Stanford's SleepFM AI predicts 130+ diseases from a single night's sleep data

SleepFM Transforms Sleep Data Into Predictive Health Insights

A restless night might reveal more than just tomorrow's fatigue. Scientists at Stanford Medicine have developed SleepFM, an AI foundation model that examines sleep data to predict risk of medical conditions years before symptoms appear1

1

. Published in Nature Medicine, this research demonstrates how AI trained on sleep data can identify hidden physiological patterns that signal future disease2

2

.

The system was trained on nearly 585,000 hours of polysomnography recordings from approximately 65,000 individuals3

3

. Polysomnography captures brain activity, heart function, breathing patterns, eye movement, leg motion, and other physiological signals during overnight sleep study sessions. While these comprehensive assessments are typically used only to diagnose sleep disorders, researchers recognized they contain vast amounts of underutilized health information.

"We record an amazing number of signals when we study sleep," said Emmanuel Mignot, the Craig Reynolds Professor in Sleep Medicine and co-senior author. "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"1

1

.

AI Foundation Model Learns the Language of Sleep

SleepFM operates as a foundation model, similar to large language models like ChatGPT but trained on biological signals rather than text. The researchers divided each sleep recording into five-second segments that function like words in a language-based system3

3

. This approach allows the model to process long nights as sequences and learn normal patterns across multiple data streams.

"SleepFM is essentially learning the language of sleep," explained James Zou, associate professor of biomedical data science and co-senior author4

4

. The team developed a training method called leave-one-out contrastive learning, which removes one type of signal at a time and asks the model to reconstruct it using remaining data. This technical advance helped harmonize different data modalities so they could work together effectively1

1

.

Disease Prediction Across 130 Medical Conditions

After initial training, researchers linked polysomnography records with long-term health outcomes from the Stanford Sleep Medicine Center, founded in 1970 by the late William Dement. The largest dataset included approximately 35,000 patients aged 2 to 96, with sleep studies recorded between 1999 and 2024 paired with electronic health records tracking some individuals for up to 25 years1

1

.

SleepFM analyzed more than 1,000 disease categories and identified 130 conditions that could be predicted with reasonable accuracy using sleep data alone3

3

. The system demonstrated particularly strong performance for cancers, pregnancy complications, circulatory diseases, and mental health disorders, achieving prediction scores above a C-index of 0.82

2

.

The C-index, or concordance index, measures how well a model ranks risk across individuals. "A C-index of 0.8 means that 80% of the time, the model's prediction is concordant with what actually happened," Zou clarified3

3

.

Strongest Results for Specific Conditions and Mortality

SleepFM achieved notably high accuracy for several specific outcomes. For Parkinson's disease, the model reached a C-index of 0.93 over a six-year prediction window2

2

. Other strong predictions included dementia (0.85), hypertensive heart disease (0.84), heart attack (0.81), prostate cancer (0.89), and breast cancer (0.87)4

4

.

The system also demonstrated the ability to predict overall mortality risk with a C-index of 0.844

4

. Researchers emphasize these predictions reflect statistical risk stratification rather than causal relationships or imminent disease onset2

2

.

Cross-Talk Between Body Systems Reveals Health Clues

The research revealed intriguing patterns in how different physiological signals contribute to disease prediction. Heart signals during sleep factored more prominently in predictions for circulatory diseases, while brain signals played larger roles in mental health disorder forecasts4

4

. However, combining all available data streams produced the most accurate predictions for early disease detection.

"The most information we got for predicting disease was by contrasting the different channels," Mignot noted. Sleep data showing misalignment between body systems—such as a brain that appears asleep while the heart shows waking patterns—seemed particularly indicative of future health problems4

4

.

What This Means for Future Healthcare

From an AI perspective, sleep has been relatively understudied compared to fields like pathology or cardiology, despite being such a vital part of life1

1

. This work represents the first application of AI to sleep data on such a massive scale, opening new possibilities for how overnight sleep study data could serve as an early warning system.

The team is now working to improve SleepFM's predictions by potentially incorporating data from wearable devices and developing better interpretation techniques to understand what the model identifies when making specific disease predictions. "It doesn't explain that to us in English," Zou said, "But we have developed different interpretation techniques to figure out what the model is looking at"4

4

.

For clinicians and patients, this research suggests sleep studies could evolve beyond diagnosing sleep disorders to become comprehensive health screening tools. The ability to predict future health conditions from a single night's physiological data could enable earlier interventions and more personalized preventive care strategies, particularly for conditions like Parkinson's disease, various cancers, and cardiovascular diseases where early detection significantly impacts outcomes.