The SleepFM AI model has learned to identify the risk of serious diseases from polysomnography — a night-time medical sleep examination. Trained on more than 585,000 hours of sleep from 65,000 people, the model predicts the development of Parkinson’s disease, cognitive disorders, strokes, cancer and even the likelihood of death with high accuracy years before symptoms appear.
Unlike previous approaches, SleepFM analyzes the full range of physiological signals, including sleep stages, breathing and eye movements. Its accuracy for a number of diagnoses reaches 0.93 on the AUROC metric, one of the main indicators of a predictive model’s quality.
The developers plan to integrate SleepFM with electronic medical records and lab test results. This will make it possible to produce even more precise and personalized forecasts, changing the very logic of medicine: not treating when it is already too late, but warning in advance.
Sleep, it turns out, is not only rest, but also a source of valuable information about future health.
