Passive Monitoring for Suicide Risk Prediction in U.S. Military Personnel: Detecting Behavioral Changes with Wearable Data

Abstract Suicide rates in U.S. military personnel exceed national averages, with crises escalating rapidly outside clinical hours. Passive wearable sensing offers a continuous and objective approach to detect proximal suicide risk that may go unnoticed by periodic clinical assessments. This secondary analysis evaluated whether behavioral changes passively captured by wearable devices could predict suicidal events in a high-risk military sample. We analyzed 28-day Fitbit data from 86 U.S. military service members and veterans, including physical activity, heart rate variability, and sleep, alongside self-reported suicidal events collected through ecological momentary assessment. Probabilistic heterogeneous mixture models were used to generate daily behavioral profiles, and Bayesian change-point detection identified instability preceding events within a 72-hour window. Source contribution analysis was then conducted to determine the most informative passive feature combinations. The model achieved AUC scores of 0.90 for suicide attempts, 0.85 for preparatory acts, and 0.82 for non-suicidal self-injury, with an average AUC of 0.79 across suicidal behaviors. Performance was lower for suicidal urge, suggesting that passive sensing may better capture observable behavioral crises than subjectively reported ideation. These findings support wearable-based change detection as a scalable strategy for real-time suicide prevention in military populations.