Sleep data: Researchers have built an artificial intelligence model that uses data from an overnight sleep study to estimate a person’s future risk of developing more than 100 different health conditions, suggesting that routine sleep recordings may contain early signals of disease long before symptoms appear.
The system, called SleepFM, was trained on nearly 600,000 hours of sleep data collected from about 65,000 people. In early testing, the model performed standard sleep-analysis tasks, including identifying sleep stages and assessing the severity of sleep apnoea. It was then applied to a broader question: whether patterns captured during sleep could forecast later illness.
Also Read | Sleep apnea types explained: Obstructive, central, and complex
To do that, the team analysed sleep recordings alongside clinical health records from a sleep clinic. They reviewed more than 1,000 diagnostic categories in those records and reported that 130 conditions could be predicted with what they described as reasonable accuracy using sleep data alone.
SleepFM is built to handle the full “multimodal” stream captured during polysomnography, a commonly used overnight test that records multiple body signals at once. These can include brain-wave activity, heart rhythms, muscle activity, pulse, breathing airflow, and related measurements. By learning how these signals interact over the course of a night, the model attempts to capture broader physiological patterns rather than focusing on one channel at a time.
A key part of the approach was a training method the researchers called “leave-one-out” contrastive learning. In this setup, one stream of sleep data is hidden, and the model is challenged to reconstruct it based on the remaining signals, a technique designed to help the system learn relationships across the body’s different systems during sleep.
The model’s strongest results were reported in several broad categories, including cancers, pregnancy complications, circulatory diseases and mental disorders. The researchers said performance for these groups reached a concordance score above 0.8, a commonly used measure that reflects how well a model can rank which of two people is more likely to experience an outcome sooner.
Across individual conditions, the team reported that a single night of sleep data was enough to predict 130 outcomes with a concordance score of at least 0.75. Examples cited included all-cause mortality, dementia, myocardial infarction, heart failure, chronic kidney disease, stroke and atrial fibrillation. The researchers also reported strong performance for Parkinson’s disease risk, where sleep disturbances can be early indicators, as well as developmental delays and disorders.
The findings point to sleep as a dense source of health information, and raise the possibility that sleep studies could one day offer more than the diagnosis of sleep disorders alone. Researchers cautioned, however, that prediction does not equal certainty, and that further work is needed to understand what biological signals the model is using and how well it performs across different settings and populations.