751 MEP162 – A preliminary study on 284 football players using a hamstring injury prediction with machine learning

AbstractBackground Hamstring injury stands as the primary diagnostic concern football. Consequently, prioritising the development and implementation of risk mitigation strategies for these injuries is of importance for both football players and their support networks. Therefore, identifying athletes with a high risk of hamstring injuries is necessary. In this context, using prognostic modelling approaches through machine learning technics emerges as a promising solution.Objectives We aim to assess the ability of several machine learning techniques, namely logistic regression, random forests and an AdaBoost method, to predict the occurrence of hamstring injuries in football players.Design Prospective cohort studySetting Screening on football players with sprints mechanical assessment in field conditions at the start of the season.Participants Male football players (n=284) from 16 professional or semi-professional teams in three countries (Japan, France, and Finland).Main Outcomes Measurements (i) Model classification metrics with the proportion of injuries among positive predictions (precision), the proportion of injury and non-injury classification (recall), the global ability to make good classification (accuracy) and two global measures (F1-Score, ROC-AUC score). (ii) Feature importance classification for each model.Results Over each 200 model iterations, logistic regression had the highest recall, F1-Score, and ROC-AUC score with 0.60 ± 0.14, 0.33 ± 0.06, and 0.68 ± 0.08 respectively. Random forest had the highest precision, specificity, and accuracy with 0.27 ± 0.12, 0.89 ± 0.05, and 0.80 ± 0.04 respectively. The most important feature where 10m time sprint and the number of hours of training per week for the logistic regression and random forest. There is a high variability in coefficient value overs the 200 iterations for each model.Conclusions This preliminary study demonstrates the feasibility of a predictive algorithm for football hamstring injuries. However, it's crucial to assess diverse metrics to describe model performance and maintain an integrated perspective of analyses.