Machine Learning for Injury Risk Prediction and Prevention in Athletics: A Data Driven Framework Integrating Wearable Sensor Metrics, Neuromuscular Screening, and Ensemble Learning for Prospective Injury Surveil

Abstract

Common and expensive injuries in college and professional sports include hamstring strains, anterior cruciate ligament tears and ankle sprains. Such injuries can significantly impact player well-being, team effectiveness and healthcare costs. Traditional injury prevention strategies may be based on individual clinical parameters which do not sufficiently address the complex and dynamic factors that contribute to injury. In this work, machine learning (ML) models for predicting non-contact musculoskeletal injuries (NC MSI) were studied. The large data set included information on training load, results of neuromuscular tests, data about the biomechanics of the wearable device, self-reported sleep and wellness data and a history of injuries for each person. Over the course of two season, 412 collegiate athletes from six different sports participated in prospective cohort research that produced 8,946 athlete-week observations. In order to prevent data leakage, four algorithms [(logistic regression, support vector machine, random forest, extreme gradient boosting, (XGBoost)] were tested using the cross validation (CV) approach, which was done in a layered fashion using the athletes as a group. The performance was assessed using several metrics such as AUC-ROC, sensitivity, specificity and F1-score. XGBoost had an AUC of 0.89, sensitivity of 0.85 and specificity of 0.88, which was significantly better than logistic regression (AUC = 0.75, p < 0.001), support vector machine (AUC = 0.81, p= 0.001) and random forest (AUC = 0.87). The acute and chronic workload ratio, past injury history and countermovement jump asymmetry were identified as some of the main prediction of performance and were consistent with known physiological effects, through the use of SHapley Additive explanation (SHAP). This was an observational comparison but there was a lower number of recruitment injuries among the athletes identified as high-risk and referred for clinical directed load adjustment during second season deployment compared with historical control. The findings agree with the use of ensemble machine learning (ML) for athlete monitoring (AM) systems to deliver personalized and comprehensible risk assessments. The results, however, suggest that technical tools are a tool to complement clinical judgement when dealing with managing an athlete’s health not replace it.

 Machine learning, Injury prediction, Sports medicine, Wearable technology, XGBoost, Workload monitoring, Athlete health.