Scientists at Stanford University have developed an artificial intelligence model capable of estimating the risk of over 100 diseases – from dementia to heart failure – based on data from just one night's sleep.
The analysis is based on physiological signals recorded during sleep, writes Nature Medicine.
SleepFM is a “fundamental“ artificial intelligence model, analogous to language models that are trained on vast amounts of text. It is trained on nearly 600,000 hours of sleep data collected from 65,000 people. The algorithm analyzed recordings at five-second intervals, identifying complex patterns in brain, heart and respiratory function.
The data was obtained using polysomnography, the “gold standard“ in sleep research. During this procedure, a person is attached to multiple sensors that record brain activity, heart rate, breathing, eye movements and leg movements.
“When we study sleep, we pick up a surprising number of signals“, said Emmanuel Mignot, a professor of sleep medicine at Stanford University and one of the study's senior authors.
To improve the reliability of the model, the researchers used a contrastive learning method: some signals were intentionally excluded and the artificial intelligence had to reconstruct the missing information based on other data. The sleep recordings were then compared with the patients' medical records, in some cases covering a period of observation of up to 25 years.
As a result, SleepFM was able to predict the risk of 130 different diseases from more than a thousand categories analyzed with acceptable accuracy. The model performed particularly well in predicting cancer, pregnancy complications, circulatory diseases and mental disorders. For several conditions, the accuracy score (C-index) exceeded 0.8, demonstrating a high correspondence between the predictions and the actual results.
According to biomedical data expert James Zhou, co-author of the study, the greatest predictive value comes from contradictory physiological signals - for example, when the brain "appears to be asleep" while the heart behaves as if the person is awake. Such discrepancies were more often associated with adverse long-term outcomes.
“If language models learn to understand human speech, then SleepFM is essentially learning to understand the language of sleep,“ concluded Zou.