Sleep is a vital biological process that involves complex interactions among the brain, heart, and respiratory systems. While polysomnography (PSG) is the gold standard for clinical evaluation, its diverse signals remain underutilized for broader health screening. To address this gap, Emmanuel Mignot and James Zou of Stanford University aimed to develop a multimodal foundation model, SleepFM, capable of learning the "language of sleep" to systematically predict a wide range of future health outcomes.
For this purpose, the researchers trained SleepFM on over 585,000 hours of PSG data from 65,000 diverse participants using leave-one-out contrastive learning (LOO-CL) to align electroencephalogram (EEG), electrocardiography (ECG), electromyography (EMG), and respiratory signals. The model’s performance was validated against electronic health records (EHR) with up to 25 years of follow-up, utilizing Harrell’s concordance index (C-index) to rank risk across 1,041 disease categories.
Key Results of the Model Include:
- Broad-Spectrum Disease Forecasting: The SleepFM model successfully predicts the future onset of 130 distinct health conditions with a C-index of at least 0.75, offering clinicians a systemic health screening tool from a single night of sleep.
- High-Precision Neurological Risk Stratification: The study demonstrates exceptional accuracy in identifying long-term risk for neurodegenerative disorders, specifically achieving a C-index of 0.89 for Parkinson’s disease and 0.85 for dementia.
- Robust Cardiovascular Event Prediction: The model provides strong prognostic value for major circulatory conditions, reaching a C-index of 0.81 for myocardial infarction and 0.80 for heart failure.
- Enhanced Mortality and Oncology Forecasting: Significant predictive capability is shown for all-cause mortality (0.84 C-index) and specific malignancies, such as prostate and breast cancer, which both achieved an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.90.
- Superiority of Multimodal Data Integration: The highest diagnostic accuracy is reached by combining all PSG modalities—including EEG, ECG, and respiratory signals—confirming that the synchronization of these diverse physiological systems is a vital health metric.
The study concludes that SleepFM transforms a single night of PSG into a comprehensive systemic screening tool that accurately forecasts the future onset of 130 different health conditions, including Parkinson’s disease (0.89) and dementia (0.85), with a C-index of at least 0.75.
The author suggests that SleepFM could serve as a non-invasive complement to existing clinical risk assessment tools, helping healthcare providers with early disease detection and longitudinal health monitoring through standard polysomnography evaluations.
While the clinical cohort’s selection bias may limit broader representativeness, the authors suggest that future research should focus on enhancing model interpretability and validating its predictive power across a wider range of diseases and populations.
Reference
Thapa, R., Kjaer, M. R., He, B., et al. (2026). A multimodal sleep foundation model for disease prediction. Nature Medicine.
Disclaimer: This website is primarily for healthcare professionals. The content here does not replace medical advice and should not be used as medical, diagnostic, endorsement, treatment, or prescription advice. Medical science evolves rapidly, and we strive to keep our information current. If you find any discrepancies, please contact us at corrections@medicaldialogues.in. Read our Correction Policy here. Nothing here should be used as a substitute for medical advice, diagnosis, or treatment. We do not endorse any healthcare advice that contradicts a physician's guidance. Use of this site is subject to our Terms of Use, Privacy Policy, and Advertisement Policy. For more details, read our Full Disclaimer here.
NOTE: Join us in combating medical misinformation. If you encounter a questionable health, medical, or medical education claim, email us at factcheck@medicaldialogues.in for evaluation.