AI analysis of heart activity during sleep predicts cardiovascular risk up to a decade in advance

ECG data from sleep studies can detect heightened risk of atrial fibrillation, stroke, heart failure and death, according to US research

By LineZotpaper

Published21 October 2026

Read Time3 min

Artificial intelligence trained on electrocardiogram (ECG) data recorded during sleep can predict a person's risk of developing atrial fibrillation, stroke, heart failure, or dying from any cause within the next 10 years, researchers at Emory University and Georgia Institute of Technology have found. The study, published in the journal SLEEP and supported by the US National Institutes of Health, suggests that subtle changes in heart activity during early sleep stages contain early warning signals currently overlooked in standard sleep tests.

Researchers in the United States have developed a deep learning system that uses overnight ECG recordings to forecast long-term cardiovascular outcomes. The model was first trained on data from nearly 16,000 patients enrolled in a sleep study at Massachusetts General Hospital in Boston, who had their heart activity measured during sleep using a single ECG lead.

Once trained, the AI was tested on two independent datasets: nearly 10,000 patients from Emory University Hospital and 13,000 patients from Beth Israel Deaconess Medical Center in Boston. The system sorted participants into low, medium, and high risk categories for a range of cardiovascular outcomes over the following decade.

For each standard deviation increase in the AI's risk score, the model found a 103 to 172 percent higher risk of atrial fibrillation (irregular heartbeat), a 19 to 40 percent higher risk of stroke, a 69 to 133 percent higher risk of heart failure, and a 65 to 82 percent higher risk of death from any cause. Only one group showed a significantly increased risk of future myocardial infarction, or heart attack.

Importantly, the predictive value held even after adjusting for established risk factors such as age, sex, diabetes, hypertension, and smoking. The AI model improved existing cardiovascular risk models by up to 6 percent for atrial fibrillation, up to 9 percent for heart failure, and up to 7 percent for mortality prediction.

"Importantly, we found that adding the score from our deep learning model improved prediction of cardiac outcomes beyond established risk factors," said Gari Clifford, study author and professor of biomedical engineering at Georgia Institute of Technology. "Adding these scores into sleep studies has great potential for improving early detection of cardiovascular disease, determining who is most at risk, and being more proactive in patient care."

The research suggests that the lightest stage of sleep, when the nervous system undergoes rapid changes, may be especially revealing. "These rapid autonomic fluctuations may amplify subtle abnormalities in heart-rate dynamics or conduction patterns," the authors wrote.

David Goff, acting director of NIH's National Heart, Lung, and Blood Institute, said the approach "has the potential to identify individuals at risk for cardiovascular disease years before clinical symptoms arise." He added that evaluating whether this information improves traditional risk prediction is an important next step.

During a polysomnography (PSG) test, the gold standard for diagnosing sleep disorders, the heart is often recorded as well. But unless the data shows a serious red flag, the findings are rarely considered. The researchers said that subtle information on cardiac function could instead be put to use.

With further development, the AI system could help health practitioners screen patients with sleep disorders who may face an elevated risk of heart failure.

Analysis

Why This Matters

Cardiovascular disease is a leading cause of death worldwide, and early detection can significantly improve outcomes. This AI approach could identify at-risk individuals years before symptoms appear.
Sleep studies (polysomnography) already record ECG data, but that information is largely unused for cardiovascular risk assessment. This method would not require additional testing, just better analysis of existing data.
The model could be integrated into routine sleep study diagnostics as a screening tool for multiple adverse cardiovascular outcomes, potentially saving lives through earlier intervention.

Background

Sleep disorders such as insomnia are known risk factors for cardiovascular disease, but the heart's electrical activity during sleep has not been widely used for long-term risk prediction. Traditional cardiovascular risk models typically consider ECG recordings taken when a person is awake or walking. This study, led by scientists at Emory University, trained a deep learning system on public health data from nearly 16,000 patients and validated it on two additional cohorts totaling more than 23,000 patients. The research was published in the journal SLEEP with support from the US National Institutes of Health.

Key Perspectives

Researchers: The authors argue that ECG data from sleep, particularly during light sleep when autonomic nervous system fluctuations occur, contains valuable predictive information that is currently overlooked. They see this as a low-cost way to improve existing risk models.
NIH: David Goff of the National Heart, Lung, and Blood Institute supports the approach but cautions that the next step is to evaluate whether this information improves traditional risk prediction in practice.
Critics/Skeptics: The absolute improvements in risk prediction are modest: up to 9 percent for heart failure, up to 7 percent for mortality. Some clinicians may question whether these gains justify changes to clinical practice without further validation in diverse, real-world populations.

What to Watch

Whether the NIH or other funding bodies support a prospective clinical trial testing the AI system in routine sleep study settings.
If the deep learning model is made available as a software tool or integrated into commercial polysomnography equipment.
How the model performs in populations outside the hospital-based cohorts studied, including community and primary care settings.

Sources

The Way You Sleep Can Reveal Your Risk of Heart Failure Up to 10 Years in Advance · ScienceAlert

Checked

12 claims in this story checked against 1 source: 12 supported.

Show each claim

SupportedAI analysis of heart activity during sleep predicts cardiovascular risk up to a decade in advance.ScienceAlert: “With that, artificial intelligence can predict someone's risk of an irregular heartbeat, stroke, heart failure, or death from any cause in the next 10 years.”
SupportedSleep-study ECG data can detect heightened risk of atrial fibrillation, stroke, heart failure and death, according to US research.ScienceAlert: “Using a deep learning algorithm, researchers in the US have now found that ECG recordings show strong predictive value for cardiovascular health outcomes when combined with sleep stages.”
SupportedGari Clifford said: “Importantly, we found that adding the score from our deep learning model improved prediction of cardiac outcomes beyond established risk factors.”ScienceAlert: ““Importantly, we found that adding the score from our deep learning model improved prediction of cardiac outcomes beyond established risk factors,” says Gari Clifford, study author and professor of biomedical engineering at Georgia Institute of Technology.”
SupportedThe authors wrote: “These rapid autonomic fluctuations may amplify subtle abnormalities in heart-rate dynamics or conduction patterns.”ScienceAlert: ““These rapid autonomic fluctuations may amplify subtle abnormalities in heart-rate dynamics or conduction patterns,” the authors explain.”
SupportedDavid Goff said the approach “has the potential to identify individuals at risk for cardiovascular disease years before clinical symptoms arise,” and that evaluating its effect on traditional risk prediction is an important next step.ScienceAlert: ““This new approach has the potential to identify individuals at risk for cardiovascular disease years before clinical symptoms arise,” says David Goff, acting director of NIH's National Heart, Lung, and Blood Institute (NHLBI). “Evaluating whether this information improves traditional risk prediction is an important next step.””
SupportedThe model was trained on data from nearly 16,000 patients enrolled in a sleep study at Massachusetts General Hospital in Boston.ScienceAlert: “The research team, led by scientists at Emory University, first trained a deep learning system on the public health data of nearly 16,000 patients. These patients were enrolled in a sleep study at Massachusetts General Hospital in Boston”
SupportedThe model was tested on nearly 10,000 patients from Emory University Hospital.ScienceAlert: “Once trained, the AI system was then put to the test on two more datasets, including nearly 10,000 patients from Emory University Hospital”
SupportedThe model was tested on 13,000 patients from Beth Israel Deaconess Medical Center in Boston.ScienceAlert: “and 13,000 patients from Beth Israel Deaconess Medical Center in Boston.”
SupportedFor each standard deviation increase in the AI risk score, atrial fibrillation risk was 103 to 172 percent higher.ScienceAlert: “For each standard deviation increase in the AI's risk score, there was a 103 to 172 percent higher risk of atrial fibrillation, which is an irregular heartbeat.”
SupportedFor each standard deviation increase in the AI risk score, stroke risk was 19 to 40 percent higher.ScienceAlert: “There was also a 19 to 40 percent higher risk of stroke”
SupportedFor each standard deviation increase in the AI risk score, heart failure risk was 69 to 133 percent higher.ScienceAlert: “a 69 to 133 percent higher risk of heart failure”
SupportedFor each standard deviation increase in the AI risk score, risk of death from any cause was 65 to 82 percent higher.ScienceAlert: “and a 65 to 82 percent higher risk of death from any cause.”