Bipolar disorder relapse detection and prediction using smartwatches. A pilot study for general and personalized machine learning models using anomaly detection methods. (Preprint)

BACKGROUND Bipolar disorder (BD) is a chronic and recurrent mental illness characterized by alternating mood episodes, with relapse affecting up to 70% of patients within two years. Early identification of prodromal symptoms preceding mood episodes is critical for timely intervention, yet remains challenging, especially in patients with impaired alexithymia, self-awareness or early trauma. Wearable technologies and digital phenotyping offer novel avenues to monitor physiological and behavioral changes in real time, potentially enabling the prediction of relapses. OBJECTIVE This pilot study aimed to assess the feasibility and performance of general and personalized machine learning models using smartwatch-derived physiological data for detecting and predicting mood relapses and prodromal phases in individuals with BD. METHODS Ten patients with BD were monitored over six months using Fitbit smartwatches continuously recording heart rate (and heart rate variability - HRstd), sleep (light sleep, REM sleep, deep sleep), and physical activity (represented by the number of steps). Monthly clinical assessments and psychometric evaluations (e.g., MADRS, YMRS) were used to label periods as healthy, prodromal, or relapse. Anomaly detection models were trained on healthy data only and evaluated on their ability to detect (relapse) and predict (prodromal) episodes using several time windows (1h, 4h, 24h) and features. Both general and personalized modeling strategies were explored using unsupervised methods such as Isolation Forest, Local Outlier Factor, and One-Class SVM. RESULTS Across 1,395 days of usable data (79% of total), 15 depressive and 8 hypomanic relapses were observed. Significant physiological changes were detected during prodromal and relapse phases. In general models, the best-performing feature combinations for detecting depression were HRstd and steps (Precision Recall Area Under the Curve - PR AUC = 0.67), while light sleep and mean HR were best for hypomania (PR AUC = 0.33). Prodromal depression and hypomania could be predicted with lower but above-chance performance (e.g., PR AUC = 0.34 for prodromal depression using light sleep). Personalized models showed improved performance for relapse detection (e.g., PR AUC = 0.99 for one patient), but not for prodromal prediction, likely due to limited data (median of 2–4 prodromal days per patient). CONCLUSIONS Smartwatch-derived physiological signals combined with unsupervised anomaly detection methods can effectively detect mood relapses and, to a lesser extent, predict prodromal states in BD. General models are more suitable for predicting prodromes, whereas relapse detection benefits from personalized approaches. These findings support the integration of wearable monitoring and AI into clinical practice to anticipate relapses and enable timely intervention. CLINICALTRIAL ClinicalTrials.gov Identifier: NCT04703972