Particle.news
Download on the App Store

Stanford’s SleepFM Uses One Night of Polysomnography to Forecast Risk for 130 Conditions

The Nature Medicine study draws on decades of linked sleep recordings with health records to forecast long-term outcomes.

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.