The model, named SleepFM, was trained on nearly 600,000 hours of polysomnography data from 65,000 participants. Polysomnography, the gold standard in sleep studies, records a wide range of physiological signals—including brain activity, heart rate, breathing patterns, eye and leg movements—while participants sleep in a lab.

“Sleep studies capture an extraordinary amount of physiological data over eight hours,” said Emmanuel Mignot, MD, PhD, Craig Reynolds Professor in Sleep Medicine and co-senior author of the study published Jan. 6 in Nature Medicine. “We realized that only a fraction of this data has been used—and with AI, we can now extract meaningful patterns across all these signals.”

Learning the language of sleep

SleepFM is a foundation model, a type of AI that trains itself on vast datasets and applies its learning to a wide range of tasks—similar to how large language models like ChatGPT learn from text. The researchers divided the 585,000 hours of sleep data into five-second increments, analogous to the “words” used in language models, allowing SleepFM to learn the intricate patterns of sleep.

“SleepFM is essentially learning the language of sleep,” said James Zou, PhD, associate professor of biomedical data science and co-senior author. The model integrates multiple data streams—EEG, ECG, EMG, pulse readings, and airflow—to understand how different signals relate to one another. A novel training approach, called leave-one-out contrastive learning, allows the AI to predict missing signals from the rest of the data, harmonizing all modalities into a coherent representation of sleep physiology.

From sleep analysis to disease forecasting

After training, SleepFM was first tested on conventional sleep analysis tasks such as classifying sleep stages and diagnosing sleep apnea. It matched or outperformed existing state-of-the-art models.

Next, researchers aimed higher: predicting future disease onset. By pairing polysomnography data with decades of electronic health records from the Stanford Sleep Medicine Center—founded in 1970 by sleep pioneer William Dement—the team linked sleep patterns to long-term health outcomes. The cohort included approximately 35,000 patients, aged 2 to 96, with up to 25 years of follow-up.

SleepFM analyzed over 1,000 disease categories and identified 130 that could be predicted from sleep data. Predictions were particularly strong for cancers, circulatory diseases, pregnancy complications, and mental health disorders, achieving a C-index above 0.8, a measure of predictive accuracy.

The model excelled in forecasting conditions such as Parkinson’s disease (C-index 0.89), dementia (0.85), hypertensive heart disease (0.84), heart attacks (0.81), prostate cancer (0.89), breast cancer (0.87), and even mortality (0.84). “We were pleasantly surprised by the breadth of conditions SleepFM can predict,” Zou said.

Interpreting AI predictions

The researchers are exploring ways to further improve SleepFM, including integrating wearable data. While the model does not provide human-readable explanations, the team developed methods to interpret which signals influence specific predictions.

“Heart signals contribute most to heart disease predictions, brain signals to mental health outcomes, but the combination of all signals gives the most accurate forecasts,” Mignot explained. “Discrepancies—like a brain that looks asleep while the heart appears active—can indicate future health risks.”

Collaborations and authorship

The study was co-led by Rahul Thapa, PhD, and Magnus Ruud Kjaer, PhD, with contributions from researchers at the Technical University of Denmark, Copenhagen University Hospital – Rigshospitalet, BioSerenity, University of Copenhagen, and Harvard Medical School. Mignot and Zou are members of the Wu Tsai Neurosciences Institute.

The research marks a significant step toward using AI to unlock the hidden potential of sleep data, potentially transforming how clinicians predict and prevent disease from the earliest signals in the body.