Machine Learning for Mental Health Episode Prediction via Wearable Devices

Nearly 970 million people worldwide live with mental health (MH) disorders, many involving recurrent episodes that may be mitigated through timely prediction. This review evaluates machine learning (ML) approaches using wearable physiological data to forecast MH episodes and examines whether device grade influences predictive accuracy. Empirical English-language studies were identified through EBSCO and screened in Rayyan.ai by two independent reviewers, with a third reviewer resolving conflicts. Of 332 records identified, 32 studies met the inclusion criteria. Most studies focused on depression or stress, and 30 applied AI/ML methods. Findings show that predictive performance was driven more by ML approach than by device grade. Reported ranges reached 76%–87% for depression, 85%–98% for stress, 75.6%–92.8% for panic-attack risk, and 97%–99% for anxiety. Key predictors included heart-rate variability, sleep, activity, circadian stability, and electrodermal activity. Multimodal and personalized temporal models performed best.