Identifying Robust Predictors of Operational Fatigue in Emergency Medical Services Providers: A Machine Learning Approach

Fatigue in emergency medical services (EMS) providers is associated with impaired performance and safety risk, yet data-driven fatigue modeling in EMS remains limited. This study investigated the feasibility of predicting operational fatigue using multimodal wearable physiological, behavioral, and contextual data obtained from 11 EMS providers during a 5-day real-world field data collection. Features were extracted from 6-hour retrospective temporal windows preceding fatigue assessments and used to train Logistic Regression, Random Forest, XGBoost, and CatBoost models. Logistic Regression achieved the best predictive performance (AUROC = 0.87), demonstrating reliable discrimination between high- and low-fatigue states. Feature importance analyses identified robust fatigue predictors across models, including time of day, cumulative work exposure, dispatched call volume, and slope-based pulse rate features. Findings suggest that operational fatigue in EMS reflects interacting circadian, workload-related, behavioral, and physiological processes and supports the feasibility of scalable wearable-based fatigue monitoring in safety-critical operational environments.