Abstract
This study aimed to identify relationships between external and internal load parameters with subjective ratings of perceived exertion (RPE). Consecutively, these relationships shall be used to evaluate different machine learning models and design a deep learning architecture to predict RPE in highly trained/national level soccer players. From a dataset comprising 5402 training sessions and 732 match observations, we gathered data on 174 distinct parameters, encompassing heart rate, GPS, accelerometer data and RPE (Borg's 0-10 scale) of 26 professional male professional soccer players. Nine machine learning algorithms and one deep learning architecture was employed. Rigorous preprocessing protocols were employed to ensure dataset equilibrium and minimize bias. The efficacy and generalizability of these models were evaluated through a systematic 5-fold cross-validation approach. The deep learning model exhibited highest predictive power for RPE (Mean Absolute Error: 1.08 ± 0.07). Tree-based machine learning models demonstrated high-quality predictions (Mean Absolute Error: 1.15 ± 0.03) and a higher robustness against outliers. The strongest contribution to reducing the uncertainty of RPE with the tree-based machine learning models was maximal heart rate (determining 1.81% of RPE), followed by maximal acceleration (determining 1.48%) and total distance covered in speed zone 10-13 km/h (determining 1.44%). A multitude of external and internal parameters rather than a single variable are relevant for RPE prediction in highly trained/national level soccer players, with maximum heart rate having the strongest influence on RPE. The ExtraTree Machine Learning model exhibits the lowest error rates for RPE predictions, demonstrates applicability to players not specifically considered in this investigation, and can be run on nearly any modern computer platform.