SleepFM is a foundation AI model built in a similar way to ChatGPT, but trained on nearly 600,000 hours of sleep recordings from 65,000 participants. While language models work with text, SleepFM analyzes short segments of sleep data obtained in clinical settings. The data were collected using polysomnography—a technique that tracks brain activity, heart function, breathing, and body movements during sleep via sensors. The researchers tested SleepFM using their own method called “contrastive learning with leave-one-out,” which forces the system to reconstruct missing data based on other biological signals. To strengthen the results, they combined polysomnography records with tens of thousands of long-term medical reports covering up to 25 years of patient follow-up.

SleepFM analyzed more than a thousand disease categories and was able to accurately predict 130 of them based solely on sleep data. The model performed best in forecasting cancers, pregnancy complications, circulatory disorders, and mental health conditions, achieving a C-index above 0.8.

“A C-index of 0.8 means that in 80% of cases, the model’s prediction matches what actually happened,” explains James Zou, a biomedical data science expert at Stanford and co–senior author of the study.

The model also demonstrated strong performance on the AUROC metric, which measures its ability to predict the likelihood of developing specific conditions over a six-year period. It outperformed existing predictive systems, particularly in forecasting Parkinson’s disease, heart attacks, strokes, chronic kidney disease, prostate cancer, breast cancer, and overall mortality. This further confirms the link between sleep disturbances and negative health outcomes, which can serve as early warning signals for a wide range of diseases.

The most accurate predictions from SleepFM came from analyzing the relationships between different physiological functions. Especially telling were cases of desynchronization—for example, when the brain shows signs of sleep while the heart remains highly active—patterns that may indicate underlying health risks.

The researchers note several limitations, including changes in clinical practices and patient populations over recent decades, as well as the fact that polysomnography data were primarily collected from individuals referred for sleep studies.