New AI model predicts disease risk during sleep monitoring sessions: Study - Video

Your sleep tonight might reveal diseases lurking years in your future. A groundbreaking study from Stanford Medicine, published in Nature Medicine, shows that artificial intelligence can analyze a single night of sleep data to predict your risk of developing over 100 health conditions—from cancer and heart disease to dementia and Parkinson's. The model, called SleepFM, was trained on nearly 600,000 hours of sleep recordings from 65,000 people, learning to read the body's hidden signals in ways human doctors never could.

Sleep clinics have long collected incredibly detailed physiological data—brain waves, heart rhythms, breathing patterns, eye movements, and leg activity—all recorded continuously through a process called polysomnography. Yet most of that rich information has gone underutilized. Modern AI changes that equation.

The Stanford team built a "foundation model," similar to how ChatGPT learns language patterns from vast text, but here the model learns the "language of sleep." They split sleep recordings into five-second chunks and used a novel training technique called leave-one-out contrastive learning, which hides one data stream and challenges the model to reconstruct it from the others. This forced the AI to understand how brain signals, heart activity, breathing, and muscle movements all relate and interact.

After training, researchers tested SleepFM on two fronts. First, it performed as well as or better than current clinical models at standard sleep tasks—identifying sleep stages and diagnosing sleep apnea severity. Then came the ambitious test: predicting future disease. Using 35,000 patients from Stanford's sleep center with up to 25 years of health records, the model analyzed over 1,000 disease categories.

Results were striking. SleepFM predicted Parkinson's disease (89% accuracy), dementia (85%), and several cancers with high precision. It even predicted heart attacks and overall mortality risk with 81-84% accuracy. The model excelled because it didn't rely on isolated signals—combinations of misaligned body rhythms (a sleeping brain paired with an alert heart, for instance) proved most predictive.

This breakthrough suggests that one night's sleep could eventually become a powerful screening tool for disease prevention.

REFERENCE: Thapa, R., Kjaer, M.R., He, B. et al. A multimodal sleep foundation model for disease prediction. Nat Med (2026). https://doi.org/10.1038/s41591-025-04133-4