velocity estimation error. The MIMU methods discussed here are limited by sensor imperfections and model assumptions. In other biomechanics contexts, some employ machine learning (ML) techniques both for classification 14 – 17 and regression. 18 – 24 Mannini and Sabatini 18 estimated running speeds less
Reed D. Gurchiek, Hasthika S. Rupasinghe Arachchige Don, Lasanthi C. R. Pelawa Watagoda, Ryan S. McGinnis, Herman van Werkhoven, Alan R. Needle, Jeffrey M. McBride and Alan T. Arnholt
Anantha Narayanan, Farzanah Desai, Tom Stewart, Scott Duncan and Lisa Mackay
but the use of cut points is still necessary when organizing these data into intensity categories. Manually defined algorithms have also been used to classify raw data into activity types with varying levels of success. 14 More recently, researchers have employed machine-learning techniques to
Arne Jaspers, Tim Op De Beéck, Michel S. Brink, Wouter G.P. Frencken, Filip Staes, Jesse J. Davis and Werner F. Helsen
Australian football (AFL) found that artificial neural networks (ANNs), a machine learning approach, more accurately predicted the RPE in response to ELIs compared with traditional statistics. 9 Other machine learning techniques could be used for this task as well, and each technique has strengths and
David Whiteside, Olivia Cant, Molly Connolly and Machar Reid
Quantifying external workload is fundamental to training prescription in sport. In tennis, global positioning data are imprecise and fail to capture hitting loads. The current gold standard (manual notation) is time intensive and often not possible given players’ heavy travel schedules.
To develop an automated stroke-classification system to help quantify hitting load in tennis.
Nineteen athletes wore an inertial measurement unit (IMU) on their wrist during 66 video-recorded training sessions. Video footage was manually notated such that known shot type (serve, rally forehand, slice forehand, forehand volley, rally backhand, slice backhand, backhand volley, smash, or false positive) was associated with the corresponding IMU data for 28,582 shots. Six types of machine-learning models were then constructed to classify true shot type from the IMU signals.
Across 10-fold cross-validation, a cubic-kernel support vector machine classified binned shots (overhead, forehand, or backhand) with an accuracy of 97.4%. A second cubic-kernel support vector machine achieved 93.2% accuracy when classifying all 9 shot types.
With a view to monitoring external load, the combination of miniature inertial sensors and machine learning offers a practical and automated method of quantifying shot counts and discriminating shot types in elite tennis players.
Alexander H.K. Montoye, Kimberly A. Clevenger, Kelly A. Mackintosh, Melitta A. McNarry and Karin A. Pfeiffer
( Crouter, Horton, & Bassett, 2012 ; Freedson, Pober, & Janz, 2005 ) to machine learning models ( Mackintosh, Montoye, Pfeiffer, & McNarry, 2016 ; Trost, Wong, Pfeiffer, & Zheng, 2012 ), which use count-based or raw data as inputs. Machine learning models have generally yielded more accurate predictions
Ryan M. Chambers, Tim J. Gabbett and Michael H. Cole
orientation criteria. The second step extracted features of the accelerometer and gyroscope signals from each event. These calculations included summary statistics using different time windows based on the event time stamp and formed the 33 variables for the machine learning process. Variable selection was
Ferdous Wahid, Rezaul Begg, Noel Lythgo, Chris J. Hass, Saman Halgamuge and David C. Ackland
Normalization of gait data is performed to reduce the effects of intersubject variations due to physical characteristics. This study reports a multiple regression normalization approach for spatiotemporal gait data that takes into account intersubject variations in self-selected walking speed and physical properties including age, height, body mass, and sex. Spatiotemporal gait data including stride length, cadence, stance time, double support time, and stride time were obtained from healthy subjects including 782 children, 71 adults, 29 elderly subjects, and 28 elderly Parkinson’s disease (PD) patients. Data were normalized using standard dimensionless equations, a detrending method, and a multiple regression approach. After normalization using dimensionless equations and the detrending method, weak to moderate correlations between walking speed, physical properties, and spatiotemporal gait features were observed (0.01 < |r| < 0.88), whereas normalization using the multiple regression method reduced these correlations to weak values (|r| < 0.29). Data normalization using dimensionless equations and detrending resulted in significant differences in stride length and double support time of PD patients; however the multiple regression approach revealed significant differences in these features as well as in cadence, stance time, and stride time. The proposed multiple regression normalization may be useful in machine learning, gait classification, and clinical evaluation of pathological gait patterns.
Kristof Kipp, John Krzyszkowski and Daniel Kant-Hull
Purpose: To use an artificial neural network (ANN) to model the effect of 15 weeks of resistance training on changes in countermovement jump (CMJ) performance in male track-and-field athletes. Methods: Resistance training volume load (VL) of 21 male division I track-and-field athletes was monitored over the course of 15 weeks, which covered their indoor and outdoor competitive season. Weekly CMJ height was also measured and used to calculate the overall 15-week change in CMJ performance. A feed-forward ANN with 5 hidden layers was used to model how the VL from each of the 15 weeks was associated with the overall change in CMJ height. Results: Testing the performance of the developed ANN on 4 separate athletes showed that 15 weeks of VL data could predict individual changes in CMJ height with an average error between 0.21 and 1.47 cm, which suggested that the ANN adequately modeled the relationship between weekly VL and its effects on CMJ performance. In addition, analysis of the relative importance of each week in predicting changes in CMJ height indicated that the VLs during deload or taper weeks were the best predictors (10%–17%) of changes in CMJ performance. Conclusions: ANN can be used to effectively model the effects of weekly VL on changes in CMJ performance. In addition, ANN can be used to assess the relative importance of each week in predicting changes in CMJ height.
Joshua Twaites, Richard Everson, Joss Langford and Melvyn Hillsdon
focusing on the raw data are becoming more commonplace. As acceleration is only a surrogate of the true PA performed, methods for converting the data into behavioral metrics are required. A technique of growing popularity is activity classification via machine learning, which creates a model, based on some
Heidi R. Thornton, Jace A. Delaney, Grant M. Duthie and Ben J. Dascombe
To investigate the ability of various internal and external training-load (TL) monitoring measures to predict injury incidence among positional groups in professional rugby league athletes.
TL and injury data were collected across 3 seasons (2013–2015) from 25 players competing in National Rugby League competition. Daily TL data were included in the analysis, including session rating of perceived exertion (sRPE-TL), total distance (TD), high-speed-running distance (>5 m/s), and high-metabolic-power distance (HPD; >20 W/kg). Rolling sums were calculated, nontraining days were removed, and athletes’ corresponding injury status was marked as “available” or “unavailable.” Linear (generalized estimating equations) and nonlinear (random forest; RF) statistical methods were adopted.
Injury risk factors varied according to positional group. For adjustables, the TL variables associated most highly with injury were 7-d TD and 7-d HPD, whereas for hit-up forwards they were sRPE-TL ratio and 14-d TD. For outside backs, 21- and 28-d sRPE-TL were identified, and for wide-running forwards, sRPE-TL ratio. The individual RF models showed that the importance of the TL variables in injury incidence varied between athletes.
Differences in risk factors were recognized between positional groups and individual athletes, likely due to varied physiological capacities and physical demands. Furthermore, these results suggest that robust machine-learning techniques can appropriately monitor injury risk in professional team-sport athletes.