Winston-Salem, N.C., August 6, 2026 — A new artificial intelligence system developed by researchers at Wake Forest University School of Medicine could help clinicians detect several forms of heart dysfunction using a routine electrocardiogram (ECG), including heart failure with preserved ejection fraction (HFpEF), a condition that frequently escapes early diagnosis. The model was also able to make useful predictions from a single ECG lead, a finding that raises the possibility of adapting the technology for more accessible screening in the future.
Heart failure affects more than 6 million people in the United States and remains a major cause of hospitalization and death. Although symptoms such as breathlessness, fatigue and swelling can signal the disease, early heart failure may develop quietly. Confirming the condition often requires an echocardiogram, an ultrasound examination that measures how the heart contracts, relaxes and fills. Such imaging is highly valuable, but it may not be immediately available in primary-care offices, rural clinics or other settings with limited resources. An AI-assisted ECG could provide a rapid way to identify people who warrant further evaluation.
The study, published in the Journal of the American Heart Association, describes an AI model designed to classify three types of left ventricular dysfunction. These include reduced ejection fraction (rEF), in which the heart’s main pumping chamber ejects substantially less blood than normal; mildly reduced ejection fraction (mEF); and HFpEF. Ejection fraction is the percentage of blood expelled from the left ventricle during each heartbeat. In HFpEF, that percentage can remain within a normal range even though the ventricle has become stiff or otherwise abnormal, preventing it from filling and functioning efficiently.
That distinction makes HFpEF particularly difficult to recognize. A patient can have significant symptoms and impaired cardiac performance without the obvious reduction in pumping strength associated with conventional systolic heart failure. The Wake Forest team’s model searches the electrical waveform of an ECG for subtle patterns associated with these different forms of dysfunction. Rather than relying on a clinician to recognize a visible abnormality, the system uses machine-learning algorithms to analyze relationships across the ECG signal that may be too complex or faint for the human eye.
“Our AI model can detect various types of heart dysfunction from a simple, single-lead ECG alone,” said Oguz Akbilgic, Ph.D., corresponding author and professor of artificial intelligence in the Department of Cardiovascular Medicine at Wake Forest University School of Medicine. The single-lead configuration is similar to the measurement captured by some smartwatches and wearable ECG devices. However, the researchers emphasized that the study did not test data collected directly from consumer wearables. The results indicate technical potential, not a currently validated smartwatch diagnostic.
Researchers trained the system using more than 1 million ECGs collected at Atrium Health Wake Forest Baptist. Training at this scale allowed the model to encounter a broad range of electrical patterns and clinical presentations. The investigators then evaluated it on a separate dataset containing more than 72,000 ECGs from the University of Tennessee Health Science Center. This external testing was important because an algorithm can perform well in the hospital where it was developed but lose accuracy when applied to patients from another institution, region or demographic background.
The researchers created two versions of the model. One analyzed the complete 12-lead ECG routinely used in clinical medicine, while the other used only a single lead. Both systems classified recordings into four categories: rEF, mEF, HFpEF or no detected dysfunction. The 12-lead model was especially effective at separating patients with reduced ejection fraction from those without it. Its performance was somewhat lower for mildly reduced ejection fraction and HFpEF, although the investigators described the results as potentially useful for clinical screening and decision support.
The single-lead system performed nearly as well as the 12-lead version, suggesting that much of the relevant information may be contained in a limited portion of the heart’s electrical signal. That result is significant because single-lead recordings can be collected more easily and inexpensively than conventional diagnostic ECGs. The model also showed strong performance in pediatric patients when identifying reduced ejection fraction, matching or exceeding earlier models, although the pediatric group was relatively small. In addition, the researchers reported that performance generalized well across different demographic populations.
The technology is now being piloted in a family medicine clinic at Atrium Health Wake Forest Baptist. This real-world evaluation will examine whether the AI can help clinicians identify patients who should receive additional heart-failure testing, as well as how its use affects clinical decisions and health-care resources. “Some of these conditions can progress without noticeable symptoms and may not be found until they become more severe,” Akbilgic said. “Our model helps fill that gap by identifying electrical patterns in the heart that humans can’t easily see.” The study was partially funded by the National Heart, Lung, and Blood Institute of the National Institutes of Health, and the authors reported no disclosures.
News Publication Date: August 6, 2026
Web References: Wake Forest University School of Medicine: https://school.wakehealth.edu/ ; Journal of the American Heart Association article: https://www.ahajournals.org/doi/10.1161/JAHA.124.041948
References: Journal of the American Heart Association, DOI: 10.1161/JAHA.124.041948
