Highly accurate prediction of mood episodes in mood disorder patients using sleep and circadian rhythm features from wearables

초록

Advances in wearable devices enable the collection of extensive data including sleep, heart rate, step count, and light exposure. These datasets have been utilized to develop mood episode prediction models for mood disorder patients. However, existing models require the collection of various types of data, limiting their applicability in the real world. Here, we develop a model that can accurately predict mood episodes using only sleep-wake data, which can be collected passively and easily through smartphones or wearable devices. Specifically, by applying mathematical modeling to longitudinal sleep-wake data of mood disorder patients (n = 168; 57 with major depressive disorder, 42 with bipolar 1 disorder, and 69 with bipolar 2 disorder, 587 days of average clinical follow-up, 267 days of average wearable data), we obtained 36 comprehensive and accurate features of sleep and estimated circadian rhythm. By applying machine learning algorithms, we accurately predicted the following day's depressive, manic, and hypomanic episodes (AUC 0.925, 0.984, and 0.985 respectively). Our findings from a prospective observational cohort study indicate that sleep-wake data alone can effectively predict mood episodes, and provide new opportunities for improved monitoring and treatment of mood disorders.