Stanford Medicine researchers and collaborators report that an artificial intelligence model called SleepFM can analyze a single overnight polysomnography study and estimate a person’s future risk for more than 100 medical conditions, including dementia, heart disease and some cancers. The team says the system learns patterns across multiple physiological signals recorded during sleep and could reveal early warning signs years before clinical diagnosis.
A single night in a sleep laboratory may contain more information about future health than clinicians typically extract today.
Researchers at Stanford Medicine and collaborating institutions have developed an artificial intelligence model, SleepFM, that analyzes data from polysomnography—the gold-standard overnight sleep study that records signals such as brain activity, heart activity, breathing, eye movements and muscle or leg movements.
The team trained SleepFM on about 585,000 hours of polysomnography recordings from roughly 65,000 people, and then linked sleep studies from a large Stanford Sleep Medicine Center cohort to long-term electronic health records. In that cohort—about 35,000 patients ages 2 to 96 with sleep studies performed between 1999 and 2024—some individuals had as much as 25 years of follow-up.
“We record an amazing number of signals when we study sleep,” said Emmanuel Mignot, the Craig Reynolds Professor in Sleep Medicine at Stanford Medicine and a senior author of the study. He described the overnight exam as “very data rich.”
To build the model, the researchers used a “foundation model” approach—more commonly associated with large language models—treating physiological sleep recordings as sequences. They split each recording into five-second segments and trained the system to learn how different channels relate to one another. “SleepFM is essentially learning the language of sleep,” said James Zou, an associate professor of biomedical data science at Stanford and a senior author. The team also reported using a training method called leave-one-out contrastive learning, in which one signal type is removed and the model learns to reconstruct it from the remaining channels.
In standard sleep-analysis tests, SleepFM performed as well as or better than existing state-of-the-art models at classifying sleep stages and assessing sleep apnea severity.
The researchers then evaluated whether one night of sleep data could help forecast longer-term medical outcomes. After reviewing more than 1,000 disease categories in linked health records, the study reported that 130 conditions could be predicted with what the authors describe as reasonable accuracy using sleep data alone. Performance was strongest for groups of outcomes that included cancers, pregnancy complications, circulatory diseases and mental health disorders.
The study used the concordance index (C-index), a measure of how well a model ranks individuals by risk. Zou said that a C-index of 0.8 means that, across pairs of people, the model correctly ranks who will experience an event earlier about 80% of the time.
Among the reported examples, the model achieved C-index values of 0.89 for Parkinson’s disease, 0.85 for dementia, 0.81 for myocardial infarction (heart attack), 0.89 for prostate cancer and 0.87 for breast cancer. The researchers also reported strong performance for outcomes including all-cause mortality.
The team said different physiological channels carried different predictive weight depending on the outcome—for example, heart-related signals were more influential for cardiovascular predictions and brain-related signals mattered more for mental health—while combining channels produced the best results. Mignot said that mismatches between systems—such as a brain that appears asleep while the heart looks more alert—were among patterns associated with higher risk.
The paper, titled “A multimodal sleep foundation model for disease prediction,” was published online in Nature Medicine on January 6, 2026. PhD students Rahul Thapa (Stanford) and Magnus Ruud Kjaer (Technical University of Denmark) are listed as co-lead authors.
Researchers cautioned that the work is an early step toward clinical use. The team said ongoing work includes improving interpretability—understanding what the model is “seeing” in the signals—and exploring whether similar approaches could incorporate data from wearable devices to broaden access beyond specialized sleep labs.
