenmayer1

1University of Massachusetts Amherst

Several data processing methods exist for youth accelerometer data. Still, these methods have not been validated against a criterion measure of behavior, such as video-recorded direct observation of free-living behaviors. Purpose: To determine the accuracy of common methods for processing youth accelerometer data into physical activity intensity categories, using direct observation as the criterion method. Methods: N=42 13-17 year-old adolescents (mean+SD; age=15.2+ 1.5 years, 50% male) engaged in their usual behaviors during four 1-hour video-recorded sessions in different contexts to capture a range of behaviors and intensity levels. Participants were video-recorded during these sessions while wearing one accelerometer on the right hip and one on the non-dominant wrist, both sampling at 80hz. Videos were annotated for all changes in behaviors (postures and activity types) lasting ≥1 sec and age-adjusted MET values were assigned to determine sedentary behavior (SED), plus light, moderate, and vigorous physical activity (LPA, MPA, VPA, MVPA) for each epoch. Accelerometer data were processed using nine hip and three wrist methods, then synchronized with the video annotations. Bias (estimate - criterion) and 95%CI were calculated for minutes spent in each intensity category. Results: No method was unbiased for all intensities, and wide confidence intervals for VPA were observed due to less behavior at this intensity. The Crouter 2-regression method (2012) overestimated SED (3.3 min[1.4, 5.2]), produced unbiased estimates for LPA, MPA, and VPA (-2.4 to 0.6 min), and was almost unbiased for MVPA (-2.7 min[-5.7, -0.1]). The Crouter wrist methods (2015) produced unbiased estimates of MPA, VPA, and MVPA (-1.0 to 2.1 min). Conclusion: There was large heterogeneity in bias estimates across the methods and intensity categories. Additional research is needed to develop and validate accelerometer algorithms that are accurate and precise across all intensity categories.

Validation of Devices in Real World Settings

Comparison of Time Spent in Activity Type From the activPAL and Video-Recorded Direct Observation

Sarah Keadle1, Cami Christopher1, Alexander Tolas1, Shreya Patel2, Pedro Saint-Maurice1, Charles Matthews1

1California Polytechnic State University, 2National Cancer Institute

The activPAL (AP) is a thigh-worn accelerometer that has been validated for measuring sedentary time, standing and stepping. The newest software release for this device expanded the categories of activity types to include cycling, sedentary travel (i.e., driving), primary and secondary lying. Our objective was to compare AP estimates of activity types compared to video-recorded direct observation (DO). Participants (N=23) were aged 18-75y (mean =37y) and an average BMI of 25.9 kg/m2. Each participant completed two, 3-hour DO sessions while wearing an AP device on their right thigh. Videos were annotated by trained coders for behavior type and posture using Noldus Observer XT software. We calculated overall time and second-by-second comparisons between DO and AP. In total, 6,557minutes of simultaneous DO and AP data were available. Overall accuracy was 79.9%. Total time observed in each category and cross-classification metrics are shown in Figure 24. Accuracy was highest for sitting (89.6%) and lower for less commonly observed activities (cycling 70% and lying 66%). Lower accuracy for lying/cycling was driven by low sensitivity, indicating the AP correctly identified most cycling and lying time, but misclassified other activities as cycling/lying. DO estimate of standing were higher than AP, while the opposite was true for stepping. The DO coding protocol indicated standing still required 10 seconds without movement for DO while the AP may have a lower threshold. AP tended to identify more time as lying as compared to DO, which classified more time as sitting, but both are broadly considered sedentary behavior during the waking day, for which accuracy was 94.4%. These findings confirm the AP provides valid estimates of sedentary versus active time, and newly confirms estimates for sedentary travel are accurate. Estimates for lying and cycling were less accurate and should be interpreted with caution given the relatively low amount of observed time in those behaviors.

Cumulative and Diurnal Change in GPS-Derived Distance as a Novel Measure of Community Mobility in Older Adults

Kyle Moored1, Breanna Crane2, Michelle Carlson2, Andrea Rosso1

1University of Pittsburgh, 2Johns Hopkins University

Background: GPS technology can objectively measure subtle changes in community mobility, including how individuals accumulate distance across the day, that may be more sensitive to functional differences in later life. We derived a measure of distance accumulation and compared its functional associations with those of traditional summary GPS metrics. We also evaluated differences in hourly distance accumulation by functional status. Methods: Our sample consisted of 149 adults (Age: M=77.1±6.5, 67% women) from a randomized trial to improve walking in community-dwelling older adults. Participants carried one of two GPS devices at baseline that passively collected real-time location data for 5-7 days. Change in inter-point distance (m) was binned by hour and summed across the day to produce a cumulative distance measure. Traditional GPS metrics included standard deviational ellipse area (km2), percent time out-of-home, and maximum distance from home (km). Spearman correlations tested associations with physical (6-Minute Walking Test, 6MWT) and cognitive (Trail Making Test Part B, TMT-B) function. Mixed effects models (unstructured covariance) examined diurnal changes in distance stratified by median functional performance. Results: Compared to traditional GPS metrics, cumulative distance had stronger associations with both 6MWT (ρ=.31 vs. .22-.25) and TMT-B (ρ=.22 vs. .14-.17). Those in the upper (vs. lower) median of 6MWT had significantly greater changes in distance from 9am-12pm (B=477m, 95% CI: 178, 776, p=.002) and from 12pm-3pm (B=667m, 95% CI: 206, 1128, p=.005), adjusted for age, sex, device, and season. Discussion: Cumulative distance measures may capture how function relates to later-life community mobility with greater sensitivity than commonly used GPS metrics. Differences by functional status appeared to be greatest in the later morning and early afternoon (9am-3pm). Future work will examine whether accumulated distance predicts incident health changes. (Figure 25)

The Acceptability of Wearing an Activity Monitor (activPAL) on the Thigh to Older Adults

Philippa Dahl1, Pedro dos Santos1, Sebastien Chastin1, Simon Cox2, Ian Deary2, Mary-Kate Hannah3, Dawn Skelton1

1Glasgow Caledonian University, 2University of Edinburgh, 3University of Glasgow

Introduction: Adherence to wearing body-worn sensors measuring physical behaviours may vary by wear location, but aspects of acceptability are rarely considered. Methods: Older adults from longitudinal cohort studies (Lothian Birth Cohort 1936; West of Scotland Twenty07 Study) wore an activPAL monitor for 7 days. The monitor was waterproofed (layflat plastic tubing), attached to front of thigh using a hypoallergenic gel pad (PALstickie), and covered with waterproof dressing (Opsite Flexifix). On monitor removal, participants were asked questions about acceptability (see Figure 26; yes/no response plus free text comments): did it go well? were there issues? was there discomfort? was the monitor removed early? and researchers removing the monitor were asked about signs of skin irritation. Answers to questions were compared between those with (7days) and without complete data, using Fisher’s exact test. Results: Participants who returned the monitor (n=771, 44% of those approached) were 52% female and aged ∼64 (n=340), ∼79 (n=300) or ∼84 (n=129), with no differences between those with (n=700) and without complete data. In the whole group, most people felt wearing the monitor went well (95%; table), reported no issues (87%), reported no discomfort (88%), had no skin irritation on inspection (85%), and did not remove the monitor early (83%). Participants without complete data (9%) were significantly more likely to feel things did not go well (p<0.001), have had issues (p<0.001), report discomfort (p=0.008), and remove the monitor early (p<0.001). However, levels of skin irritation observed by researchers on removal were similar (p=0.843), perhaps due to high levels of early removal. Comments suggested that most skin irritation was mild (e.g. “slight redness”). Discussion: In this group of older adults, acceptability of the activPAL monitor was very high. Acceptability was lower in those with data excluded from analysis in the main study, which may introduce confounding.

Validation of Previous-Day Recall for Estimates of Duration and Context in Comparison to activPAL and Direct Observation

Charles Matthews1, David Berrigan1, Pedro Saint-Maurice1, Cami Christopher2, Jeffrey Huang2, Joshua Freeman1, Shreya Patel1, Sarah Keadle2

1National Cancer Institute, 2California Polytenchnic State University

Activities: Completed over Time in 24-hours (ACT24) is an internet-based previous-day recall instrument that provides valid estimates of active and sedentary time. The participant interface was recently updated, which may influence instrument validity. ACT24 estimates of time spent in specific behavioral domains have never been evaluated. Our objective was to validate the new ACT24 for active and sedentary time as well as estimates of time spent in domain-specific behaviors. We enrolled 47 adults (40 years; 63% women) who wore an activPAL for 7-days and during this period, completed up to three ACT24 recalls. A subsample (n=23) allowed video-recorded direct observation of daily living (two, 3-hour sessions). Participants reported means of 9.1 hrs/d sedentary time and 6.5 hrs/d physically active time via ACT24, which were not significantly different from activPAL values of 9.3 and 6.3 hrs/d, respectively. Spearman correlations between measures were 0.61 for sedentary and 0.65 for active time, and there was no evidence of systematic bias in Bland-Altman analysis. Correlations between ACT24 and direct observation were 0.73 and 0.76 for sedentary and active time, respectively. Classification accuracy of total time in behavioral domains via ACT24 compared to direct observation ranged from 62% to 90%. Kappa (K) values were substantial (K=0.61-0.80) for work and shopping/errands; moderate (K=0.41-0.60) for leisure-time and transportation; and fair (K=0.21-0.40) for personal care and household activities. Results for domain-specific estimates stratified by sedentary and active time were similar. Overall, the updated ACT24 provided accurate group-level estimates of total sedentary and active time with no evidence of systematic bias in comparison to activPAL. ACT24 also provided useful domain-specific estimates of time use, particularly the more commonly observed domains, which included work, shopping, leisure-time, and transportation-related behaviors.

Validation of Two Deep Learning Methods to Estimate Aspects of Physical Activity/Inactivity From Accelerometers

John Staudemayer1, John Sirard1, Robert Marcotte1, Evan Ray1, Tom Cook1, Yujian Wu1

1University of Massachusetts Amherst

Methods: Participants (n = 57, mean age 20.3 ± 1.4 years, BMI 23.4 ± 3.9) participated in four 1-hour sessions that took place in four environments: home, school/learning, community, and exercise. Participants wore ActiGraph GT3X-BT accelerometers on the right hip and non-dominant wrist. Research assistants video recorded participants using a GoPro camera as they went about their free-living activities. Video data were then annotated using a direct observation (DO) annotation protocol (Cox et al, 2020) to denote exactly when participants’ MET levels, postures, and activities changed. All DO labels and transitions were synchronized with the ActiGgraph time stamps. The waist and wrist accelerometer data were processed separately using two published convolutional neural network (CNN) models to estimate sitting time (Greenwood-Hickman et al, 2021, HIP-CNN) and behavior relative intensity (sedentary, light, and moderate to vigorous, Nawaratne et al, 2020, WRIST-CNN), respectively. Results: The DO and HIP-CNN classifications of sitting agreed 91.7% of the time, and the DO and WRIST-CNN classifications of intensity categories agreed 66.7% of the time, across all categories. Both of those performance metrics (and others) were similar to the cross-validated estimates published in the CNN papers. Conclusions: Independent validation shows that the HIP-CNN method for classifying sitting events and the WRIST-CNN method for classifying intensity category performs very to moderately well respectively. Both methods should improve if given more training data, and this is a promising area of future research.

Poster Presentations

Applications

24-Hour Compositions of Physical (In)activity Among Office Workers During the COVID-19 Pandemic: A Comparison Between Brazil and Sweden

Luiz Augusto Brusaca1, Leticia Bergamin Januario2, Svend Erik Mathiassen2, Dechristian França Barbieri3, Rafaela Veiga Oliveira1, Marina Heiden2, Ana Beatriz Oliveira1, David M. Hallman2

1Federal University of São Carlos, 2University of Gävle, 3Clemson University

Background: Sedentary behavior (SED) has generally increased during the COVID-19 pandemic in people working from home, and physical activity (PA) has therefore decreased. However, it is unclear whether the pandemic has affected office workers in different countries in the same way. We aimed to compare the 24-hour time-use compositions of physical behaviors between Brazilian and Swedish office workers at working and non-working days during the pandemic. Methods: Physical behaviors were monitored over 7 days using thigh-worn accelerometer in 73 Brazilian and 202 Swedish workers. Daily time-use compositions were exhaustively described in terms of SED in short (<30min) and long (≥30min) bouts, light PA (LPA), moderate-to-vigorous PA (MVPA), and time-in-bed (TIB). Following a compositional data analysis, isometric log-ratios (ilr) were calculated to express the ratio of TIB to time spent awake, SED (short and long bouts) relative to LPA and MVPA, SED in short relative to long bouts, and LPA relative to MVPA. We examined differences between countries using MANOVA, followed by univariate post-hoc tests of pairwise differences. Results: Both groups spent most of their time SED and in bed. On working days, Brazilian workers spent 294 min in SED in short bouts, 477 min in SED in long bouts, 157 min in LPA, 50 min in MVPA and 461 min TIB; Swedes spent 274, 365, 257, 86 and 458 min, respectively. During non-working days, results were 279, 359, 237, 61 and 504 min among Brazilians and 263, 251, 305, 93 and 529 min among Swedes. Brazilians and Swedes differed significantly in the set of ilrs as a whole during working (p<0.001) and non-working days (p<0.001) and in all pairwise comparisons, except for the ratio of TIB to time spent awake during non-working days. Conclusions: During the COVID-19 pandemic Brazilian and Swedish office workers behaved differently. Whether this relates to restrictions being different or to differences even before the pandemic is not clear.

A 12-Week Cycling Workstation Intervention Improves Cardiometabolic Risk Factors in Healthy Office Workers: The REMOVE Study

Terry Guirado1, Lore Metz1, Bruno Pereira2, Carole Brun1, Anthony Birat1, Audrey Boscaro1, David Thivel1, Martine Duclos2

1University Clermont Auvergne, 2Clermont-Ferrand University

Objective: The present study aimed to evaluate the effects in a randomized multicentric study of a cycling workstation intervention (60-min per working day) for 12 weeks among healthy tertiary employees on body composition, biological parameters, physical fitness, physical activity (PA) and sedentary time (ST). We hypothesized that 60-min of a cycling workstation per working day during 12 weeks will decrease sedentary behaviors (SB) during weekdays and improve cardiometabolic parameters. Methods: Forty healthy office workers have been recruited from six tertiary worksites in Clermont-Ferrand, France. Thirty-two participants completed the study. Subjects has been randomly assigned to either: i) an intervention group (INT) who performed 60-min of cycling workstation per working day (n=17; 44.9±8.6 years; 23.7±3.5 kg/m2), or ii) a control group with no intervention (CTRL) (n=15; 43.7± 9.7 years; 23.3± 3 kg/m2). Every following outcome was assessed at baseline (T0) and 3 months (T1): 7-days PA and SB (3-D accelerometers-inclinometer), body composition (bioelectrical impedance), physical fitness (aerobic fitness, upper and lower limb strength) and metabolic outcomes (glycemia, insulinemia, lipids profile and inflammatory cytokines). Results: Intervention enabled to reduce significantly waist to height ratio (-0.01± 0.04, p≤0.05) and improve aerobic fitness during a submaximal exercise (157.3±19.2 vs 153.1±18.5 bpm, p=0.012). INT group improved significantly Δus-CRP (-0.11±1.3, p=0.008), Δtotal cholesterol (-0.06±0.1, p=0.028) and ΔLDL cholesterol (-0.09±0.3, p=0.048) compared to CTRL group. The use of cycling workstation during weekdays, reduced significantly ST (63.2±6.9 vs 59.7±7.7 %, p≤0.01) and increase both light intensity (201.2±56 vs 227.1±54.2 min/day, p≤0.01) and moderate-to-vigorous (22.9±12.1 vs 31.9±20.4 min/day, p≤0.01) PA during the 3-month intervention. Conclusion: Implementation of PPM in worksite among normal-weight individuals has shown several positive effects on PA, SB, anthropometric measurement, physical fitness and cardiometabolic parameters. The REMOVE study emphasizes that this strategy is efficient in primary prevention to attenuate the deleterious effects of SB and physical inactivity without generating compensation during weekend days.

A Comparison of Methods for Analyzing Wrist Worn Actigraph Data Among Older Adults With Dementia

Karl Brown1, John Ostrander1, Sarah Payne1, Adrienne Jankowski1, Andrew Shutes-David1, Katie Wilson1, Edmund Seto1, Debby Tsuang1

1Seattle Institute for Biomedical and Clinical Research

Background: Actigraphy devices have the potential to generate data in between clinic visits to help improve the measurement of physical activity in participants with dementias. Objectives: We sought to identify optimal actigraphy analytic methods and to compare activity levels in individuals with dementia with Lewy bodies (DLB) and Alzheimer’s disease (AD). Methods: Participants with DLB (n=5) and AD (n=4) wore actigraphy devices on their nondominant wrists for 2 weeks. Activity data were processed and scored using a variety of activity cut points with Actilife (the ActiGraph analytic platform) and the GGIR package (version 2.5-0) of R (version 1.4.1717). Step counts and time spent in moderate to vigorous physical activity (MVPA) were compared across disease groups and scoring methods. Results: Mean (SD) percent time spent in MVPA across all participants ranged from 3.97 (3.73) using Troiano cut points to 24.5 (8.87) using Nero Parkinson’s cut points. Mean (SD) step count for the entire wear period across all individuals was 109,261 (49,048) when scored by Actilife and 67,811 (41,953) when scored by GGIR. Comparing the average number of minutes in MVPA for DLB individuals vs. AD individuals, Freedson cut points suggest DLB individuals were more active (mean [SD] = 984.7 [1140.2] for DLB and 772.6 [217.9] for AD) whereas Nero Parkinson’s cut points suggest AD individuals were more active (mean [SD] = 3,266.6 [2,225.1] for DLB and 3,956.9 [640.3] for AD). Conclusion: Our findings suggest that the standard analytic software and cut points generate inconsistent values and overestimate key outcome measures related to actigraphy in older adults with dementia. New cut points are thus necessary that can more accurately and consistently score wrist worn actigraphy data in older adults with dementia. (Figure 27)

A Data-Informed Approach to Choosing Measures of Prolonged Sedentary Accumulation Patterns

Elisabeth Winkler1, Paddy Dempsey2, Bronwyn Clark1, David Dunstan3, Genevieve Healy1

1University of Queensland, 2University of Cambridge, 3Baker Heart and Diabetes Institute

Objective: Patterns of accumulating sedentary behaviour in long bouts are associated with mortality, chronic disease risk, and biomarkers of cardiometabolic health. Typically, this is tested via many indicators, or few indicators and limited rationale for selection. This study presents a data-informed approach to choosing a small set of indicators that is adequately comprehensive but minimises multiple testing. Methods: ActiGraph GT3X+ and activPAL data collected in the AusDiab study 2012 were processed via diary-based and automated methods (n=684, 55% female, age 36-89 y with ≥4 valid days and biomarker data). Nine indicators of the sedentary bout-duration distribution were tested in relation to each other (Spearman’s correlations; rs) and clustered cardiometabolic risk (CMR) scores (regression models with restricted cubic splines; R2). Results: Pattern indicators were non-linearly interrelated. Redundant measures with perfect or near-perfect correlations were: alpha/geometric mean; arithmetic mean/fragmentation index (FI); and, usual bout duration (UBD)/% prolonged (% sitting in ≥30 min bouts). Alpha/geometric mean showed |rs|≥0.8 (all devices) with all indicators except average hazard, gini, and UBD/% prolonged. Arithmetic mean/FI showed |rs|≥0.8 with indicators except gini and median. Average hazard showed |rs|≥0.8 with all indicators except alpha/geometric mean, gini and median. The indicators explaining most variation in CMR (Model R2) were alpha/geometric mean, plus median (activPAL) or arithmetic mean (GT3X+). Conclusion: Possible sets of pattern indicators to minimise multiple testing while omitting only indicators correlated at |rs|≥0.8 include: alpha/geometric mean + gini + UBD /% prolonged or average hazard and arithmetic mean/FI + median + gini. Sets with alpha/geometric mean may have advantages in detecting relationships of interest, based on R2 with CMR. (Figure 28)

A Systematic Scoping Review on the Application of Latent Class Analysis Applied to Accelerometry-Assessed Physical Activity and Sedentary Behavior

Michael Kebede1, Annie Howard1, Yumeng Ren1, Chongzhi Di2, Melissa Troester1, Blake Anuskiewicz3, Kelly Evenson1

1University of North Carolina, 2Fred Hutchinson Cancer Research Center, 3University of California

Background: While latent class analysis (LCA) has been increasingly used to identify heterogeneous groups in a population, its application to accelerometry has not been systematically explored. We conducted a scoping review summarizing the use of LCA applied to accelerometry. Methods: Comprehensive searches included studies published through 2021 in PubMed, Web of Science, CINHAL, SPORTDiscus, and Embase. Using Covidence, two researchers independently evaluated inclusion criteria based on title/abstract and full text. Studies with accelerometry or combined accelerometry/self-reported measures were selected for the review. Data were extracted based on study characteristics, accelerometer use, and application of LCA. Results: Of 2555 articles found, 66 full-text articles were screened, and 12 articles from 8 unique studies were included in the review (11 cross sectional and 1 cohort). Study sample sizes ranged from 217- 7931 with a mean 2248.9 (standard deviation 2780.4). The latent class variables included measures of physical activity (100%) and sedentary behavior (75%). When developing the classes, 63% of studies used accelerometry only and 38% used both accelerometry and self-report. The accelerometer wear protocol included “waking hours” (63%) or a “24-hour protocol” (25%); 13% of studies did not report this information. The physical activity and sedentary behavior variables were explored in the LCA daily over 7 days (63%), weekday and weekend (25%), daily variation (13%), and hour-by-hour (13%). Criteria to guide selection of the final number of classes and model fit varied, with the most common being Bayesian Information Criterion (63%), sample size of classes (50%), interpretability (50%), entropy (50%), Akaike Information Criterion (50%), and Bootstrap Likelihood Ratio Test (38%). Conclusions: This review spotlights methods used in applying LCA to accelerometry and identifies areas of difference in order to facilitate future work.

Accelerometer-Measured Physical Behavior as an Indicator of Perceived Work Ability

Pauliina Husu1, Kari Tokola1, Henri Vähä-Ypyä1, Harri Sievänen1, Tommi Vasankari1

1The UKK Institute for Health Promotion Research

Introduction: Perceived work ability (WA) reflects a combination of an individual’s resources, demands of the work, and related environment. For example, individual health-related behaviors may affect WA. Previous studies have shown that higher physical activity (PA) is associated with good WA. The purpose of the present study was to investigate associations between 24/7 physical behavior and WA. Methods: In the FinFit 2017 -population-based study physical behavior of 20-69-year-old working Finns was measured in terms of PA, standing, and sedentary behavior (lying, reclining, sitting) using MAD-APE-algorithms. During waking hours, a tri-axial accelerometer was worn on an elastic belt on the hip. During the time in bed (TIB), the device was worn on a wristband on the non-dominant wrist. TIB categories were based on wrist movements. WA was assessed by the sum of four questions excerpted from the Work ability index. Statistical analyses were conducted by GLM adjusted for age, sex, education, and physical demands of the work. Results: Participants (n=1668, mean age 46.6, SD=10.9, 57% women) scored on average 23.3 on the summary score of WA (range 6-27), a higher score indicating better WA. They spent on average 9 hours per day sedentary, 2 hours standing, 3.9 hours in light PA, and 49 minutes in moderate-to-vigorous PA. TIB covered 8.3 hours per day, 17% of which was categorized as high-movement time. More minutes in standing (p=0.001) and in moderate (p=0.004) and vigorous PA (p<0.001) were associated with a higher summary score of WA. Increased time spent lying down during waking hours (p<0.001) and in high-movement- (p<0.001) and total TIB (p=0.001) was associated with a smaller summary score. Discussion: Higher accelerometer-measured PA and standing indicated better perceived WA, whereas higher amounts of lying down and TIB were associated with poorer WA. Detailed analysis of 24/7 physical behavior can be utilized in identifying individual-related indicators of WA.

Accuracy of Consumer Grade Wearable Activity Monitors for Step Count and Heart Rate Recovery following Aerobic Exercise

Lindsay Toth1, Cristal Benitez1, Andrew Gomez1

1University of North Florida

Objective: To determine the step count and heart rate error from consumer grade activity monitors following exercise on aerobic fitness equipment at moderate (mod) and vigorous (vig) intensities. Methods: Participants (n=14, 23.8±3.9 yrs) completed the protocol over 2 days. They wore a Polar T31 heart rate (HR) transmitter (criterion HR) on the chest and 2 to 3 randomly assigned wrist-worn activity monitors (maximum of two monitors/wrist; Apple Watch Series 5, Fitbit (FB) Sense, FB Inspire HR, Garmin (GM) Vivosmart 4, FB Charge 4). Participants completed 3-min of mod and 3-min of vig exercise separated by at least 60-sec, on 6 aerobic exercise machines (step-like (SL) e.g., elliptical; non-step-like (NSL) e.g., row). HR was recorded immediately following exercise and after 60-sec of rest (heart rate recovery, HRR) for all bouts. Steps were recorded from the monitor screens pre and post exercise; criterion steps were hand counted from videos of the lower body. HR and SL data were converted to pct criterion and compared to 100%. Steps accumulated during NSL activities were compared to 0 (0 steps were taken during NSL activities). Results: All monitors significantly underestimated steps by 24.3-61.5% during SL activities (P<.001). During NSL activities all monitors significantly overestimated steps at mod (36-48 step/min) and vig (49-64 step/min) intensities (P<.001). Immediately post-exercise, HR was underestimated by 3.5%-28.8% (P<.03). Most monitors underestimated HRR for both conditions (0.9%-2.3%, P≥.05). HRR following mod exercise was significantly overestimated by FB Sense (2.5%, P=.008) and underestimated by GM Vivosmart 4 (8.7%, P<.001). Following vig exercise, HRR was significantly underestimated by GM Vivosmart 4 (13.4%, P<.001) and FB Charge 4 (4.1%, P=.03). Conclusion: Low HR error is promising for tracking intensity and recovery with monitors, but users tracking steps/day should be aware of error associated with aerobic fitness equipment.

Agreement Among ActiGraph, activPAL, and Diary Measured Time in Bed in University Students

Michael Schmidt1, Benjamin Boudreaux1, Ginny Frederick2, Patrick O’Connor1, Ellen Evans1

1University of Georgia, 2Mercer University

Objective measures, such as time in bed (TIB), have potential advantages over self-reports in characterizing the 24-hour activity cycle. However, the extent to which TIB estimates may differ by measurement method has not been fully explored. Purpose: To compare TIB estimated from the ActiGraph GT9X, the activPAL3, and diaries in university students. Methods: Students at a large U.S. university (n=93, aged 18-28 years, 61% female) continuously wore an ActiGraph GT9X on their non-dominant wrist and an activPAL3 on their dominant thigh while completing a daily diary for at least 3 days. Sleep timing and TIB were identified using the ActiLife software automated sleep detection feature (GT9X data), the PALanalysis software via the CREA algorithm (activPAL3 data), and the self-reported time getting into and out of bed. Results: Estimated TIB (mean ± SD) was highest for the activPAL (8.86±1.04 hr) and lowest for the GT9X (7.31±1.03 hr). Compared to the diary (8.24±0.95 hr), the mean absolute error (MAE) was lower for the activPAL (MAE=0.76 hr; 95% CI 0.65-0.87 hr) than for the GT9X (MAE=1.17 hr; 95% CI 0.99-1.35 hr). Pearson correlations were also higher for the activPAL vs. diary (r=0.76; 95% CI 0.65-0.83) than for the GT9X vs. diary (r=0.34; 95% CI 0.15-0.51). Compared to the diary, time getting into bed was, on average, 17.9 min (95% CI 11.7-24.2 min) earlier for the activPAL and 22.9 min (95% CI 14.0-31.9 min) later for the GT9X. Compared to the diary, time out of bed was 18.1 min (95% CI 11.9-24.3 min) later for the activPAL and 34.1 min (95% CI 24.5-43.8 min) earlier for the GT9X. Conclusions: TIB estimates from the activPAL more closely approximate self-reported estimates than do ActiGraph GT9X estimates derived using the ActiLife automated sleep detection feature. ActiGraph and activPAL estimates of TIB may differ substantially and lead, in turn, to differences in 24-hour activity cycle characteristics.

Association Between Accelerometer-Measured Physical Activity and Motor Skills in Preschool-Aged Finnish Children

Janne Kulmala1, Tuomas Kukko2, Harto Hakonen1, Anette Mehtälä1, Piritta Asunta1, Arja Sääkslahti3, Tuija Tammelin1

1Jamk University of Applied Sciences, 2Liikunnan ja kansanterveyden edistämissäätiö LIKES, 3University of Jyväskylä

Objective: This study aims to provide knowledge of the association between accelerometer-measured physical activity and motor skills among preschool-aged boys and girls. Methods: The study population consisted of 613 children (54 % girls) aged 4-6 years (5.61 ± 0.85 years) who participated. PA was assessed for seven consecutive days using a wrist-worn accelerometer (Axivity AX3, 100 Hz, ± 8 g). Three concurrent response variables describing the quantity and intensity of PA were calculated: daily amount of moderate to vigorous physical activity (MVPA), average acceleration, and intensity gradient. Three tests representing balance, locomotor, and manipulative skills (throwing-catching combination, two-legged jumps from side- to-side and standing long jump) were chosen to form a standardized composite variable for motor skills. Boys and girls were divided into three groups (low/intermediate/high) based on their motor skills scores. The PA of children belonging to the three groups was compared using Student’s t-test. Results: Children with low motor skill scores had significantly lower levels and intensities of PA compared to children with intermediate or high motor skill scores (boys: int. vs. low p <0.05, high vs. low p < 0.001; girls: int. vs. low p < 0.001, high vs. low p <0.001). The results were similar for different PA variables: MVPA (min/day), average acceleration (mg) and intensity gradient. Conclusions: Based on this study, children with lower motor skill scores are less active and have weaker intensity profiles than their peers. Children with lower motor skill scores may need support to enhance their participation in daily activities that include a sufficient amount and intensity of PA. (Figure 29)

Can Wearable Sensors Provide Insight Into Delirium in Inpatients With Parkinson’s Disease? A Feasibility Study

Gemma Bate1, Sarah Richardson1, John-Paul Taylor1, David Burn1, Louise Allan2, Alison Yarnall1, Yu Guan1, Silvia Del Din1, Rachael Lawson1

1Newcastle University, 2University of Exeter

Background: Delirium is an acute neuropsychiatric syndrome characterised by altered levels of consciousness, impaired cognition and inattention. Patients with Parkinson’s disease(PD) are at increased risk of delirium but delirium may be missed due to overlapping symptoms. Sleep-wake cycle disruptions and motor-activity alterations are core features of delirium and may guide identification. Wearable sensors provide a unique opportunity to continuously and objectively monitor activity and sleep patterns. For the first time, this study aimed to assess the feasibility of using wearable sensors in PD inpatients with/without delirium and to quantify clinically relevant digital outcomes. Methods: The study is nested within DELIRIUM-PD, a study aiming to establish optimal criteria to identify delirium in PD. In parallel to the longitudinal delirium assessments, all eligible inpatients have the option to wear sensors on their lower back and wrist for 7 days to monitor movement and sleep patterns. The location and duration the sensor was worn was recorded, along with reasons for removal and tolerability. A participant compliance rating was collated (0-non-compliant to 10-fully compliant). Walking activity outcomes and sleep episodes(night awakenings and rest-activity measures) will be derived. Results: Data collection is ongoing. Twenty-seven patients were recruited comprising 46 separate admissions (Figure 30). The mean age of the participants is 76.7±10.5 years. Mean duration for wearing the sensors was 3.5±2.3 days. Of the 46 admissions (39.1%) wore both lumbar and wrist sensor, (58.7%) wore the wrist sensor only. Mean compliance rating for the lumbar sensor is 8.8±2.2 and for the wrist sensor 8.5±2.9. The sensor was removed without re-securement during 5 admissions. Conclusion: This objective approach appears feasible and future work will determine if digital outcomes of activity and sleep may be useful as complementary tools for diagnosis of delirium in inpatients with PD.

Characterizing Free-Living Physical Behaviors in Chronic Post-Stroke Adults With Aphasia

Albert Mendoza1, Jennifer Sherwood1, Michelle Gravier1

1California State University

Introduction: Stroke survivors often live with high disability which limits physical activity (PA). Approximately one in four stroke-survivors also have aphasia, an acquired communication disorder. Although an estimated 2.5 million people in the U.S. have aphasia, limited research exists on PA behaviors of chronic post-stroke adults with aphasia (AWA). The present study aims to characterize free-living PA behaviors in chronic post-stroke AWA participating in a virtual, aphasia-friendly exercise class. Methods: Six participants (two female); age mean (SD): 68.6 (8) years.; Time post stroke: 145.5 (112) months; were recruited and assessed for cognitive, linguistic, and PA measures. Average Western Aphasia Battery score at entry 44.75 ± 23.5 [range: 7 (very severe) to 74.8 (moderate)]. Free-living PA data was collected using activPAL 4+ (thigh-worn on most active leg) for 7-consecutive days at baseline. An average of 6.8 valid wear days (≥4 days with ≥10 hrs. of wear) were examined. Results: Average daily wear time was 1440 minutes. On average, AWA accumulate fewer daily steps, spend less time stepping and have fewer sit-to-stand transitions, while standing and sitting (including bouts ≥30-minutes) more as compared to post-stroke adults (Figure 31). Conclusions: AWA may face additional barriers to engaging in PA due to their communication disability and may benefit from aphasia-friendly interventions to: (1) increase PA behaviors, such as daily steps, stepping, sit-to-stands and standing, (2) decrease sedentary behaviors (SB), such as, daily sitting (including bouts ≥30-minutes). Future research may also examine the relationship between PA and SB in AWA.

Comparability of Free-Living Physical Activity Classified Using Counts, Open-Source Counts, Euclidean Norm Minus One, and Mean Amplitude Deviation in Adults

Katherine McKee1, Karin Pfeiffer1, Amber Pearson1, Kimberly Clevenger2

1Michigan State University, 2National Cancer Institute

ActiGraph accelerometers are often used to collect free-living physical activity (PA) data. Lack of uniformity in selection of metrics to obtain time spent in PA intensities resulted in limited comparability across studies. To compare PA estimates from four accelerometer metrics: mean amplitude deviation (MAD), Euclidean norm minus one (ENMO), vector magnitude counts (VMc), and open-source vector magnitude counts (VMo) for characterizing free-living PA of adults. Participants (n=154, age=54.7) wore an ActiGraph wGT3X-BT at the right hip for one week (one 10-h wear-day required for inclusion). Metrics (MAD, ENMO, VMc, VMo) were calculated in 60-s epochs and time spent in sedentary/light, moderate, vigorous, or moderate-to-vigorous (MVPA) intensities were categorized using metric-associated cut points. Metrics and time spent in each activity intensity were compared using Pearson’s correlations, weighted kappa, mean absolute differences, and two one-sided equivalence testing. On average, participants spent 3.3% of time in MVPA. Epoch-level correlations (r) between metrics ranged from 0.40 (MAD/VMc) to 0.94 (VMc/VMo). Weighted kappa (k) for intensity classifications ranged from 0.38 (VMo/MAD) to 0.90 (VMo/VMc). Mean absolute differences in percent of time spent in MVPA were 0.4±0.3 (VMo/VMc), 1.2±3.7 (MAD/ENMO), 1.8±2.6 (VMc/MAD), 1.8±2.9 (VMc/ENMO), 1.9±2.9 (VMo/ENMO), and 2.0±2.9 (VMo/MAD). Time spent in sedentary/light and vigorous activity were equivalent across all metrics, but only VMo and VMc classifications were equivalent across all intensities. There was lack of statistical equivalence for MVPA as classified using VMc or VMo versus MAD or ENMO with mean absolute differences that translated to 7.2 - 2-mins per 10-h wear-day. Use of different metrics and activity intensity classification approaches results in differing estimates of free-living PA. However, open-source counts can be used interchangeably with counts which may be an avenue for harmonization.

Comparison of Self-Reported and Accelerometer Measured Daily Sitting Time in Cancer Survivors

Emma Gomes1, Mary Hidde2, Kate Lyden3, Heather Leach1

1Colorado State University, 2Medical College of Wisconsin, 3KAL Consulting LLC

Study’s Objective: Sitting is a concerning health behavior for cancer survivors (CS), with higher amounts of sitting time associated with poorer health and cancer-related outcomes. The aims of this study are 1) to compare self-reported vs accelerometer-measured sitting time, and 2) explore differences between measurements based on how sitting time is accumulated (i.e., bouts of 1-2 hours.). Methods: CS who had completed active cancer treatments (N=56; Mage=51.75±12.98) self-reported sitting time over the last 7-days using the International Physical Activity Questionnaire-Short (IPAQ) and wore an activPAL accelerometer, 7-days, 24 hours per day. Absolute and relative agreement between the activPAL (criterion) and IPAQ for average daily sitting time was compared using two-way mixed model (single class) interclass correlation coefficient (ICC) and Pearson correlations. Results: Sitting time was M= 493.39±246.45 mins/day and M=680.84±111.47 mins/day as measured by the IPAQ and activPAL, respectively. ICC was low [ICC=.064, 95% CI = -.111-.263], and Pearson’s correlation showed no linear relationship between the two measurements [r = .071, p=.602]. Among CS (n=18, 32%) who accumulated ≥25% of their daily sitting time in bouts of 1-2 hours, correlations were higher [ICC=.183, 95% CI = -.224-.567, r =.345, p=.161] compared to participants who did not [ICC=.006, 95% CI = -.158-.218, r = .012, p=.943] but still not significant. Conclusion: Self-reported sitting time in CS does not agree with an accelerometer-derived measure of sitting time, regardless of how the sitting time is accumulated. Reliable and valid measures of sitting time are important to understand how sedentary behaviors impact cancer-relevant outcomes (e.g., fatigue, physical function, and quality of life), and to determine where/when behavioral interventions might be necessary. Additional studies are needed to further explore how sitting patterns impact CS ability to accurately recall sitting time.

Context-Matched Gait Variability Measures Capture Longitudinal Change in Real Life Walking in Degenerative Cerebellar Ataxia

Jens Seemann1, Winfried Ilg1, Martin Giese1, Matthis Synofzik1

1University of Tübingen

Background: Measures of gait variability and body sway have proven sensitivity to ataxia mostly in cross-sectional studies by correlating measures from lab-based assessments with clinical scores. However, to serve as ecologically valid biomarkers in intervention studies for degenerative cerebellar ataxia (DCA), gait measures have to prove (i) sensitivity to short-term longitudinal change and (ii) the capability to quantify ataxia in lab-based assessments and real-life walking. This study aimed to unravel performance outcomes in real-life walking which are sensitive for longitudinal change in DCA. Method: Longitudinal study in subjects with DCA (N=26). Gait was assessed by 3 body-worn inertial sensors (APDM) attached to the feet and the lower back in (a) laboratory assessment and (b) real-life monitoring. Out of the real-life data we extracted walking bouts of length with 20-50 strides. Since gait measures are influenced by context, e.g. indoor vs. outdoor, we performed a longitudinal matching procedure where context is approximated by gait speed and bout length. We focused on the analysis of step variability and body sway as well as on compound measures derived from these. Results: Lab-based assessment showed changes after 2 years (stride length CV: p=0.008, rprb=0.66; lateral sway CV: p=0.013, rprb=0.62). Longitudinal analysis of real-life data revealed changes between baseline and 1-year follow-up (compound measure: p=0.037, rprb=0.53). 2-years analyses confirmed results with higher effect sizes (compound: rprb=0.974). We found correlations between lab-based and real-life measures, and between gait measures and subjective balance confidence. Conclusion: Descriptors of gait context were used to match walking trials. After matching, variability measures capture ataxic-specific changes in real-life walking even within 1 year. Context-matched measures of gait variability thus present promising ecologically valid performance outcomes for intervention studies.

Determining Clinically Relevant Gait Parameters Measured From Load Monitoring Insole Worn During Tibial Fracture Rehabilitation Using Fuzzy Inference Systems

Kylee North1, Grange Simpson1, Robert Hitchock1, Amy Cizik1

1University of Utah

Despite lower extremity fractures being common injuries, little is known about how patient weight-bearing behavior during rehabilitation contributes to long-term outcomes. Monitoring patient weight-bearing behavior using wearable sensors would allow clinicians to develop data-driven rehabilitation protocols. The objective of this study was to categorize gait parameters based on their ability to differentiate between patients with excellent and average long-term outcomes using Fuzzy Inferences System (FIS). Methods: Patients with closed tibial or bimalleolar ankle fractures were recruited in this 3 year observational study. An insole load sensor continuously monitored patient weight-bearing during rehabilitation. Longitudinal data was reduced to 93 gait parameters. Using the 1 year physical function outcome score patients were divided into two groups: Excellent Outcomes, and Average Outcomes. A FIS classified gait parameters based on their ability to differentiate between the two outcomes. Results: Of the 42 patients enrolled, 17 had both 1 year physical function outcome score (9 Average, 8 Excellent) and complete insole data (33.7 + 14.5 y/o, 60% female). The FIS revealed that gait parameters related to step count and active walking time best differentiated the two outcome groups. Weight-bearing magnitude moderately differentiated the two groups, and cadence and static loading variables did not have strong differentiation. All metrics with strong FIS classification had statistically significant two-tailed T-test results (P-value < 0.03), while weak FIS differentiated groups did not. Conclusion: FIS proved to be a powerful tool for automated gait parameter classification due to its ease of implementation, adaptability, and intuitive graphical inputs. Although the data came from a pilot study with small patient size, FIS implementation indicated what gait patterns to focus on when designing higher-powered future clinical trials to produce data-driven protocols.

Determining the Locations of Physical Activity of Community-Dwelling Older Adults: A Global Positioning System-Based Study

Shiho Amagasa1, Yutaka Fukuoka2, Shigeru Inoue2, Hiroshi Murayama3, Takeo Fujiwara4, Yugo Shobugawa5

1Teikyo University, 2Kogakuin University, 3Metropolitan Institute of Gerontology 4Tokyo Medical and Dental University, 5Niigata University

Background: Spatial data from portable global positioning systems (GPS) have the potential to allow objective monitoring of people’s mobility and the locations at which physical activity occurs in the community. However, few studies have identified residents’ daily behavior under free-living conditions. This study aimed to determine the locations visited by older residents using GPS data. Methods: This study used data form the Neuron to Environmental Impact across Generations (NEIGE) study. A randomly-selected sample of 511 community-dwelling Japanese older adults (47% men; aged 65-84 years) living in a rural city wore a GPS logger (i-gotU GT-600) and accelerometer (Active style Pro HJA-750C) for seven consecutive days in 2017. GPS measure longitude and latitude approximately every 10 minutes. We applied the k-means clustering algorithm (k=1800) to identify the locations the participating older adults visited during their daily lives. The program for the analysis was developed using Python. Results: We identified 56 clusters (places) visited by more than 40 participants in one week and most (89.3%) of them were within the city. The most visited locations were supermarkets in the city center, followed by civic centers, home improvement stores, and the agricultural cooperative society. On the other hand, some clusters were also observed at intersections and in privately owned areas such as homes. Discussions and Conclusions: GPS data can be used to identify the locations visited by residents, with many of them relevant to their daily lives. Utilizing GPS data or combining GPS data with accelerometer data may be helpful for better understanding of people’s physical activity in the future.

Digital Measures of Gait and Turning Increase Discriminative Ability to Predict Future Falls in People With Parkinson’s Disease

Vrutangkumar Shah1, James McNames2, Graham Harker1, Patricia Carlson-Kuhta1, John Nutt1, Mahmoud El-Gohary2, Kristen Sowalsky2, Martina Mancini1, Fay Horak1

1Oregon Health and Science University, 2APDM Wearable Technologies

Background and Aim: Although much is known about the multifactorial nature of falls in Parkinson’s disease (PD), it remains unclear whether gait and turning measures can improve prediction of future falls over fall history alone. We investigated if gait and turning digital measures increase discriminative ability to predict future fallers from non-fallers when combined with falls history. Methods: We recruited 34 subjects with PD (17 fallers and 17 non-fallers), age: 68 ± 6 years, MDS-UPDRS III ON: 31 ± 9. Participants were classified as fallers (at least one fall) or non-fallers based on self-reported fall history in the past six months. Subjects wore three inertial sensors (Opals) on feet and waist for a week of passive monitoring. We derived 55 gait, turning, and activity measures averaged across all strides. To investigate the ability of gait and turning measures to predict future fallers, we computed the area under the receiving operating curve (AUC). To check for the overfitting, we used a randomized 5-fold cross-validated AUC (CV-AUC). Participants were followed over the year after obtaining gait and balance measures for self-reported falls. Results: Twenty-five subjects reported falls in the follow-up year. Activity measures were similar in future fallers and non-fallers. The 5-fold CV-AUC to discriminate future fallers from non-fallers using history of falls alone was 0.77 (95% CI: [0.50-1.00]), and using pitch angle of the foot during mid-swing when combined with falls history was 0.90 [0.77-1.00]. In addition, coronal trunk range of motion (0.86 [0.74-0.98]), variability of gait speed (0.86 [0.74-0.98]), variability of step duration (0.86 [0.72-1.00]) and variability of heel strike angle (0.86 [0.75-0.97]) increased 5-fold CV-AUC when used in combination of falls history. Conclusions: The results suggest that digital measures of gait and turning can improve the ability to predict future falls when combined with fall history.

Duration of Medium Cadence Stepping Bouts is Linked With Disability and Functional Mobility for People With End-Stage Knee Osteoarthritis

Rashelle Hoffman1, Hope Davis-Wilson2, Katherine Balfany2, Paul Kline3, Elizabeth Juarez-Colunga4, Edward Melanson2, Cory Christiansen2

1Creighton University, 2VA Eastern Colorado Healthcare System, 3High Point University, 4University of Colorado

Individuals with knee osteoarthritis (OA) typically have slower walking cadence, a risk factor for joint cartilage damage. Total weekly time accumulated in higher walking intensities, measured by step cadence, is also linked to mobility, disability, and quality of life. However, it is unknown whether the cadence of daily stepping bouts is related to disability, function, and other patient-specific health factors for people with end-stage knee OA. This study aimed to identify patient factors that explain variability in duration of higher intensity walking. Individuals with end-stage knee OA awaiting total knee arthroplasty were included in this cross-sectional analysis. Stepping activity was assessed over 10 days with a thigh-mounted accelerometer (activPAL). The average maximum daily bout duration of at least medium cadence (MCBD; ≥80 spm) was calculated with a custom MATLAB code. A backward stepwise linear regression determined the most parsimonious model to explain the variability of MCBD. OA-related outcomes of disability and functional mobility, age, sex, ethnicity, work status, comorbidities, depression, social support, cognition, and education were assessed as candidate variables. Variables that met a P<0.2 simple linear regression association with MCBD were included in the backward model. Ninety-four individuals (83 males, age: 66.9 years [SD=7.1]) were included in this analysis. The average MCBD for the sample was 3 min 25 secs [SD=4 min 16 sec]. Individuals with lower self-reported disability (standardized parameter estimate [β]=-0.25, P=0.01) and better functional mobility (longer 6-min walk test distances; β=0.26; P=0.009) had longer MCBD. These variables explained 15% of the variability of time spent in MCBD. The duration of higher intensity stepping activity is best explained by lower disability and better functional mobility. Self-reported symptoms and limited functional mobility should be considered for walking interventions for people with knee OA.

Effect of Accelerometer Epoch Length on Physical Activity and Sedentary Time in Toddlers

Elyse Letts1, Sara King-Dowling2, Natascja Di Cristofaro1, Joyce Obeid1

1McMaster University, 2The Children’s Hospital of Philadelphia

Introduction: Accelerometers have the capacity to record accelerations at very high resolution, yet many still use longer epochs (30s or greater) for analysis. Given that toddler movement often occurs in short, sporadic bursts, choice of epoch length may have important implications for quantifying activity. The objective of this study is to determine the effect of epoch length on the activity intensity time distribution during a semi-structured play session in toddlers. Methods: A total of 116 toddlers (ages 12-35 months) were invited to complete two ∼1 hour guided activity assessment sessions while wearing an ActiGraph accelerometer around their waist. Accelerometer data were downloaded and reintegrated to 1s, 3s, 5s, 15s, and 30s epochs using ActiLife. The Trost (2012) cut-points for toddlers were used to determine time (min/hr) spent sedentary (SED), and in light (LPA) and moderate-to-vigorous (MVPA) physical activity, which were then summed across the two play sessions. One-way repeated measures ANOVAs were conducted separately for SED, LPA, and MVPA using R v4.1.1. Results: Complete data were available for 112 participants (52% girls, mean age ± SD = 21.3 ± 7.0 months). Time spent in SED decreased with longer epochs (F(1.3, 144.1) = 3809.3, p = < 0.001). Conversely, LPA increased with longer epochs (F(1.2, 132.9) = 2266.2, p = < 0.001). There was a main effect of epoch for MVPA (F(1.2, 135.4) = 29.4, p = < 0.001), with the 15s epoch yielding the greatest time in MVPA (12.7 ± 4.7 min/hr) and the 1s epoch the lowest (11.2 ± 3.0 min/hr). See Figure 32 for detailed results. Discussion: Even during a short 2-hour session, our results show that epoch length is a critical consideration for measuring activity, particularly SED and LPA, in toddlers. Future analyses will validate epoch-based activity intensity to direct observation to determine the most appropriate epoch length for toddlers.

Effectiveness of a Personalised 3-mo eHealth Intervention on Daily Steps Among Patients of Elective Cardiac Procedures: A Randomised Controlled Trial

Tommi Vasankari1, Jari Halonen2, Ville Vasankari3, Sini Vasankari4, Kari Tokola1, Henri Vähä-Ypyä1, Pauliina Husu1, Harri Sievänen1, Juha Hartikainen2

1UKK Institute for Health Promotion Research, 2Kuopio University, 3Helsinki University, 4Turku University

Background: We investigated effectiveness of a personalized accelerometer - and smartphone - based eHealth intervention (PACO) to increase PA and decrease SB among patients recovering from cardiac procedures. Methods: This prospective clinical trial was scheduled for elective cardiac patients: patients of coronary artery bypass grafting, valvular surgery or coronary angiography/percutaneous coronary intervention (n=210) were randomly assigned either to a 12-wks interactive PA guidance (interactive accelerometer-application (ExSed)-cloud) or standard care. Participants received personalized, increasing goals for daily steps, which were examined with an interactive accelerometer (app for participant and cloud for physiotherapist at Heart Center). The physiotherapist encouraged patients to increase daily steps using short phone calls twice per month. Data of effectiveness was collected before operations, after 3-mo intervention and after 9-mo f-up. This trial is registered with ClinicalTrials.gov (NCT03470246) and is ongoing. Results: We pooled all patients groups that participated in PACO study. Concerning intervention effectiveness, mean MVPA and daily steps increased in patients of intervention group by 46% and 21% after 3-mo intervention and by 29% and 17% at 12-mo f-up compared to control group (5% and 10% decrease at 3-mo and 10% and 11% decrease at 12-mo, respectively) (p=0.028 and p=0.015 for MVPA and steps). The PACO intervention increased MVPA by 16 min and daily steps by 1250 after 3-mo intervention, while MVPA (-4 min) and number of steps (-680) decreased in controls at 3-mo. The differences between the groups were also seen at 12- mo f-up. Conclusions: The used personalized, interactive accelerometer-application-cloud based eHealth intervention with increasing goals for daily steps in patients after cardiac procedure was found effective and the positive effect persisted during follow-up when compared to patients of usual care.

Engagement With an Integrated Two-Way Communication Near-Real-Time Mobile Health Intervention to Motivate Adults >60y to ‘Move More and Sit Less’

Quinn Anderson-Song1, Diego Arguello1, Ethan Rogers1, Grant Denmark1, Gregory Cloutier1, Carmen Sceppa1, Charles Hillman1, Arthur Kramer1, Dinesh John1

1Northeastern University

Mobile health technologies have shown promise in motivating older adults (OA) to ‘sit less and move more.’ However, such technologies rarely integrate meaningful and intelligent two-way communication, which is necessary to enable complex and interactive problem-solving. This may improve participant engagement and thus lead to sustainable physical behavior (PB) change. Purpose: This descriptive study aimed to quantify intervention engagement using a semi-automated, behavior-aware, text-based virtual “Companion” that utilizes a “human-in-the-loop” approach integrated with wearable wireless-sensing and a mobile phone to motivate overweight/obese OA to ‘sit less and move more.’ Methods: Fifteen OA (age: 67.5±5.9y; BMI: 29.1±4.3kg?m-2) from the treatment-arm (n=23/46) of an ongoing pilot trial completed 16-weeks of exercise training supplemented with Companion to effect PB change outside training sessions in near-real-time using knowledge of recent/ongoing behavior and its context. Engagement with Companion was measured as overall response rate (RR), just-in-time (JIT) RR (response within 10 min), conversive engagement rate (CER) (>1 response within 10 min), median response delay (RD), and unsolicited engagement (UE). Overall and JIT-RR were also stratified by week. Results: Companion communication (641 ± 321 texts/participant) had a mean RR of 65.1% (95% CI: 55.8, 74.5), JIT-RR of 41.3% (95% CI: 30.4, 52.2), CER of 30.8% (95% CI: 22.2, 39.3), and a median RD of 17.2 min (95% CI: 0.9, 33.5); 8.2% (95% CI: 3.8, 12.5) of participant communications (375 ± 250 texts/participant) were UE. Overall RR and JIT-RR fluctuated on a weekly basis (see Figure 33). Conclusions: Existing static communication technologies demonstrate a wide JIT-RR (∼21% to 39%). Intelligent adaptive PB coaching may be more successful at engaging OA than such static communication-based approaches. More research is needed to understand those communication elements that cause improved engagement.

Full-Day Spontaneous Leg Movement Quantity in Infants at High Risk for Cerebral Palsy

Beth A. Smith1, Federico Gennaro2, Thubi H.A. Kolobe3, Laura A. Prosser4

1University of Southern California, 2Children’s Hospital Los Angeles, 3University of Oklahoma, 4Children’s Hospital of Philadelphia

Paucity of movement is the hallmark of physical disability in individuals with cerebral palsy (CP) with research suggesting early onset. Little is known about the quantity of movement that infants at high risk for CP produce and how this may affect how they learn (or fail to learn) to move. Here we present preliminary analyses of spontaneous leg movement data for 10 infants with brain injury at risk for CP. Data were collected across 2 full days from both legs using wearable sensors at 1, 2, 3, and 4 months of age (adjusted for prematurity when applicable). Using custom, validated MATLAB software (Smith et. al, 2015; Trujillo-Priego et. al, 2017) we calculated average number of leg movements per hour awake time for each leg. We also calculated the ratio of left to right leg movements per day. Infants averaged 7.4 hours per day of awake time (SD = 1.6, range = 3.3 - 11.1), produced an average of 428 movements per hour awake on the right leg (SD = 198, range = 123-1011) and 459 on the left leg (SD = 210, range = 141-906). Figure 34 shows average number of left leg movements per hour awake for individual infants for each day of data collection, across 4 months. These numbers represent around 1/3 of movements observed in infants with typical development (Smith et. al, 2015). The ratio of left to right leg movements produced per day ranged from 0.47 to 1.76. In 70% of data collections, this symmetry ratio was outside of the reference range for infants with typical development (Smith and Lang, 2019), which may be representative of hemiparesis. Quantifying leg movements across full days can provide fundamental missing information about how infants at high risk for CP move in daily life, which may affect learning motor skills.

Gait Detection From a Wrist-Worn Sensor Using Machine Learning Methods: Daily Living Study in Older Adults and Patients With Parkinson’s Disease

Yonatan Brand1, Dafna Schwartz1, Eran Gazit2, Aron Buchman3, Jeffrey Hausdorff1

1Tel Aviv University, 2Center for the Study of Movement Cognition and Mobility, 3Rush University

Objective: Altered gait patterns and reduced daily-living physical activity are common in older adults and people with movement disorders, such as Parkinson’s disease (PD). Accurate measurement of daily-living gait promises to augment the ability to monitor changes in function. We aimed to validate and assess algorithms for automatically detecting the presence or absence of gait from a wrist-worn accelerometer collected from people with PD and older adults. Methods: 18 individuals with PD (mean age=72.7 yrs, 8 women; mean disease duration=5.6 yrs) and 12 older adults (mean age=76.1 yrs, 9 women) wore a tri-axial accelerometer on their wrist and the lower back for 7 days. Labeling the dataset was conducted based on gold-standard annotations from the lower back sensor. An additional dataset that is used for the model’s validation is a 'PhysioNet’ dataset accelerometry signals from 32 healthy participants (mean age=39 yrs, 19 women). We compared 3 different algorithms: 1. Unsupervised model using autoregressive infinity hidden Markov model (Y Raykov et al. 2021) 2. Deep Convolutional Neural network (Q Zou et al. 2020) 3. Deep residual Bidirectional Long Short-Term Memory (Y Zhao et al. 2018). Results (see Figure 35): Figure 1a shows the association between the unsupervised model’s output and the gold standard, and figure 1b,c shows the ROC and precision-recall curves for the deep learning models. Conclusion: While the models perform well on the 'PhysioNet’ dataset, there is a significant trade-off between the precision and the sensitivity of the model on daily living data. To identify all walking bouts during daily living, the current methods need additional work. Still, for some purposes, the current results may already be adequate. For example, if the goal is to evaluate gait quality, it might be sufficient to use a model with high precision for correctly identifying more than 50% of the gait bouts, even though not all the bouts are identified.

Gait Patterns During Daily Life Differ by Frailty Status in Older Men Treated With Androgen Deprivation Therapy for Prostate Cancer

Deanne Tibbitts1, Martina Mancini1, Sydnee Stoyles1, Christopher Palmer1, Ramyar Eslami1, Mahmoud El-Gohary2, Fay Horak1, Kerri Winters-Stone1

1Oregon Health and Science University, 2APDM Wearable Technologies

Objective: Androgen deprivation therapy (ADT) can increase the risk of frailty, a weakened physiological state associated with falls, disability, and early mortality. Gait impairment may signal the early onset of frailty; therefore, sensitive and ecologically valid measures of gait are needed. Using passive gait monitoring, we evaluated relationships between daily life gait characteristics and frailty in men treated with ADT. Methods: Cross-sectional analysis of men treated with ADT for prostate cancer. We classified men as Frail, Pre-frail, or Robust using the FRAIL scale, a 5-item self-report measure (fatigue, weakness, slowness, illness, weight loss). Men self-reported falls in the past year. Gait was collected using instrumented socks (APDM Wearable Technologies) that housed inertial sensors on the feet, and were worn at home during waking hours for up to 7 days. We derived measures of quantity and quality of gait by averaging across all gait bouts from at least 5 days of passive monitoring. Linear models were used to investigate differences among groups. Results: Of 64 available cases (mean age, 72y), 36% of men were Pre-frail, 3% Frail, and 36% reported a fall in the past year. Gait analysis showed that men who were Pre-frail/Frail had a significantly slower gait speed (beta=-0.17, p<0.01), shorter stride length (beta=-0.13, p<0.05), lower cadence (steps/min; beta=-2.6, p<0.05), fewer strides per gait bout (beta=-9.1, p<0.05), lower foot angle at heel strike (beta=-2.7, p<0.05), lower foot angle at toe-off (beta=-2.0, p<0.05), and higher double support time (beta=2.9, p<0.01) than men classified as Robust. Conclusions: Daily life gait characteristics derived from instrumented socks can differentiate between frail and non-frail men treated with ADT. Gait patterns in frail men are comparable to patterns associated with falls and disability. Future work should examine if quantity or quality of gait can predict development of frailty during ADT.

Genetic and Environmental Influences on Features of Objectively Measured Physical Activity, Sleep and Circadian Rhythmicity in Adolescent Twins

Wei Guo1, Victoria O’Callaghan2, Andrew Leroux3, Vadim Zipunnikov4, Margie Wright5, Ian Hickie6, Kathleen Merikangas1

1National Institute of Mental Health, 2Queensland Brain Institute, 3University of Colorado, 4Johns Hopkins University, 5University of Queensland, 6University of Sydney

Background: The growing use of mobile tools has facilitated our ability to examine daily patterns of sleep and physical activity (PA) in real time. Gaining insight into the extent to which these patterns are driven by genetic factors and environmental context can inform our understanding of their role in mood and other disorders. Here, the Brisbane Adolescent Twin Study (BATS) was used to estimate heritability for accelerometry derived features from the domains of physical activity, sleep, and circadian rhythmicity. Methods: The BATS data (58% female) included 55 identical (MZ) and 81 nonidentical (DZ) twin pairs, and 35 unpaired twins aged 12-18 years, with up to two weeks of accelerometry data. Polygenic heritability estimates, defined as the proportion of the total phenotypic variance explained by additive genetic effects were implemented in the SOLAR package, with age and sex adjustment. Heritability was estimated separately from the domains of physical activity, sleep, and circadian rhythmicity, respectively. Results: Estimated heritability was greater than 0.6 (se<0.08) for all features across the three domains. Sleep duration had heritability of 0.64 (se=0.07) and MVPA had heritability of 0.82 (se=0.04). Heritability differed on weekends vs weekdays and by sex for some of the features examined. Discussion: These findings suggest that behavioral interventions in sleep and PA may target those with increased susceptibility to unhealthy patters of sleep and physical activity that may influence both mental and physical health.

Locations of Women’s Physical Activity Before and During the COVID-19 Pandemic

Katelyn Holliday1, Michael Zimmerman1, Laura Fish1, Daniela Sotres-Alvarez2, Truls Østbye1

1Duke University, 2University of North Carolina at Chapel Hill

Objective: The COVID-19 pandemic changed many facets of daily life. We sought to understand how often women used various locations for physical activity before and during pandemic-related restrictions. Methods: We recruited women aged 20-40 who received primary care at an academic family medicine clinic and a federally qualified health center in an urban North Carolina town from November 2020-May 2022. Women completed surveys in English or Spanish about their physical activity behaviors before and during pandemic restrictions and a subset wore accelerometer and GPS units. Ordinal measures of association were used to test for differences in usage frequency of physical activity locations by sociodemographic characteristics. Results: Surveys were completed by 134 women (median age 30; 32% Hispanic Latina, 40% Non-Hispanic Black, 28% Non-Hispanic White). Before the pandemic, women most regularly used homes, roads, and paid fitness facilities for physical activity, with 27.7%, 27.8%, and 20.9% reporting use at least 3 days/wk, respectively. During the pandemic, the percent of women completing regular physical activity at home and on roads increased to 45.7% and 33.9% whereas the percent regularly using paid fitness facilities decreased to 7.8%. Before the pandemic, 32.0% and 37.7% of women reported never using outdoor public spaces like trails and parks. These percentages increased during the pandemic to 40.2% and 56.7%, respectively. Non-Hispanic Black women used roads (p=0.01) and trails (p<0.01) less often than other groups, controlling for household income (e.g. 15.4% of Non-Hispanic Black vs 20.0% of Hispanic Latina vs 54.0% of Non-Hispanic White women reported using roads at least 3 days/wk). Conclusions: Surprisingly, women reported less frequent use of public outdoor spaces like trails and parks during the COVID-19 pandemic. These survey results will also inform interpretation of accelerometry/GPS data collected for a subset of these women during the pandemic.

Monitoring Postures and Motions in Hospitalized Patients; A Review on Methodological Approaches

Marlissa Becker1, Henri Hurkmans1, Jan Verhaar1, Johannes Bussmann1

1Erasmus University

Mobilization is an important task of hospital staff, as adverse hospital outcomes are caused by bedrest and inactivity. Unfortunately, hospital staff has limited time and resources available to monitor and assists physical behavior. Wearable sensor technology can be a solution, specifically to monitor postures and motions. However, previous studies showed different approaches in study design when using these wearables. Therefore, this review aims to provide an overview of methodological approaches of monitoring postures and motions in hospitalized patients with sensor technology. A systematic search of Embase, Medline (Ovid), Web of Science and Google Scholar was conducted for the period of 2010 to 2021. Eligible articles were screened by two researchers independently. Data extraction focused on the methodological approaches of included articles. The screening of 10.718 articles resulted in 34 included studies with a large variety in patient populations (e.g. older adults, surgical or stroke patients). The most common goals of included studies were to gain insight into behavioral patterns (n=23), to examine intervention effect (n=6), to investigate predictors (n=3) or to examine new measurement methods (n=2). Included studies showed different approaches in their study design and lacked integrality in reported details, e.g. in their duration of measurement. Especially information on device settings, data analysis and algorithms was underreported. However, similarities were found in the type of sensor, wear location, attachment method, device outcomes and wear time a day. More uniformity and transparency in methodological approaches are needed to improve the comparability of studies and to establish the benefits of monitoring postures and motions using sensor technology in clinical practice. Using objective sensor technology is feasible in a wide variety of hospital populations and is currently most applied in observational or cohort studies. Future studies should carefully consider clinical challenges in their study design, to measure relevant and reliable outcome values.

Moving From Intention to Behavior: First Results of an App-Based Physical Activity Intervention With a Randomized Controlled Trial Design (i2be)

Lili Kókai1

1Erasmus University

Objective: To evaluate the efficacy and processes of an 8-week app-based physical activity intervention using wearables in women with a prior hypertensive pregnancy disorder. The intervention is based on the integrated behavior change model, which outlines the motivational, volitional and automatic processes that lead to physical activity. Methods: The primary outcome is weekly minutes of moderate-to-vigorous physical activity (MVPA) measured by Fitbit Inspire 2. Secondary outcomes are weekly average Fitbit-measured daily resting heart rate, and self-reported body mass index, waist-hip ratio, cardiorespiratory fitness and well-being. Tertiary outcomes are self-reported variables representing motivational, volitional and automatic processes. Outcomes will be assessed at baseline, immediately post-intervention and at 3 and 12 months post-intervention. Efficacy will be determined by available case analysis. A process evaluation will be performed immediately post-intervention using user engagement statistics and self-reported quantitative and qualitative measures of appropriateness and appeal. Participants (n=661) were randomly allocated to 1 of 3 conditions (start date Sept '21). The information condition received information. The motivation condition received information and content targeting motivational processes. The action condition received information, content targeting motivational processes and content targeting volitional and automatic processes. Results: Linear regression analysis will be used to assess differences between all 3 groups at each timepoint. Descriptive statistics and qualitative methods will be used to assess intervention processes. The current submission will evaluate immediately post-intervention efficacy and processes. Conclusion: The current submission will present novel findings, analyzing 1) wearables data to evaluate intervention efficacy and 2) wearables use and perception of participants in an app-based physical activity intervention. (Figure 36)

Physical Activity and Sedentary Behaviour in Children With Neck or Back Pain: An Observational Study

Anna Clark1, Anna Cooper-Ryan1, Alexandra Clarke-Cornwell1, Tamara Brown1, Stephen Preece1

1University of Salford

Objective: Being physical active and reducing sedentary behaviour is vital to children’s health and welfare. Levels of these behaviours in healthy children and children with different muscle pain is still relatively unknown. The aim of this study was to calculate levels of activity and inactivity in children with neck or back pain and to test whether these differed from their healthy peers. Methods: An observational study recruited participants (7-17 years) via social media: 10 with neck pain, 10 with back pain, and 10 healthy children. Habitual time spent moving, sedentary time, and steps were measured with the activPAL for seven days (week and weekend). Comparisons were made between the groups with neck or back pain and healthy participants. Results: There was no significant differences between number of steps between the three groups during the week and at the weekends. Healthy participants spent more time sitting compared to children with back pain throughout the week (Cohen’s D: 1.79 weekdays, 0.85 weekend). Those with back pain spent more time in a lying position when awake compared to healthy participants during the week. Those in the pain groups had a greater total sedentary time throughout the week compared to healthy participants; however, this was reversed at the weekend. Neck pain participants spent more time in a lying position when awake compared to healthy participants, both during the week and at weekends (Cohen’s D: 0.61 weekdays, 0.80 weekend). On average, the healthy participants slept less than both pain groups at the weekends and during the week. Conclusions: Physical activity (steps) did not differ between the three groups; however, sedentary behaviours and sleep differed between the pain groups compared to that of their healthy peers. Further investigation is required to determine whether these behaviours are part of the causal pathway or as a consequence of musculoskeletal pain.

Physical Activity in Community-Dwelling Older Adults: Which Accelerometry Measures are the Most Robust? A Structured Review

Khalid Abdul Jabbar1, Ríona Mc Ardle2, Sue Lord3, Ngaire Kerse1, Silvia Del Din2, Ruth Teh1

1University of Auckland, 2Newcastle University, 3Auckland University of Technology

The aim of this structured review was to understand the validity, reliability and responsiveness of physical activity (PA) measures derived from accelerometry of community-dwelling older adults (≥ 65 years) in real-world conditions. Eight electronic databases were searched. Step count, duration of walking, lying, sitting and standing were key PA outcomes reported. Criterion validity was established either as agreement between a gold standard reference and accelerometry measures or as percentage of agreement between video observation and accelerometry measures. Construct validity was examined by looking at agreement between step counts and moderate-to-vigorous PA. Inter-rater reliability was reported as intraclass correlation coefficients [ICC], criterion validity as either ICC [95% CI] or as percentage of agreement, and construct validity as Spearman’s Rho. Sixty-eight papers were screened with seven meeting the inclusion criteria. Two additional papers were retrieved from reference lists, with six from the former seven included in the review for a total of eight papers. Three reported inter-rater reliability (n = 40). Seven studies reported criterion validity (n = 164) and one reported construct validity (n = 30). Inter-rater reliability results were as follows: for walking duration (0.94 and 0.95), lying duration (0.98 and 0.99), sitting duration (0.78 and 0.99) and standing duration (0.98 and 0.99). Results for criterion validity ranged from 0.83 (95% CI 0.59, 0.93) to 0.96 (0.91, 0.99) for step counts, and from 63.6% to 93.5% agreement for walking, 35.6% to 100% for lying, 79.2% to 94.5% for sitting, and 38.6% to 80.1% for standing. Construct validity of step counts and moderate-to-vigorous PA was reported as r = 0.68 and 0.72. Inter-rater reliability and criterion validity of step count, duration of sitting, standing, walking, and lying in community-dwelling older adults derived from accelerometry in real-world conditions is tentatively established. (Figure 37)

Physical Activity, Sedentary Time, and Wear Time Recorded by Accelerometer in a Nationwide Sample – Results From MoMo Wave 3 (2018–2020)

Simon Kolb1, Alexander Burchartz1, Leon Klos1, Steffen Schmidt1, Alexander Woll1

1Karlsruhe Institute of Technology

Objective: Since 2015, accelerometers have been used in the MoMo study to record physical activity, among other things. The last survey wave started in August 2018 and was stopped in March 2020, when COVID restrictions were put in place. The current article gives a first insight into the results. Methods: Participants were asked to wear an ActiGraph accelerometer on the right side of the hip for one week during waking hours. Data sets including at least four weekdays and one weekend day (4+ 1 rule) with more than 8h of wear time per day were considered valid and interpreted as the waking phase. Results: 1,160 (76.1%) participants fulfilled the 4+ 1 rule. Participants with valid data wore the sensor for 5 (5.1%), 6 (26.9%), or 7 (68.0%) days, respectively for a total of 7690 valid days in the sample. Mean wear time averaged 833 min (CI95% = 829, 836) per day and varied by day of the week. From Monday to Thursday, average wear time lies between 847 min (CI95% = 839, 856) and 852 min (CI95% = 843, 861). Friday was the longest day on average with 882 min (CI95% = 872, 891) and on the weekend, the lowest wear times with 809 min (Saturday, CI95% =798, 819) and 733 min (Sunday CI95% = 723, 744) were recorded. On 2,640 (34%) days, 60 min or more moderate-to-vigorous physical activity (MVPA) was logged. As a result, only 31.4% of the participants averaged ≥60 min of MVPA per day. Meanwhile, 4,998 days (65%) with sedentary time ≥ 480 min were recorded and 69% of the participants had an average sedentary time ≥ 480 min/day. Conclusion: The present study shows that the duration of the waking phases differs on Saturday and Sunday. We assume that PA evaluation may be influenced by the type of weekend day included in the analysis. Since shorter waking periods limit the possible absolute active time on weekends, in addition to absolute minutes, relative proportions of MVPA should be reported in future studies.

Physiological and Perceived Responses to a Prolonged Moderate Intensity Walking Bout in Older Adults

Laura Karavirta1, Timo Aittokoski1, Timo Rantalainen1, Antti Löppönen1, Olli-Pekka Mattila1, Lotta Palmberg1, Kirsi Keskinen1, Taina Rantanen1

1University of Jyväskylä

Objective: Physiological responses to physical activity contribute to health and performance adaptations. Current physical activity guidelines emphasise that every minute counts rather than advocating for sustained bouts of activity. We examined physiological responses and perceived exertion to a prolonged moderate intensity walking bout in older adults. Methods: Participants from the ongoing AGNES follow-up study (n=42, mean age 79.5 (SD 2.3) yrs, 41% women) performed first a 6-minute walk at self-selected moderate intensity (RPE 13 on Borg’s 6-20 rating of perceived exertion) and thereafter continued walking at the same speed by following pacer lights until RPE reached 16 (vigorous) or the total walking duration reached 30 minutes, whichever occurred first. RPE, blood lactate, heart rate and respiratory frequency were recorded before, after 6 minutes and at the end of the test. Results: Although the walking speed was kept constant (1.40 (0.21) m/s), RPE increased during the prolonged walking (p<0.001) and reached 16 in 76.2 % of the participants. For them, the duration to reach RPE 16 was 14.9 (4.1) min. Heart rate and respiratory frequency increased from 79 (13) /min and 19 (3) /min pre-test to 109 (18) and 25 (5) after 6 minutes of walking and continued to increase towards the end of the test to 118 (20) and 28 (5), p<0.001. Blood lactate increased from pre-test 1.2 (0.3) to 1.9 (0.7) mmol/l after 6 minutes but remained at the same level after that. Discussion: Perceived exertion increased from moderate to vigorous during a prolonged walking bout at constant speed in a majority of older adults within 30 minutes. While the external work rate did not change, heart rate and respiratory frequency continued to increase revealing an increased physiological response per time unit to the continued walking bout. The present results may have implications for the quantification of free-living physical activity by emphasising the importance of bout duration.

Predictors of Physical Activity Up to One Year After Hospitalization for COVID-19; Results From the CO-FLOW Study

Johannes Bussmann1, Julia Berenschot1, Gijs Broeren1, Martine Bek1, Ruben Regterschot1, Merel Hellemons1, Joachim Aerts1, Gerard Ribbers1, Majanka Heijenbrok-Kal1, Rita van den Berg-Emons1

1Erasmus University

Objectives: To identify predictors of accelerometer-assessed physical activity (PA) in patients with COVID-19 up to 1 year after hospital discharge. Design: Multicenter prospective cohort study. Setting: Recruitment in academic and general hospitals. Participants: 336 adult patients who had been hospitalized for COVID-19 (age 60±10 years, 68.8% were male, body mass index (BMI) 29.3±5.7 kg/m2), with at least 1 measurement during follow-up. Intervention: n.a. Outcome measure: PA was measured using a tri-axial accelerometer device (GENEActiv), worn around the dominant wrist for 7 consecutive days, at 3 or 6 and 12 months after hospital discharge. A multivariable linear mixed model analysis was performed to identify predictors of minutes of moderate-to-vigorous PA (MVPA). Potential predictors included time after hospital discharge, age, sex, BMI, comorbidity, self-reported pre-COVID physical activity level (PAL), Intensive Care Unit (ICU) admission, and length of stay (LOS) in hospital. Results: Data were collected from 195 patients at 3 months, 140 at 6 months, and 156 at 12 months. The estimated mean time MVPA was 71.5±3.9 min/d, 75.4±4.1 min/d, and 76.2±4.0 min/d at 3, 6, 12 months, respectively. Significant predictors for lower MVPA were older age (beta -1.3 [95%CI -1.7 to -0.9], p<0.001), higher BMI (-1.5 [-2.3 to -0.7], p<0.001), comorbidity (-11.5 [-20.3 to -2.8], p=0.01), inactive pre-COVID PAL (-25.7 [-38.6 to -12.7], p<0.001), longer LOS (-0.4 [-0.6 to -0.1], p=0.009), whereas time, sex, and ICU admission did not independently predict MVPA. Conclusion: Time MVPA did not change over 1 year follow-up after hospitalization for COVID-19. In addition to the classical risk factors (age, BMI, comorbidity, inactive pre-COVID PAL) LOS in hospital for COVID-19 was an independent predictor of MVPA. Further analysis is required to establish the effect of daily PA on the recovery of prevalent symptoms, such as fatigue and dyspnea, after COVID-19 infection.

Reliability of Sleep Midpoints Assessed Over 7-Days Using ActivPAL and Sleep Logs

Joshua Freeman1, Pedro Saint-Maurice1, Shreya Patel1, Sarah Keadle2, Charles Matthews1

1National Cancer Institute, 2California Polytechnic State University

Sleep midpoint, the median time between sleep onset and awakening, is an important indicator of circadian disruption and has been associated with poor health. Sleep midpoint is usually assessed via a single self-report of typical sleep onset, awakening, and time in bed. Activity monitors, such as activPAL (AP), can estimate sleep midpoint over multiple nights, but how well AP estimates sleep midpoint and how reliable these estimates are compared to self-report is unclear. Our objective was to evaluate the reliability of sleep midpoint measured using AP and self-report in sleep logs (SL). We evaluated sleep among n=48 participants using SLs and thigh-worn accelerometers (AP) for 7-days. We calculated weekday, weekend, and mean sleep midpoints and tested group mean differences using t-tests. We evaluated reliability in sleep midpoints using intraclass correlation coefficients and 95% confidence intervals (ICC [95% CI]). On average 6.1 nights of sleep were assessed. Time in bed was longer when estimated by AP (Mean (M): 9.2 hours, Standard Deviation (SD): 1.5) vs. SL (8.3 hours, SD: 0.9; p-value (p)<0.01). Mean sleep midpoints were highly correlated (r=0.82, p<0.01) and were not significantly different when measured by AP (M: 3.0 AM, SD: 1.2) or SL (M: 3.1 AM, SD: 1.0; p=0.18). Weekend sleep midpoints were later than weekday sleep midpoints. AP estimates of weekend sleep midpoints (M=3.5 AM, SD=1.3) were not significantly different from SL (M=3.5 AM, SD=1.0; p=0.84). Weekday sleep midpoints were also similar. Sleep midpoints were more reliable when measured by SL (ICC=0.68, [0.56, 0.77]) vs. AP (ICC=0.42, [0.28, 0.56]). Similar results were observed for weekday sleep midpoints (SL: ICC=0.78, [0.68, 0.85] vs. AP: ICC=0.45, [0.30, 0.61]), and weekend sleep midpoints (SL: ICC=0.75, [0.57, 0.87] vs. AP: ICC=0.62, [0.38, 0.81]). In summary, AP and SL group estimates of sleep midpoints were similar with AP estimates having less stability over the 7-day protocol.

Resting Heart Rate as Biomarker for Tracking Change in Cardiorespiratory Fitness: The Fenland Study

Tomas Gonzales1, Justin Jeon2, Timothy Lindsay1, Kate Westgate1, Ignacio Perez-Pozuelo1, Stefanie Hollidge1, Katrien Wijndaele1, Kirsten Rennie1, Nita Forouhi1, Simon Griffin1, Nick Wareham1, Soren Brage1

1University of Cambridge, 2Yonsei University

Resting heart rate (RHR; beats per minute, bpm) may be a suitable biomarker of cardiorespiratory fitness ('fitness'; ml/min/kg), but few population studies have examined this relationship in large samples. In a UK population-based cohort study (The Fenland Study), we examined the cross-sectional and longitudinal relationships between RHR and fitness, and use a smartphone application to demonstrate the utility of RHR for population surveillance. In cross-sectional analyses (5,143 men, 5,722 women), RHR was measured while seated, supine, and during sleep. Fitness was estimated from a submaximal test. Mean (SD) RHR while seated, supine, and during sleep was 67.6 (9.8), 63.5 (8.9), and 56.9 (6.9) bpm, respectively. Age- and sex-adjusted associations with fitness were -0.26 (95%CI -0.27; -0.24), -0.31 (95%CI -0.33; -0.29), and -0.31 (95%CI -0.34; -0.29) ml/min/kg/beat. In longitudinal analyses (6,589 participants), RHR and fitness were reassessed after a median (interquartile range) of 6 (5-8) years. Each 1-bpm increase in supine RHR was associated with 0.23 (95%CI 0.20; 0.25) ml/min/kg decrease in fitness. Using a smartphone application (1,914 participants), we remotely captured serial measurements of RHR from August 2020 to April 2021, during which national lockdowns were implemented in the UK as a mitigation strategy against the COVID-19 pandemic. We examined RHR dynamics prior to and during the UK national lockdown periods by sex-specific fitness tertiles (low, mid, high). RHR was stable in those with mid and high fitness, however in low fitness groups RHR progressively increased from the 3rd UK national lockdown onwards (0.025 bpm/week increase, 95%CI 0.0076, 0.042) which may indicate the least fit getting worse when opportunities for free movement are restricted. These findings position RHR as a biomarker of fitness in the general population, in clinical care, and research, with application in personal goal setting and remote patient monitoring.

Scoping Review of Observational Studies of Adults With Accelerometry Measured Physical Activity and Sedentary Behavior

Kelly Evenson1, Elissa Scherer2, Kennedy Peter1, Carmen Cuthbertson1, Stephanie Eckman2

1University of North Carolina – Chapel Hill, 2RTI International

Objective: This scoping review identified observational studies of adults that utilized accelerometry to assess physical activity and sedentary behavior. Key elements on accelerometry data collection were abstracted to describe current practices and completeness of reporting. Methods: We searched three databases (PubMed, Web of Science, and SPORTDiscus) for articles published up to June 1, 2021. We included studies of non-institutionalized adults with an analytic sample size of at least 500. Results: The search returned 5686 unique records. After reviewing 1027 full-text publications, we identified and abstracted accelerometry characteristics on 155 unique observational studies (154 cross-sectional/cohort studies and 1 case control study). Five of these studies were distributed donor studies, where participants connected their accelerometers to an application and voluntarily shared data with researchers. Data collection occurred between 1999 to 2019, and 12.9% included multiple waves of accelerometry data collection. Most studies used one accelerometer (94.8%), but 7 studies (4.5%) used 2 accelerometers and 1 study (0.6%) used 4 accelerometers. Accelerometers were more commonly worn on the hip (48.5%) as compared to the wrist (21.8%), thigh (5.5%), other locations (13.9%), or not reported (10.3%). Overall, 12.7% of the accelerometers collected raw accelerations and 44.2% were worn for 24 hours/day throughout the collection period. Conclusion: The review identified 155 observational studies of adults that collected accelerometry, utilizing a wide range of accelerometer data processing methods. Researchers inconsistently reported key aspects of the process from collection to analysis, which needs addressing to support accurate comparisons across studies.

Self-Report Versus Accelerometer-Derived Measurement of Physical Activity in Metastatic Breast Cancer: How Do They Compare?

Mary Hidde1, Patricia Sheean2, Lauren Matthews, Kathleen Jensik1, Whitney Morelli1, Melinda Stolley1

1Medical College of Wisconsin, 2Loyola University

Background: Women with metastatic breast cancer (MBC) are often excluded from behavior interventions, resulting in limited data regarding their lifestyle habits. Self-report (SR) measurements of physical activity (PA) are often utilized in studies due to their simplicity and low cost, however, the congruent validity of SR PA is often questioned. Therefore, the aim of this study is to compare SR PA to accelerometry-derived PA using two different cut points for classifying MBC’s as active or inactive. Methods: Clinically stable women with MBC were recruited to participate in a randomized, 12-week diet and PA intervention. Participants completed the Godin Leisure Time Physical Activity Questionnaire (GLTPAQ) and wore an GT3X accelerometer (Actigraph) for 7-days. Troiano 2008 and Freedson 1998 cut points were used to measure moderate-vigorous PA (MVPA). Activity was dichotomized as “active” or “inactive” with active defined as leisure score index (LSI) of ≥24 compared to ≥150 minutes of MVPA/week measured by the Actigraph. Fisher’s exact test compared active vs. inactive between the GLTPAQ, Troiano, and Freedson cut points. Statistical significance was set at p<0.05. Results: Thirty-five women (55.49±12.71 years, BMI 29.70±6.80 kg/m2) provided GLTPAQ and Actigraph data pre-intervention. On average, LSI scores were 9.00±9.57. MVPA was 115.08±70.69 and 123.81±74.89 min/week by Troiano and Freedson, respectively. GLTPAQ categorized 4/35 as active compared to Troiano cut point which categorized 9/35 as active (p=0.34). For the Freedson cut point compared to the GLTPAQ, 10/35 were considered active (p=0.13). Discussion: It appears women with MBC may underestimate their PA utilizing the GLTPAQ when compared to accelerometry, regardless of cut point utilized. To better support validity and reliability, MBC patient perceptions of PA intensity and PA descriptions may require improved clarification and description going forward in studies relying solely on SR questionnaires.

Sleep Trajectories in Preschool Aged Children After 6 Months Participating in a Health Promotion Study

Hannah J. Coyle-Asbil1, Bridget Coyle-Asbil1, David W.L. Ma1, Jess Haines1, Lori Ann Vallis1

1University of Guelph

Objective: This study compared the sleep of preschoolers enrolled in the longitudinal intervention study, the Guelph Family Health Study (GFHS) at baseline (BL) and 6-months (6M). Methods: A subsample of preschoolers (N=43; 4.41 ± 0.37; 23 M) enrolled in the full study cohort of the GFHS were examined in this preliminary analysis. In brief, the GFHS intervention consisted of bi-weekly emails and four home visits from a health educator. Of the 43 children included in the current analysis, 15 were enrolled in the intervention group. At both BL and 6M children were instructed to wear an ActiGraph wGT3X-BT accelerometer (100Hz) on their right hip for 7 consecutive days of 24 hours. To verify compliance, parents were asked to track their child’s bedtime and wake-up times using a logbook. Upon return of the accelerometers, the data was downloaded and exported to 60 second epoch.agd files. Subsequently, the Sadeh (1994) algorithm was used for sleep scoring. Following which, the Tudor-Locke (2014) algorithm was applied to calculate the sleep metrics, including, total sleep time (TST), sleep efficiency, wake after sleep onset, sleep onset and wake-up times. To determine the effect of time-point, sex and intervention, a 3-way mixed-model ANOVA was performed, and Bonferroni adjustments were used when appropriate. Results: The ANOVA indicated a significant main effect of time-point for TST [F(1, 39)=4.217 (p=0.047)] and wake-up time [F(1, 39)=10.394 (p=0.003)]. The pairwise comparison revealed that compared to BL, TST was significantly reduced at 6M, and wake-up time was significantly earlier. Furthermore, a significant main effect of sex for TST [F(1, 39)=6.176 (p=0.017)] and wake-up time [F(1, 39)=5.382 (p=0.026)] were found. TST was significantly increased, and wakeup times were significantly later in females compared to males. Conclusion: The current results suggest that with age, preschoolers sleep patterns shift to earlier wake-up times and decreased TST.

Step Test Assessment Using Markerless Motion Capture in a Virtual Reality Setting

Andre Freligh1, Kevin Abbruzzese1, Vincent Alipit1, Sally LiArno1 Kevin Abbruzzese1, Vincent Alipit1, Sally LiArno1

1Stryker Orthopaedics

Unity3D gaming software can be utilized to create controlled and scalable virtual environments for remote assessments. Sensors such as Markerless Motion Capture (MMC) cameras can be integrated into Virtual Reality (VR) platforms to provide a systematic method to remotely assess accuracy and repeatability of physical measurements in virtual space. The objective of this study was to validate virtual object scale and determine accuracy of virtual object height during a step task. An Azure Kinect MMC camera was integrated with Unity3D using RF Solutions Unity package to quantify activities of daily life. A VR step task was designed to assess virtual step height measurements in the sagittal plane with MMC. Various step heights were measured in a VR interface in Unity3D and confirmed with a tape measure. Eighteen measurements were recorded for six different object heights assessed over three trials. A single participant was required to step onto a physical object and virtual object height was adjusted until the virtual object contacted the avatar. Virtual and analog step height measurements were assessed for significance. An Anderson-Darling test was performed to assess normality (p<0.005). A Mann-Whitney U Test was performed to assess significance between virtual and analog measurements. No significant differences in measurements were detected with average virtual step height of 10.77in±5.08in and average physical step height of 11.83in±4.99in, (p=0.666). Average accuracy and repeatability in virtual step height measurements was 1.06in±0.707in. This work demonstrates the potential to monitor patient progress in a virtual environment during a step test with MMC. The results confirmed various virtual heights can be used to assess step activities in the sagittal plane with low error. A VR interface with Azure Kinect camera demonstrated comparable results to analog measurements and may represent a modality to perform step assessments to remotely monitor patient progress. (Figure 38)

The Arm Activity Tracker: A Wearable System Measuring and Providing Feedback on Paretic Arm Activity in Stroke Patients. - Preliminary Results

A.J. Langerak1, G.R.H. Regterschot1, R.W. Selles1, G.M. Ribbers1, J.B.J. Bussmann1

1Erasmus University

Introduction: To stimulate daily life arm use in stroke patients, we developed the Arm Activity Tracker (AAT), a system based on wrist-worn accelerometers measuring arm activity and providing direct visual and tactile feedback on arm activity. This study aims to evaluate the feasibility of the AAT and secondarily assess its efficacy. Methods: A randomised, cross-over within-subjects study was conducted in sub-acute stroke patients admitted to a rehabilitation centre. Participants wore the AAT for five weeks: one week to determine baseline arm use, a two-week intervention condition where AAT provided feedback on arm activity, and a two-week control condition without feedback. Regarding feasibility, adherence was evaluated as the drop-out rate and the number of participants with successful AAT data collection for > 4 days per week, > 8 hours a day. Acceptance was measured with the System Usability Scale (SUS, range: 0-100) and the Technology Acceptance Model (TAM, range: 0-112). Efficacy was evaluated by estimating the difference between the intervention and control conditions for 1) activity of the paretic arm, and 2) the ratio between activity of the paretic and reference arm. Results: Seventeen stroke patients were included. A 29% drop-out rate was observed, reasons for drop-out were a high mental burden (n=2), nickel allergy (n=1) or early discharge (n=1). Ten participants met the requirements for successful AAT data collection. Participants who adhered to the study acceptance of the AAT was good (TAM median (IQR): 81 (72-104); SUS median (IQR): 80 (75-87.5)). Although the arm activity results were beneficial for the intervention period (+ 8.84% arm activity, + 21.25% ratio), these effects were not significant (p=0.08, p=0.05, respectively). Conclusion: Preliminary results show good feasibility of the AAT in patients adhering to the study. Reasons for drop-outs should be further evaluated, and a sufficiently powered trial should be performed to analyse efficacy.

The Associations Between Patterns and Changes in Regular Exercise Behavior and the Changes in Clinical Biomarkers Related to Cardiometabolic Diseases

JooYong Park1, Jaesung Choi1, Ji-Eun Kim1, Miyoung Lee2, Ji-Yeob Choi1

1Seoul National University, 2Kookmin University

This study examined whether the effects on risk factors related to cardiometabolic diseases (CMD) and their relationships were different according to the changes in exercise behavior. A community-based cohort study in Korea was used and 2,668 subjects were included in the final analyses. Participation in regular exercise was investigated by questionnaire from baseline to second follow-up. Changes in exercise behavior were defined into four groups “No exercise consistently (N-N)”, “Exercise consistently (Y-Y)”, “Change to exercise behavior (N-Y)”, and “Change to no exercise behavior (Y-N)”. Fourteen clinical biomarkers (CB) related to CMD including glucose level, lipid profile, and obesity indices were used in this study. The relative change in CB from baseline to second follow-up was calculated. The average of changes in CB was calculated in each of the four groups. The association between changes in regular exercise behavior and the changes in CB was estimated by sex using a generalized linear regression model, adjusting for age. Network of CB was constructed in each group based on partial correlation adjusting for age and topological comparison was conducted between networks. Y-N associated with more increased fasting blood sugar (β = 2.87, p=0.0036) and insulin level (β = 13.89, p=0.0376) in men and more increased total cholesterol (β = 3.85, p=0.0059) and LDL cholesterol (β = 4.61, p=0.0368) in women. Meanwhile, N-Y was associated with body fat % (β = -4.39, p=0.0012), visceral fat % (β = -0.49, p=0.0079), fasting insulin (β = -16.28, p=0.0177), and triglyceride (β = -15.71, p=0.0043). Obesity-related CB, especially waist circumference played a central role in most networks. More edges were found in N-Y or Y-Y group of men than in N-N or Y-N, while N-Y or Y-Y group of women had more edges than N-N or Y-N in women. Exercise behavior had favorable effects on the CMD related clinical factors although their inter-relationships were different by sex.

The Impact of Anti-Hypertensive Medication on the Relationship Between Daily Step Count and Blood Pressure

Craig Speirs1, Mark Dunlop1, Marc Roper1

1University of Strathclyde

Objective: There is extensive research about the relationship between physical activity and blood pressure. Despite that fact that pharmacological intervention can significantly reduce blood pressure, medication is seldom considered when investigating this relationship. We investigated the relationship between mean daily step count, anti-hypertensive treatment and blood pressure. Methods: We identified 3609 individuals from the BCS70 study with 7 days of valid activity data, blood pressure and medication data. Individuals were classified as hypertensive if they had blood pressure above 140/90mmHg. NICE prescribing guidelines were used to identify individuals taking one or more classes of drugs used to manage hypertension. Per individual mean daily step counts were calculated. For each blood pressure/anti-hypertensive medication combination the mean daily step count was calculated. ANOVA was used to identify significant differences in mean step count between the groups. Results: In total 257 (7.7%) individuals were taking anti-hypertensive drugs. A further 103 (2.9%) individuals had diagnosed high blood pressure but were not taking anti-hypertensive drugs. Mean daily step count was significantly higher for the non-medicated normal blood pressure group compared to the other groups. Conclusion: In this population, the treatment group whose blood pressure was effectively lowered took more steps on average than those whose blood pressure remained elevated. Further investigation is warranted to look at the role of increased daily stepping in potentiating the effects of anti-hypertensive medication. We believe this approach can identify other clinical populations with standard treatment profiles, including diabetes and angina, allowing for the confounding effect of drug treatment to be considered in physical activity research. (Figure 39)

The Preliminary Impact of a Family-Based Health Behaviour Intervention on Physical Activity and Adiposity in Toddlers in the Guelph Family Health Study (GFHS)

Lisa Wedel1, Hannah Coyle-Asbil1, Becky Breau1, David W Ma1, Jess Haines1, Andrea Buchholz1, Lori Ann Vallis1

1University of Guelph

Objective: Explore longitudinal impact of a family-based health behaviour intervention on toddler’s physical activity (PA) and possible associations with measures of adiposity. Methods: For this sub-analysis, data for children who were toddlers (<3 years) at both time points and met the PA criteria (360 minutes/day wear time for minimum 3 days) were analyzed (N=31 out of 322 children in GFHS). Families were randomized to control (N=16 toddlers) or intervention (N =15); the later group received 4-home visits from a health educator who assisted in setting family-based health-behaviour goals. At Baseline (BL) and 6-months (6M), children wore an Actigraph wGT3X-BT accelerometer (100 Hz; hip) for 7 days, 24 hours/day. Non-wear times were removed, and daily mean time (mins) spent engaged in sedentary behaviours (SED), moderate-to-vigorous physical activity (MVPA) and light physical activity (LPA) were calculated (Trost et al 2012). Adiposity measures included body mass index (BMI) z scores, % fat mass (%FM; via body impedance analyses) and waist-to-height ratios (WHtR). Results: A mixed model ANOVA revealed a significant (p <.05) interaction effect between Timepoint and Group for SED F(1,29) = 5.86 (p=0.022), LPA F(1,29) = 4.52 (p=0.04) and MVPA F(1,29) = 4.435 (p=0.04). The pairwise comparison revealed that in the intervention group, daily mean mins of MVPA and SED both increased at 6M compared to BL (by 15 and 44 mins, respectively). Also, the intervention group at 6M engaged in a mean 21 mins per day more of MVPA and a mean 18 mins per day more of LPA than control. Further analyses revealed significant (p<.05) negative correlations at both Timepoints between MVPA and WHtR (BL R= -.507, 6M R= -.629), and LPA and WHtR (BL R= -.483, 6M R= -.639). Conclusions: Preliminary findings suggest that a family-based health intervention may increase physical activity leading to lower WHtR in toddlers. Additional testing with a larger sample is needed to confirm these results.

The Validity of Using Smart Glasses to Measure Spatiotemporal Gait of Patient With Parkinson’s Disease

James Fang1, Jacob Sosnoff1, Kelly Lyons1, Rajesh Pahwa1

1Kansas University

Wearable technology is an emergent field for remote monitoring in the health care domain. Smart glasses wearable has the advantage of compliance and the ability to collect head movement information. However, previous studies on wearable application were done with sensors placed at lower back. Therefore, the objective of this study is to examine the validity of smart glasses wearable to measure spatiotemporal gait in patients with Parkinson’s disease. 32 patient diagnoses with Parkinson’s disease participated in this study. Spatiotemporal gait was measured by two devises: the smart glasses device is a eye glasses with 1 IMU sensor implanted and sample at 25 Hz(Ellcie health Inc, FR); the gold standard device is a full body wearable system with 6 IMU sensors placed at the feet, wrists, lumbar, and sternum(APDM Inc, USA)) and sample at 100 Hz. Both devices were worn by the participants simultaneously while they perform two standard walking test: 3-meter Timed Up&Go(TUG) and 7-meter Stand and Walk(SAW)). The IMU data were transferred to a personal computer and processed using a customized MATLAB code following the protocol of [1] (see Figure 40). The major outcomes were step duration, cadence, step length, swing percentage, turn duration and velocity, and step in turn for SAW task, and duration, turn duration and turn velocity for TUG task. Inter-class correlation coefficient (ICC) and Pearson correlation were performed to examine the reliability and correlation between the two devices. Absolute error and relative error rate were calculated to examine the difference between the two devices. The smart glasses wearable shown great validity in measuring step duration, TUG duration, and cadence. The poor validity on measuring step length and swing percentage may result from the sensor location is different from the placement (lower back) that the current applied algorithm was designed for. Further investigation on more suitable algorithm for smart glasses wearable is needed.

Use of Wearable Sensors to Classify Activities of Amputees in the Real-World for Improved K Level Assessment

Matthew Wassall1, Sibylle Thies1, Malcolm Granat1, Saeed Zahedi2

1University of Salford, 2Blatchford Prosthetics

The prosthesis components a lower limb amputee receives are determined by their assigned K level. K levels are assigned based on the activity levels of the patient. There are known issues with the reliability of the assessments and the consistency of K level allocations, especially when deciding between a K2 and K3. We are creating a sensor-based system to assess a patient’s activity levels in the real-world to reduce the issues with reliability and consistency during K level assignment. To fully understand the requirements of the system, interviews were conducted with clinical experts. The ability of the patient to vary their cadence, traverse different terrain, walk without a walking aid and the distance they can walk were emphasised as the main differences between a K2 and K3 patient, and would constitute the data that would be required from the proposed sensor-based system. A review was conducted to identify the specification of the system. Cadence has been measured with a shank mounted IMU, and an estimate of distance travelled can be calculated from the same data. A combination of IMUs and pressure sensors have been shown to accurately identify between flat ground, stairs, and ramps. Uneven terrain has only previously been measured on able-bodied participants. No studies could be found that identified walking aid use using body or prostheses mounted sensors. We will be creating a IMU and pressure sensor based system, and validating in controlled conditions over different terrain. A real-world study which will show how the systems data can improve the reliability of K level assignment. A group of prosthetists will assess a number of patients using traditional methods, then the patients will wear the system for a month to record their real-world activity data and another group of prosthetists will assess the same patients with the help of the data. The inter-reliability of assigned K levels will be used to assess the effectiveness of the system.

Validation of Low-Cost Measurement Tools for Assessing Habitual Physical Activity in Pre-Schoolers: The SUNRISE Study

Tawonga Mwase-Vuma1, Xanne Janssen1, Anthony Okely2, Mark Tremblay3, Catherine Draper4, Alex Antonio Florindo5, Chiaki Tanaka6, Denise Koh7, Guan Hongyan8, Hong Tang9, Kar Hau Chong2, Marie Löf10, Mohammad Sorowar Hossain11, Penny Cross2, PW Prasad Chathurangana12, John Reilly1

1University of Strathclyde, 2University of Wollongong, 3CHEO Research Institute, 4University of the Witwatersrand, 5University of Sao Paulo, 6Tokyo Kasei Gakuin University, 7Universiti Kebangsaan Malaysia, 8Capital Institute of Pediatrics, 9Pham Ngoc Thach University of Medicine, 10Karolinska Institutet, 11Biomedical Research Foundation, 12University of Colombo

Objective: This study aimed to validate parent-reported habitual total physical activity (TPA) against accelerometry and the three existing step-based TPA thresholds in a sample of pre-schoolers from geographically and culturally diverse contexts. Methods: We used data involving 352 3-4-year-olds (49.1% girls, mean age 4.4 years) from 13 middle- and high-income countries in the SUNRISE pilot study. Pre-schoolers wore an activPAL accelerometer continuously for 3-5 days (24h/day). Parents completed a brief survey and reported on their child’s TPA. Validity of parent-reported habitual TPA against accelerometry was assessed using Spearman’s rank-order correlation and Bland-Altman plots; classification ability of parent reports was assessed using Kappa statistics (κ). Receiver Operating Characteristic Area Under the Curve (ROC-AUC) analyses were used to validate three existing step-based TPA thresholds: the De Craemer 11,500 steps/day, the Vale 9,000 steps/day, and the Gabel 6,000 steps/day. Results: Spearman’s rank coefficients between parent-reported and accelerometer-measured TPA were weak though statistically significant (r: 0.140; P: 0.009). There was slight agreement between parent-reported and accelerometer-measured TPA (κ: 0.030). Parents over-estimated their child’s habitual TPA by 69 min/day compared to the activPAL data (SD: 126.3; 95% limits of agreement: -170, 316). ROC analyses revealed an AUC of 0.945 for the De Craemer threshold (sensitivity: 100.0%; specificity: 88.9%), 0.773 for the Vale threshold (sensitivity: 100.0%; specificity: 54.6%), and 0.577 for the Gabel threshold (sensitivity: 100.0%; specificity: 15.1%). Conclusions: The findings show that parent reports may have limited validity for assessing the level of habitual TPA in young children. The De Craemer threshold of 11,500 steps/day provided a valid low-cost alternative (e.g., using pedometers) and may potentially be used for population-surveillance of physical activity in young children.

Big Data & Statistical Analysis

Activity Recognition Using Body-Worn Sensors and Load-Dependent Injury Risk in Swiss Armed Forces Recruits

Regina Oeschger1, Rahel Gilgen-Ammann1

1Swiss Federal Institute of Sport Magglingen

The Swiss Armed Forces conduct basic military training (BMT) for various occupational specialties (OS). The purpose of this study was to quantify the daily physical demands of these OS, examine their impact on musculoskeletal injuries and whether certain OS are at greater risk for musculoskeletal injuries. During 18 weeks of BMT, all musculoskeletal injuries were continuously recorded from four OS (communications intelligence, n=93; armored infantry n=156; rescue technicians n=119; fusilier infantry n=111). Injuries were recorded when a recruit sustained physical damage and visited a physician at the medical center. Of each OS, 20 recruits wore a body-wearable sensor system during BMT week 2-9, which included 2 triaxial accelerometers (Axiamo PADIS 2.0) and a heart rate monitor (Rhythm24). From 479 subjects (20.0±1.2 y, 178.3±6.5 cm, 73.6±10.4 kg) 258 injuries (111 acute, 147 overuse injuries) with an injury incidence rate (IIR; injuries per 100 recruits/week) of 3.0 (1.3 acute and 1.7 overuse injuries) and an injury incidence proportion (number of injured recruits) of 39.2% (19.2% acute, 24.4% overuse injured) were recorded. Post-hoc analysis showed that the IIR of the communications intelligence was significantly (p<0.02) lower from all other OS (0.6±1.1 vs. 4.6±2.7 to 6.1±2.6). Significant correlations between the acute IIR and intense activity (p=0.02, r=0.43), marching with backpack (p=0.02, r=0.45) and running (p<0.05, r=0.37), respectively were observed. The overuse IIR showed no correlation to any activity class. A total of 87 (18.2%) recruits left the assigned OS before the end of BMT. Time spent in marching with a backpack, intense activity and running have an impact on acute injuries during the BMT. A more detailed analysis with the inclusion of the recruits’ physical performance would give further indications about the occurrence of overuse injuries. This data will be presented at ICAMPAM 2022.

An Open-Source and Automated Data Processing and Reporting Pipeline for Continuous Wearable Data in Adaptive Interventions

Ethan Rogers1, Diego Arguello1, Grant Denmark1, Gregory Cloutier1, Carmen Castaneda-Sceppa1, Charles Hillman1, Arthur Kramer1, Dinesh John1

1Northeastern University

Objective: Adaptive interventions involve continuous sensor data gathered over several months and simultaneous remote engagement with multiple participants to adapt and deliver an intervention. This requires powerful backend capabilities for storage and analyses. Here, we describe an automated system to integrate independent software tools to seamlessly process high-resolution sensor data into meaningful metrics for near-real-time use. Methods: The system is flexible and interchangeable for specific project needs. It accepts multiple formats of raw data, which is indexed into a simple file system that can then be used to derive desired metrics. The system uses data analyses libraries to integrate machine learning algorithms and is optimized for local or high-performance cloud computing. Results: The system is being piloted in a 4-month clinical trial investigating the effect of tailored health coaching to improve daily activity in adults>60 y (N=46). The backend receives continuously worn wrist sensor data each hour and pipeline processing (Figure 41) detects anomalous signals, non-wear, sleep, and waking behavior. It creates a database of labeled behaviors by participant for the study duration, which can be queried for visualization. Deployability on the cloud is a major advancement in facilitating near-real-time adaptive intervention delivery. For e.g., processing 24h data on a local computer and the cloud takes 11 and 8min, respectively. However, the latter enables efficient and simultaneous data processing from a much higher number of participants as tasks can be spread across multiple nodes. This also bypasses inefficiencies due to local storage limitations; our study has generated 16 person-years of raw sensor data. Conclusions: Our system automates multiple steps of data processing, thus increasing the efficiency of delivering adaptive interventions. Future goals involve system refinement and optimization to visualize behavior and communication history.

Functional, Distributional, and Dynamic Modelling for Mobile Digital Health Data

Vadim Zipunnikov1

1Johns Hopkins University

Wearable and smartphone data is a rich source of information that can deepen our understanding of links between human behaviors and health. Existing modelling approaches typically use data summarized at subject-level via scalar measures. We present recently developed novel statistical techniques suitable for modelling more comprehensive subject-level summaries such as temporal (time-of-day/diurnal) curves via functional regression, distributions of wearable-estimated quantities such as bout duration or activity counts via distributional regression, and dynamic interactions between actigraphy-estimated sleep and motor activity and within-day mood and energy reported in smartphone apps via dynamic structural equation modelling. We illustrate these approaches using data from Baltimore Longitudinal Study of Aging and NIMH Family Study.

Population Activity Profiles: Comparison of Standard Time to Relative Time

Malcolm Granat1, Emmanuel Stamatakis2, Mark Hamer3, Ben Griffiths1

1University of Salford, 2University of Sydney, 3University College London

Objective: Profiles of free-living physical behaviour over a day, for a population, show how physical activity levels vary over time by the hour of the day. These population profiles show a gradual increase in activity at the beginning of the day, taking several hours to reach a maximum. Whilst this shows the variation of activity over the hours of a day, it might not reflect the behaviour of individuals. The study aimed to look at these population profiles when activity is aligned to each individual’s time out of bed. Methods: Data used were from 4,979 individuals in the 1970 British Cohort collected between 2016 and 2018. For these people, a thigh-mounted triaxial accelerometer (activPAL) was used to collect objective physical activity data over 7 days. For all days, time out of bed was determined from the event files, and activity (steps per hour and upright time) was referenced to time out of bed. The average number of steps and percentage time upright, per day per individual, for each 30-minute period, was then determined. Mean and standard errors, across all days, for each 30-minute bin calculated, separately for both weekdays and weekends. Results: 22,088 week days and 8,950 weekend days were used. The weekday and weekend mean steps per day were 4,725 and 4,549, respectively. The weekday and weekend mean upright times were 388 minutes and 371 minutes. Profiles of steps per hour and percentage time upright for both standard time and time relative to getting out of bed are shown in Figure 42. Conclusions: Population activity profiles for both stepping and upright time were very different when referenced to time out of bed. Peak values for upright time and number of steps per hour occurred shortly after getting out of bed within 30 to 60 minutes.

Using Deep Learning in Stroke Rehabilitation - A Step Towards Individualized Patient Care

David Sina1, Sonja Georgievska2, Yang Lui2, Cunliang Geng2, Michiel Punt3

1Vrije Universiteit Amsterdam, 2Netherlands eScience Center, 3Hogeschool Utrecht

Purpose: This study aimed to explore the potential of variational autoencoders (VAEs) to allow an holistic view on the individual without the need to preselect parameters for gait evaluation. Methods: 422 files containing the sagittal lower-limb joint angles of 31 stroke patients (age 59.3 ± 12.1 years) and 42 healthy controls (age 42.6 ± 14.1 years) collected during treadmill walking were used to train a VAE in Python. The VAE consisted of an encoder, used to extract the latent variables from the original joint angle data and a decoder, trained to reconstruct output similar to the original joint angle data using the two latent variables from the encoder. The reconstruction accuracy of the decoder was assessed using the Pearson correlation coefficients and the (normalized) RMSE, calculated between the input and the reconstructed time series. The latent space characteristics were explored by comparing the joint angles of selected participants on the 2D map. Results: The stroke patients are located along a diagonal axis on the 2D latent space. The Pearson’s correlation coefficient between the input and the reconstructed angles was 0.84 ± 0.19 (range 0.93 - 0.7). The RMSE was 7.4 ± 0.9 degrees (normalized RMSE: 21.9 %). The outer edges of the latent space are characterized by reduced ankle plantar flexion and gait speed and increased left-right asymmetry. Comparing the upper half of the map to the lower one results in an decrease of hip extension. Discussion and Conclusion: VAE are a promising tool for gait analysis. There is no need for preselection of parameters and the whole movement can be considered as the interaction between the joint angles. By reconstructing the angles from the latent space, the characteristics of patients can be analyzed and used for individualized patient care. (Figure 43)

Using Machine Learning to Classify Sitting and Sleep History From Raw Accelerometry Data During Simulated Driving

Georgia Tuckwell1, Charlotte Gupta1, James Keal2, Sally Ferguson1, Jarrad Kowlessar3, Grace Vincent1

1Central Queensland University, 2University of Adelaide, 3Flinders University

Breaking up sitting and adequate sleep are important for optimal performance of tasks such as driving. Objective knowledge of a driver’s recent sitting and sleep history could help manage safety risks. However, methods for classifying these behaviours during a driving task can be expensive and labour intensive. Using machine learning (ML), it is possible to classify breaking up sitting and sleep history from wearables such as accelerometers. The aim of this study was to apply ML to raw accelerometry data collected during a simulated driving task to classify recent sitting and sleep history. Participants (n=84, aged 18-35 years, 49% female) participated in a 7-day laboratory study. The study was a between-subjects 2 × 2 design with participants allocated to one of four conditions (sitting or breaking up sitting) and (5-h or 9-h sleep opportunity). Raw accelerometry data was collected from a thigh-worn accelerometer during two 20-min simulated driving tasks performed twice per day (8:10 h and 17:30 h). Data were analysed using two convolutional neural networks (CNN’s) (ResNet-18 and DixonNet) to determine classification accuracy of the four conditions, and an 80/20 splitting procedure was performed using 5-fold cross validation. ResNet-18 produced higher accuracy scores for both sleep opportunity (Mean ± SD, 87.0± 2.1%) and sitting/breaking up sitting classification (86.3±2.4%), compared to DixonNet, which produced accuracy rates of 78.1±4.2% and 78.0±2.7% for sleep opportunity and sitting/breaking up sitting respectively. This study demonstrates the suitability of CNN’s in classifying sitting and sleep history via thigh-worn accelerometers during a simulated driving task. This approach could be used in real world scenarios for the identification of fatigued and impaired drivers using any wearable sensor containing an accelerometer. This feasible and cost-effective approach to identify at-risk drivers could have implications for future road safety research.

What Does It Mean to Use the Mean? The Impact of Different Data Handling Strategies on the Percent of Children Classified as Meeting the 24-hr Movement Guidelines

Christopher Pfledderer1, Sarah Burkart1, Roddrick Dugger1, Hannah Parker1, R. Glenn Weaver1, Bridget Armstrong1, Elizabeth Adams1, Michael Beets1

1University of South Carolina

Despite the widespread endorsement of 24-hour movement guidelines (24hrG), there is no standardized way to process movement data prior to classifying participants as meeting guidelines. Variability in data handling procedures (averaging data across multiple days; classifying as meeting/not meeting guidelines for each monitored day) may produce vastly different estimates of guideline compliance. The purpose of this study is to illustrate the impact of different data handling strategies on the estimated proportion of children meeting 24hrG. A subset of 524 children (ages 5-12) from an observational cohort with 24-hour behavior measures on at least 10 days was used to compare different data handling strategies’ impact on estimates of meeting 24hrG. Physical activity and sleep were measured via wrist-worn actigraphy. Screentime was measured daily via parent self-report. Comparisons of meeting 24hrG were made using 1) average of behaviors across all measured days (AVG-24hr), 2) classifying each separate day as meeting/not meeting the guidelines and evaluating the percentage meeting the guidelines from 10-100% of their measured days (DAYS-24hr), and 3) the average of a random sample of 4 days (3 weekdays and 1 weekend day) across 10 iterations (RAND-24hr). AVG-24hr resulted in 14.7% of participants meeting 24hrG. DAYS-24hr resulted in 63.5% meeting 24hrG on at least 10% of measured days, 13.9% meeting 24hrG on at least 50% of measured days, and <1% meeting 24hrG on 100% of days. RAND-24hr resulted in 15.9% of participants meeting 24hrG. Across all 10 iterations, 63.6% of participants never met 24hrG regardless of the days sampled, 3.4% always met 24hrG, and 33.0% were classified as meeting 24hrG for at least one of the 10 iterations of days sampled. Different data handling strategies and the number of days sampled produce varying estimates of children meeting 24hrG. To compare estimates across studies, standardization of data handling strategies is necessary. (Figure 44)

Technology & Algorithm Development

Can a Perceptual Threshold be Identified to Distinguishing Walking on Flat Ground From Uphill and Downhill?

Philippa Dall1, Anna Iveson1, Brian Ellis1, Malcolm Granat2

1Glasgow Caledonian University, 2Salford University

Introduction: Body-worn sensors (e.g. GPS) can distinguish between walking uphill and downhill, but it is unclear how to identify when someone is walking on flat ground as this is unlikely to be precisely zero. This study explored the feasibility of identifying a threshold based on human perception to distinguish between flat ground and a slope. Methods: The perception of walking was conducted on two routes in the UK, selected for convenience. Researchers (n=3) stated their perception of gradient in their own words and identified the location of changes in perceived gradient on a map. The actual gradients of segments were calculated from Open map. A threshold for flat walking was selected by identifying where perceptions changed between flat and slopes. Finally, the GPS data from second route was used to calculate gradient between successive points (smoothed with a moving average window of 7) to assess concordance with observation. Results: Route 1 had fifteen segments identified as having different gradients, ranging from -3.8% to + 4.9%. Six segments were perceived as 'down’ or 'slightly down', four segments were perceived as 'flat', one as 'very gradually up', and four segments as 'up’ or 'slightly up'. The segment perceived as 'very gradually up’ had a gradient similar to other segments perceived as 'flat', and was grouped with those. Route 2 had eleven segments, three perceived as 'down', one as 'slightly down or flat’ and seven as up or 'slightly up'. A gradient threshold of ±1.3% distinguished between flat, downhill and uphill segments, and the GPS signal provided concordance. Discussion: Perceptions of walking on flat ground versus uphill and downhill were relatively similar, and changed at a gradient of approximately ± 1.3%, which could be used in future studies to explore differences in walking. The threshold should be confirmed using the perceptions of a wider group.

Can We Use a Wearable Sensor to Determine the Locus of a Person’s Activity?

David Loudon1, Douglas Maxwell1, Craig Speirs1

1PAL Technologies Ltd

Objective: Community activity has many benefits for health, so to quantify the proportion of the day spent in the primary (household) locus could be an important clinical outcome. This study describes an approach to activity data analysis which identifies and characterises individual loci of activity and the transitions between them. Methods: The overnight locus is defined as the primary locus. In previous work, upright bouts with all stepping bouts <1m were found to be associated with the household locus. In addition, we identify that the number of steps that can be taken between turns is constrained by the physical layout of indoor spaces. Locus transitions are defined as: -Upright bouts which include stepping bouts with straight line walking >= 1m -Cycling -Transportation. To demonstrate the potential of this approach, an example day was selected from free-living data from three different people: an older adult, a working age adult and a high school student. Results (see Figure 45): Examples A and B show similar constrained loci for most of the day typical of a household locus. In Example A, there is a change of locus identified, but via short car journeys, leading to a low total step count of 3602. However, in Example B, there are two periods of outdoor walking, defined by several minutes of straight-line walking accounting for 5324 out of the total of 8132 steps for the day. Example C shows several loci, including home, school, and after school. Crucially, the transition between loci is by walking, accumulating 13380 steps for the day. Conclusions: The use of heading information provides a means to characterise the locus of stepping behaviours which is not easily discriminated using acceleration alone. This approach offers the potential to identify time spent in the home vs community locus as an important clinical outcome measure. Further, the separation of the stepping data into loci has the potential to characterise outdoor stepping and its relation to MVPA.

Classification of Daily Physical Behavior in Older Adults Using Machine Learning

Astrid Ustad1, Stine Trollebø1, Aleksej Logacjov1, Kerstin Bach1, Beatrix Vereijken1, Nina Skjæret-Maroni1

1Norwegian University of Science and Technology

Background: Physical activity monitoring combined with machine learning (ML) methods can contribute to more detailed knowledge about physical behavior. However, extant models are typically developed and validated on datasets from young, healthy adults. It is an open question to what extent such models accurately identify daily living activities in heterogenous older adults. This study 1) evaluates the performance of an existing activity type recognition ML model, based on data from healthy adults and classifying activity categories with high accuracy (Bach et al., 2022), in a sample of fit-to-frail older adults, and 2) uses the sample dataset to further develop the ML model. Methods: The sample included 18 older adults aged 70-95 years (79.6±7.6 years; 50% female) with a wide range of physical function, including 5 participants using walking aids. Activity was recorded using two Axivity AX3 accelerometers (thigh and lower back) and a chest-mounted camera pointing downwards during a semi-structured free-living protocol that included repetitions of walking, standing, sitting, and lying. Video recordings were labelled according to pre-defined activity definitions and were used as gold standard. Results: The overall accuracy of the original model on this dataset was 85% (87% without walking aids and 80% with walking aids), which improved to 94% (94% without walking aids and 92% with walking aids) after data of older adults was included in the training. Conclusions: Classification of daily physical behavior in older adults was considerably more accurate when the model was trained on data from older adults, especially for the frailest participants using walking aids. This improvement illustrates that it is necessary to use training data sets that are representative for the population of interest. The resulting validated ML model for fit-to-frail older adults may contribute to accurate and detailed knowledge of daily physical behavior that is essential for future research. (Figure 46)

Classifying Activity Intensity in Children With Spina Bifida Based on Wrist-Worn ActiGraph

Zijian Huang1, Andrea Moosreiner2, Michele Polfuss3, Dan Ding1

1University of Pittsburgh, 2Medical College of Wisconsin, 3University of Wisconsin

Background: Obesity prevalence is 40% higher in children with spina bifida (SB), partly due to decreased physical activity (PA). Assessing PA using wearable devices can be challenging in this population because of their varied mobility impairments, resulting in ambulation with difficulty, ambulation with a walker or crutches, or using a wheelchair for mobility. PA intensity classification thresholds for wrist-worn ActiGraph devices in this population are missing. Existing thresholds such as Chandler’s wrist-worn ActiGraph count-based thresholds for children with normal walking capacity and Shwetar’s raw acceleration-based mean absolute deviation (MAD) thresholds for adult manual wheelchair users (MWU) have not been evaluated in children with SB. Objective: To evaluate the validity of the Chandler ActiGraph vertical axis count (VAC) and the Shwetar MAD activity intensity thresholds in children with SB. Method: The thresholds were applied to wrist-worn ActiGraph accelerometer data from 15 ambulatory children and 11 MWU children during four structured activities. Criterion intensities were obtained as sedentary, light, and moderate to vigorous PA (MVPA) using Butte’s and Conger’s physical activity compendium for children and MWUs, respectively. Threshold performance was evaluated using accuracy, balanced accuracy, recall, precision, and specificity. Result: The Chandler VAC achieved an overall accuracy of 63.5%, with a sedentary classification accuracy of 82.0% and an MVPA classification accuracy of 80.0%. The Shwetar MAD achieved an overall accuracy of 83.2%, with a sedentary classification accuracy of 95.0% and an MVPA classification accuracy of 88.3%. Figure 47 contains detailed results. Conclusion: The Shwetar MAD can reliably classify sedentary, light, and MVPA activities in children with SB who use a wheelchair. The Chandler VAC lacks overall accuracy to classify PA into three intensities, but can classify sedentary activities in ambulatory children with SB.

Comparison of an Automated Algorithm Applied to activPAL Data for Estimating Time in Bed With Polysomnography

Tatiana Plekhanova1, Alex Rowlands1, Tom Yates1, Andrew Hall1, Melanie Davies1, Charlotte Edwardson1

1University of Leicester

Objective: Thigh-worn accelerometers such as the activPAL are increasingly used in large-scale studies and can capture the whole 24-hour physical behaviour cycle. However, despite capturing the sleep period, studies are not reporting sleep data due to the difficulty in identifying this behaviour within the data. This study aimed to compare estimates of time in bed (TIB), bedtime, and waking times from an automated algorithm (within Processing PAL) applied to activPAL data with concurrent polysomnography (PSG). Methods: Twenty-seven healthy volunteers (age 31.1±6.8 years, BMI 25.1±3.5 kg/m2) wore an activPAL accelerometer on the thigh during one-night laboratory-based PSG assessment. TIB, bedtime, and waking time estimates were generated using an automated algorithm in 'Processing PAL’ previously validated against diaries. Agreement between Processing PAL and PSG was determined using pairwise 95% equivalence tests (±10% equivalence zone), intra-class correlation coefficients (ICCs), and limits of agreement (LoA). Results: TIB, bedtime, and waking time were within the proposed ±10% equivalence zone of the PSG. Reliability between Processing PAL and PSG was moderate for TIB (ICC=0.68) and good for bedtime (ICC=0.73) and waking time (ICC=0.82). The mean bias were -30.9, 8.4 and 13.8 minutes for TIB, bedtime, and waking time, respectively. Wide 95% LoA were observed for TIB (±61 minutes) and bedtime (±66 minutes). Conclusions: The automated algorithm within Processing PAL provides comparable estimates of TIB, bedtime, and waking time with PSG. Future studies should continue to improve algorithms for estimating TIB for thigh-worn accelerometers and develop methods that allow discrimination of periods of sleep and wakefulness.

Comparison of Physical Activity Intensity Estimated By Direct Observation to Whole Room Indirect Calorimetry

Julian Martinez1, John Staudenmayer2, Edward Melanson3, Ann Swartz1, Scott Strath1

1University of Wisconsin, 2University of Massachusetts Amherst, 3University of Colorado

Objective: To validate video-recorded direct observation (DO) physical activity intensity estimates to whole room indirect calorimetry (IC). Methods: Participants (n = 10, mean age 45.5 ± 21.5 years, mean BMI 24.4 ± 4.5) stayed within an IC room for 12 waking hours while being video recorded. Participants were asked to perform a variety of daily activities without instruction on duration and timing. DO videos were annotated using the Noldus Observer XT 14 for activity behavior and sedentary, light and moderate to vigorous physical activity (MVPA) intensity. MET cutoffs of < 1.5 for sedentary, 1.5 - 2.9 for light and ≥ 3.0 for MVPA were used for intensity classifications of behaviors ≥ 1 minute to compare DO and IC. Overall percent agreement was calculated and a confusion matrix was computed to determine agreement and misclassification by intensity category. Bias was computed as the mean of DO METs minus IC METs within each intensity category. A t-distribution was used to make 95% confidence intervals for the bias estimates. Activity behaviors were examined to see where intensity misclassifications occurred. Results: 5967.2 minutes of DO video were analyzed. The average duration for each intensity from IC was 336.7, 201.5 and 64.5 minutes for sedentary, light and MVPA, respectively. Overall percent agreement was 78.3% whereas within intensity agreement was 98.9%, 41.9% and 81.5% for sedentary, light and MVPA intensities, respectively. DO bias was 100.5 (-64.4, 265.5), -106.2 (-252.7, 40.3) and 0.34 (-28.0, 27.4) minutes for sedentary, light and MVPA, respectively. The majority of overestimation of sedentary and underestimation of light intensities occurred when participants were observed sitting/lying using electronics or an inactive leisure activity such as reading. Conclusions: Intensity estimated from DO is accurate in estimating MVPA intensity. Future work is warranted to delineate the accuracy of DO light intensity estimations for non-sitting/lying activities.

Comparison of Raw Accelerometer Data of Three Different Devices Using a Mechanical Orbital Shaker

Theresa Schweizer1, Rahel Gilgen-Ammann1

1Swiss Federal Institute of Sport Magglingen

To monitor fatigue, gait patterns extracted from raw acceleration (Araw) are promising markers (Baghdadi, A., et al., 2021). ActiGraph wGT3X-BT (AG), GENEActiv original (GAO) and Axiamo PADIS 2.0 (AXP) are accelerometers with Araw outputs. This study aimed to compare the devices’ Araw to evaluate accuracy and reliability. An AG, GAO and AXP were placed on a mechanical orbital shaker (MOS). The MOS oscillated on a fixed radius of 35.5mm at 21 frequencies ranging from 0.42 to 4.00Hz for 60s each. The devices recorded Araw at their maximal dynamic range: AG ±8g, 100Hz sampling rate (SR); GAO ±8g, 100Hz SR; AXP ± 8g, 64Hz SR. Araw was summarized into unfiltered normalized magnitude (1-s epochs). The differences between lowest and highest sinusoidal peak acceleration values in each revolution were averaged per frequency to calculate mean peak to peak acceleration (PPA) within each device. For the criterion, centrifugal acceleration was calculated: ((ω2*r+ 9.81)+(ω2*r-9.81))/9.81[g]. Criterion validity was assessed using mean difference (MD), mean absolute error (MAE) and percentage error (MAPE) as well as Pearson correlation. Intra-instrument reliability was assessed via the coefficient of variation (CV). Due to missing values in the AG, the lowest frequency was removed. All devices correlated significantly with the MOS (p<.001): r=.94 (AG), r=.97 (GAO), r=.98 (AXP). MD in the PPA to the criterion were -0.31g, 0.51g, 0.30g, MAE (MAPE) were 0.36g (14%), 0.60g (122%), 0.48g (102%), and mean CV were 4.05%, 6.07%, 7.21% for AG, GAO, AXP, respectively (Figure 48). All devices showed a high correlation with the MOS. Only the AG recorded accurate Araw until 2.68 Hz, yet showed difficulties in higher frequencies. AG revealed the lowest variation. Low CV are crucial for gait pattern analysis to detect the pattern changes related to e.g. fatigue. Thus, we recommend the application of filters to possibly reduce variability. Also, inter-device comparability cannot be recommended.

Criterion Validity of Activity Monitors and Processing Methods to Assess Daily-Life Walking Bouts

Adrien Chanteau1, Antoine Meliand1, Muriel Pressigout1, Thomas Bourgoin1, Romane Clouet1, Angéline Magois1, Alexis Le Faucheur1

1University of Rennes

Introduction: A paradigm shift in physical activity (PA) research and recommendations have occurred with the latest guidelines, since now “each PA bout counts”, even short ones. Accurately assessing these daily-life intermittent walking bouts of different durations and speeds relies on the validity of both activity monitors and processing methods. However, most of the available validation studies were conducted in standardized contexts that were not fully representative of daily-life context. We aimed to determine the criterion validity of different activity monitors and processing methods in the assessment of daily-life walking bouts. Methods: Ten healthy young adults wore an activPAL4 (thigh level), the new StepWatch4 (ankle level), an ActiGraph wGT3X-BT (hip level) accelerometers, and a video camera (hip level) during free-living periods (3x ∼1h30) embracing different PA domains. From the video recordings, actual steps and then walking bouts were identified by an experimenter using a self-made software for video analysis and labelling (PANAMA software). From activity monitors data, the following bouts detection methods were tested: the Orendurff method (2008) from steps, the watershed algorithm (Taoum et al., 2021) from raw acceleration, counts and steps, and the activPAL proprietary algorithm. The walking bouts detection rates [95% IC] obtained by each method were calculated and compared to the actual bouts using the McNemar test. Results: Experiments were on-going at the time of submission; results will be presented at the conference. Discussion: These preliminary data are part of an on-going larger research project including additional activity monitors and bout-level analysis methods in the assessment of daily-life walking bouts in a larger population (aged 20-80). From an epidemiological perspective, a better assessment of short and incidental intermittent PA bouts would contribute to a better knowledge of their potential health enhancing properties effect.

Definition of Activity Counts Using ActiGraph Devices

Tyler Guthrie1, Ali Neishabouri1, Joe Nguyen1, John Samuelsson2, Tyler Guthrie1, Matt Biggs1, Jeremy Wyatt1, Doug Cross1, Marta Karas3, Jairo Migueles4, Sheraz Khan2, Christine Guo1

1ActiGraph, 2Massachusetts General Hospital, 3Harvard University, 4Karolinska Institutet

Research Objectives: To define and provide documentation for a method of converting the raw acceleration data using ActiGraph devices to “counts”. This has previously been held as proprietary company information. This lack of transparency has hindered comparability between studies using different devices and limited their broader clinical applicability. Methods: We reviewed internal documentation to assess the different processing steps for each device model. The ActiLife algorithm that calculates raw to epoch data was translated into Python code. The results of the Python implementation were compared to the output of ActiLife and CentrePoint. We have published this implementation under an open source license. The code was tested for all model specific sampling frequencies and varying epoch length durations. Results: We present flowcharts outlining the high-level differences between device models around the accelerometer, device processing, and cloud processing steps. The Python code was validated to ActiLife software and CentrePoint outputs. No differences were found between any of the three implementations. Conclusions: The term “counts” is used by multiple different device manufacturers, but the methods in which they are derived are not consistent between those manufacturers. Allowing researchers to have access to data processing steps will allow them to better compare data between studies using different devices, as well as historical datasets. We believe that this material will provide a useful resource for the research community, accelerate digital health science, and facilitate clinical applications of wearable accelerometry.

Evaluation of Range of Motion During Daily Activities Through Knee-Worn Wearable Sensor System After TKA Surgery

Emily Hampp1, Ricardo Antunes1, Paul Jacob2, Robert Marchand3, Andrew Meyer1, Elaine Justice2, Kelly Taylor3, Kelly Luttazi3, Matthias Verstraete1

1Stryker, 2Oklahoma Joint Reconstruction Institute, 3Ortho Rhode Island

Introduction: Remote patient monitoring through wearable devices and connected patient engagement platforms provides clinicians with additional visibility to patients’ recovery, such as knee angle metrics during activities of daily living. This study presents range of motion measurements made using a novel knee-worn wearable system used for remotely monitoring patients’ recovery after surgery and published reports. Methods: A total of 92 patients were enrolled in an observational study two to four weeks prior to undergoing unilateral total knee replacement surgery. Patients were instructed on the use of a wearable sensor system which consists of two wireless inertial sensor nodes that regularly measure the patient’s knee flexion angles. Hourly 5-degree bins were assessed and compared with published TKA recovery progression studies. Results: Sensor reported mean minimum and maximum knee flexion angles show a clear recovery trend following surgery (Figure 49). The wearable maximum flexion trend aligns very well with the Kittelson (2020) recovery model throughout the postoperative 90 days considered, and with both Zhou et al (2015) and Mutsuzaki et al (2017) at 30 days postoperative. The latter study reported ∼20° less flexion than the wearable mean at day one after surgery. With respect to extension, mean wearable measurements tended to be below published reports, but showed a similar increase in extension range following surgery, particularly in the early post-operative period. Conclusion: We have shown that range of motion monitored remotely through a wearable sensor shows variation consistent with the recovery expectations, and thus it is useful in supporting clinical decisions and facilitate the identification of patients with an abnormal recovery. While our trend was similar to some published reports, it differed from others, but the magnitude of these differences falls within the differences among the considered studies themselves.

Evidence for Accelerometry as a Surrogate Measure of Infant Energy Expenditure

Emily Flanagan1, Nicholas Broskey2, Samuel LaMunion3, Abby Altazan1, Leanne Redman1

1Louisiana State University, 2East Carolina University, 3National Institute of Diabetes and Digestive and Kidney Diseases

In young infants actigraphy data is rare and there are no published studies of actigraphy in combination with total daily energy expenditure (TDEE). Together such data can be used to understand the contribution of physical activity (PA) to EE in this population and to develop algorithms for PA EE and TDEE that can be used by the field. The aim of this analysis is to explore the relationship between TDEE in infants and actigraphy measured at two distinct anatomical locations. This dual monitor approach was chosen to help understand the unique movement patterns that occur due to passive movement through caregiver handling and prolonged periods of sleep. Fourteen infants (8 M/8 F) aged one to three months completed a 7-day doubly labeled water study while wearing wGT3X-BT on the hip and ankle. 7-day data were cleaned to remove activity counts not feasibly produced by the infant (vector magnitude; VM >400 for waist and >2250 for ankle) and non-wear time. VM (counts per minute) were calculated individually to produce average daily VM for 1) ankle, 2) hip, and 3) the difference between ankle and hip (A-Hdiff). Regression analyses were used to determine if accelerometry together with easily obtained infant anthropometrics significantly predicted infant TDEE. Average TDEE was 435±157 kcals/day. Mean 7-day ankle and hip VM were correlated (r=0.86, p<0.001) but significantly differed (Ankle: 155.8±35.8 and Hip: 29.7±13.7, p<0.001). There were no relationships between TDEE and ankle or A-Hdiff VM. There was trending correlation for TDEE and hip VM (r=0.50, p=0.06). Hip VM in conjunction with length (TDEE=2849 + 2.5WaistVM + -46.9length) explained 54% of the variance in TDEE (F(2,14)=7.13, p=0.009). Age, sex, or weight were not significant predictors of TDEE in combination with VM. Hip-worn actigraphy predicted TDEE greater than ankle or A-Hdiff. Further analysis is warranted, yet these data provide an intriguing first glimpse to the use to infant actigraphy to predict TDEE.

Exploratory Analysis: Number of Days Required to Reliably Estimate Workplace Physical Behaviours and Sedentary Time Using Three Weeks of activPAL3 Objective Accelerometry

Aidan Buffey1, Brian Carson1, Alan Donnelly1

1University of Limerick

Study’s Objective: Previous research has examined the number of days required to reliably estimate habitual physical behaviours, such as Light Intensity Physical Activity (LIPA) and time spent sedentary (ST). However, to the authors knowledge, no previous studies have examined the number of days required to reliably estimate ST and LIPA during working hours (WH). Methods: Adults (>30 years) working from home (n = 28), with desk-based occupations, wore an activPAL 3M accelerometer, using a 24-hour wear protocol for three weeks. Participants completed a diary of their WH, allowing the differentiation of WH vs. non-WH. ST and LIPA were expressed as a percentage of the participants WH. Using 10-working days of accelerometer data we calculated single day intraclass correlations (ICC) values (Model: Two-way mixed effects - absolute agreement (ICC: 2,1)) to determine the reliability of using any single day to estimate workplace LIPA and ST. The Spearman-Brown Prophecy formula was then used to calculate the number of days required to reach a desired ICC of 0.8, estimated days = [0.8 * (1 −SingledayICC))/SingledayICC * (1 −0.8)]. Results: Day-to-day variability was assessed using a repeated measures ANOVA, which found a significant difference between days for ST (see Figure 50). No difference between days were observed for LIPA. The Spearman-Brown Prophecy Formula calculated that to achieve a reliability of 0.8, a minimum of 7 days is required to estimate workplace ST (Single Day ICC: 0.386) and a minimum of 5 days is required to estimate workplace LIPA (Single Day ICC: 0.448). Conclusions: The minimum days required to estimate workplace LIPA is less than ST (5 vs. 7 working days). These findings imply that researchers who are focused on workplace physical behaviours should record a period longer than a 7-day week to capture 7 full working days to produce reliable results.

Feeling Better After TKA: Reference Chart for Remotely Collected Pain Scores

Emily Hampp1, Ricardo Antunes1, Paul Jacob2, Robert Marchand3, Elaine Justice2, Kelly Taylor3, Kelly Luttazi3, Andrew Meyer1, Matthias Verstraete1

1Stryker, 2Oklahoma Joint Reconstruction Institute, 3Ortho Rhode Island

Introduction: Remote patient monitoring, using wearable devices and connected patient engagement platforms, has the potential to improve timely clinical decisions. There is a lack of normative references for pain scores after total knee arthroplasty (TKA). This paper presents a normative recovery model for pain scores collected remotely from TKA patients using a remote data. Methods: The VAS pain scores on a 10-point Likert scale were analyzed for TKA patients enrolled in an observational study. Scores from 65 patients reporting at least five scores in the 90 days following surgery were used to produce a normative recovery model. The model was fitted using a state-of-the-art Bayesian approach allowing for estimation of the fitted parameters’ uncertainty, including the neural network weights. Model fitting was carried out using Stochastic Variational Inference (maximizing the evidence lower bound, ELBO) with custom scripts in Python. Results: The population mean trend shows an increase in pain in the first few days following surgery, higher than the preoperative mean, with wide dispersion showing scores ranging throughout the 10-point scale (Figure 51). After the first week, the expected pain score steadily decreased, resulting in a score no higher than two in 50% of the population beyond 90 days after surgery. The fitted model allows referencing individual patient’s pain scores at different stages of recovery, against the model’s predicted distribution. One patient presented with pain scoring outside the model’s expectation coinciding with infection, illustrating the potential use of normative models to alert the occurrence of possible complications during recovery. Conclusion: We produced a useful reference curve for TKA patient pain scores collected remotely through engagement platforms. This can support early detection of patients that significantly deviate from the reference model and be a useful integration into clinical decision support software tools.

Free-Living Validity of Energy Expenditure Estimates From Wrist-Worn ActiGraph Monitors: A Doubly Labeled Water Study

Paul Hibbing1, Gregory Welk2, Robin Shook1

1Children’s Mercy Kansas City, 2Iowa State University

Wrist accelerometry is now a mainstay of physical behavior research, including assessments of energy expenditure (EE). However, there is limited research validating wrist-based free-living EE estimates against criterion values from doubly labeled water (DLW). Purpose: To perform a DLW validation of 8 EE models for wrist-worn ActiGraph devices. Methods: 24 adult participants (58% female; 45% overweight/obese; age 21-52 years) wore an ActiGraph GT9X on the non-dominant wrist throughout two 14-day DLW assessments. Mean daily EE (kcal/day) was compared between DLW and the following ActiGraph-based models: Hildebrand equations (linear and non-linear), Hibbing two-regression models (left wrist and right wrist), Staudenmayer wrist models (linear model and random forest), and Montoye neural networks (left wrist and right wrist). The focal performance metrics were mean bias and mean absolute percent error (MAPE), which were tested using null-formulated mixed effects models to account for repeat measurements. Results: Summary statistics are shown in the Figure 52. Mean bias for the Hildebrand and Hibbing models was within 109-167 kcal/day of DLW (p = 0.05-0.18), versus 448-586 kcal/day for the others (all p < 0.001). Similarly, MAPE was 18.1%-19.3% for the Hildebrand and Hibbing models, versus 29.5%-34.7% for the others (all p < 0.001). Conclusions: Accuracy was variable across the different models as well as across individuals. However, handedness did not appear to influence accuracy, as shown by the similar performance of left- versus right-wrist models for both the Hibbing and Montoye methods. Models need to be carefully selected in future studies of free-living EE. Refinements are also needed to make these models easier to use, including frameworks for accomplishing common tasks in free-living EE assessment (e.g., imputing EE values during non-wear periods).

Improving Energy Expenditure Estimation With Wearables Measuring Physiological Signals

Wouter Bijnens1, Kenneth Meijer1, Guy Plasqui1

1Maastricht University

Regular physical activity (PA) is proven to help prevent and treat several non-communicable diseases such as heart disease, stroke, and diabetes. Intensity is a key characteristic of PA that can be assessed by estimating energy expenditure (EE). However, the accuracy of the estimation of EE based on accelerometers are lacking. It has been suggested that the addition of physiological signals can improve the estimation. How much each signal can add to the explained variation and how they can improve the estimation is still unclear. Therefore, the goal of the current study is twofold. Firstly, to explore the contribution of heart rate (HR), breathing rate (BR) and skin temperature to the estimation of EE. Secondly, develop and validate a statistical model to estimate EE in simulated free-living conditions based on the relevant physiological signals. For this study, 56 healthy subjects, aged 18-64 will be recruited. Accelerometers (MOX1, Maastricht Instruments), placed on the upper leg, hip, wrist, and chest will assess PA. Additionally, a research grade ECG and bio-impedance logger will be used for the collection of HR and BR. iButtons on the shin, thigh, shoulder, and chest will be used to measure average skin temperature. Next to that, body composition will be measured with deuterium dilution according to the Maastricht Protocol. A respiration chamber (Room Calorimeter, Maastricht Instruments) will be used as a criterion measure for EE. Subjects will spend approximately 8 hours in this chamber. After the measurement of the resting metabolic rate, subjects will perform a simulated free-living protocol consisting of activities of daily living. Next to that, they will perform a progressive treadmill protocol to assess higher intensities. With this data, the explained variation can be assessed for every variable, in order to identify relevant physiological signals as well as to determine the optimal wearing location.

Objective Classification of Huntington’s Disease Chorea Severity During Walking Using a Gait Anomaly Detection Algorithm

Dafna Schwartz1, Monica Busse2, Lori Quinn3, Ran Gilad-Bachrach4, Jeffrey M Hausdorff1

1Tel Aviv Sourasky Medical Center, 2Cardiff University, 3Columbia University, 4Tel Aviv University

Background: Chorea is a major, disabling symptom of Huntington’s disease (HD). A method for measuring chorea continuously in daily living settings is needed. Gait disturbances in HD patients occur in all stages of HD and gait variability has been suggested to be a biomarker of disease severity. We hypothesized that gait abnormality is correlated to chorea severity and can be used to detect chorea severity automatically and continuously during walking. Objective: To automatically detect chorea using a wrist-worn accelerometer and to classify chorea severity in walking to assess the chorea manifestation over time. Methods: Wrist-worn accelerometer signals from 27 HD patients (age: 53.3±11.6 yrs, 40.7% women) and 10 healthy controls (HC) (age: 55.2±10.1 yrs, 50% women) were collected. Subjects completed a clinical assessment that was videotaped. Offline review labeled the video in a continuous manner; chorea scores of the right hand were labeled in a continuous manner using the Unified Huntington’s Disease Rating Scale (UHDRS) (Figure 53, 1a). Walking signals were then extracted and evaluated using a gait anomaly detection algorithm. The gait anomaly detection algorithm is based on an encoder-decoder LSTM structure that was trained using a dataset of 32 healthy adults from Physionet. The anomaly levels of the walking signals of both HD and HC were scored and the association with the clinical chorea level was examined. Results: The anomaly scores and the chorea levels were correlated. Wilcoxon signed-ranks tests indicated that the anomaly score ranks of a certain chorea level were significantly higher than the anomaly detection score ranks of a lower chorea level (see Figure 53, 1b). Conclusions: A major challenge in chorea detection methods is collecting and labeling data. Here we describe the results of a promising method for detecting chorea severity from a wrist-worn sensor that is based on unsupervised learning from an independent database of healthy adults.

Patient Compliance With Remote Monitoring: Findings From a Multi-Center Study

Emily Hampp1, Ricardo Antunes1, Robert Marchand2, Paul Jacob3, Andrew Meyer1, Elaine Justice3, Kelly Taylor2, Kelly Luttazi2, Matthias Verstraete1

1Stryker, 2Ortho Rhode Island, 3Oklahoma Joint Reconstruction Institute

Introduction: There has been an increase in the use of technology to remotely administer rehabilitation and monitor recovery after total knee arthroplasty (TKA). However, the implementation and adoption of these devices is not well-characterized. The purpose of this study is to review the preliminary findings related to patient compliance from a multi-center study for a newer generation wearable device after TKA. Methods: A total of 101 patients (mean age of 65 y/o, range 50 to 79 y/o) were enrolled at two centers. Device use compliance was characterized by the mean worn time, per each 30-day period pre- and post-surgery, normalized by each patient’s total expected worn days in that period (to normalize for any dropped patients). The mean hours per day the wearable sensors were worn by the patient were also characterized for each of these 30-day periods. Results: On average, patients wore the device for 80%, 70%, 70% and 40% of the expected days for the pre-operative period, 1-30 days post-operative, 31-60 days post-operative, and 61-90 days post-operative, respectively. When the device was worn, the mean worn time for the pre-operative period, 1-30 days post-operative, 31-60 days post-operative, and 61-90 days post-operative, was 9.8, 9.8, 9.8, and 10.3 hours per day, respectively (Figure 54). Conclusion: The average device worn time demonstrated high compliance in terms of the percent of expected days that the device was worn (pre-operatively and up to 60 days post-operatively), and the number of hours the device was worn per day throughout recovery. A limitation in the expected worn time is the assumption that the mean percent is constant within the 30-day period if a patient drops out of the study. These preliminary findings demonstrate the potential for remote therapeutic monitoring reimbursement.

Validation of Wearable Sensors for Functional Assessment of TKA Patients in a Clinical Setting

Kevin Abbruzzese1, Jenna Lyon1, Vanessa LoBasso1, Jayishni Maharaj2, David Llyod2, Price Gallie3

1Stryker Orthopaedics, 2Griffith University, 3Coast Orthopaedics

Introduction: Total Knee Arthroplasty (TKA) is a commonly undertaken procedure in the treatment of knee osteoarthritis. TKA by any method is typically assessed qualitatively using patient reported outcome measures (PROMS) or more advanced and time-consuming quantifiable methods like 3D clinical gait analyses (3DCGA). This study examined the use of a two sensor-IMU system to assess knee motion after robotically assisted TKA during activities of daily living (ADL) compared to 3DCGA. Method: The knee joint angles during a range of ADL were assessed for twenty patients (N=20) one-year post TKA. Two IMUs (Notch Interfaces Inc, NY) were attached to the subject where one was attached to the thigh and one to the shank of the surgically operated limb. Seventeen reflective markers were added to the same lower limb for optical motion capture. Multiple ADL (N=5) were assessed for all subjects. The ADL were sit-to-stand, walking, squatting, and stair ascent and descent. The kinematic output from the IMU and 3DCGA was compared and assessed for correlations and error. Results: A coefficient of multiple correlation (CMC) analysis was conducted to assess sagittal plane movement for each of the activities where the knee motion was compared between systems. It was found that the CMC between systems for sagittal plane knee motion was 0.98±0.02, 0.98±0.02, 0.87±0.07, 0.88±0.02, and 0.90±0.06 for sit-to-stand, squatting, walking, stair ascent and stair descent, respectively. RMSE for the activities were computed and were 7.11±4.08°, 7.73±3.59°, 7.11±2.68°, 11.97±2.39°, 12.47±3.11°, respectively. Conclusion: The two-IMU sensor system has the potential to reliably capture knee motion in sit-to-stand, squatting, and walking to objectively assess outcomes after robotically assisted TKA, compared to 3DCGA. These static and low velocity conditions would be ideal for clinical evaluation. There is opportunity to develop kinematic modelling approaches to improve the quantitative measurements.

SYMPOSIA

Continued Use of Established Approaches to Analyzing Accelerometer Data for the Measurement of Physical Activity: How and Why to Keep it Simple

Kimberly Clevenger1, Karin Pfeiffer2, Alexander Montoye3

1National Cancer Institute, 2Michigan State University, 3Alma College

Summary: When accelerometers were first being used to characterize physical activity (PA) participation, cut-point or regression based approaches emerged as useful methods for classifying PA intensity. However, there were limitations of propriety activity count metrics and the proliferation of cut-points or models led to confusion over which method to use (e.g., the 'cut-point conundrum'). More recently, the availability of raw acceleration data in research-grade accelerometers and increased focus on techniques like machine learning have given rise to increasingly complicated methods that may require resources like computing clusters, dedicated computer science collaborators, and coding knowledge. While these advances are clearly promising, researchers who are not PA measurement experts continue to overwhelmingly use more feasible approaches like count cut-points. Thus, there is a clear division in PA research that will only grow as the number of available methods proliferates. In this symposium (chaired/moderated by Dr Cheryl Howe, Ohio University), Dr Pfeiffer will provide the rationale for continued use of more established, albeit more 'simple' approaches to analyzing accelerometer data. In addition to examining the accuracy and use of these approaches in practice, she will include a focus on how use of more complex methods can be a diversity, equity, and inclusion issue. Dr Clevenger will discuss how to overcome two barriers to using established approaches. First, she will highlight the use of open-source counts which allow researchers employing any accelerometer brand to make use of the methods developed using ActiGraph counts. Second, she will describe a consensus method which pools estimates from multiple cut-points or classification approaches to provide more stable estimates of PA outcomes and allow for improved comparability across past and future studies. Dr Montoye will discuss the future of cut-points or other established methods. He will describe how researchers can further develop existing models instead of creating new models, for example through cross-validation. He will describe a framework for robust development of new 'simple' methods using the Monitor-Independent Movement Summary (MIMS) unit as an example. Robust development includes initial work on validity and reliability, validation and cross-validation of methods using adequate sample sizes and heterogeneity, and dissemination that ensures proper use in the field. The moderated scientific exchange will include how the field can future-proof and past-proof our research as new models are developed, whether cut-points methods are the issue or if it is our development approaches (small, homogenous samples), and the tradeoffs between using or developing more simple or complex models. We will discuss potential opportunities for pooling data to improve validation or cross-validation of methods and solicit feedback on how to make data analysis easier for general PA researchers.

Using open-source counts and a consensus approach to facilitate continued use of established approaches to analyzing accelerometer data

Kimberly Clevenger1, Karin Pfeiffer2, Alexander Montoye3

1National Cancer Institute, 2Michigan State University, 3Alma College

Background and Aim: Numerous methods for characterizing physical activity participation using accelerometry have been developed and implemented in prior research. A barrier to continued use of these methods in future studies, particularly those employing other device brands, is the frequent reliance on ActiGraph counts, which until recently were generated using a proprietary algorithm. A second, and well-established issue is the 'cut-point conundrum' in which the number of available methods makes it difficult for researchers to select which approach is best to use, further limiting comparability across studies. Our purpose is to address these issues through the use of open-source activity counts and a consensus method which pools estimates from multiple classification approaches. Method: First, to illustrate the use of open-source counts, we calculated activity counts using data from 30 participants who wore a GENEActiv and ActiGraph GT9X on their left wrist during two laboratory visits (one structured and one simulated free-living). Second, to illustrate application of the consensus method, we used hip-worn ActiGraph GT9X data from the same 30 adults. Nine methods were used to estimate minutes of moderate-to-vigorous physical activity (MVPA), including cut-point, two-regression, and machine learning approaches using both count and raw inputs and several epoch lengths. Results: At the epoch-level, open-source counts were highly correlated between the two wrist-worn monitors (r=0.96) with mean absolute differences of 764.8 ± 1229.3 counts per minute. Once collapsed to the participant level, total vector magnitude counts and minutes of MVPA were highly correlated between devices (r=0.82-0.85) with a mean absolute difference of 6.6 ± 9.2 min. For the hip-worn data, the consensus estimate was 37.3 min (Figure 55), with mean MVPA for the sample ranging from 32.2 to 45.4 min across the different methods. Compared to a criterion (observed activity), the consensus method had a smaller mean absolute difference (5.7 min MVPA) compared to individual methods (6.2 to 12.0 min). Additionally, the consensus method allowed for estimation of variance at the participant level; the average standard deviation across methods for an individual was 7.4 min. Conclusions: Differences between device open-source counts may be exaggerated due to the high degree of MVPA in this protocol (GENEActiv: 70.0 min, ActiGraph: 73.6 min). Recent release of information regarding ActiGraph's proprietary counts may further improve open-source count algorithms. The consensus method enables the addition/removal of methods depending on data availability or field progression while limiting variability due to convergence between estimates. We propose standardized ways of deciding which methods to include in consensus approaches to further reduce variability. Together, use of open-source counts and the consensus approach may allow us to “past proof” and “future proof” research studies.

Rationale for continued use of more established and simple approaches to analyzing accelerometer data

Karin Pfeiffer1, Kimberly Clevenger2, Alexander Montoye3

1Michigan State University, 2National Cancer Institute, 3Alma College

Background and aim: Cut-point or regression-based approaches were the initial methods established for classifying physical activity (PA) intensity when accelerometers were first started being used to characterize PA participation. Over time, use of these methods revealed issues including 1) using propriety activity count metrics, 2) encouraging generation of population-specific methods, and 3) sparking creation of methods for multiple epoch lengths, among other issues. This proliferation of cut-points or models also led to confusion over which method to use (e.g., the 'cut-point conundrum'). More recently, the availability of raw acceleration data from research-grade accelerometers and increased focus on techniques like machine learning have given rise to increasingly complicated methods that may require resources like computing clusters, dedicated computer science collaborators, and coding knowledge. However, not all end users of accelerometer devices possess the expertise or funding to support these complicated methods. The access to and ease of use of cut-point or regression-based approaches seems to continue to drive their utilization. Further, there remains to be consensus regarding the best procedures for collecting, reducing, and analyzing data, which applies to both simpler and more complex methods. Questions remain regarding how well existing methods have been developed (e.g., if have they been validated in an independent sample) and the added benefit more complex methods might have over the simpler methods. Methods and Results: This presentation will address the accuracy and use of these simple and complex approaches in practice. Another key element of this presentation will highlight how use of more complex methods can be an issue that does not adequately address attention to diversity, equity, and inclusion. Conclusions: This presentation establishes why simple approaches to analyzing accelerometer data are necessary and sets the stage for suggesting different methods to do so.

Refining existing methods and developing robust new methods to analyze accelerometer data

Alexander Montoye1, Karin Pfeiffer2, Kimberly Clevenger3

1Michigan State University, 2National Cancer Institute, 3Alma College

Background and Aim: Technological advances in accelerometer devices have opened possibilities for data collection, analysis, and interpretation that were unthinkable only a few years ago, and this trend is likely to continue. The research community has followed suit, exploring the capabilities of new technologies and developing innovative ways to analyze such data in order to continue to advance our understanding of how physical behaviors affect health and function. However, the development and proliferation of new analytic approaches/models has far exceeded their adoption. Recent review studies have shown the stark reality that most analytic approaches are not cross-validated, made accessible, nor used by others in research or clinical settings to analyze accelerometer data. Further development and use of existing methods, including approaches like cut-points, should be prioritized over development of new methods, unless these new methods fill a gap or demonstrate marked improvement over existing approaches and have reasonable potential for implementation. This further development includes independent sample cross-validation, validation using other device brands, settings, or populations, or improved usability for general physical activity researchers (e.g., development of software, vignettes). In the long-term, this will improve comparability across studies and make it easier for researchers to identify which approach to use. In addition to lack of usability or additional development/validation, many existing methods were developed on small, homogenous samples in laboratory-based or structured settings using a single brand/model of device. Therefore, we also posit that some of the issues that have arisen surrounding use of cut-points or regression-based models is due to the employed development protocols. Methods and Results: We describe a framework for robust development of new methods, using the Monitor-Independent Movement Summary (MIMS) unit as an example. Robust development includes initial work on validity and reliability, validation and cross-validation of methods using adequate sample sizes and heterogeneity, and dissemination that ensures proper use in the field. Conclusions: As a research community, we have the potential to leverage technology and substantial technical and behavioral expertise to positively impact the research community and the broader population. Proper attention to balancing technological and analytic advances, robust development and testing, and user-friendly deployment will help us to fully realize this potential.

Harmonisation methods of accelerometery and linkage with prospective health data in the ProPASS Consortium: pooling international cohorts for individual participant meta-analyses

Matthew Ahmadi1, Andy Atkins2, Magnus Svartengren3, Emmanuel Stamatakis1

1University of Sydney, 2University of East Anglia, 3Uppsala University

Summary: Individual participant data involves harmonisation of individual level data and represents one of the gold-standards in meta-analytic research of health effects. ProPASS is the largest international accelerometry consortium compromising over 20 international cohorts with over 135,000 participants in 33 countries. In 2021 ProPASS and ISMPB began a collaborative partnership with the scientific objective to directly inform the next generation of evidence on physical behavior and long-term disease outcomes using objective measures of physical activity, sedentary behaviour, and sleep. ProPASS has established industry partnerships with DataSHIELD and Malestrom Research; both are global leaders in the development of harmonisation methodology and software. Prospective harmonisation is a powerful tool that can overcome heterogeneity and selection bias which are two of the largest obstacles to rigorous evidence synthesis. This symposium will cover three primary challenges to data processing and harmonisation and resolutions that are broadly applicable to current and future accelerometry-based meta-analytical health studies: 1) Introduction of the development and validation of a specialised accelerometry software tailored to ProPASS current and future research objectives. The foundation of this software is also being used in the SurPASS Consortia and will be made freely available to all researchers. 2) A federated data analysis infrastructure to pool cohort data and abide by data confidentiality and protection laws such as the European Union's General Data Protection Regulation Law. 3) Harmonisation of metabolic, anthropometric, demographic, and behavioural data across studies that have implemented different measures and quality controls. Prof. Stamatakis will moderate a discussion that extends from the individual presentations: 1) Evidence-based guideline priorities on objective measures of physical activity, sedentary behaviour, and sleep as critical tools for public health policy, health surveillance, and community and clinical medicine. 2) Pearls and pitfalls of collaboration with industry partners to improve the scalability of data harmonisation procedures and linkage with administrative health and mortality records 3) Ethico-legal implications of individual level data sharing and how a federated data platform addresses the most basic challenges in facilitating access for researchers and health-care professionals. Each discussion point will provide practical implications to facilitate an open-forum discussion with the audience on research methodology applications, collaborative efforts with research and industry partners, and global data sharing. Secondarily, this will provide the audience with strategies to facilitate pooling of studies implementing accelerometry. We will also leave key takeaways for harmonisation of objective physical behaviour and health data.

Harmonisation methods of accelerometery and linkage with prospective health data in the ProPASS consortium: pooling international cohorts for individual participant meta-analyses

Matthew Ahmadi1

1University of Sydney

A federated data platform provides a novel technological solution that can address some of the most basic challenges in facilitating the access of researchers and other health care professionals to individual level data. Federated data analysis can be used in research environments where data must be analysed but cannot physically be shared with researchers. The presentation will include information on ProPASS' collaboration with DataSHIELD, an industry partner who had developed a federated software infrastructure. The open-source structure of the platform facilitates research in settings where: 1) co-analysis of individual level data from several studies is necessary but governance restriction prevents the release of required data or renders data sharing unacceptably slow, 2) governance concerns hinder access to a single dataset, 3) researchers wish to actively share information held in their data with others but do not wish to cede control of the governance and/or intellectual property.

The harmonisation of non-accelerometer data in ProPASS: Where we’ve been and where we’re going

Andrew Atkin1

1University of East Anglia

This two-part presentation will (1) summarise the methods and outcomes of the harmonisation of metabolic, anthropometric, demographic and behavioural data in ProPASS to date and (2) outline future developments to this process. Part one will describe the process and timeline for harmonisation of the non-accelerometer data, provide illustrative examples of some of the variables that have been harmonised thus far and offer some critical reflections on the process as it was implemented. Part two will outline future plans for data harmonisation in ProPASS, addressing some of the challenges outlined in part one. This will include preliminary details on a collaboration with Maelstrom Research, global leaders in the development of retrospective harmonisation methodology and software.

ActiPASS - A software for processing thigh worn accelerometer data in ProPASS

Magnus Svartengren1, Peter Johansson1, Pasan Hettiarachchi1, Matthew Ahmadi2, Emmanuel Stamatakis2, Andreas Holterman3, Patrick Crowley3

1Uppsala University, 2The University of Sydney, 3National Research Centre for the Working Environment, Denmark

ActiPASS - A Software for processing thigh worn accelerometer data in ProPASS. Background and Aim: The ProPASS consortium consists of several cohorts which have used different brands of thigh worn accelerometers. To pool data between these cohorts there is a need for a transparent, validated and harmonized data processing procedure, that produces variables according to the ProPASS 24/7 construct of physical behaviour. The Acti-4 algorithm, that has been developed by the National Research Center of Working Life in Copenhagen, is a validated algorithm that can be used to process raw data from several brands of accelerometers. Acti-4 identifies the physical behaviours: sitting, standing, moving, walking, running, stairwalking and bicycling with high precision, but identification of sleep is lacking. We have now further developed Acti-4 to also identify lying down time and sleep from thigh worn accelerometers. These new features has been validated in field studies. Methods: An already existing algorithm that uses information of thigh rotation to to differentiate lying down from sitting, developed by Lyden et al, was combined with the Acti-4 algorithm and refined. This was validated in a dataset where 47 participants wore two Axivity-AX3 devices for 7 days, one on the thigh and one on the back as a reference. The sleep algorithm was developed and optimized on a dataset consisting of 23 single-night polysomnography registrations (PSG), from 15 asymptomatic adults. Then this algorithm was validated on another dataset, in which, 71 adult males (age 57 ± 11 years) wore ambulatory PSG equipment and one Axivity-AX3 on the thigh simultaneously, while sleeping one night in their homes. Results: Lying down time was identified with a sensitivity of 0.95, specificity of 0.94 and accuracy of 0.94 compared to lying down time, identified by the back accelerometer. The mean difference between the total identified lying down time/day, between the refined algorithm and the back accelerometer was +2.9 (95% limits of agreement; -135 to +141) minutes per day. Sleep was identified with a mean sensitivity of 0.84, specificity of 0.55 and accuracy of 0.80 compared to PSG. Sleep intervals were underestimated by -21 (95% limits of agreement -86 to +44) minutes. Total sleep time was underestimated by -32 (95% limits of agreement -148 to +85) minutes. Conclusions: Acti4 and the added functionality to identify lying down time and sleep is now integrated into ActiPASS, that is a new streamlined, automated software for processing raw accelerometer data in large batches that fits the need for the ProPASS consortium.

Measuring Sleep with Wearables: The ABC’s of Measuring Z’s

Seth Creasy1, Kong Chen2

1University of Colorado, 2National Institute of Diabetes and Digestive and Kidney Diseases

Summary: Aspects of sleep health including sleep duration and sleep quality are important determinants of health. Thus, being able measure and quantify these parameters of sleep are of interest to clinicians and researchers. Current wearable devices estimate sleep quality and physical activity using a variety of sensors (accelerometer, light-sensor, inclinometer, etc.). The gold-standard for measuring sleep is polysomnography (PSG) which includes multiple measurement inputs and is typically conducted in laboratory for clinical purposes. While data from PSG are considered accurate and precise, the in-laboratory environment can be disruptive to an individual's sleep compared to their in-home environment. In contrast, wearable devices (especially accelerometry-based wrist actigraphy) can be utilized to measure parameters of habitual sleep over longer periods of time. However, device-based measurements of sleep are more prone to inaccuracies and misclassification and do not capture the same physiological signals that PSG captures. This session will outline the history of sleep measurement, the significance of measuring sleep related to epidemiological research, clinical perspectives on wearables and sleep measurement, as well as the accuracy of consumer-grade device-based estimates of sleep.

History and significance of sleep measurement

John Chase1, Rebecca Spencer1

1University of Massachusetts Amherst

Sleep is critical for physical, cognitive, and psychological health. Sleep is simultaneously influenced by confounding life factors such as development, aging, and disease. Accurate and precise sleep measurement is crucial for our understanding of the relationships between health outcomes and life factors. Technological advancements in sleep measurement have preceded an era when sleep measurement is widely portable and accessible in clinical, research, and commercial platforms alike. In this talk, we will review the historical progression of sleep measurement from early self-report questionnaires to contemporary sleep measurement tools, including polysomnography and wearable technology (e.g., accelerometers). We will explain why the question of interest dictates what type of sleep measurement device is needed, while highlighting the strengths and limitations of common device-platform combinations. Finally, we will discuss how burgeoning technological advancements, such as the incorporation of biometric signals in portable devices, can improve our understanding of the relationships between sleep and health outcomes across the lifespan.

Sleep measurement in research & clinical settings

Stacey Simon1

1University of Colorado

Background and Aim: Sleep health is a multidimensional concept consisting of a variety of factors such as duration, timing, quality, and satisfaction. Poor sleep health is endemic in individuals across the lifespan: nearly 35% of adults and 78% of adolescent report sleeping less than the recommended amount per night, and sleep complaints are one of the most common parental concerns for pediatricians. Sleep disorders such as obstructive sleep apnea and insomnia are also increasingly prevalent. Thus, the aim of this presentation is to describe measures of sleep health and discuss pros and cons, indications for use, and consideration for special populations. Methods: A review of objective and subjective measures of sleep health frequently used in research and clinical settings will be provided. Results: Laboratory-based polysomnography is the gold standard for objective sleep evaluation but is expensive, burdensome, requires trained staff to administer and score, and captures only a single night of sleep in an atypical environment. Alternative devices such as accelerometer-based wrist actigraphy, dry-EEG headbands, and peripheral arterial signaling finger-worn devices can be used in the home environment over extended periods of time but may also be costly or less accurate. Conclusions: Accurate assessment of sleep health is important for both researchers and clinicians and a number of assessment tools are available for different populations, settings, and outcomes.

Integrating physical activity and sleep measurements in epidemiological research

Charles Matthews1

1National Cancer Institute

The application of accelerometry in large scale epidemiologic studies has accelerated the interest among physical activity researchers to investigate the health benefits and risks associated with the full range of behaviors occurring in the 24-hour day, including sleep, physical activity, and sedentary behavior. There are many similarities in studying sleep and physical activity using ambulatory monitors, but there are also important differences that should be considered. This presentation will describe the parallels in measuring the two behaviors as well as the important differences. Current state of the art applications of monitor-based measures of physical activity and sleep in large epidemiologic studies will be discussed, with a particular focus on key etiologic questions related to risk for developing cancer and how better assessments of sleep and physical activity may advance our cancer prevention efforts.

Accuracy and utility of consumer-grade devices for measuring sleep

Evan Chinoys1

1Naval Health Research Center

Recent advances in technology and demand for biometric data have led to the creation of a variety of personal consumer devices that track physiological signals and behavioral patterns, including sleep. Such devices help meet the important need for long-term, automated, real-time sleep tracking, with the added benefits of being less expensive and burdensome than standard research methodologies. Although such technologies have widespread use among the general population for everyday sleep tracking, the algorithms are often proprietary and the claims made by technology companies regarding device accuracy and utility are debated by researchers and clinicians. A related concern is that the ability of researchers to formally evaluate the validity of devices is much slower than the pace of new devices being released onto the consumer market. Despite this research gap, the number of high-quality validation studies have increased recently, helping elucidate the strengths and weaknesses of many new and popular consumer sleep-tracking devices. This includes our lab which, over the past 5 years, has conducted a series of validation studies testing many of the latest consumer sleep-tracking devices, to systematically evaluate their performance under different conditions. In general, our findings show that many, but not all, devices can track sleep-wake patterns on most nights as well as (or slightly better than) the mobile sleep assessment standard methodology, research-grade actigraphy, in healthy individuals under fixed sleep conditions in a controlled laboratory setting, as well as at home with ad libitum sleep schedules and environments. However, consumer devices still display some of the performance limitations inherent to research-grade actigraphy devices, such as low epoch-by-epoch specificity and bias toward underestimating true periods of wake - indicating that device accuracy may be lower on nights with disrupted sleep patterns. We also found that consumer devices are inconsistent in their ability to accurately classify individual sleep stages (i.e., light, deep, or rapid eye movement sleep) and to track irregular sleep schedules (e.g., naps, split sleep). Additionally, our lab has started implementing sleep-tracking devices into real-world operational military environments to evaluate their feasibility for everyday use and utility of their sleep data as inputs into fatigue management platforms to identify potential sleep issues and reduce operational risks. The continued improvement and versatility of new consumer devices strengthens their potential use cases as beneficial alternatives to standard methodologies for tracking real-world sleep patterns, though with some important considerations and limitations.

Measuring the interrelationships between dietary intake and physical activity in free-living settings

Sarah Purcell1, Danielle Ostendorf2, Derek Hevel3, Edward Sazonov4, Krista Leonard5

1University of British Columbia - Okanagan, 2University of Colorado, 3University of North Carolina - Greensboro, 4University of Alabama, 5Arizona State University

Summary: Co-chairs: Sarah A. Purcell, PhD (University of Colorado and University of British Columbia) and Danielle Ostendorf, PhD (University of Colorado). A growing body of evidence suggests that physical activity (PA) and dietary intake (DI) relate to one another via complex mechanisms. Intuitively, these behaviors do not exist in isolation; they differentially and synergistically impact each other and health outcomes. However, current understanding of energy balance is limited by 1) data collected in discrete time intervals, 2) inaccuracies in subjective recalls, and 3) a general focus on one behavior/outcome. Some of these weaknesses may be addressed by integrating real-time PA and DI data. For example, ecological momentary assessment (EMA) uses smartphones to collect self-reported behaviors, contexts, and/or perceptions in naturalistic settings. Others have developed novel, user-friendly wearable devices that collect objective data proximal to the time and place of the behavior (i.e., PA, DI); they can also reduce recall errors and improve validity of self-reported data. However, there are unique challenges in integrating these methods to measure health behaviors in different contexts and populations. This symposium will highlight emerging technology and methodologies that can capture different aspects of DI and PA and integrate these constructs for a more holistic understanding of health. We have invited three external speakers that will provide unique perspectives including: 1) practical and statistical methods to assess DI, appetite, and PA via EMA (Mr. Derek Havel), 2) novel technologies such as wearable eyeglasses and shoe insoles for objective measurement of DI and PA, respectively (Dr. Edward Sazonov), and 3) energy balance concepts in general and during an intensely-adapted prenatal intervention using mobile health devices (Dr. Krista Leonard). Our goal is for attendees to gain a better perspective of methods to integrate DI within PA data through collaboration with scientists in multiple disciplines. This will be executed through a standard format for scientific symposia. We will begin our session with an introduction on the rationale and need for integrating PA and EI data. The introduction will also include a 'case study' example from one of the session co-moderators (Dr. Sarah Purcell) on integration of PA, DI, and appetite in the context of a weight loss intervention with yoga. Then, each invited speaker will present their work for 20 minutes, with 20 minutes of interactive questions and discussion at the end of the symposium. The co-chairs will prepare questions and discussion points ahead of time to encourage discussion among the audience, if needed.

Physical activity and dietary intake measurement via ecological momentary assessment: Practical considerations and potential statistical analyses

Derek Hevel1, Jaclyn Maher1

1University of North Carolina - Greensboro

Background and Aims: Physical activity (PA) and dietary intake (DI) are repeat occurrence health behaviors that have mental and physical health implications. Yet, traditional measures of PA and DI have often been limited in the past with the use of retrospective and infrequent assessments which are prone to recall biases and often lack ecological validity. Further, patterns of PA and DI likely change across short timescales (e.g., hours), vary across different contexts (e.g., environment), and co-occur with other behaviors. Limitations of traditional measures of PA, DI, and correlates may contribute to reductions in the predictive power of theories and techniques of health behavior engagement. Methods: Ecological Momentary Assessment (EMA) can overcome previous limitations by intensively capturing PA, DI, and correlates to elucidate how behaviors unfold across time. Results: The collection of PA and DI via EMA brings many practical considerations including how to adequately capture PA and DI, the selection of assessments, participant burden, and the pairing of EMA data with other data (e.g., accelerometers). New statistical analyses can use EMA data to address new questions including how individuals differ from one another and how they differ from their usual levels. Conclusion: Studies of emerging and older adults' PA and emerging adults' DI behaviors will be discussed to highlight practical considerations and potential statistical analyses.

Monitoring of energy intake and expenditure with Automatic Ingestion Monitor

Edward Sazonov1

1The University of Alabama

Background and Aim: The Automatic Ingestion Monitor (AIM) is a passive food intake sensor requiring no self-report of eating episodes, just compliance with wearing the device. This talk will present our ongoing work on using the AIM for monitoring of energy intake, diet, physical activity, and energy expenditure. Methods: Results from several completed and ongoing studies will be presented, including 1) An overview of the sensors and operation of the AIM device; 2) Online (real-time) and off-line (postprocessing) models for accurate detection of food intake in free-living and capture of images of the foods being eaten with privacy preservation; 3) Use of AIM data for estimation of energy intake in respect to weighed food records; 4) A novel method for joint recognition of physical activity and energy expenditure from the AIM data. Results: The accuracy of food intake detection in free-living varied from 81.8% to 96% F1-measure in various studies. The AIM was successfully deployed in several studies, including studies in rural and urban Africa, providing reliable data on food consumption. The difference in daily energy intake estimated using sensor and food image data with respect to weighed food records were (Mean±SD) -0.45±2.60 MJ/d and 3.30±2.87 MJ/d, respectively. The accuracy of physical activity classification was 97%, while the model for energy expenditure produced a 10% mean absolute error. Conclusions: The AIM sensor shows promise as a tool for joint assessment of diet, energy intake, physical activity, and energy expenditure. Further studies are needed to refine the models used in the estimation of energy intake and expenditure.

Methodological considerations in measuring physical activity, energy intake, and resting energy expenditure in the context of an adaptive prenatal weight gain intervention

Krista Leonard1, Danielle Symons Downs2

1Arizona State University, 2The Pennsylvania State University

Challenges associated with measuring prenatal energy balance (e.g., feasibility, misreporting) have limited our understanding of the complex interrelations of the components of prenatal energy balance and its impact on gestational weight gain (GWG) regulation in pregnant women with overweight or obesity (PW-OW/OB). PW-OW/OB are at risk for excessive GWG (i.e., >11.5 kg for overweight and >9.0 kg for obese), which is an independent predictor of adverse maternal (e.g., gestational diabetes) and infant (e.g., macrosomia) outcomes and long-term development of obesity. Evidence suggests that excessive GWG is a result of behavioral factors (i.e., high energy intake; to a lower extent, low physical activity). As such, GWG regulation trials have primarily focused on the combined effects of promoting physical activity and moderating energy intake. However, many PW-OW/OB experience unique psychosocial and physical challenges, which can make health behavior changes and subsequent regulation of GWG difficult. Our prior work as well as others have suggested that in addition to energy intake and physical activity, another component of energy balance that is physiologically regulated and contributes to GWG is resting energy expenditure (REE). The lack of evidence regarding the interrelations between the components of energy balance and GWG may party be attributed to methodological challenges such as a lack of feasible measures that can assess daily physical activity, energy intake, and REE over time, the absence of gold standard protocols for wearable devices, and inaccuracies associated with self-reported measures (e.g., overreporting of physical activity, underreporting of energy intake). The objective of this presentation is to recommend measurement strategies that aim to address these methodological issues to improve the collection of prenatal physical activity, energy intake, and REE data. Incorporating these novel measurement strategies can help future researchers answer the question of how the components of prenatal energy balance are interrelated and predict GWG regulation in PW-OW/OB in order to support long-term health for mothers and children. These measurement strategies will be discussed within the context of a longitudinal, adaptive prenatal GWG regulation intervention, Healthy Mom Zone. Dr. Leonard will provide an overview on the importance of understanding components of prenatal energy balance for predicting GWG regulation in PW-OW/OB. She will also discuss data from her research and others that use novel, practical, and cost-effective methods to improve the accuracy of measuring prenatal physical activity, energy intake, and REE via mobile health devices and validated equations. Lastly, Dr. Leonard will provide recommendations for how these measurement techniques can be utilized in future studies aimed at understanding energy balance to prevent excessive GWG.

Mobility outcomes for clinical trials in cerebellar ataxia: the route from the clinic to daily life

Winfried Ilg1, Fay Horak2, Vrutangkumar Shah2

1University Tübingen, 2Oregon Health & Science University

Chair: Winfried Ilg, Hertie Institute for Clinical Brain Research, University Tübingen. Many rare neurological diseases that affect mobility (e.g. cerebellar ataxia) have no established treatment but now have exciting, novel drugs appearing in the pipeline. These clinical trials are hampered by clinical scale outcomes that have inadequate effect size. Objective measures of mobility (e.g. gait, balance, turning and physical activity) have the potential to provide better effect size and to add ecological validity. Wearable technology has become feasible for large clinical trials but the most sensitive and reliable balance and gait metrics to serve as performance outcomes need to be determined. Global initiatives are currently underway to unify assessment protocols to enable longitudinal, multicentric studies. In this symposium, we will present recent approaches of quantifying ataxic gait and balance during clinical assessments and in patient's daily life for quantifying disease progression and response to treatment. We will discuss the necessary steps for regulatory approval and outline standardized protocols of clinical assessment and real-life gait recordings. In the interactive discussion, we will generalize the topic of regulatory approval of balance and gait outcomes to other movement disorders. F. Horak will introduce the clinical picture of ataxic balance and gait and the recent challenge of new potential therapies for this rare disease, without sensitive clinical outcomes. She will present challenges in determining the optimal set of balance and gait outcomes from body-worn, inertial sensors for multi-site clinical trials. What scientific evidence is needed, how to relate concepts of interest and specific outcomes to meaningful measures of health, and what are further steps needed for regulatory approval of gait and balance outcomes, both in prescribed tests and in daily life monitoring. V. Shah will demonstrate how quantitative assessment of ataxia-specific gait and balance impairments from wearable technology could provide sensitive performance outcomes with high face validity. He will present a novel approach, MCDA, to develop a composite posture and gait score and will describe the clinemetric properties of this composite score to test its validity and reliability. W. Ilg will present specific measures of ataxic gait and turning in daily life which demonstrate high sensitivity to small differences in disease severity. Moreover, measures of turning capture changes of dynamic balance with sensitivity to longitudinal changes and to the prodromal stage. Thus, daily life measures present as promising ecologically valid biomarkers, even in the most treatment-relevant early stages of disease. We will start the scientific exchange (moderated by W. Ilg, F. Horak) with the question: Do you think that active, performance tests or passive, daily life monitoring will provide the most useful outcomes for future clinical trials?

Towards ecologically valid biomarkers: real-life walking and turning assessment captures subtle longitudinal and preataxic changes in cerebellar ataxia

Winfried Ilg1

1Hertie Institute for Clinical Brain Research

Background and aim: While manifold targeted molecular treatments for cerebellar ataxias are on the horizon, clinical and regulatory acceptance will depend on their proven effects on subject's ataxia using quantitative biomarkers. Thus, sensitive biomarkers with high relevance for patients' daily life are highly warranted. Moreover, it is hypothesised that real-world gait is more sensitive to disease-specific signatures compared to clinical settings, due to the complexity of the environments as well as the larger amount of gait data captured by wearable inertial sensors. Measures of spatiotemporal variability have been shown to allow the quantification of disease severity and capturing treatment-related improvements in ataxic gait. The transfer of variability measures to real life is hereby complicated by the fact that real-life gait is inherently far more variable and that patients are free to use various compensation strategies, thus increasing heterogeneity of walking patterns. Thus, variability measures may lose their accuracy for characterizing ataxic changes in real life. Methods: We performed a combined cross-sectional and longitudinal (1-year interval) study in degenerative cerebellar disease including pre-ataxic mutation carriers. Gait and turning movements were assessed by three body-worn inertial sensors in (1) laboratory assessment, and (2) unsupervised real-life movements. We focused on measures of step variability in gait and measures quantifying dynamic balance during turning. Results: We identified measures that allowed not only to capture the variability inherent in ataxic gait in real life, but also demonstrate high sensitivity to small differences in disease severity. Lateral step deviation and a compound measure of spatial step variability (i) categorized patients against controls with high accuracy (ii) both were highly correlated with clinical ataxia severity, with highest effect sizes in real life (r=0.76). Moreover, the turning measure LVC (lateral velocity change) allow to capture changes on dynamic balance in real life, with sensitivity to the preataxic stage (δ=0.53) and high effect size of 1-year longitudinal change (rprb=0.66). Together with good test-retest reliability (ICC=0.91) this results in low sample sizes for detecting a 50% reduction of progression by a hypothetical intervention (n=66). Conclusions: Our results prepared steps towards regulatory approval of digital-motor biomarkers as endpoints for future trials, demonstrating (i) power as ecologically valid biomarkers, (ii) correlation with clinical ataxia severity and patient-reported balance confidence outcomes, (iii) sensitivity to subtle changes longitudinally, and (IV) test-retest-reliability in real-life recordings.

The challenge of identifying balance and gait digital health outcomes for clinical trials in ataxia

Fay Horak1

1Oregon Health & Science University

I will introduce the challenges in determining the best, objective balance and gait outcomes for multi-site clinical trials for patients with rare neurological diseases. What scientific evidence is needed, how specific does the outcome need to be for each type of disease (ie; over 50 types of cerebellar ataxia), how to relate concept of interest and specific outcomes to meaningful measures of health, and what are further necessary steps for regulatory approval of these gait and balance biomarkers in clinical trials. Rare neurological diseases that affect balance and gait, such as degenerative cerebellar ataxias, currently have no established treatment but now have exciting, novel drugs appearing in the therapeutic pipeline. Unfortunately, these clinical trials are hampered by clinical scale outcomes that have inadequate effect size for the size of the population with the disease. Objective measures of balance and gait from body-worn, inertial sensors that are now feasible for multisite clinical trials, have the potential to provide better effect size for clinical trials with smaller number of subjects. Key challenges in using wearable technologies are an excessive number of measures and a lack of consensus on the most useful measures for each neurological disorder. It is also not clear whether prescribed, observed standing and walking tasks in the clinic will provide better or worse outcomes for clinical trials than measuring mobility passively, in daily life. I will introduce the clinical characteristics of ataxic gait and standing balance and recent success in identifying quantitative impairments with face validity for static and dynamic balance disorders spinocerebellar disorders. Clinical trial digital outcomes need to demonstrate they are valid (sensitive and specific) for a particular cohort, reflect severity of disease, are reliable, and sensitive to change or progression. Recently, we have shown that digital measures of ataxia can be identified in prodromal patients, before neurological exams are abnormal. To get regulatory approval for digital health outcomes, however, it is also important to demonstrate the outcome is meaningful to patients and that change in the outcome reflects a meaningful change to patients and that is difficult without effective treatments.

How to select the balance and gait measure for spinocerebellar ataxia

Vrutangkumar Shah1

1Oregon Health & Science University

Recently, we demonstrated how quantitative assessment of the severity of ataxia-specific gait and postural sway impairments from wearable technology appropriate for multi-site clinical trials could provide sensitive performance outcome measures with high face validity to power clinical trials. We tested standing balance and gait characteristics in 150 people with spinocerebellar ataxia and 50 control subjects to identify the most sensitive and specific measures for ataxia. The ataxic patients included 40 with SARA scores <3, that is prodromal ataxia, without clinically observable balance or gait disorders in genetically determine patients with SCA 1,2,3 or 6. Standing for 30 seconds with eyes open and with feet together or apart provided the best measures of balance and gait variability from a 2-minute, natural pace walk the best gait measures. I will show how quantitative assessment of the severity of ataxia-specific gait and postural sway impairments from wearable technology could provide many potential performance outcome measures with high face validity to power clinical trials. In this talk, I will focus on how to select several balance and gait outcomes based on expert opinion on the most important clinimetrics for a clinical trial. This novel approach to selecting the best objective measure of balance and gait for cerebellar ataxia can be applied to any digital outcome for any disease.

Physical Behaviors and Health: New Methods and Insights from Large Epidemiologic Studies Using Accelerometry

Sarah Keadle1, Pedro Saint-Maurice2, Qian Xiao3, Kelly Evenson4, Amanda Paluch5

1California Polytechnic State University, 2National Cancer Institute, 3The University of Texas, 4University of North Carolina Chapel Hill, 5University of Massachusetts Amherst

Topic and rationale: The application of wearable devices in large epidemiologic studies are rapidly advancing our understanding of how many steps per day are associated with disease risk, how sleep quality and timing may affect health, and how patterns of behavior within and between days influence health risk. The availability of new and rich data sources in large health studies are also driving refinement and development of new analytic methods that can help answer questions that have been difficult to resolve previously. This Symposium will feature a diverse group of speakers who are addressing critical questions related to physical behaviors and health in large epidemiologic investigations using a range of study designs, behavioral metrics, and analytic methods. It features researchers presenting new evidence regarding step counts, sleep, and health in prospective cohort studies as well as the application of statistical methods to characterize patterns of physical activity and sedentary behaviors and rest-activity rhythms to further investigate the impact of our daily behavioral patterns on disease risk. Contributions from each speaker Sarah Keadle, California Polytechnic State University will chair the session and moderate the discussion. Pedro Saint-Maurice from the National Cancer Institute will report new findings regarding actigraphy-based estimates of the duration, quality and timing of sleep and risk for all-cause and cardiovascular mortality in the United Kingdom Biobank Study. Qian Xiao from the University of Texas Health Science Center at Houston will present new results from the 2011-14 National Health and Nutrition Examination Survey which applied cosinor models and functional principal components analysis to characterize 24-hour rest-activity patterns in relation to health outcomes using wrist-worn accelerometers. Kelly Evenson from the University of North Carolina - Chapel Hill will describe her team's work to identify multicomponent patterns of physical activity and sedentary behavior using latent-class analysis of the accelerometer-based measures in the Women's Health Initiative's Objective Physical Activity and Cardiovascular Health (OPACH) Study. Amanda Paluch from the University of Massachusetts Amherst will describe work from the Steps for Health Collaborative to close the gap between common knowledge and scientific evidence regarding the health benefits of taking 10,000 steps per day derived from meta-analyses of step counts and mortality and cardiovascular disease across 15 international studies with nine different monitors.

Sleep duration, quality, timing, and mortality risk

Pedro Saint-Maurice1

1National Cancer Institute

Background & Aim: Most evidence describing the amount of sleep associated with a lower mortality risk comes from studies that used self-reported measures of sleep and includes limited information about other sleep dimensions like sleep quality and timing. This study examined associations between accelerometer-derived sleep duration, quality, timing, and mortality. Methods: Data are from the UK Biobank cohort of adults aged 40-69 years (2006-2010). Approximately 6 years post baseline, 103,712 adults participated in an activity monitoring sub-study and wore an Axivity AX3 wrist-worn triaxial accelerometer over 7-days. Monitor data were processed using the R package GGIR to generate sleep duration (hours/day), sleep quality (wake after sleep onset, sleep efficiency), and sleep timing (onset, offset, midpoint) exposures. Data were linked to mortality outcomes including all-cause, cardiovascular disease (CVD), and cancer mortality assessed via National Health Service registries in UK with follow-up up to 12/31/19. We first estimated Hazard ratios (HRs, 95% CI) for sleep duration and mortality outcomes using cubic splines. Next, we computed HRs for quartiles of the sleep quality and timing exposures in relation to mortality. All models were adjusted for age, sex, race-ethnicity, education, Townsend deprivation index, employment status, lifestyle factors, chronic conditions, functional pain, and general health rating. Sensitivity analysis included examinations of heterogeneity in our sleep duration-mortality associations by demographic and lifestyle variables. Results: Over an average of 5.1 years 1,762 deaths occurred (1,108 cancer, and 338 CVD deaths). Participants slept on average from 23:41 to 7:12, for about 6:42 hours/day, and were awake for 46 minutes. When compared to sleeping 7.0 hours/d, sleeping less than 6 hours per day was associated with a 14-33% higher risk for all-cause mortality (p<0.01; e.g., HR5 hrs/d: 1.23 [0.95, 1.61]); 28-56% higher risk for CVD mortality (p=0.05; e.g., HR5 hrs/d: 1.41 [0.78, 2.56]), with no clear associations for cancer mortality (p>0.05). Sleeping less than 6 hours/day on 3+ nights in a week was associated with a 20% increased risk for all-cause mortality (HR=1.20 [1.06, 1.36]) when compared to individuals with 0 nights of short sleep. Measures of sleep quality and timing were not associated with mortality risk (p>0.05). Our examinations of heterogeneity showed that sleeping < 6 hours/day was consistently associated with all-cause mortality across demographic and lifestyle subgroups except across quartiles of moderate-vigorous physical activity (pheterogeneity=0.02). Conclusions: Accelerometry measured sleep duration, but not the quality or timing of sleep were associated with mortality. These findings suggest that sleeping less than 6.0 hrs/d can increase mortality risk among men, women, young, and older adults.

24-hour rest-activity patterns and health

Qian Xiao1

1University of Texas

Physical activity, sedentary behaviors and sleep are fundamental human movement behaviors organized in a 24-hour rhythmic cycle. These behaviors are orchestrated by the internal circadian timing system, and influenced by common environmental exposures (e.g., light, daily schedules and social interactions). The conventional approach to study diurnal movement behaviors focuses on measures of individual components such as physical activity intensity and volume, duration of sitting, and sleep duration and efficiency. However, However, there's been little focus on the timing and rhythmic profiles of these behaviors and movement over the 24-hour day. The highly interconnected nature of these behaviors requires an integrated and holistic approach to study the overall patterns of the 24-hour rest-activity cycle. There are various methods that have been developed for characterizing 24-hour rest-activity patterns, including both parametric and nonparametric methods. The former assumes a cosine or cosine-like shape of daily activity patterns and produces rhythmic measures such as amplitude, mesor, acrophase and overall rhythmicity. In contrast, the nonparametric methods have no underlying assumption about activity patterns and derive metrics that measure specific aspects of the rest-activity cycles, such as stability, variability/fragmentation. More recently, an alternative approach to overcome these limitations is the functional principal component analysis (fPCA), which applies flexible algorithms to fit activity data with no a priori assumptions and is able to identify overall rest-activity profiles. In this section, we will discuss different methodology for characterizing rest-activity patterns using 24-hour actigraphy data, and present two recent studies in the National Health and Nutrition Examination Survey (NHANES), focusing on 1) the associations between cosinor-based rest-activity characteristics and metabolic health; and 2) fPCA-derived rest-activity profiles among US adults. These studies demonstrate the utilization of different methodology for rest-activity measurement, highlight the importance of rest-activity rhythms in health, and identify sociodemographic and socioeconomic correlates of rest-activity patterns in the US population.

Identifying multicomponent patterns of accelerometry-assessed physical activity and sedentary behavior: The Objective Physical Activity and Cardiovascular Health Study

Kelly R. Evenson1, Fang Wen1, Chongzhi Di2, Michael Kebede1, Andrea Z. LaCroix3, Michael J. LaMonte4, I-Min Lee5, Lesley Fels Tinker2, Chris Moore1, Annie Green Howard1

1University of North Carolina - Chapel Hill, 2Fred Hutchinson Cancer Research Center, 3University of California San Diego, 4University at Buffalo - SUNY, 5Brigham and Women's Hospital

Background and Aim: Latent class analysis (LCA) is a useful statistical tool to describe patterns of physical behavior (e.g., physical activity (PA) and sedentary behavior (SB)). Single component LCA has been previously applied to accelerometry to provide unique class assignments for SB and the various intensities of PA. The objective of this study was to explore multi-component LCA to integrate the full spectrum of physical behavior among women age 64 and older in a unique LCA model. Methods: Participants were from the United States and enrolled in the Women's Health Initiative Objective Physical Activity and Cardiovascular Health Study. Overall, 6,126 women 64 to 97 years wore an ActiGraph GT3X+ accelerometer at their hip for 4-7 days of adherent wear (defined as >=10 hours/day). Using accelerometry data, we assessed time spent in SB (0-18 VM/15-s), light low (19-225 VM/15-s), light high (226-518 VM/15-s), and moderate to vigorous (MVPA) (>=519 VM/15-s). Multi-component LCA classified women based on all four metrics across time of day in 1-hour windows during time awake, averaging across adherent days. Results: Mean (SD) physical behaviors in minutes/day were: 556 (99) SB, 189 (50) light low, 98 (36) light high, and 50 (34) MVPA. Optimally, 6 classes were identified for the full spectrum of physical behavior, including SB, light low, light high, and MVPA. Class assignments ranged from the highest SB and lowest MVPA (class 1) to the lowest SB and highest MVPA (class 6), both averaged across all 1-hour windows. The percent (n) from the lowest to highest class were 13.1% (805), 28.0% (1713), 21.7% (1330), 17.1% (1045), 14.0% (858), and 6.1% (375). Slower self-reported walking speed was associated with a lower class assignment (p<0.0001). Conclusions: Unique multi-component physical behavior patterns in free-living older women were observed using novel analysis of accelerometry. By identifying heterogenous patterns which capture a profile encompassing a range of physical behaviors, these methods can be used to find new insights into habitual patterns and intensities for targeted interventions aimed to improve health outcomes, such as enhancing aging resiliency and independence, among older women.

10,000 steps per day? Closing the gap between common knowledge and scientific evidence

1Amanda Paluch

1University of Massachusetts Amherst

The simplicity of steps/day as a metric makes it appealing for physical activity promotion in clinical and population settings. Summarizing the association of steps and health can advance health promotion guidelines. The Steps for Health Collaborative has compiled data from cohort studies for a meta-analysis with device-measured steps and prospective health outcomes. This presentation will discuss the process of the consortium effort and conducting a harmonized meta-analysis. Results on the associations of steps and all-cause mortality will be discussed. This meta-analysis included 15 studies, of which seven were published and eight were unpublished, including nine different step counting devices. The total sample included 47,471 adults, among whom there were 3013 deaths (10•1 per 1000 participant-years) over a median follow-up of 7•1 years ([IQR 4•3-9•9] (297,837 person-years). Quartile median steps per day were 3553 for quartile 1, 5801 for quartile 2, 7842 for quartile 3, and 10 901 for quartile 4. Compared with the lowest quartile, the adjusted HR for all-cause mortality was 0•60 (95% CI 0•51-0•71) for quartile 2, 0•55 (0•49-0•62) for quartile 3, and 0•47 (0•39-0•57) for quartile 4. Restricted cubic splines showed progressively decreasing risk of mortality among adults aged 60 years and older with increasing number of steps per day until 6,000-8,000 steps per day and among adults younger than 60 years until 8,000-10,000 steps per day. Taking more steps per day was associated with a progressively lower risk of all-cause mortality, up to a level that varied by age. The findings from this meta-analysis can be used to inform step guidelines for public health promotion of physical activity.

Spatial analyses with behavioral data

Jasper Schipperijn1, Jordan Carlson2, Marta Jankowska3, Jing-Huei Huang4, Aaron Hipp1, Jasper Schipperijn1

1University of Southern Denmark, 2Children’s Mercy Kansas City, 3City of Hope, 4North Carolina State University

All activity behavior takes place in a spatial context and adding this contextual information to device-based measures of physical behavior can provide many novel insights. Conducting spatial analysis with behavioral data can be done in many different ways, from relatively simple to very complex, and many physical behavior researchers would benefit from an increased understanding of the different methods, including their benefits and drawbacks. At this symposium we have brought together a group of researchers who have numerous years of experience conducting research that involves spatial analyses of behavioral data and they will each share their methods, lessons learned, as well as what dissemination opportunities and stories the chosen methods provided. The symposium will consist of four presentations followed by a moderated discussion. In presentation 1, Dr Carlson will present approaches for defining and investigating pre-determined activity locations, such as at home, at school, at work, in parks, and in the home neighborhood. The presentation will include an overview of existing data processing systems, recent validity results for easy-to-employ trip detection algorithms, and analyses demonstrating how location-specific information can help uncover location-specific correlates of physical activity for informing multi-level and multi-location interventions. In presentation 2, Dr Jankowska will present her work on time-weighted spatial averaging (TWSA) approaches, which account for the amount of time spent in specific locations. She will show how various TWSA methods for calculating exposures associate with physical activity outcomes and explain what types of research questions these approaches are useful for. In presentation 3, Dr Huang will present her team's work on systematically identifying children's play episodes through a density-based clustering method using park-based physical activity data recorded by accelerometer and GPS devices. This approach could provide valuable information for practitioners by identifying and mapping natural play patterns, characterizing children's free play, correlating play episodes with specific structures and layouts of playspaces, and characterizing the use of playspaces. In presentation 4, Dr Schipperijn, will present on the work he and his team have been doing on detecting hotspots for physical activity for various population groups, and in various settings, using accelerometer and GPS data. He will show how this type of analyses can inform future interventions. After the four presentations, Dr Hipp will provide a brief summary of the different approaches to spatial analysis with behavior data before posing a series of debate statements to the audience, each of which will then be discussed as a group with presenters weighing in on audience comments.

Detecting hotspots for physical activity using accelerometry, GPS and GIS

Jasper Schipperijn1, Pulan Bai2, Thea Amholt1, Hayley Christian2

1University of Southern Denmark, 2Telethon Kids Institute

Background and Aim: Daily physical activity is not one behavior that takes place in one location; it consists of many different behaviors occurring in different locations. To get a better understanding of the correlates and determinants of physical activity behavior, knowing in which context it occurs can add valuable additional information with the emergance of methods to combine accelerometer and global positioning system (GPS) The aim of this presentation is to explain how the process of identifying physical activity hotspots works, and demonstrate the method using examples from several studies conducted in Australia and Denmark. Methods: Data were collected among schoolchildren in Denmark and preschool children in Australia using an accelerometer (ActiGraph GT3X or Axivity) and a GPS (Qstarz BT-Q1000X) for 7 days (5 week days, 2 weekend days) to determine their level of activity and movement patterns. The GPS position was recorded every 15 seconds and their activity level was recorded and 100Hz and compiled into 15 second epochs. Data were merged and processed using HABITUS, an online tool available via the University of Southern Denmark. The processed data-points were imported into the geographical information software ArcGISpro, where optimized hot-spot analyses were conducted to identify the statistically significant spatial clusters of GPS points with higher or lower physical activity levels. For each hotspot, we identified the type of area, revealing the built environment characteristics of places with a significantly higher level of physical activity. Results: Physical activity hotspots were identified in the outdoor areas of early care and education centers (ECEC), schoolyards, as well as neighborhoods. In neighborhoods, for schoolchildren, activity hotspots primarily consist of schoolyards, sports facilities and shared backyards between multistory social housing complexes. For preschool children, neighborhood activity hotspots were primarily in private yards, ECECs, public parks, and shopping areas. In schoolyards, activity hotspots were primarily at a ball-game areas, climbing areas, and open spaces. For ECECs, activity hotspots were in many different types of areas, but more often in open spaces and areas with large fixed-play-equipment. Conclusions: Collecting and processing accelerometer and GPS data is time-consuming, but in combination with the optimized hot-spot analysis tool in ArcGISpro, the data provides unique possibilities to identify locations where the activity level is significantly higher (or lower) than the average. Classifying built environmental characteristics of these locations reveals which type of environments are most important for physical activity, for different age groups and genders, at different geographic scales.

Basic integration of GPS and accelerometer data to address a range of spatially based research questions

Jordan Carlson1, Adrian Ortega2, Chelsea Steel1, James Sallis3, Brian Saelens4, Jacqueline Kerr3, Jasper Schipperijn5, Ann Davis2, Bethany Forseth2

1Children's Mercy Kansas City, 2niversity of Kansas, 2University of California San Diego, 3Telethon Kids Institute, 4Seattle Children’s, 5University of Southern Denmark

Background: Using Global Positioning Systems (GPS) trackers in research provides additional details about a person's activity patterns that can inform multilevel interventions. The contextual information provided by GPS can improve understanding of where activity occurs and how correlates of activity may differ by setting. This presentation covers basic tools for integrating GPS and accelerometer data and provides example data and findings related to various research questions aided by GPS. Methods: Data were from three observational studies that included concurrent wear of ActiGraph accelerometers and GPS trackers (e.g., QStarz). In Study 1, GPS-based trip detection algorithms and consumer wearables were tested for their validity for detecting pedestrian, cycling, and vehicle trips in 34 youth and adults. In Studies 2 and 3, pre-determined activity locations (e.g., at home, at school, in parks, in the home neighborhood) were investigated in 55 children living in rural communities and 472 young adolescents living in high and low walkable urban/suburban neighborhoods, respectively. Study 3 also involved assessing location-general and locations-specific environmental and psychosocial correlates of physical activity to inform their relative role in interventions. A description of the GPS processing systems used will be provided, including ArcGIS, HABITUS (Human Activity Behavior Identification Tool and data Unification System), PALMSplusR (R package), and post-processing tools. Results: The trip detection algorithms identified and correctly classified the mode of 75.6%, 94.5%, and 96.9% of pedestrian, cycling, and vehicle trips (F1s=0.84 and 0.87) and were superior to Fitbit's SmartTrack and Garmin's Move IQ. Post-processing strategies for improving GPS-based pedestrian trip classification were identified. Although about half of adolescents' overall physical activity occurred at school, when accounting for time spent in each location urban/suburban adolescents were least physically active at home (2.5 min/hour of wear time) and school (2.9 min/hour of wear time) compared to "other" locations (5.9 min/hour of wear time). Analyses for the rural children (Study 2) are pending. In Study 3, no location-general psychosocial factors were related to activity in all locations. Most location-specific environmental and psychosocial factors were associated with activity in the matching location(s) only. Conclusions: Several GPS data processing tools exist that can be implemented by researchers with introductory to intermediate geospatial expertise. Available trip-detection algorithms for GPS data have good validity in children and adults. Understanding how much time people spend in active trip modes and in physical activity in various locations can inform intervention targets for supporting overall activity. The findings regarding correlates of physical activity suggest that both environmental and psychosocial correlates of activity are often location specific.

Comparing time-weighted spatial averaging derived measures of environmental exposures and associations with physical activity

Marta Jankowska1, Jay Yang1, Calvin Tribby1, Nivedita Nukavarapu1, Tarik Benmarhnia2

1City of Hope, 2University of California San Diego

Background and Aim: Time-weighted spatial averaging approaches (TWSA) for deriving environmental exposures are growing in use as deployment of Global Positioning Devices (GPS) is becoming more common in health-related studies. TWSAs measure mobility based environmental exposure while also accounting for time spent in locations, however their utility for relating environmental exposures to physical activity (PA) is unknown. Greater spatiotemporal accuracy in measurement of environmental exposures may prove important for detecting and understanding associations between PA and built environments. Methods: Participants (N = 596; mean age = 59 years; 56% female; 42% Hispanic) from the Community of Mine study in San Diego County, USA wore hip ActiGraph GT3X+ accelerometers and Qstarz GPS devices for two weeks. Accelerometer cut points with cpm were used to classify weekly light PA (100-759 cpm) and moderate to vigorous PA (MVPA) (>759 cpm). Two TWSA activity spaces were computed for each participant's total GPS wear time (kernel density estimation - KDE, and density ranking - DR). TWSA activity spaces were used to measure exposure to three activity-related environments (walkability, recreation opportunities, and greenness). OLS regression measured TWSA exposure associations with PA outcomes, controlling for sex, age, ethnicity, and total device wear time. As a comparison, OLS regressions were also run for 1000m buffer from home exposures to the three environments. Results: Participants had a weekly average of 26.8 hours of light PA and 12.5 hours of MVPA. DR measured exposure to recreation opportunities was associated with decreased MVPA (β=-17.3, 95% CI[-28.1, -6.4]), as was DR measured walkability (β=-2.4, 95% CI[-3.8, -1.1]) and greenness (β=-57.7, 95% CI[-114.5, -0.9]). DR measured exposures were not associated with light PA. KDE measured walkability exposure was associated with decreased light PA (β=-23.5, 95% CI[-45.6, -1.3]). No other associations were detected in this sample between exposures and light PA. No home buffer measured exposures were associated with PA outcomes. Conclusion: TWSA exposure results show a counterintuitive, but consistent relationship between increased time spent in green, walkable, and recreation opportune places with reduced PA time. In comparison, no relationships were found between PA time and home buffer exposure measures. By accounting for both the total exposure of individuals as well as the time they spend in locations, we may be better able to detect relationships between environmental exposures and physical activity through more sensitive and accurate measures of exposure. Further work will need to be done to understand the counterintuitive associations found in this study.

Identifying children´s play episodes using density-based clustering methods

Jing-Huei Huang1, Scott Ogletree2, Oriol Marquet3, Jasper Schipperijn4, Claudia Alberico, Myron Floyd1, Aaron Hipp1

1North Carolina State University, 2University of Edinburgh, 3Universitat Autònoma de Barcelona, 4University of Southern Denmark

Identifying children´s play episodes using density-based clustering methods Jing-Huei Huang¹, Scott Ogletree², Oriol Marquet³, Jasper Schipperijn&sup4;, Claudia Alberico, Myron Floyd¹, J. Aaron Hipp¹ ¹North Carolina State University, ²University of Edinburgh, ³Universitat Autònoma de Barcelona, &sup4;University of Southern Denmark Background and Aim: Play is essential to children´s physical, cognitive, and social skill development. Understanding behaviors in playspaces will inform design and management that encourages the variety and enjoyment of play across communities. Accelerometers and the global positioning system (GPS) have been adopted to investigate children´s play patterns. However, it is challenging to analyze play patterns as children's free play is spontaneous, creative, changes over time and across spaces, and could vary by individual. This study aims to systematically identify and characterize play episodes using density-based clustering methods, which detect children's movements that cluster together in space and time. Methods: 324 children (5-9 years) were recruited in 12 neighborhood parks in New York City and Raleigh/Durham, NC, in spring/summer 2017-2018 to wear accelerometer and GPS for an average of 25 minutes, recording location and activity intensity of play. Caregivers reported demographic information through surveys. The dataset consisted of 38,792 points of accelerometer and GPS data joined at 15-second epochs, along with associated individual characteristics. The density-based clustering method, Multi-scale (OPTICS), identified clusters (i.e., play episodes) that consisted of at least 5 data points (≥1 minute). Identified clusters were mapped to playspaces in parks, including play areas (e.g., play structures), sport pitches (i.e., courts and fields), in-between features, and areas surrounding parks (e.g., sidewalks). Results: 1,723 play episodes were identified from collected data. On average, a child´s play consisted of five play episodes with a 2.94-minute duration and 17 meters/minute velocity. For each play episode, a child maintained moderate to vigorous intensity physical activity (MVPA) for 28% of the time. Of the 1,723 episodes, 20% were solely in play areas, 6% in sports pitches, 22% strictly in-between features, and 3% were outside of parks while 49% were across multiple areas in parks. Average time spent across spaces in/around parks varied by individual characteristics. Children maintaining an accelerometer average above the MVPA threshold (>573) spent more time in areas designated for play (+6%) and less time in spaces between features (-7%), compared to children less active. Girls spent more time in play areas (+5%) and between features (+4%) whereas boys spent more time in sports pitches (+10%). Conclusions: Results demonstrate characteristics of play episodes and how spaces in parks are used for children´s play. Findings highlight that children´s free play occurs across spaces, and not necessarily concentrated in areas designated for play, which implies the importance of spatial arrangement of various park features to the diversity and intensity of play. Advancing this methodology could provide valuable information for practitioners to better design play features and their layout that support active and meaningful play.

The CNN Hip Accelerometer Posture (CHAP) Suite: Leveraging deep learning to close the gap between thigh and hip accelerometry in the free-living measurement of sitting behavior

Mikael Anne Greenwood-Hickman1, Jordan Carlson2, Marta Jankowska3, Paul Hibbing4, Dori Rosenberg1, Loki Natarajan5

1Kaiser Permanente Washington Health Research Institute, 2Children’s Mercy Kansas, 3City of Hope, 4Children’s Mercy Kansas City, 4University of California San Diego

Summary: In this symposium (chair/moderator: Dr. Dori Rosenberg, Kaiser Permanente Washington Health Research Institute), we showcase the development and application of the novel CNN Hip Accelerometer Posture (CHAP) suite. CHAP models give highly accurate posture readings from hip accelerometry data using a combination of convolutional neural networks (CNN) and bi-directional long short-term memory networks, trained against ground truth labels from concurrently worn activPAL inclinometers. Thus far, CHAP models have been used to predict sitting bouts and postural transitions from hip-worn ActiGraph GT3X+ data. Cross-validation results have consistently shown strong convergence with activPAL output to identify sitting bouts and postural transitions (93% balanced accuracy for older and younger adults; 88% for children), representing a major improvement over existing cut-point methods and machine learning models. Thus, CHAP represents a significant step forward in the accurate measurement of sitting behavior in large cohorts and could facilitate improved single-monitor measurements across the full spectrum of 24-hr waking behavior (e.g., including both posture and intensity of movement). We will guide attendees through CHAP's development, rigorous validation and refinement, and application in two major cohorts as a way of demonstrating its potential for long-term impact on health-related physical behavior research. Speaker 1, Ms. Mikael Anne Greenwood-Hickman, will outline the critical need for approaches such as CHAP and describe its development and initial validation using concurrently collected hip-worn Actigraph and thigh-worn activPAL data from the Adult Changes in Thought cohort (age 65+). Speaker 2, Dr. Jordan Carlson will share how CHAP was further refined and validated for a broader age range in the AusDiab (age 35+) and Patterns of Habitual Activity Across Seasons (PHASE; age 8-11) cohorts, demonstrating the method's validity across cohorts and age groups. Speaker 3, Dr. Marta M. Jankowska, will then share findings from an application of CHAP in an independent cohort of 602 adults age 35+ to compare CHAP- and cutpoint-derived sitting measures and their associations with metabolic syndrome. Finally, Speaker 4, Dr. Paul Hibbing, will share an additional example of CHAP's utility for pediatric obesity research, using data from an international cohort to compare how CHAP-derived versus cut-point-derived sitting variables are associated with obesity markers (e.g., waist circumference and BMI z-score). Following individual presentations, Dr. Rosenberg will share a brief synthesis statement and then moderate audience discussion on 2 key questions: 1) How can CHAP facilitate improved measurement and study of the full spectrum of physical behavior and the 24-hour day? and 2) When should CHAP be used instead of cutpoints, and what are the tradeoffs, limitations, and synergies of each approach?

The CNN Hip Accelerometer Posture (CHAP) method for classifying sitting patterns from hip accelerometers: development and initial validation in a sample of older adults

Mikael Anne Greenwood-Hickman1, Supun Nakandala2, Marta Jankowska3, Dori Rosenberg1, Fatima Tuz-Zahra2, John Bellettiere2, Jordan Carlson4, Paul Hibbing4, Jingjing Zou2, Andrea LaCroix2, Arun Kumar2, Loki Natarajan2

1Kaiser Permanente Washington Health Research Institute, 2University of California San Diego, 3City of Hope, 4Children’s Mercy Kansas City

Background & Aim: There is growing interest in using a single wearable device (e.g., hip-worn accelerometer) to measure the full spectrum of 24-h physical behavior, from sitting time and patterns to vigorous physical activity. Traditional cutpoint methods, useful for measuring activity intensity, lack the ability to accurately detect postures and postural transitions, often overestimating these transitions and underestimating prolonged sitting bouts. To overcome this limitation, we developed the Convolutional Neural Network (CNN) Hip Accelerometer Posture (CHAP) classification method. Methods: CHAP combines a CNN with a bi-directional long short-term memory network (BiLSTM) and a Softmax output layer to predict sitting or non-sitting posture from raw hip-worn acceleration data. Initial development of CHAP leveraged data from 709 free-living older adults (age 65+ y) in the Adult Changes in Thought (ACT) study who concurrently wore hip-based ActiGraph GT3X+ and thigh-based activPAL devices for ~7 days. Non-overlapping 10 s epochs of input ActiGraph data and ground truth sitting vs. non-sitting labels from activPAL data were compiled, and first fed into CHAP's CNN layer, which automatically learned unique features of the data through repeated iterative processing in each 10 s epoch independently. Next, CNN output features were smoothed with the BiLSTM layer, which overcame the CNN's assumption of temporal independence between each 10 s epoch to automatically learn temporal features of the data. Finally, all learned features were processed by a Softmax output layer, which assigned final output behavioral classification labels by converting the refined output features from the BiLSTM into probabilities of each 10 s epoch belonging to either sitting or non-sitting behavior and selecting the label with the highest probability. CHAP-derived sitting measures, along with those from cutpoints (<100 counts/min) and an alternative machine learned algorithm (Two Level Behavior Classification [TLBC]) were validated against activPAL data. Models were developed on a training set and evaluated on a held-out test set. Results: At the minute level, CHAP had higher mean classification agreement than other methods (93% vs. 74%-83%). Detection of sit-to-stand transitions was also better, with sensitivity of 83% (vs. 26% for TLBC and 72% for cutpoint) and precision of 83% (vs. 30% for cutpoint and 71% for TLBC). At the day level, CHAP predicted similar mean sitting bout duration to activPAL (15.7 versus 15.4 min) with no significant difference, whereas other methods differed considerably and significantly (9.4 min for cutpoint and 49.4 min for TLBC). Conclusion: CHAP showed outstanding validity for classifying sitting and non-sitting posture in a free-living sample of older adults. This dramatically increases the potential of hip-worn devices to assess sitting time, patterns, and 24-h physical behaviors. Future work will refine the CHAP method in broader age groups.

The CHAP data processing tools for estimating sit-to-stand transitions and sitting bout patterns from hip ActiGraph data among children and adults

Jordan Carlson1, Nicola Ridgers2, Supun Nakandala3, Arun Kumar3, Rong Zablocki3, Fatima Tuz-Zahra3, John Bellettiere3, Jingjing Zou3, Andrea LaCroix3, Lindsay Dillon3, Marta Jankowska4, Paul Hibbing1, Chelsea Steel1, Dori Rosenberg5, Mikael Anne Greenwood-Hickman5, Chongzhi Di6, Elizabeth Winkler7, Genevieve Healy7, David Dunstan8, Neville Owen8, Loki Natarajan3

1Children's Mercy Kansas City, 2University of South Australia, 3University of California San Diego, 4City of Hope, 5Kaiser Permanente Washington Health Research Institute, 6Fred Hutchinson Cancer Center, 7University of Queensland, 8Baker Heart and Diabetes Institute

Background: Sedentary variables are commonly estimated from hip-worn accelerometer data using counts-based cut-points (e.g., 100 counts per minute [cpm]). However, cut-points do not accurately measure sit-to-stand transitions and sitting bout patterns. Improved processing/classification methods would enrich the evidence base and inform the development of more effective public health guidelines. This presentation will cover the development and evaluation of the CHAP (CNN Hip Accelerometer Posture) data scoring/classification method in children and adults. Methods. Data were from 278 children (up to 4 time points each) ages 8-11y from the Patterns of Habitual Activity Across Seasons (PHASE) study and 1397 adults ages 35-90y from the Australian Diabetes, Obesity and Lifestyle (AusDiab) and Adult Changes in Thought (ACT) studies. Assessments involved ~7d of concurrently wearing a thigh-worn activPAL (ground truth) and hip-worn ActiGraph (test measure). Separately for children and adults, data from two-thirds of the participants were used to train a CHAP deep learning model that classified each 10-second epoch of raw ActiGraph acceleration data as sitting or not sitting, creating comparable information with the ground truth measure (activPAL). In the remaining one-third of participants, the two CHAP models (child and adult) were evaluated alongside the standard 100cpm method for hip-worn ActiGraph monitors. Performance was tested for each 10-second epoch and for participant-level total sitting time and five sitting bout variables (e.g., mean bout duration). Results: CHAP-child correctly classified 10-second epochs as sitting or not sitting with a mean balanced accuracy of 87.6% (SD=5.3%) across participants. Sit-to-stand transitions were correctly classified with a mean sensitivity of 76.3% (SD=8.3). For most participant-level variables, CHAP-child estimates had a mean absolute percent error (MAPE) of ≤11% compared to activPAL, and very large correlations with activPAL (r>0.80). For the 100cpm method, most MAPEs were >30% and most correlations were small or moderate (r≤0.60). CHAP-adult showed similar performance as CHAP-child and to the previously developed older adult algorithm (CHAP-OlderAdult). Balanced accuracy for CHAP-adult was 92.6% and sensitivity for sit-to-stand transitions was 74.4%. MAPE for mean sitting bout duration was 12.2% (vs. 10.6% in children). All correlations were r≥0.78. Error was generally consistent across age, sex, and BMI groups. Conclusions: There was strong support for the validity of the CHAP-child and CHAP-adult data scoring/classification methods, which allow researchers to derive activPAL-equivalent measures of sitting time, sit-to-stand transitions, and sitting bout patterns from hip-worn triaxial ActiGraph data. Applying CHAP to existing datasets may accelerate the development of more specific public health guidelines around sitting patterns. CHAP is freely available at https://github.com/ADALabUCSD/DeepPostures.

A comparison of the CHAP versus cut point method for measuring accelerometry derived sitting patterns as associated with metabolic syndrome in adults

Marta Jankowska1, Calvin Tribby1, Dorothy Sears2, Dori Rosenberg3, Fatima Tuz-Zahra4, John Bellettiere4, Jordan Carlson5, Paul Hibbing5, Jingjing Zou4, Andrea LaCroix4, Mikael Anne Greenwood-Hickman3, Supun Nakandala4, Arun Kumar4, Loki Natarajan4

1City of Hope, 2Arizona State University, 3Kaiser Permanente Washington Health Research Institute, 4University of California San Diego, 5Children's Mercy Kansas City

Background and Aim: There is growing interest in assessment of how sitting behavior patterns (SPs) are associated with metabolic syndrome (MetS). However, study of associations between SPs and health outcomes may be limited by hip-based accelerometry cut point methods, which measure sedentary time (e.g., sitting and standing) rather than postural transitions (e.g., sit to stand). We used the Convolutional Neural Network Hip Accelerometer Posture (CHAP) algorithm to overcome this limitation and compare CHAP to cut point-derived SPs and their associations with MetS. Methods: Participants (N = 583; mean age = 59 years; 56% female; 42% Hispanic) from the Community of Mine study wore hip ActiGraph GT3X+ accelerometers for two weeks and completed anthropometric measurements and blood draw. We utilized the CHAP algorithm as a measure of sitting compared with the sedentary time cut point (≤ 100 counts/min), and generated three SP measures: median bout duration (mins), time in bouts ≥ 30 mins (hrs), and daily number of breaks. MetS was defined as having at least three of five clinically measured metabolic risk factors per NCEP ATP III: increased waist circumference, elevated triglycerides, low HDL cholesterol, hypertension, and impaired fasting glucose. Binary logistic regression was used to assess SP associations with MetS, controlling for sex, age, education, ethnicity, MetS related medication use, and device wear time. Results: A total of 153 participants (26%) had MetS. Cut point measured sedentary time was 8.7 hours per day, while CHAP measured sitting time was 8.6 hours per day. CHAP SPs measured fewer breaks (44.5 vs. 83.1 per day), longer median bouts (5.0 vs. 2.5 min), and more hours of time spent in bouts ≥ 30 mins (4.5 vs. 2.8). We found a significant increase in the odds of having MetS per one hour increase sedentary time (OR = 1.20, 95% CI [1.03, 1.40]) and sitting time (OR = 1.22, 95% CI [1.07, 1.38]). Increase per minute of median bout duration was associated with significant increase in odds of MetS for both sedentary (OR = 1.43, 95% CI [1.08, 1.90]) and sitting time (OR = 1.20, 95% CI [1.09, 1.30]), as was hours spent in bouts ≥ 30 mins: sedentary OR=1.16 95% CI [1.01, 1.34], sitting OR=1.17 95% CI [1.05, 1.30]. Number of daily sedentary breaks was not associated with MetS for either measure. Conclusions: In this population, CHAP measured less fragmented SPs with longer bouts and more time spent in prolonged bouts compared to SPs using cut points. Significant increase in odds of MetS was found using cut point and CHAP measures of total sitting/sedentary time, median bout duration, and hours spent in bouts ≥ 30 mins, indicating that both sedentary and sitting time are important predictors of MetS. Differences in association magnitudes, particularly for median bout length, points to behaviorally relevant intervention opportunities for increasing fragmentation of sitting bouts.

Deep-learned sedentary patterns and obesity in the International Study of Childhood Obesity (ISCOLE): Results from the CHAP-child model

Paul Hibbing1, Jordan Carlson1, Chelsea Steel1, Mikael Anne Greenwood-Hickman2, Supun Nakandala3, Marta Jankowska4, Dori Rosenberg2, Fatima Tuz-Zahra3, John Bellettiere3, Jingjing Zou3, Andrea LaCroix3, Arun Kumar1, Peter Katzmarzy5, Loki Natarajan3

1Children's Mercy Kansas City, 2Kaiser Permanente Washington Health Research Institute, 3University of California San Diego, 4City of Hope, 5Louisiana State University

Objectives: Sedentary behavior (SB) is associated with obesity in adults, but evidence is mixed regarding its role in pediatric obesity. The discrepant findings can potentially be resolved with the improved measures available through the newly released CNN Hip Accelerometer Posture suite, particularly the child-specific model (CHAP-child). The purpose of this study was to examine associations of SB-related metrics (derived from CHAP-child versus a traditional cut-point) with obesity-related outcomes in the International Study of Childhood Obesity, Lifestyle, and the Environment (ISCOLE). Methods: Accelerometer data were analyzed from 5880 children in 12 countries (54% female; age 9-12 y; 129-860 per country). Participants wore an ActiGraph GT3X+ on their right hip for a median of 7 days. Data were processed using a cut-point (≤ 100 counts per minute) and the CHAP-child model. For both methods, total SB time was extracted along with mean and median SB bout duration. Standardized linear mixed effects models were fitted to compare each variable with waist circumference, body fat percentage, and body mass index z-score (BMI-z), while accounting for participant nesting within schools and countries. Model 1 was adjusted for age, sex, ethnicity, parental education, and maturity offset. Model 2 was adjusted for the same variables, plus percent of time spent in moderate-to-vigorous physical activity (MVPA%). P-values were adjusted using the Bonferroni method. Results: Summary statistics (mean ± SD) were 64.1 ± 8.8 cm for waist circumference, 20.9% ± 7.6% for body fat, and 0.44 ± 1.24 units for BMI-z. In general, all standardized regression coefficients were small, and statistical significance was mixed (see Figure 56). In Model 1, total SB time (8.7 ± 1.8 hr/day for the cut-point; 11.9 ± 1.5 hr/day for CHAP-child) had slightly stronger associations with all outcomes when using the cut-point versus CHAP-child. The opposite was generally true for mean SB bout duration (5.5 ± 2.8 min for the cut-point; 8.2 ± 3.5 min for CHAP-child) and median SB bout duration (2.1 ± 0.3 min and 2.1 ± 0.5 min, respectively). In Model 2, associations were non-significant with the cut-point, whereas with CHAP-youth the associations were significant for total SB time (body fat percentage only; β = -0.06) and median SB bout length (all outcomes; β = 0.06-0.09). Conclusions: Compared to SB pattern variables from the cut-point, CHAP-child variables showed marginally stronger associations with obesity outcomes. The strongest associations were seen for median SB bout duration, but more research is needed to examine other SB pattern variables (e.g., usual bout duration, Gini index). Accounting for MVPA% attenuated the associations for most SB-related variables. Longitudinal studies may be needed to fully characterize the impact of SB on obesity and related health outcomes.

Figure 1
Figure 1

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 2
Figure 2

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 3
Figure 3

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 4
Figure 4

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 5
Figure 5

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 6
Figure 6

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 7
Figure 7

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 8
Figure 8

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 9
Figure 9

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 10
Figure 10

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 11
Figure 11

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 12
Figure 12

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 13
Figure 13

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 14
Figure 14

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 15
Figure 15

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 16
Figure 16

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 17
Figure 17

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 18
Figure 18

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 19
Figure 19

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 20
Figure 20

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 21
Figure 21

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 22
Figure 22

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 23
Figure 23

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 24
Figure 24

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 25
Figure 25

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 26
Figure 26

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 27
Figure 27

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 28
Figure 28

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 29
Figure 29

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 30
Figure 30

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 31
Figure 31

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 32
Figure 32

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 33
Figure 33

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 34
Figure 34

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 35
Figure 35

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 36
Figure 36

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 37
Figure 37

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 38
Figure 38

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 39
Figure 39

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 40
Figure 40

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 41
Figure 41

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 42
Figure 42

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 43
Figure 43

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 44
Figure 44

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 45
Figure 45

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 46
Figure 46

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 47
Figure 47

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 48
Figure 48

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 49
Figure 49

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 50
Figure 50

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 51
Figure 51

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 52
Figure 52

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 53
Figure 53

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 54
Figure 54

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 55
Figure 55

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

Figure 56
Figure 56

Citation: Journal for the Measurement of Physical Behaviour 5, 4; 10.1123/jmpb.2022-0032

  • Collapse
  • Expand
All Time Past Year Past 30 Days
Abstract Views 245 245 0
Full Text Views 702 702 323
PDF Downloads 280 280 123