Week-ahead prediction of depressive episodes using wearable-derived circadian biomarkers: A multicenter deep learning...

Early detection of depressive episodes is important because shorter duration of untreated illness is associated with better outcomes, yet routine care remains largely reactive and dependent on retrospective self-report. Wearables enable passive, continuous monitoring, and circadian disruptions in sleep-wake, activity, and diurnal heart rate (HR) patterns may precede mood episodes. However, prior wearable-based models often relied on short prediction horizons, simplistic circadian proxies, and limited clinical interpretability. We evaluated a deep learning approach that integrates two complementary circadian markers derived from a single wearable: an HR-derived cosine acrophase and a sleep-timing-based proxy measure of circadian phase related to dim-light melatonin onset (DLMO), since melatonin was not directly measured. Performance was summarized across risk-based operating thresholds. In a multicenter real-world mood disorder cohort (N = 144; 90,281 person-days), we trained a time-series model using 7-day wearable histories to predict the presence of a clinician-verified depressive episode one week ahead, with participant-level splits (6: 2:2) and discrimination consistency summarized across 10 repeated holdout runs. The combined circadian model achieved a mean area under the receiver operating characteristic curve (AUROC) of 0.772 (95% confidence interval [CI], 0.750-0.794). On the held-out test set, the top 20% of risk-ranked patient-days captured 63.5% of depressive-episode days; at a base prevalence of 5.3%, positive predictive value (PPV) was 16.7% (3.2-fold prognostic lift), with a review burden of 20 per 100 patient-days. These findings support the feasibility of wearable-based week-ahead depressive-risk stratification in mood disorders, although external validation, probabilistic calibration, transportability assessment, and prospective episode-onset evaluation are required before clinical implementation.