Comparing machine learning and conventional statistical approaches for injury prediction in young professional soccer players

Frequent injuries pose a problem in professional soccer that is being tackled with preventive measures. Consequently, injury prediction and prevention are also increasingly addressed from a statistical perspective. In a pilot study, several machine learning algorithms and conventional statistical approaches have been compared regarding their potential to predict time-loss non-contact lower-body injuries in professional youth soccer players, using data from a prospective cohort study with 56 players of which 22 were injured. The covariates considered here include basic soccer-related as well as neuromuscular and biomechanical features derived from physical testing. Lasso regularized logistic regression, naive Bayes, linear discriminant analysis, k -nearest neighbors, classification trees, random forests, XGBoost, and support vector machines are considered for binary classification and prediction of an injury occurrence. The prediction results from a cross-validated procedure are compared regarding multiple quality measures. Post Lasso logistic regression with a reduced penalty gives the best results with an accuracy of 0.625, a predictive likelihood of 0.593, and a Brier score of 0.228. The respective sensitivity and specificity are 0.773 and 0.529, with an AUC of 0.672. Moreover, an XGBoost model slightly outperforms the Lasso model in terms of accuracy (0.661), while for the other performance measures it is dominated by the Lasso. In addition to the specific results on the available injury data set, the proposed comparison procedure of several models for binary prediction provides a generally applicable analysis guideline. This roadmap can also be applied in other contexts where similarly structured small but rich data sets are available.