AI can flag risks for more than 100 health conditions using a single night’s sleep, study shows

A new artificial intelligence (AI) model can tell whether a person is at risk of developing over 100 health conditions, based on how well they sleep.

SleepFM, a large language model (LLM) developed by researchers at California’s Stanford University, reads a user’s brain activity, heart rate, respiratory signals, leg movements, and eye movements while they’re sleeping to evaluate the risk of disease.

In a new study published in Nature, researchers trained the AI model using over 580,000 hours of sleep data collected from 65,000 patients from 1999 and 2024.

The data came from sleep clinics, medical facilities that evaluate sleep patterns overnight, and were split into five-second increments, which acted like words for LLMs to train on.

“SleepFM is essentially learning the language of sleep,” said James Zou, associate professor of biomedical data science at Stanford and co-author of the study.

The researchers complemented this data with the individual health files of the sleep clinic patients, in order to train SleepFM to predict future diseases.

The AI model was right at least 80 percent of the time when predicting whether a patient would get Parkinson’s disease, Alzheimer’s disease, dementia, hypertensive heart disease, heart attack, prostate cancer, and breast cancer. It also correctly predicted a patient’s death 84 percent of the time.

The model was less accurate in predicting patients with chronic kidney disease, stroke, and arrhythmia, considered an irregular heartbeat, which it detected in at least 78 percent of cases.

“We record an amazing number of [health] signals when we study sleep,” said Emmanuel Mignot, professor in sleep medicine at Stanford. “It’s a kind of general physiology that we study for eight hours in a subject who’s completely captive. It’s very data rich.”

The authors of the study said the combination of all the data helped the model achieve the most accurate predictions. For example, body signals that were out of sync, such as a brain that looks asleep but a heart that looks awake, spelled trouble.

Stanford said that they will add data from wearables to SleepFM’s database next, to further improve the models’ predictions.

The researchers also noted that their study only used people who already suspected existing health problems due to their participation in the sleep clinic trials, meaning that their sample isn’t representative of the AI’s ability to detect disease in the general public.

​