Stanford researchers have built an AI model that can predict your risk of developing over 130 different diseases - including cancer, dementia, and heart disease - from a single night of sleep data. The study, published in Nature Medicine in January, demonstrates that sleep contains far more diagnostic information than previously understood.
How It Works
SleepFM is what researchers call a foundation model - a large AI system trained on massive amounts of data that can then be applied to many different tasks. The team trained it on 585,000 hours of polysomnography (PSG) recordings from approximately 65,000 participants across multiple research cohorts.
Polysomnography is the clinical gold standard for sleep analysis. Patients spend a night in a sleep lab connected to sensors that record brain activity (EEG), heart rhythm (ECG), eye movements, respiratory signals, and muscle activity. It’s typically used to diagnose conditions like sleep apnea.
The researchers fed all this multimodal data through a contrastive learning framework. Each signal type (brain, heart, respiratory) gets encoded separately, then combined with patient age and sex to generate disease risk predictions.
The Results
SleepFM achieved clinically significant accuracy (C-index above 0.75) across 130 conditions. The C-index measures how well a model ranks patients by risk - a score of 0.5 is random chance, 1.0 is perfect prediction.
The strongest predictions:
- Parkinson’s disease: 0.89 C-index
- Prostate cancer: 0.89
- Breast cancer: 0.87
- Dementia: 0.85
- All-cause mortality: 0.84
- Hypertensive heart disease: 0.84
- Heart attack: 0.81
- Heart failure: 0.80
- Chronic kidney disease: 0.79
- Stroke: 0.78
What stands out is the range of conditions. Sleep data predicted not just sleep disorders or brain diseases, but cancers, pregnancy complications, and kidney problems.
What This Means
The key finding wasn’t just that sleep predicts disease. It’s that combining all the different signals - brain, heart, respiratory - works better than any single one alone. Heart signals contribute most to heart disease predictions, brain signals to mental health predictions, but the ensemble outperforms either in isolation.
“The most information we got for predicting disease was by contrasting the different channels,” Emmanuel Mignot, one of the study’s authors, told Stanford News.
This suggests that diseases leave subtle signatures across multiple physiological systems during sleep, even years before diagnosis. The body’s overnight recovery processes may reveal dysfunction that daytime measurements miss.
For clinical practice, the implications are significant. Sleep studies are already routine for certain conditions. If SleepFM or similar models prove reliable in prospective trials, every sleep study could become a comprehensive health screening - one test generating risk scores for dozens of conditions.
The Fine Print
This is a retrospective study. The researchers trained and tested on historical data where outcomes are already known. Prospective validation - predicting disease in patients before it happens - is needed before clinical deployment.
The data came from sleep lab patients, not the general population. People who get polysomnography often have sleep complaints, which may not represent typical disease risk profiles. The model’s performance on healthier populations is unknown.
Polysomnography requires an overnight stay in a specialized facility with trained technicians. It costs $1,000-$3,000 in the US. Home sleep tests are cheaper but record fewer signals. Whether SleepFM works with consumer wearables that track heart rate and movement remains untested.
The study also raises questions about what exactly the model learned. Is it detecting early disease processes? Or demographic and lifestyle factors that correlate with both sleep patterns and disease risk? Mechanistic understanding lags behind predictive performance.
Still, a 0.85 C-index for dementia prediction from sleep data alone is remarkable. If validated, this could change how we think about routine health screening.