Overview
- Pretrained on roughly 585,000 hours of polysomnography from about 65,000 people, the foundation model learned cross-signal patterns from five‑second segments and later matched or exceeded state-of-the-art performance on sleep staging and apnea severity.
- The system reached strong concordance for many outcomes, including Parkinson’s disease (C-index 0.89), dementia (0.85), hypertensive heart disease (0.84), myocardial infarction (0.81), prostate cancer (0.89), breast cancer (0.87) and all-cause mortality (0.84).
- A leave-one-out contrastive learning strategy enabled the model to harmonize heterogeneous physiological channels and reconstruct missing modalities.
- Researchers paired sleep studies with up to 25 years of electronic health records to assess future disease risk and reported external validation on the Sleep Heart Health Study.
- Authors caution about selection bias from referred sleep-clinic populations and limited interpretability, and they plan to add wearable data and pursue broader validation before clinical use.