Researchers at Stanford University have developed an artificial intelligence model capable of predicting whether patients are at risk of developing more than 100 health conditions by analyzing their sleep patterns.
The model, named SleepFM, uses data from a patient’s brain activity, heart rate, respiratory signals, leg movements, and eye movements during sleep to assess disease risk. The system is trained to detect subtle physiological signals that may indicate future health problems.
A study published in Nature details how the AI was trained using more than 580,000 hours of sleep data collected from 65,000 patients between 1999 and 2024. The data, obtained from sleep clinics and medical facilities, was divided into five-second increments, which researchers treated as “words” for the model to learn patterns in a manner similar to language processing.
“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 model was trained alongside patients’ health records, allowing it to predict the likelihood of developing specific conditions.
The AI demonstrated strong accuracy, correctly predicting Parkinson’s disease, Alzheimer’s disease, dementia, hypertensive heart disease, heart attacks, prostate cancer, and breast cancer in at least 80 percent of cases. It also successfully predicted a patient’s death 84 percent of the time.
Its performance was slightly lower for conditions such as chronic kidney disease, stroke, and arrhythmia, where the model achieved accuracy of at least 78 percent. Researchers noted that out-of-sync body signals—such as a brain appearing to be asleep while the heart remains active—were strong indicators of potential health issues.
“We record an amazing number of [health] signals when we study sleep,” said Emmanuel Mignot, professor of sleep medicine at Stanford. “It’s very data rich, providing a detailed picture of physiology over eight hours in a subject who’s completely monitored.”
The researchers highlighted that the combination of multiple signals was critical to the model’s success. They plan to expand SleepFM’s database to include data from wearable devices, which could allow the model to monitor and predict health risks in a broader population outside clinical settings.
However, the study’s authors cautioned that the initial dataset focused on patients already suspected of having health problems, meaning the model’s predictive accuracy for the general public is not yet fully understood. Additional research will be needed to evaluate SleepFM’s performance in healthy individuals.
Stanford researchers said the AI could offer a new approach to early disease detection and preventive care by transforming sleep data into actionable insights, potentially allowing clinicians to identify high-risk patients before symptoms emerge.
