Stanford Medicine researchers have developed SleepFM, the world’s first multimodal foundation model that analyzes physiological signals from one night’s sleep study to predict risks for over 100 health conditions including heart failure, dementia, stroke, cancer, and Parkinson’s years before symptoms appear. Trained on nearly 600,000 hours of sleep data from 65,000 people, it offers a noninvasive window into future disease vulnerability.
Glimpse:
Published in Nature Medicine (January 2026), SleepFM uses polysomnography recordings (brain waves, heart rate, breathing, oxygen levels) to forecast 130+ conditions with high accuracy, especially neurodegenerative and cardiovascular diseases. It outperforms single-signal models, highlighting sleep’s role as a vital health biomarker. This could enable early interventions via routine sleep studies, shifting medicine toward prediction and prevention.
Stanford Medicine has introduced SleepFM, a groundbreaking AI foundation model that transforms a single night’s sleep data into powerful predictions of future disease risk. This multimodal system trained on vast polysomnography datasets decodes subtle disruptions in sleep architecture to reveal hidden vulnerabilities for over 100 conditions, often years in advance.
Key highlights:
- Data Scale: Nearly 600,000 hours from 65,000 individuals, capturing brain activity, heart rhythms, breathing patterns, muscle tone, and oxygen saturation.
- Predictive Power: Identifies risks for 130+ diseases, with strongest signals for dementia, Parkinson’s, heart failure, stroke, and chronic pain.
- Accuracy Edge: Outperforms models using isolated signals (e.g., only brain waves or heart rate).
- Early Warning: Detects patterns linked to future onset, even in seemingly healthy sleepers.
Unlike traditional sleep studies focused on disorders like apnea, SleepFM treats sleep as a comprehensive health mirror—reflecting systemic disruptions from inflammation, autonomic imbalance, or neurodegeneration.
Lead researcher Emmanuel Mignot, director of the Stanford Center for Sleep and Circadian Sciences, noted: “We record an amazing number of signals during sleep. SleepFM reveals how these signals, in combination, can predict health outcomes.”
The model builds on foundation approaches like large language models but for physiological time-series data. It could integrate into clinical sleep labs, home monitors, or wearables democratizing risk assessment without invasive tests.
Limitations: Current validation relies on sleep clinic cohorts (often with suspected disorders); broader population testing is needed. Future work aims at consumer devices for longitudinal tracking.
This innovation underscores sleep’s underappreciated role in longevity, potentially guiding lifestyle changes, screenings, or therapies before irreversible damage.
“One night of sleep may contain hidden clues that predict major diseases years before they strike.”
By
HB Team
