Bipolar disorder relapse detection and prediction using smartwatches. A pilot study for machine learning models using...

Bipolar disorder (BD) is a chronic mental illness with recurrent mood episodes, with up to 70% of patients relapsing within two years. Early detection of prodromal symptoms is critical for timely intervention but remains challenging. Wearable devices and digital phenotyping allow monitoring of physiological and behavioral changes in real time. This pilot study evaluated the feasibility of using machine learning models on smartwatch-derived data to detect and predict mood relapses and prodromal phases in BD.

Ten BD patients were monitored over six months using smartwatches recording heart rate, heart rate variability (HRstd), sleep stages, and physical activity. Monthly clinical and psychometric assessments labeled periods as healthy, prodromal, or relapse. Unsupervised anomaly detection models were trained on healthy data to detect relapses and predict prodromes across multiple time windows.

Across 1395 days of usable data, 15 depressive and 8 hypomanic relapses occurred. Physiological changes were detected during prodromal and relapse phases. HRstd and steps best detected depressive relapses (Precision-Recall Area Under the Curve [PR AUC] = 0.67), while light sleep and mean HR were optimal for hypomanic relapses (PR AUC = 0.33). Prodromal phases were predicted with lower but above-chance performance (PR AUC = 0.34 for prodromal depression).

Smartwatch-derived physiological signals combined with unsupervised anomaly detection can detect mood relapses and, to a lesser extent, prodromal states in BD. These results support the integration of wearable monitoring and AI into clinical practice to anticipate relapses and enable timely intervention.