← Back to Home

Heart Clues from a Night’s Sleep: AI Identifies Long-Term Risk from Sleep-Study ECGs

ECG signals recorded during sleep studies may open another window for cardiovascular screening. A study validated their ability to predict atrial fibrillation, heart failure, and mortality risk in two external datasets, but whether this can improve actual care still requires prospective research.

By SURL BioNews

An overnight examination to identify sleep problems may also provide clues about future heart health. A study published in *SLEEP* used deep learning to analyze electrocardiograms recorded during sleep and identify which patients were more likely to develop atrial fibrillation or heart failure, or die from any cause, giving signals already being collected a new use.

The study was supported by the National Heart, Lung, and Blood Institute, part of the U.S. National Institutes of Health. According to an explanation released by NIH on October 9 and published by Medical Xpress, the study used single-lead ECGs from polysomnography, together with expert-annotated sleep-stage data, to explore health risks over the next ten years.

The research team first fine-tuned the model using data from 15,809 patients at Massachusetts General Hospital, then tested its performance using data from two other hospitals: 9,810 patients at Emory University Hospital and 12,576 at Beth Israel Deaconess Medical Center, for a total of 22,386 patients in external validation. Subsequent disease and mortality outcomes were obtained from electronic health records. This design allowed the researchers to examine whether the model could maintain its predictive ability beyond the development data.

Patients with higher model scores also had higher long-term risks. The study still found predictive value after adjusting for age, sex, body mass index, diabetes, and hypertension. The signal persisted even after further accounting for sleep apnea severity, arousals during sleep, and sleep efficiency. This suggests that the information provided by the model may be more than another representation of existing risk factors.

However, predictive performance varied across outcomes. Currently supported uses center on atrial fibrillation, heart failure, and all-cause mortality; the researchers noted that prediction of myocardial infarction and stroke requires further optimization. All-cause mortality also includes causes of death beyond cardiovascular disease and cannot be directly interpreted as the risk of death from heart disease.

The practical appeal of this approach lies in its use of single-lead ECG signals already recorded during sleep studies. If subsequent validation succeeds, sleep studies could also help identify people who need further cardiovascular evaluation. However, this model also used expert-annotated sleep-stage data, so these findings do not establish that ordinary wearable devices or home ECG patches already have the same capability.

There is still some distance between risk stratification in the data and clinical screening. External validation increases confidence in the findings, but it has not yet demonstrated that physicians can improve patient outcomes by ordering tests or interventions based on model scores. The next step requires prospective studies to clarify whether adding this information provides benefits sufficient to change care decisions, and the costs of additional testing.

References

  1. Medical Xpress / National Institutes of Health
  2. National Institutes of Health