Overview
- Published in Nature Medicine, SleepFM was built from roughly 65,000 participants’ polysomnography, with a core Stanford cohort of about 35,000 paired to up to 25 years of electronic health records.
- The model reached strong predictive performance for many outcomes, including Parkinson’s disease (C-index 0.89), dementia (0.85), myocardial infarction (0.81), prostate cancer (0.89), breast cancer (0.87) and all-cause mortality (0.84).
- Beyond forecasting disease, SleepFM matched or exceeded current systems on standard sleep tasks such as staging and assessing sleep apnea severity.
- A leave-one-out contrastive learning approach enables the model to harmonize heterogeneous, multimodal signals and infer missing channels from the remaining data.
- Researchers note selection bias and limited interpretability as key limits, and they are pursuing explanation methods, adding wearable-device data and external validation, including results reported on the Sleep Heart Health Study dataset.