SNAPSHOT: Modeling Sleep and Mental Health in Social Networks

Link: SNAPSHOT Study

SNAPSHOT is a large-scale, NIH-funded longitudinal study investigating how sleep, stress, mood, and performance emerge and evolve within real-world social networks. The project combines continuous multimodal sensing, machine learning, and physiologically grounded modeling to understand how everyday behavior and social interaction shape health and wellbeing over time.

Multimodal Longitudinal Sensing

Conducted jointly by the MIT Media Lab (Affective Computing & Collective Learning groups) and Harvard Medical School / Brigham & Women’s Hospital, SNAPSHOT followed socially connected cohorts of college students across multiple semesters. Over four years, the study collected continuous wearable, mobile phone, survey, and laboratory data from 300+ participants, yielding thousands of days of high-resolution longitudinal data.

Participants were monitored using wearable sensors (electrodermal activity, skin temperature, motion), mobile phone–based behavioral sensing (communication patterns, mobility, screen usage), and self-reported mood, stress, and health measures, alongside clinical sleep assessments. Unlike traditional sleep studies, SNAPSHOT explicitly captured both individual physiology and social network structure, enabling the study of sleep and mental health as networked, dynamic phenomena rather than isolated behaviors.

Machine Learning for Sleep and Mental Health Forecasting

A core scientific contribution of SNAPSHOT was the development of machine learning models for ambulatory sleep, mood, and stress detection and prediction. Using multimodal wearable and phone data, the team introduced LSTM-based recurrent neural network models that significantly outperformed traditional actigraphy-based approaches for sleep detection in real-world settings. These models achieved 96.5% sleep–wake classification accuracy, with sleep onset and offset detection errors of approximately five minutes, demonstrating robust, person-independent performance at population scale.

Beyond sleep detection, SNAPSHOT showed that daily behaviors and social context can forecast next-day mood, stress, and mental health states, with predictive accuracies in the 78–82% range using personalized and multitask learning approaches. The study also identified behavioral and physiological signatures associated with declining mental health over the course of a semester, including reduced mobility, late-night communication patterns, and altered autonomic activity during sleep.

Complementing the data-driven models, SNAPSHOT contributed to the first multi-scale statistical and physiologically grounded frameworks linking individual sleep and circadian regulation to population-level dynamics in social networks. Together, these results established a foundation for early detection, forecasting, and targeted intervention in sleep and mental health, illustrating how continuous sensing and modeling can support proactive, preventative health strategies.