Overview
- SleepFM analyzed multimodal sleep signals split into five‑second segments from about 65,000 participants to learn patterns linked to future illness.
- The model integrates EEG, ECG, EMG, pulse and airflow using a leave‑one‑out contrastive learning approach to align information across modalities.
- In tests on standard tasks such as sleep staging and sleep‑apnea severity, SleepFM matched or exceeded current state‑of‑the‑art performance.
- Linked to long‑term outcomes, the system identified 130 disease categories with predictive signal, with C‑indices up to 0.89 for conditions including Parkinson’s, prostate cancer and dementia.
- The largest cohort included roughly 35,000 Stanford Sleep Medicine Center patients with recordings from 1999–2024 and up to 25 years of follow‑up, and researchers are now focusing on interpretability and potential integration of wearable data.