A ground breaking AI model created by Stanford University researchers and their collaborators may one day predict a person’s risk for more than 100 different health conditions — all without the individual needing to be awake.
According to a newly published study, the Sleep FM AI model examines a wide range of physiological signals to estimate future risks of dementia, heart failure, and all-cause mortality, using data from just one night of sleep.
Sleep FM: A Foundation Model Trained on Massive Sleep Data
Sleep FM is a foundation model, similar in concept to Chat GPT, but instead of learning from text, it learns from sleep. The system was trained on nearly 600,000 hours of sleep recordings collected from about 65,000 individuals.
Rather than words or sentences, Sleep FM processes 5-second segments of sleep data gathered from multiple sleep clinics, allowing it to recognize complex biological patterns.
How Sleep Data Was Collected Using Polysomnography
The sleep information was gathered using polysomnography (PSG), a comprehensive and widely trusted method often considered the gold standard of sleep studies. This approach relies on multiple sensors to monitor brain activity, heart rhythms, breathing patterns, eye movement, and leg motion during sleep.
“We record an amazing number of signals when we study sleep,” says Emmanuel Mignot, a Stanford sleep medicine professor and co-senior author of the study.
Teaching AI to Learn From Missing Biological Signals
To evaluate Sleep FM, researchers introduced a new training strategy known as leave-one-out contrastive learning. In this method, one type of physiological signal — such as airflow or pulse data — is intentionally removed, requiring the AI to infer missing information from the remaining biological signals.
The team then paired this sleep data with tens of thousands of long-term patient health records, spanning up to 25 years of follow-up across a wide age range.
Predicting Over 130 Diseases From Sleep Patterns
After reviewing more than 1,041 disease categories, SleepFM successfully predicted 130 health conditions with notable accuracy based solely on sleep data.
The model showed particularly strong performance in forecasting cancers, pregnancy-related complications, cardiovascular diseases, and mental health disorders, achieving a C-index above 0.8.
High Accuracy Across Multiple Prediction Models
SleepFM also performed well under the AUROC classification framework, which measures how accurately the model distinguishes between individuals who do or do not experience specific health events within a six-year prediction window.
Overall, Sleep FM outperformed existing predictive models, especially in identifying risks for Parkinson’s disease, heart attacks, strokes, chronic kidney disease, prostate cancer, breast cancer, and overall mortality.
Why Mismatched Body Signals Matter Most
While some sleep stages and data types proved more predictive than others, the strongest insights came from imbalances between bodily systems.
In particular, physiological signals that appeared out of sync — such as a brain showing sleep patterns while the heart appeared alert — were among the most reliable indicators of future disease risk.
“A brain that looks asleep but a heart that looks awake seemed to signal trouble,” Mignot explains.
Limitations and Future Potential of Sleep-Based AI
The researchers acknowledge several limitations, including changes in clinical practices over time and the fact that the data came from patients already referred for sleep studies, meaning some segments of the general population were underrepresented.
Despite these challenges, the study highlights the life-saving potential of AI in healthcare. Future applications could integrate Sleep FM with wearable sleep devices, enabling continuous, real-time health monitoring.
AI Learning the Language of Sleep
Just as large language models learn human communication by analyzing text, Sleep FM is learning to interpret sleep itself.
