Stanford AI Model Predicts Risk for 130+ Diseases From a Single Night's Sleep Data

A restless night is usually treated as a short-term nuisance that guarantees a bleary-eyed morning. But according to new research, the physiological signals our bodies emit while unconscious contain a hidden, highly detailed map of our long-term health.[1]

Scientists at Stanford Medicine have developed a first-of-its-kind artificial intelligence system capable of forecasting a person's risk for more than 130 different medical conditions based entirely on a single night of sleep data.[1][2]

The model, dubbed SleepFM (Sleep Foundation Model), was detailed in the journal Nature Medicine. It successfully predicted the onset of severe illnesses—including Parkinson's disease, dementia, heart failure, and multiple types of cancer—years before patients showed clinical symptoms.[1]

Historically, sleep medicine has been narrowly focused. When patients undergo overnight sleep studies, doctors typically look for immediate, mechanical issues like obstructive sleep apnea or insomnia, leaving the broader implications of the data largely unexplored.[3]

"We record an amazing number of signals when we study sleep," said Emmanuel Mignot, a professor of sleep medicine at Stanford and co-senior author of the study. "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][2]

Despite this richness, the vast majority of the data collected during these overnight sessions has traditionally been ignored. Human doctors simply cannot process the millions of subtle, overlapping data points generated by a sleeping brain and body simultaneously.[4]

To unlock this information, the Stanford team turned to a "foundation model"—the same underlying AI architecture that powers large language models. But instead of learning the patterns of human language, SleepFM was trained to understand the complex language of human physiology.[2]

The researchers fed the AI nearly 600,000 hours of sleep recordings from 65,000 individuals. This data came from polysomnography, the clinical gold standard for sleep testing, which uses an array of sensors to continuously track brain waves, heart rhythms, breathing patterns, blood oxygen levels, and muscle movements.[1][3]

The training process relied on a technique called "leave-one-out contrastive learning." The AI was given data in five-second chunks, but researchers would intentionally hide one data stream—such as the heart rate—and force the model to reconstruct it based solely on the brain waves and breathing patterns.[4]

By repeating this process millions of times, SleepFM learned exactly how a healthy body's systems interact. It learned, for example, that a specific type of brain wave should correspond with a specific dip in heart rate and a certain rhythm of respiration.[3]

Once the AI understood normal physiological harmony, it became exceptionally good at spotting microscopic misalignments. The researchers discovered that these subtle mismatches—such as a sleeping brain paired with an unusually alert cardiovascular system—are powerful early indicators of systemic disease.

To test the model's predictive power, the team linked the sleep data to up to 25 years of electronic health records from the Stanford Sleep Medicine Center. They asked the AI to predict which patients would eventually develop various conditions, and then checked the historical records to verify the outcomes.[1][2]

The results were striking. SleepFM evaluated more than 1,000 disease categories and found 130 that could be predicted with high accuracy using only the overnight sleep data, fundamentally changing how researchers view the diagnostic potential of sleep.[1][3]

The model's accuracy is measured using a "C-index," which evaluates how often the AI correctly ranks which of two patients will develop a condition first. A score of 0.8 means the model is correct 80% of the time, a highly significant threshold in medical forecasting.[1][2]

SleepFM achieved a C-index of 0.89 for Parkinson's disease and prostate cancer, 0.87 for breast cancer, 0.85 for dementia, and 0.81 for heart attacks. It even predicted overall all-cause mortality with an impressive 84% accuracy.[1][4]

"We were pleasantly surprised that for a pretty diverse set of conditions, the model is able to make informative predictions," noted James Zou, an associate professor of biomedical data science at Stanford and co-senior author of the research.[1][2]

The breakthrough represents a fundamental shift in preventive medicine. It suggests that overnight sleep studies, currently viewed as specialized diagnostic tools for breathing disorders, could eventually be repurposed as broad, non-invasive screening mechanisms for a wide array of life-threatening illnesses.[3]

However, researchers caution that the technology is not yet ready for consumer smartphones or smartwatches. SleepFM relies on the dense, multi-channel data provided by clinical polysomnography, which requires a patient to sleep in a lab hooked up to medical-grade sensors.[2][3]

The next frontier for the Stanford team is determining whether these predictive capabilities can be adapted for consumer wearables. If foundation models can learn to extract similar insights from the simplified sensors on a smartwatch, continuous, passive disease screening could become a reality for millions.[4]