Using Interpretable Machine Learning to Predict Injury Risk Among Collegiate Male Basketball Players

Purpose: Current predictive methodologies for injury risk are inadequate for the individualized management of athletes. This study introduces a machine learning (ML) framework tailored to evaluate injury risk among Chinese University Basketball Association (CUBA) participants. Method: We conducted a longitudinal analysis of 681 CUBA athletes (average age: 19.75 ± 1.78 years) over three basketball seasons. Academy coaches and medical personnel undertook continuous surveillance of player exposure and injuries. We documented anthropometric measurements and physical tests assessing motor coordination and physical fitness attributes. Five ensemble learning algorithms including Random Forest (RF), Light Gradient Boosting Machine (LightGBM), categorical boosting (Catboost), Gradient Boosting Decision Tree (GBDT), and Extreme Gradient Boosting (XGBoost) were used to forecast and categorize injury risk based on its optimal performance. Results: The study identified that the XGBoost model provided the highest accuracy (0.7668). It showed the largest AUC value in the 10-fold cross-validation (0.8393; 95% CI: 0.8273-0.8512) and the smallest Brier score (0.1643; 95% CI: 0.1585-0.1700). Notably, SHAP value analysis revealed that sex, match location, sprint 30 m, age at PHV, and CMJ were significant predictors of evaluating the injury risk. Conclusion: The ensemble learning model, particularly the XGBoost algorithm, demonstrates a robust capacity for predicting and classifying injuries in basketball players. These findings suggest that integrating ML models into athletic health management can significantly enhance injury surveillance and prevention strategies. The model's reliance on physical performance as predictors underscores the multifaceted nature of sports-related injuries. This paper is supported by the "Tang Scholar" project of the Cyrus Tang Foundation in China