CNSPulse

SLEEP ESTIMATION FROM LOW-FREQUENCY SMARTPHONE SENSORS VIA BAYESIAN HIDDEN MARKOV MODEL

Jin Soo Byun — John Torous1 1Beth Israel Deaconess Medical Center and Harvard Medical School Jin Soo Byun, Beth Israel Deaconess Medical Center and Harvard Medical School

2026 ASCP Annual Meeting

Objective

Sleep disturbances are recognized as transdiagnostic markers of psychiatric illness, yet objective monitoring is rare in large-scale research due to the infrastructure burdens of PSG and the cost of dedicated wearables. While smartphones offer a scalable solution for passive sensing, current approaches often lack validation in diverse psychiatric populations. This study introduces and validates a probabilistic Bayesian Hidden Markov Model (HMM) that integrates low-frequency accelerometer, GPS, and screen data to infer nightly sleep states and extract behavioral metrics without requiring additional hardware.

Methods

We designed a modular pipeline that processes passive smartphone data (accelerometer, GPS context, screen state) through a Bayesian HMM. The model utilizes a ‘sticky’ transition prior to enforce state persistence and a participant-specific Empirical Bayes initialization to adapt to individual chronotypes without imposing arbitrary timing constraints. We evaluated performance through a multi-tiered validation strategy: 1) We generated synthetic datasets based on empirical parameters to stress-test the model against varying signal-to-noise ratios and fragmentation. 2) We benchmarked estimates against concurrent research-grade wrist actigraphy (processed via the GGIR algorithm) in the AMPSCZ dataset (N = 169). 3) We compared model outputs with daily Ecological Momentary Assessment (EMA) self-reports across a diverse sample (N = 516) including individuals at Clinical High Risk for psychosis, depression, and healthy controls. Finally, we applied unsupervised clustering to 15,530 nights from 868 participants to identify latent behavioral phenotypes.

Results

Simulation analysis demonstrated high robustness, with mean accuracy reaching 0.98 in linear drift scenarios. In real-world validation, the smartphone model showed strong agreement with wrist actigraphy, with Bland-Altman analysis indicating minimal systematic bias. Comparison with EMA self-reports revealed significant population-level alignment for bedtime (r = 0.68), waketime (r = 0.65), and sleep duration (r = 0.48). Importantly, filtering out a small subset of high-noise participants (RMSE over 1.2 hours) improved correlations to r = 0.89 for bedtime and r = 0.90 for waketime, demonstrating high precision for the majority of users. Discrepancies between model and EMA estimates were not systematically related to psychiatric symptom severity, supporting validity across clinical states. Unsupervised clustering revealed five distinct phenotypes, such as ‘High symptom burden with short sleep’ and ‘Moderate symptoms with fragmented sleep,’ highlighting that similar clinical profiles map to vastly different underlying sleep behaviors.

Conclusions

This study presents a robust, validated framework for large-scale, non-invasive sleep monitoring using standard smartphones. By effectively capturing sleep-wake dynamics and revealing heterogeneity in sleep-symptom coupling, this approach offers a scalable tool for digital phenotyping and anomaly detection. This presentation is critical for researchers seeking to overcome the scalability limits of actigraphy; attendees will gain a validated methodology for implementing zero-cost, objective sleep monitoring in large cohorts, alongside new insights into how passive behavioral phenotypes can map distinct clinical trajectories.

Learning Objective 1: Demonstrate how low-frequency passive smartphone data (accelerometer, GPS, screen state) can be utilized via a Bayesian Hidden Markov Model to accurately estimate sleep metrics without the need for dedicated wearable devices.

Learning Objective 2: Evaluate the heterogeneity of sleep-symptom coupling in psychiatric populations and identify how distinct behavioral phenotypes can be derived from unsupervised clustering of objective sleep and symptom data.

References

  1. Byun AJS, Li Y, Cong S, et al. Sleep Estimation from Low Frequency Smartphone Sensors via Bayesian Hidden Markov Model. Research Square. 2025;rs.3.rs-7217304.
  2. Langholm C, Byun AJS, Mullington J, Torous J. Monitoring sleep using smartphone data in a population of college students. Npj Ment Health Res. 2023;2:3.