SIRP-600: An Interpretable Machine Learning Framework for Sports Injury Risk Prediction Using SHAP-Enhanced Ensemble Methods

Abstract

Sports injury prevention remains a critical challenge in athletic training, where early risk identification can significantly improve athlete safety and performance. This study introduces SIRP-600, a comprehensive dataset of 600 athlete samples with 15 validated risk indicators spanning demographic, behavioral, physiological, and historical factors. We systematically evaluate seven machine learning algorithms for binary injury risk classification, demonstrating that ensemble tree-based methods achieve superior predictive performance. XGBoost attains the highest test set AUC of 0.946, with Extra Trees achieving perfect precision (100%) and recall (100%) on the validation set. Confusion matrix analysis reveals that ensemble methods maintain low false positive (≤4%) and false negative (≤1.8%) rates, critical for practical injury prevention applications. SHAP interpretability analysis identifies injury history, training intensity, and sleep hours as the three most influential predictors, providing actionable insights for targeted interventions. Our results demonstrate that machine learning combined with explainable AI offers a robust, interpretable framework for proactive sports injury risk assessment, enabling personalized prevention strategies based on modifiable risk factors.

Ähnliche Arbeiten