Sleep Estimation from Low Frequency Smartphone Sensors via Bayesian Hidden Markov Model

Sleep disturbances are recognized as transdiagnostic markers and potential mechanistic contributors to psychiatric illness, yet objective sleep monitoring remains rare in large-scale psychiatric research due to infrastructure and methodological barriers. While smartphones enable scalable, real-world behavioral sensing, most current approaches are limited by single-sensor thresholds, proprietary algorithms, or lack of validation in diverse populations. Here, we introduce a probabilistic Bayesian hidden Markov model that integrates accelerometer and screen state data to infer nightly sleep states and extract a set of behavioral sleep metrics. Model performance was evaluated using both empirically derived simulation and real-world self-reported sleep logs. Analyzing 5,888 nights from 516 participants, we identified substantial heterogeneity in individual sleep-symptom coupling, with unsupervised clustering of sensor-derived sleep and symptom dynamics revealing five distinct phenotypes that were consistent with independent clinical assessments. Our approach provides a robust framework for large-scale, non-invasive sleep monitoring, with direct applications in digital psychiatry and individualized intervention.