AI spots hidden disease warnings during sleep from one night of testing

Researchers report an AI called SleepFM that was trained on about 585,000 hours of polysomnography and can use one night of sleep data to predict risk for many future conditions, including some cancers, dementia and heart disease.

Scientists have developed an artificial intelligence system, SleepFM, that analyses detailed signals recorded during one night of polysomnography to identify patterns linked to later health outcomes. The researchers trained the model on roughly 585,000 hours of sleep recordings from patients evaluated at sleep clinics and linked those records with long-term medical outcomes from the same clinic. In routine practice much of the physiological data from sleep studies is not deeply analysed, and the team reports the AI found signals across brain, heart and breathing measures that relate to future disease risk. The work is described as an application of a foundation model approach to physiological data rather than text. Key findings: - SleepFM was trained on about 585,000 hours of polysomnography from around 65,000 individuals and used five-second segments of recording to learn patterns. - The model integrates multiple streams of information including brain activity, heart rhythms, muscle activity, pulse measures and breathing airflow. - Researchers linked sleep recordings to decades of medical records from the same sleep clinic to evaluate future disease outcomes. - SleepFM identified about 130 conditions that could be predicted from sleep data with reasonable accuracy, and showed especially high prediction scores for Parkinson's disease, dementia, hypertensive heart disease, heart attack, prostate and breast cancer, and all-cause death. Summary: The study indicates that a single night of detailed sleep testing can contain signals associated with a wide range of later health outcomes, suggesting sleep recordings contain information beyond standard sleep diagnoses. Undetermined at this time.