Sleep-study ECG model predicts later heart risks in hospital cohorts

An artificial intelligence model that analyzes heart recordings collected during sleep studies identified patients at higher risk of later atrial fibrillation, heart failure and death, according to research published October 9 in SLEEP. The findings suggest that information already collected during overnight testing could contribute to cardiovascular risk assessment, a potential route toward earlier prevention.

The National Institutes of Health highlighted the work the same day. Its practical appeal is straightforward: sleep laboratories already record the heart’s electrical activity, but those recordings may contain information beyond the immediate assessment of a sleep disorder. Whether extracting that information improves patients’ health remains an unanswered question.

### What researchers tested

The retrospective study combined single-lead electrocardiograms, or ECGs, with sleep stages annotated by experts. Researchers developed the model using 15,809 Massachusetts General Hospital patients and evaluated it in separate cohorts of 9,810 patients at Emory University Hospital and 12,576 at Beth Israel Deaconess Medical Center.

The intended prediction targets were atrial fibrillation, stroke, myocardial infarction, heart failure and death over a ten-year horizon. Outcomes came from electronic health records. This was an analysis of existing recordings and subsequent outcomes; patients were not assigned to care guided by the algorithm.

Higher model scores remained associated with cardiovascular risk after accounting for conventional factors and sleep characteristics. NIH’s account emphasized the findings for atrial fibrillation, heart failure and mortality, while identifying heart attack and stroke prediction as areas requiring further work.

### The limits of a risk signal

The paper’s conclusions distinguish those outcomes more explicitly: the model did not substantially improve prediction of myocardial infarction and stroke beyond established risk factors. That distinction matters because a model can show a statistical association without adding enough information to improve a clinical decision.

The researchers also acknowledge that cardiovascular diagnoses relied on billing classification codes rather than direct clinical confirmation. Testing at separate hospitals strengthens the evidence that the signal travels beyond the development dataset, but it does not establish performance in everyone who might eventually be screened.

For a healthspan application, the relevant question extends beyond who receives a higher score. A useful tool would need to improve decisions and ultimately reduce illness or disability. This study does not demonstrate that using its predictions prevents cardiovascular events or extends life.

### A wider challenge for sleep AI

A separate perspective published in SLEEP on October 8 provides relevant methodological context. Its authors examined commonly used sleep-research cohorts and reported that relatively simple models based on age and sex, or traditional cardiovascular risk factors, could closely match results from more complex AI approaches.

That perspective concerns other models and datasets; it is not a direct rebuttal of the October 9 hospital study. It nevertheless highlights an important comparison: whether an algorithm discovers useful additional information or largely reconstructs risk already captured by familiar patient characteristics.

The new study’s adjusted analyses address part of that question. They do not settle whether deploying the model would produce a worthwhile change in care, how many unnecessary follow-up investigations it might generate, or how well it would work across different patient populations.

The immediate advance is evidence that routinely collected overnight ECGs merit further study as a source of prognostic information. Prospective validation and assessment of clinical usefulness are still needed before this approach can support routine screening.