Published October 9, 2026. A deep-learning model that analyzes electrocardiogram signals collected during overnight sleep studies could help identify people at higher long-term risk of atrial fibrillation, heart failure and death, according to research highlighted by the National Institutes of Health.
The results are promising, but they do not establish a new clinical test. The NIH said the method still needs further evaluation against traditional risk prediction and additional optimization for heart attack and stroke.
How the study worked
Overnight polysomnography is the standard test used to diagnose sleep disorders such as sleep apnea and insomnia. The tests often record a single-lead ECG, but that signal is not routinely used to estimate a patient’s long-term cardiovascular risk.
The research team combined single-lead ECG data with expert-labeled sleep-stage information. The model was fine-tuned using data from 15,809 patients at Massachusetts General Hospital, then assessed using two external datasets: 9,810 patients from Emory University Hospital and 12,576 from Beth Israel Deaconess Medical Center. Outcomes were drawn from electronic health records.
The model sorted patients into groups with different levels of 10-year risk. Its score remained predictive after the researchers adjusted for factors including age, sex, body-mass index, diabetes, hypertension, sleep-apnea severity, nighttime arousals and sleep efficiency, NIH reported.
What the model predicted—and what it did not
- The model showed predictive value for atrial fibrillation, heart failure and death from any cause.
- It added information beyond several established cardiovascular and sleep-related risk factors in the datasets studied.
- Researchers said more optimization is needed for predicting myocardial infarction and stroke.
- The study does not show that using the score in routine care improves patient outcomes.
- The results apply to people already undergoing sleep studies and should not be interpreted as a recommendation for consumer ECG devices or self-diagnosis.
The study was published in the journal Sleep. Its multi-center validation is a strength, but retrospective electronic-health-record research can still be affected by differences in patient populations, clinical practice and the quality of recorded outcomes. Prospective testing would be needed before a model like this could guide treatment.
Why existing sleep-study data could matter
The approach uses one ECG lead rather than a conventional 12-lead setup. Because many sleep studies already collect that signal, a validated system could potentially extract additional information without adding a separate overnight procedure. NIH said low-cost patches might also make single-lead data easier to collect, though that remains a future possibility rather than an approved clinical use.
Sleep apnea itself is associated with cardiovascular risk, so the researchers adjusted for apnea severity and other sleep measures. The model’s continued predictive value after those adjustments suggests the ECG signal may contain additional information, but it does not prove why a particular patient is at risk.
What patients should know
The research does not change current screening or treatment guidance. Patients should discuss symptoms and conventional risk factors—such as blood pressure, cholesterol, diabetes, smoking and family history—with a clinician. Anyone receiving a sleep study should follow the interpreting physician’s report rather than expecting an AI cardiovascular score from a standard test today.
For related coverage, visit AskNovus’s Artificial Intelligence section and the latest Health News.
Primary sources
Read the NIH’s October 9 news release and the peer-reviewed study in Sleep.
Featured image: Jair Lázaro via Unsplash. The photo is illustrative.