Reducing Video Verification Burden: Machine Learning Classification of Head Acceleration Events in Youth Football.

This study aimed to develop and evaluate a machine learning pipeline to classify head acceleration events (HAEs) in youth American football and reduce the manual burden of video verification. An eXtreme Gradient Boosting (XGBoost) classifier was trained on three seasons of instrumented-mouthguard data from three organizations, including athletes aged 11-14 years, using a comprehensive set of kinematics-derived features. Model performance was evaluated under a highly imbalanced outcome distribution using precision, recall, and F1 score, and confirmed via permutation testing to ensure results exceeded chance performance. The proposed feature set and model achieved comparable or superior performance relative to previously published support vector machine approaches, while operating under more heterogeneous, real-world field conditions. To quantify operational burden, we compared the manual effort required to review all sensor-recorded events versus only those flagged as potential HAEs across resultant linear acceleration review thresholds. Total video review time was reduced from approximately 90 hours (all events reviewed) to under 8 hours (classifier-flagged events only), a 91% reduction in workload, while maintaining a misclassification rate of 2.5-3% of all events. When applied in conjunction to the model, thresholds in the 10-15 g range cut review time by more than half relative to reviewing all events and retained a low error in the model (4%). These findings show that an appropriately tuned classification and review-threshold pipeline can make large-scale video-verified monitoring of youth football HAEs substantially more feasible for research and surveillance applications.