A joint modeling framework for time-to-fatigue prediction with a single wearable sensor biomarker

Physical fatigue is a prevalent issue in the U.S. workforce and has resulted in reduced productivity and a higher incidence of workplace accidents. Being able to estimate time to fatigue facilitates proactive interventions before fatigue onset. Existing research has primarily focused on the fatigue state, intensity, and wearable sensor characteristics while there remains a gap in predicting time to fatigue and identifying indicative biomarkers that can dynamically signal impending fatigue risks. In this study, we utilize simulated manufacturing work tasks and employ a joint modeling framework that links a longitudinal biomarker to predict individual time-to-fatigue outcomes. Across two case studies with differing task demands and durations, the joint model with a single sensor trajectory consistently outperformed baseline methods, including Kaplan-Meier estimates and Cox proportional hazards models, demonstrating its effectiveness for dynamic fatigue prediction in simulated manufacturing tasks while accounting for variations across tasks and individual sensor responses. Although predictive features may vary by task, our results suggest that, within the contexts examined, jerk-based features–identified in previous studies, including our own, as indicators of fatigue states–may also serve as predictors of time to fatigue. These findings advance the understanding of fatigue progression and offer foundational insights for supporting timely interventions in future applied settings. This work highlights the potential of online wearable monitoring and interpretable statistical methods to inform proactive occupational safety strategies and worker well-being initiatives.