An artificial intelligence model trained on hundreds of thousands of hours of overnight sleep recordings can flag future risk for more than 100 different diseases, according to research published in Nature Medicine on January 6, 2026.
The Stanford-led team behind the model, called SleepFM, reports that a single night in a sleep lab holds signals linked to conditions ranging from heart attack, stroke and heart failure to dementia, Parkinson's disease, and several cancers. All-cause mortality was also among the outcomes the model could rank with reasonable accuracy.
The signals a sleep specialist never reads
A standard sleep study, called polysomnography, records brain waves, heart rhythm, breathing, blood oxygen, eye movements and leg movements through the night. In routine clinical care, most of that data is used to score sleep stages and to grade the severity of sleep apnea.
That leaves a lot on the table. Only a fraction of the recorded physiology is used in current sleep medicine, the researchers write, and SleepFM was built to look at all of it at once. The model integrates EEG, ECG, EMG and respiratory channels through a single architecture.
Also read: Five everyday items you should be cleaning more often
"SleepFM is essentially learning the language of sleep," said James Zou, associate professor of biomedical data science at Stanford and a co-senior author of the study.
What the numbers actually show
The team curated more than 585,000 hours of sleep recordings from roughly 65,000 people, with the main cohort drawn from about 35,000 Stanford Sleep Medicine Center patients aged 2 to 96, followed for up to 25 years, according to Science Daily.
Out of more than 1,000 disease categories tested, 130 could be predicted with reasonable accuracy from a single night of data, and many of those exceeded a concordance index of 0.8, a level clinicians typically treat as strong discrimination. The strongest predictions included Parkinson's disease and prostate cancer at 0.89, breast cancer at 0.87, dementia at 0.85, and heart attack at 0.81.
A concordance index of 0.8 means that in 80 percent of head-to-head comparisons, the model correctly ranks which of two patients develops the disease sooner. Models with lower scores, often around 0.7, are already used in some areas of clinical medicine.
Also read: How long can humans actually live? A new model puts the ceiling far higher than most expect
Not a diagnosis, and not yet a doctor's order
SleepFM is a research tool, not a clinical test. The model was trained on people referred for sleep studies, which the authors flag as a source of selection bias, and it does not yet explain its own reasoning in a way clinicians can audit, Inside Precision Medicine reports.
"It doesn't explain that to us in English," Zou said of the model's inner workings. His group is developing interpretation techniques to trace which parts of the overnight signal drive a given disease prediction.
The next steps include tightening accuracy, adding data from consumer wearables, and testing how the model behaves outside a sleep lab. For anyone already scheduled for a polysomnography, the practical point is unchanged. The study is done for the reason it was ordered, and anyone worried about sleep problems, snoring, unusual daytime tiredness or a family history of the diseases above should raise it with their own doctor.
Also read: How creatine can affect your kidneys if you take it every day
Also read: How much omega-3 actually eases chronic pain, according to a review of 41 trials