Sleep patterns may predict risk of dementia, cancer and stroke

Stanford researchers trained an AI called SleepFM on nearly 600,000 hours of polysomnography from over 60,000 participants and reported it could predict risk for more than 100 health conditions, including dementia, cancer and stroke; the team says the tool is experimental and not yet validated for clinical use.

Researchers at Stanford Medicine developed an AI model named SleepFM and trained it on nearly 600,000 hours of polysomnography collected from over 60,000 people. The team paired those sleep recordings with electronic health records that in some cases covered up to 25 years. The study reported the model could predict risk for more than 100 diseases, and the results were published in Nature Medicine with partial funding from the National Institutes of Health. The authors and an external physician emphasized that the work is research-only and that the tool is not ready for routine clinical use. Key findings: - SleepFM was trained using polysomnography, a comprehensive sleep test that records brain, heart, respiratory and movement signals. - The researchers matched sleep recordings to participants' electronic health records spanning as much as 25 years. - The model identified about 130 diseases it could predict with "reasonable accuracy" and reported strong signals for cancers, pregnancy complications, circulatory conditions, mental disorders, dementia, heart disease and stroke. - The study appeared in Nature Medicine and was partly funded by the National Institutes of Health. - Investigators noted limitations, including reliance on multi‑modal sleep recordings and the need for testing outside research settings; an outside expert said a strong signal does not equal a ready medical tool. Summary: The study suggests patterns in detailed sleep recordings can carry signals linked to long‑term disease risk, and SleepFM produced predictive associations for many conditions. Researchers plan further work to test the approach outside the lab and to explore whether wearable device data can reproduce the findings; until then the model remains experimental and not available for consumer or clinical use.