First results of a study carried out as part of the Master Degree Mention Training and Optimization of Sports Performance, Sciences and Techniques of Physical Training Course of the University. Collaborative work between the EuroMov laboratory and the ITM Mines d’Alès under the direction of Stéphane PERREY (PR, Univ. Montpellier) and Nicolas SUTTON-CHARANI (MA, IMT Mines Alès).
On the front of the study described below, work is currently being carried out on a thesis, the aim of which is to increase the degree of reliability of injury risk prediction using « machine learning » methods by including more parameters taking into account different factors of the person’s internal and external loads. The effect of predictive planning on the potential risk of injury will also be analyzed.
Introduction. In professional soccer, many studies (e.g., Ehrmann et al., 2016) focused on the control of training and match loads and their potential effects on athletes’ injuries. Most of the relevant methods in the literature assume simple linear links between injury risk and training performance in professional soccer. Recently, industry-led machine learning methods have been proposed for evaluating more complex relations (Rossi et al., 2018). Indeed, the latter study entirely used GPS tracking technology (external load) for describing the training workload of players in a professional soccer club during a season. Besides objective physical workload, it is also possible to use subjective measures (internal load) to predict injury. Therefore, the purpose of this study was to disambiguate the multitude of relationships between several factors related to internal and external loads that contribute to predict injuries. A variety of machine learning models were applied / compared in predictive accuracy terms.
Methods. Twenty-five players (age: 28.7± 6.2 years; height: 178.1± 4.2 cm; body mass: 76.9± 9.2 kg) from the same French Ligue 2 team were observed during 245 training sessions, 38 Domino’s Ligue 2 matches, 2 ‘Coupe de la Ligue’ matches and 3 ‘Coupe de France’ matches during the 2017-2018 season. Predictive models were fed with data collected by GPS (Optimeye S5, Catapult Innovation, 10 Hz), subjective perception questionnaires (wellness, perceived exertion) and heterogeneous training data. The GPS data and questionnaires respectively represent the external and internal training loads and training data contain injuries information.
Machine learning models such as decision trees, random forest, SVM and neural network were applied in order to predict injuries from the two types of training load variables. Decision trees were finally interpreted in order to get a better understanding of the relation between the external and internal training loads and injuries.
Results and discussion. Preliminary analyses show encouraging results in terms of accuracy for injury prediction (i.e. 94% and 74% for the next week and month, respectively) with error rates decreasing between 50% and 20% compared to randomness. Using injury predictive decision trees, first nodes are almost always associated with subjective variables. Individualization seems a promising track, since interpretable models such as decision trees enable players’ personality representation, which can improve in turn training programs’ efficiency by adapting them to players’ profile.
Conclusion. The internal load might play an important role in injuries occurrence.
References.
Ehrmann, F. E. (2016). J. Strength Cond. Res. DOI: 10.1519/JSC.0000000000001093.
Rossi, A., et al. (2018). Plos One. DOI:10.1371/journal.pone.0201264
Mémoire Master 2 – Entraînement et Optimisation de la Performance Sportive :
Memoire_MASTER-EOPS-Emmanuel_VALLANCEPoster ACAPS 2019 – Paris – Octobre 2019 :
Machine-learning-approach-for-predicting-injury-in-soccer-from-external-and-internal-training-loads-EMV