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 Surveillance in Collegiate Athletes

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