One Size Doesn't Fit All: Supervised Machine Learning Classification in Athlete-Monitoring

Abstract

Athlete movement data are integral for optimizing athlete-performance and can lead to reduced fatigue and in turn can mitigate injury risk. There is a substantial amount of scientific literature, which investigates the ability of computer-vision and inertial sensor technologies to classify sport-specific movements. The coupling of automatic sport action labeling and athlete-monitoring data can significantly enhance athlete work-load monitoring. Two recent systematic reviews of the literature, pertinent to sport-specific movement classification, revealed that the majority of journal articles use athlete-dependent classification model training and evaluation methods. These methods can significantly enhance model classification performance, particularly with movements which have high interathlete technique variation. This is because it enables models to learn features distinctive to all athletes during training. This letter details the training and evaluation of supervised machine learning models to automatically classify running surface (athletics track, hard sand, and soft sand) using features extracted from an upper-back inertial measurement unit sensor. Possible classification performance enhancement is demonstrated by comparing athlete independent and athlete dependent supervised machine learning methods. Using athlete dependent methods significantly increased the classification performance in terms of weighted average precision, recall, F1-score, and accuracy (p <; 0.05).