AI model turns sleep data into early disease predictions
Researchers show how sleep data from a single night can reveal long-term disease risk.
Stanford Medicine researchers have developed an AI model that can analyse a single night of sleep to predict long-term disease risk. Known as SleepFM, the system uses physiological signals recorded during overnight sleep studies to identify early indicators of future health conditions.
The model was trained on nearly 600,000 hours of polysomnography data from 65,000 participants. Polysomnography captures brain activity, heart rhythms, breathing patterns, eye movements, and muscle signals, creating one of the most data-rich assessments used in medicine.
SleepFM was designed as a foundation model that learns how multiple biological signals interact during sleep. By reconstructing missing data streams, the system identifies patterns across different physiological systems rather than analysing signals in isolation.
After training, the model matched or outperformed existing tools in standard sleep data assessments, including sleep stage classification and sleep apnoea severity. Researchers then linked sleep data with long-term health records to evaluate its ability to predict future disease onset.
The model demonstrated strong predictive performance across 130 conditions, encompassing various diseases, including cancers, cardiovascular disease, and neurological disorders. Researchers say the findings position sleep data as an early warning signal, while further work will focus on interpretation and real-world clinical use.
Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!