AI unmasks hidden heart, brain risks in routine sleep studies

Clinical takeaway: High-risk patients in this cohort turned up at every level of apnea severity, including normal. A low score should not end the conversation about cardiac or cognitive risk in a patient with disrupted sleep.

An overnight sleep study records brain activity, eye movement, chin muscle tone, airflow, chest effort, cardiac tracing, and oxygen saturation across a full night. Nearly all of that detailed data goes unused by the time the report reaches the ordering clinician, where the reading comes down to one count of breathing events per hour.

The apnea-hypopnea index (AHI) decides the diagnosis, sets the severity label, and often determines whether insurance covers treatment. It counts how often the airway gives out. It was never built to say what the night is doing to the rest of the body, and in epidemiologic studies it has consistently failed at that job.

Sleep medicine has spent years documenting that AHI predicts treatment need better than it predicts outcomes. The American Thoracic Society made finding a replacement an official research priority in 2025, and earlier attempts, including hypoxic burden and heart-rate response measures, each captured one slice of physiology. None used the full recording.

With one million to four million studies performed annually in the US, an unused measure is an unused measure at scale. Researchers trained a model on more than 9,600 overnight recordings from about 9,300 adults referred for in-lab testing at a large US health system between 2012 and 2022, each linked to a decade or more of medical records.

Risk rose across each of five tiers established across nearly every outcome measured. In the highest tier, atrial fibrillation ran more than twice as common as in the lowest and cognitive impairment nearly twice, with myocardial infarction, heart failure, and major cardiovascular events all elevated. Mortality followed the same gradient, more than doubling in the top tier. AHI severity categories showed no increase in mortality at all.

Absolute gaps were narrower than the ratios imply: roughly 67% of top-tier patients stayed free of major cardiovascular events, against 81% in the lowest. The model transferred, too. Applied without retraining to more than 6,000 community-based adults, it split high- from low-risk patients on death and heart failure in both sexes. Earlier AHI-based work in that cohort found associations only in men with severe disease.

This was a retrospective analysis, and the risk gradient persisted after adjustment for AHI itself over a mean 4.9 years after the sleep study. What the model reads is not a refined breathing score: zeroing out the brain and cardiac channels scrambled the tiers, meaning the pattern rests partly on physiology no home sleep test records.

The open question is not whether the tiers predict, but whether acting on them can help. No trial has tested whether flagging a top-tier patient for closer monitoring changes what happens next, and prospective work would settle that question.

"For decades we have distilled an overnight sleep study into a handful of summary measures," said Reena Mehra, MD, professor of medicine at the University of Washington and the study's senior clinical author. "AI gives us the opportunity to move beyond those summaries and learn from the full richness of sleep physiology."

Source: Bilal E, et al (2026 Aug 3) Nat Commun. A foundation model for sleep-based risk stratification and clinical outcomes