Explainable Prediction of Non-Contact Injury Risk Using Machine Learning

Abstract

Injuries not only hinder players’ competitive performance and increase team medical expenses, but also affect players’ long-term or even lifelong physical activity. The use of machine learning algorithms to model the relationship between an player’s training data and risk of injury can help to assess an player’s risk of injury and provide a basis for decision making on training load adjustment. However, existing research reports mostly focus on the prediction accuracy improvement of the model while ignoring the explainability and reliability issues of the model, which makes the prediction model in practical application rate reduced. Therefore, this study collected data on the training load, subjective perceived wellness, menstrual status, athletic ability test and injury information of 18 young female basketball players in Fujian Province during their preparation for the 2021 National Games of the People’s Republic of China. Multiple machine learning algorithms were used to construct non-contact injury risk prediction models for young female basketball players. Experiments show that the model based on eXtreme Gradient Boosting (XGBoost) has the best performance and can detect approximately 89.4% of non-contact injury risk with a prediction of 66.6%, outperforming other counterparts. Four factors influencing the risk of non-contact injury were screened by the model’s importance variable analysis, and the results were highly consistent with research reported in sports science, demonstrating the reliability of the model. The application of the model can provide a reliable basis for decision making in sports injury prevention practice.