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

Area of Science:

• Digital phenotyping in mental health research.
• Application of machine learning for psychiatric disorder subtyping.
• Objective measurement of behavioral patterns in major depressive disorder.

Background:

• Depression diagnosis is challenged by symptom heterogeneity and reliance on subjective self-reports.
• Digital phenotyping offers objective, real-time behavioral and physiological data.
• Previous subtype discovery was limited by predefined clinical categories and supervised models.

Purpose of the Study:

• To identify depression subtypes using objective sleep and activity data via unsupervised learning.
• To analyze temporal transitions between identified behavioral subtypes in major depressive disorder.

Main Methods:

• Analysis of longitudinal Fitbit data from 623 participants with recurrent major depressive disorder.
• Application of Gaussian mixture models and hidden Markov models for subtype identification.
• Robust model selection using grouped cross-validation and seed selection.

Main Results:

• Consistent identification of three activity subtypes (high, light, low) and four sleep subtypes (efficient early/late, disrupted, variable late).
• Identified subtypes correlate with known depression-behavior associations.
• Transition modeling indicated stable behavioral phenotypes over time, not just momentary fluctuations.

Conclusions:

• Wearable sensor data can identify reproducible, clinically relevant sleep and activity subtypes in major depressive disorder.
• These objective subtypes may reduce phenotypic heterogeneity and aid research stratification and personalized monitoring.
• Further validation in independent cohorts is needed to assess predictive utility for depression outcomes.