Sleep and Activity Patterns in Depression From Wearable Data: Unsupervised Clustering Study

J Med Internet Res. 2026 Jun 10;28:e86900. doi: 10.2196/86900.

ABSTRACT

BACKGROUND: Efforts to advance our understanding of depression have long been constrained by the disorder’s vast symptom heterogeneity and by the reliance on self-report, which offers only a partial view of phenotypic expression. Digital phenotyping provides an opportunity to address these core challenges by generating real-time, objective data on behavior and physiology, offering new perspectives on understanding depression phenotypes. Yet, prior efforts to identify such objectively derived subtypes have relied on predefined diagnostic labels or supervised models, limiting discovery to existing clinical categories.

OBJECTIVE: This study aimed to identify subtypes of depression based on objective sleep and activity data using an unsupervised learning method and to explore how participants transition between these subtypes over time.

METHODS: We analyzed longitudinal Fitbit data from 623 participants with recurrent depression enrolled in the Remote Assessment of Disease and Relapse in Major Depressive Disorder study. To identify our subtypes, we applied Gaussian mixture models and hidden Markov models, incorporating a thorough model selection approach that combined grouped cross-validation and seed selection to ensure robustness.

RESULTS: Three activity subtypes (high, light, and low activity) and 4 sleep subtypes (efficient early sleepers, efficient late sleepers, disrupted sleepers, and variable late sleepers) were consistently identified. These subtypes align with known associations between depression and behavioral patterns. Transition modeling revealed stability within individuals over follow-up, further suggesting the presence of behavioral phenotypes rather than momentary fluctuations.

CONCLUSIONS: The results demonstrate that wearable-derived features can identify reproducible and clinically relevant behavioral subtypes of sleep and activity in individuals with major depressive disorder. These subtypes reflect known behavioral correlates of depression and may offer a data-driven framework for reducing phenotypic heterogeneity, improving research stratification, and supporting personalized patient monitoring. Further work is needed to validate these findings in independent cohorts and evaluate their potential use in reducing noise when using sleep or activity data to predict depression outcomes.

PMID:42268589 | DOI:10.2196/86900