Abstract / Summary
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. A total of 92,832 sensor-triggered events from 84 male athletes across three seasons and three organizations were analyzed, including 11,706 true HAEs and 81,126 false HAEs, reflecting a 6.93:1 class imbalance. Multiple classifiers were trained on a comprehensive set of kinematics-derived features, and XGBoost provided the best overall performance in this dataset (F1-score = 0.765, recall = 0.713, precision = 0.825, AUC = 0.931). Feature ranking suggested that performance improved as biomechanical, time-domain, frequency-domain, and time-frequency features were added, with the strongest balance of discrimination and parsimony observed using the top 20 features. Permutation testing indicated that model performance was greater than expected by chance. To assess operational utility, review time was evaluated across resultant linear acceleration thresholds and low-confidence probability windows. Under the proposed workflow, reviewing only model-flagged events reduced estimated video review time from approximately 90 hrs to under 8 hrs, and thresholds in the 10–15 g range reduced review time by more than half while maintaining a low misclassification burden. These findings suggest that combining machine learning classification with threshold-based review prioritization may reduce manual workload in large-scale youth football impact monitoring while preserving oversight of higher-magnitude events.