Multi-sensor fusion outperforms single indicators for fatigue prediction in university soccer players: a machine learning approach.

Background: Collegiate football players face unique challenges balancing academic and athletic demands, yet research on multi-sensor training load monitoring for this population remains limited. Objective: To evaluate multi-sensor wearable devices for training load monitoring and fatigue prediction in collegiate football players. Methods: Forty-eight male collegiate football players were monitored over 12 weeks using GPS devices, heart rate monitors, and subjective questionnaires. External and internal load indicators were collected during 536 training sessions and 24 matches. Fatigue status was defined using countermovement jump, heart rate variability, wellness scores, and RPE. XGBoost, random forest, and logistic regression models were developed and validated. Results: Strong correlations existed between external and internal load indicators (Player Load vs. TRIMP: r = 0.81). The XGBoost model achieved optimal performance (AUC = 0.895), significantly outperforming single-indicator models. Wellness score (18.5%), ACWR (16.2%), and morning HRV (13.8%) were the most important predictive features. Position-specific load patterns were observed, with midfielders covering greatest distances and forwards showing highest sprint outputs. Conclusion: Multi-sensor fusion combined with machine learning (XGBoost, AUC = 0.895) significantly outperforms single-indicator models for fatigue prediction in university soccer players, with wellness score, ACWR, and morning HRV identified as the most important predictive features.