The aim of this study was to quantify and predict relationships between rating of perceived exertion (RPE) and GPS training-load (TL) variables in professional Australian football (AF) players using group and individualized modeling approaches.
TL data (GPS and RPE) for 41 professional AF players were obtained over a period of 27 wk. A total of 2711 training observations were analyzed with a total of 66 ± 13 sessions/player (range 39–89). Separate generalized estimating equations (GEEs) and artificial-neural-network analyses (ANNs) were conducted to determine the ability to predict RPE from TL variables (ie, session distance, high-speed running [HSR], HSR %, m/min) on a group and individual basis.
Prediction error for the individualized ANN (root-mean-square error [RMSE] 1.24 ± 0.41) was lower than the group ANN (RMSE 1.42 ± 0.44), individualized GEE (RMSE 1.58 ± 0.41), and group GEE (RMSE 1.85 ± 0.49). Both the GEE and ANN models determined session distance as the most important predictor of RPE. Furthermore, importance plots generated from the ANN revealed session distance as most predictive of RPE in 36 of the 41 players, whereas HSR was predictive of RPE in just 3 players and m/min was predictive of RPE in just 2 players.
This study demonstrates that machine learning approaches may outperform more traditional methodologies with respect to predicting athlete responses to TL. These approaches enable further individualization of load monitoring, leading to more accurate training prescription and evaluation.
Bartlett, O’Connor, Pitchford, and Robertson are with the Inst of Sport, Exercise and Active Living, Victoria University, Melbourne, Australia. Torres-Ronda is with the Dept of Health and Kinesiology, Exercise and Sport Nutrition Laboratory, Texas A&M University, College Station, TX.