AI Reading Heart Signals During Sleep Studies Predicts 10-Year Cardiovascular Risk

The work was published in the journal SLEEP in 2026. It is a retrospective study of patient records from three U.S. hospitals, not a trial. The model was fine-tuned on 15,809 patients at Massachusetts General Hospital. It was then tested on two separate groups: 9,810 patients at Emory University Hospital and 12,576 at Beth Israel Deaconess Medical Center. Outcomes came from electronic health records, using diagnosis codes.

The recordings come from polysomnography, the overnight test used to diagnose sleep disorders such as sleep apnea. During the test, a single-lead electrocardiogram (ECG) tracks the heart's electrical activity. According to the authors, these ECGs are recorded but rarely analyzed. Sleep disorders are recognized as risk factors for heart disease, so the team asked whether the data could do more.

The researchers used a deep learning model, a type of artificial intelligence, that had first been trained to detect arrhythmias, or abnormal heart rhythms. They combined its ECG analysis with information on sleep stages. The goal was to predict 10-year risk of atrial fibrillation, stroke, heart attack, heart failure and death from any cause.

The model's output was strongly associated with long-term risk (p < .0001). This held after the team adjusted for age, sex, body mass index, smoking, hypertension and diabetes. It also held after adjusting for sleep measures, including the apneahypopnea index, which counts breathing pauses, as well as arousals, limb movements, time in each sleep stage and sleep efficiency.

The sizes of the associations are reported as hazard ratios. A hazard ratio above 1 means a higher rate of the outcome. For each one-standard-deviation increase in the model's output, the hazard ratio for atrial fibrillation was 2.03 at Emory and 2.72 at Beth Israel Deaconess. For heart failure, it was 1.69 and 2.33. For death from any cause, it was 1.82 and 1.65.

Results for stroke and heart attack were weaker. The stroke hazard ratios were 1.19 and 1.40. For heart attack, they were 1.38 at Emory and 1.10 at Beth Israel Deaconess, and the second figure had a confidence interval (0.86 to 1.40) that includes no effect.

What does this mean for someone referred for a sleep study? Nothing changes today. The authors conclude that building such predictions into polysomnography could improve detection and support personalized clinical decision-making. That is the authors' proposal, not something this study has tested in patients.

There are limits. The outcomes came from health records and diagnosis codes, which can miss or misclassify cases. The study shows association, not that the score causes or changes anyone's outcomes.

The next test is whether acting on such a score leads to better outcomes for patients, and whether the model performs as well for stroke and heart attack.

England and Wales Mortality Hits Record Nine-Month Low, Pressuring Annuity Pricing Parkinson's Stem Cell Developer Aspen Neuroscience Files for Nasdaq IPO IMF Says AI Productivity Gains Could Outweigh Asia's Aging-Driven Workforce Decline UCSF Details How FDA's Bone Density Endpoint Will Shrink Osteoporosis Trials Tiny Health Raises $33 Million Series B as Adult Microbiome Testing Overtakes Pediatric