Deep learning analysis of sleep-study ECGs may forecast future cardiovascular events

October 9, 2026

National Institutes Of Health

Overview

A deep learning model using single-lead ECGs recorded during overnight polysomnography, together with sleep-stage information, stratified patients by subsequent cardiovascular risk. In the reported study, predictive value persisted after adjustment for established cardiovascular and sleep-related factors; the authors identified further optimization as necessary to establish prediction of myocardial infarction and stroke.

Background

Polysomnography is the standard diagnostic test for sleep disorders such as insomnia and obstructive sleep apnea, which are recognized cardiovascular risk factors. ECG signals are collected during these studies but, according to the source, are not routinely analyzed for cardiovascular risk prediction. The study examined whether sleep-study ECGs and sleep-stage data could predict 10-year risk of atrial fibrillation, stroke, myocardial infarction, heart failure, and all-cause death. Related guidance and research place this work within established sleep-testing workflows and broader evaluation of AI tools in cardiovascular care.

Data Highlights

Study element	Reported details
Model inputs	Single-lead ECGs recorded during sleep studies and expert-annotated sleep-stage data
Fine-tuning cohort	15,809 patients treated at Massachusetts General Hospital
Evaluation cohorts	9,810 patients from Emory University Hospital and 12,576 from Beth Israel Deaconess Medical Center
Outcomes assessed	10-year risk of atrial fibrillation, stroke, myocardial infarction, heart failure, and death from any cause
Reported predictive findings	Risk stratification and prediction of atrial fibrillation, heart failure, and all-cause death; further optimization was considered necessary for myocardial infarction and stroke prediction
Adjustment factors	Age, sex, BMI, diabetes, hypertension, sleep-apnea severity, arousal frequency, and sleep efficiency

Key Findings

The model used single-lead sleep-study ECG recordings alongside expert-annotated sleep-stage information.
Fine-tuning used data from 15,809 patients; evaluation used separate cohorts of 9,810 and 12,576 patients, with outcomes determined from electronic health records.
Model scores separated patients into groups by long-term cardiovascular risk, with higher scores indicating greater risk.
Predictive value remained after adjustment for the reported cardiovascular risk factors and sleep-related characteristics.
The model predicted atrial fibrillation, heart failure, and death from any cause; the study team said further optimization was needed to establish prediction of myocardial infarction and stroke.
The NIH source describes the single-lead approach as simpler to collect than conventional 12-lead ECG and notes that low-cost patches could provide access to ECG data during a night of sleep. The AHA science advisory cited in the web context addresses evaluation and monitoring of AI tools in cardiovascular care.

Clinical Implications

The study authors stated that incorporating model scores into sleep studies could aid earlier cardiovascular disease detection, help identify individuals at greatest risk, and support more proactive care. The NIH source identifies validation and assessment of whether the data improve conventional risk prediction as necessary next steps; the AHA advisory provides broader context on evaluating and monitoring AI tools in cardiovascular care.

Conclusion

Sleep-study ECGs paired with sleep-stage information showed potential for stratifying risk of selected future cardiovascular outcomes in the reported cohorts. Prediction of myocardial infarction and stroke requires further model optimization, and the source identifies validation as a next step.