Overview
- The multimodal model was trained on roughly 585,000–600,000 hours of polysomnography from about 65,000 people, combining brain, cardiac, respiratory and muscle signals.
- Validation linked overnight sleep studies to long‑term medical records, with follow‑up extending to as much as 25 years in some cohorts.
- Across more than a thousand diagnostic categories, the system showed meaningful predictive signal for around 100–130 conditions, with stronger concordance for Parkinson’s, dementia, hypertensive heart disease, myocardial infarction, and some cancers such as prostate and breast.
- SleepFM learns latent representations of night‑time physiology that rank relative risk between individuals, yet it does not provide human‑readable explanations of its predictions.
- The authors emphasize it is not a diagnostic tool, calling for interpretability work, validation in more diverse populations, and exploration of wearable data, with the study involving institutions in the United States and Europe.