A Data-Driven Machine Learning Framework For Detecting Fatigue-Related Injury Risk Using Physiological And Motion Metrics

Injury prevention in sport requires early identification of physiological fatigue and maladaptation before performance decline or injury occurs. However, traditional monitoring approaches often rely on isolated indicators, limiting their ability to capture the multifactorial nature of injury risk. This study aimed to investigate the factors associated with elevated injury risk in athletes.

For this study, data from the publicly available Real Time Physiological Monitoring in Sports dataset (Kaggle) was used. (colabsss, 2025). The dataset contained 23,400 observations from 234 athletes across five sport disciplines. To ensure independence of observations and reduce within-athlete information leakage, a controlled athlete-level sampling approach selected one representative observation per athlete, resulting in a dataset of 234 samples. A LightGBM classification framework was trained using physiological and motion based metrics.

The model obtained ROC-AUC of 0.705, recall of 0.52, accuracy of 0.64, precision of 0.67, and F1 score of 0.59. The biggest predictors of injury risk, according to SHAP analysis, were age, heart rate, motion-based variables and oxygen saturation. Severe departures from normal physiological patterns were linked to increased injury risk.

These findings suggest that combining physiological and motion-based data can help detect early signs of fatigue and increased injury risk. Although challenges such as personalization and defining optimal thresholds remain, wearable-driven analytics shows promise for supporting better training and injury prevention decisions.