Multimodal Model Uses Overnight Sleep to Predict Future Disease Risk

A new artificial intelligence model suggests that a single night of lab‑recorded sleep may contain early clues to a person’s long‑term health. The model learns from polysomnography, an overnight test that tracks brain activity, heart rhythms, breathing, movements and other physiological signals, and is trained on five‑second segments of these recordings treated like “words” so it can, as James Zou, senior author of the study published in Nature Medicine, put it, learn “the language of sleep.”

Polysomnography has long been the gold standard for diagnosing sleep disorders, but only a small portion of its rich signals is typically used. To tap into that complexity, the team built a foundation model capable of integrating multiple data streams, including electroencephalography, electrocardiography, electromyography, pulse readings and airflow measurements, and then relating them to one another. To harmonize these different modalities, the researchers created a training strategy called leave‑one‑out contrastive learning, in which one type of signal is hidden and the model must reconstruct it from the remaining channels.

Once trained, the model was first tested on familiar tasks such as sleep staging and assessing sleep apnea severity, where it matched or exceeded current top‑performing models. The researchers then asked whether the same model could forecast future disease, drawing on clinical records from the Stanford Sleep Medicine Center and pairing polysomnography data from about 35,000 patients, collected between 1999 and 2024, with up to 25 years of follow‑up health information.

By examining over 1,000 disease categories, the model identified 130 conditions it could predict with reasonable accuracy, particularly cancers, pregnancy complications, circulatory problems and mental disorders, often reaching a concordance index above 0.8. It was especially accurate in anticipating Parkinson’s disease, dementia, hypertensive heart disease, heart attack, prostate cancer, breast cancer and risk of death.

The group is now exploring how to refine predictions, possibly by adding wearable data, and is developing interpretation techniques to see which features drive specific forecasts.

“It doesn’t explain that to us in English,” Zhu added. “But we have developed different interpretation techniques to figure out what the model is looking at when it’s making a specific disease prediction.”