AI model helps diagnose often undetected heart disease from simple EKG

Doctors may soon be able to diagnose an elusive form of heart disease within seconds by using an AI model developed at University of Michigan, according to a recent study.

Researchers trained the model to detect coronary microvascular dysfunction, a complex condition that requires advanced imaging techniques to diagnose, using a common electrocardiogram.

Their prediction tool significantly outperformed earlier AI models in nearly every diagnostic task, including predicting myocardial flow reserve, the gold standard for diagnosing CMVD.

The results are published in NEJM AI, a monthly journal from the New England Journal of Medicine family.

“Our model creates a way for clinicians to accurately identify a condition that is notoriously hard to diagnose — and often missed in emergency department visits — using a 10-second EKG strip,” said senior-author Venkatesh L. Murthy, M.D., Ph.D., associate chief of cardiology for translational research and innovation at the U-M Health Frankel Cardiovascular Center and the Melvyn Rubenfire Professor of Preventive Cardiology at U-M Medical School.

Around 14 million people visit either the ER or an outpatient clinic each year for chest pain.

Unlike coronary artery disease, which occurs due to a blockage in the heart’s large blood vessels, CMVD affects the tinier vessels.

It also causes chest pain and increases the risk of heart attack, but diagnosing CMVD requires advanced methods such as PET myocardial perfusion imaging.

How the AI model works

These high value scans are both expensive and rarely accessible outside of specialty centers.

The limited available scans posed a challenge for Murthy and his research team as they looked for data on which to train their AI model.

They solved this problem with self-supervised learning, or SSL.

They began by pre-training a deep learning model called a vision transformer on more than 800,000 unlabeled EKG waveforms and fine-tuned it on a smaller, labeled dataset of PET scans.

“Essentially, we taught the model to ‘understand’ the electrical language of the heart without human supervision,” Murthy said.

Once trained on the basics, researchers taught the model to accurately break down advanced PET data using 12 different demographic and clinical prediction tasks, including several that are not possible using current EKG-AI models.

The model not only succeeded at predicting CMVD across different databases, but it consistently improved diagnostic accuracy of prediction tasks for more common cardiac conditions compared to previous state-of-the-art models.

Four of the diagnostic tasks the model uses often involve electrocardiograms taken during exercise stress tests.

However, the results showed only a minimal increase in performance when using stress EKGs compared to resting EKGs.

Future of cardiac AI

Several groups have successfully developed AI tools to interpret EKGs by training them on large EKG databases.

Those models, however, are used for more general tasks, such as automatic interpretation of heart rhythm and detection of left ventricular systolic dysfunction.

By using the less accessible “gold standard” data from PET MPI scans to train its model, Murthy’s team believes it can extend an EKG’s ability to predict a tougher-to-spot microvascular condition like CMVD.

“People who come to the ER for chest pain might have CMVD, but their angiogram will show up as ‘clear,’” said co-author Sascha N. Goonewardena, M.D., associate professor of internal medicine-cardiology at U-M Medical School.

“In hospitals with limited resources or non-specialty centers, using our EKG-AI model to predict myocardial flow reserve and CMVD will be an easy, cost-effective and non-invasive way to identify when a patient would benefit from advanced testing for a serious condition.”

Additional authors: Jonathan B. Moody, Ph.D., Alexis Poitrasson- Rivière, Ph.D., Jennifer M. Renaud, Ph.D., Tomoe Hagio, Ph.D., Michael D. Vanderver, and Edward P. Ficaro, Ph.D., all of INVIA, Fares Alahdab M.S., of University of Missouri, and Mouaz H. Al-Mallah, M.D., M.S., of Houston Methodist DeBakey Heart and Vascular Center.

Funding/disclosures: This study was partially supported by the National Heart, Lung, and Blood Institute (HL150392, HL178315, HL136685) the National Institute of Diabetes and Digestive and Kidney Diseases (DK123013) and the National Institute on Aging (AG059729) of the National Institutes of Health.

The study also received partial support from the Department of Veterans Affairs MERIT grant (1I01CX002560), the Taubman Medical Research Institute Wolfe Scholarship and the Melvyn Rubenfire Professorship in Preventive Cardiology.

The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH or the Department of Veterans Affairs.

Murthy is an associate editor for NEJM AI but was recused from his editorial duties with respect to this work and had no role in the acceptance for publication, peer review, or other editorial processes related to this paper.

Murthy has a paid relationship with INVIA.

Paper cited: “A Foundation Transformer Model with Self-Supervised Learning for ECG-Based Assessment of Cardiac and Coronary Function,” NEJM AI. DOI: 10.1056/AIoa2500164

Sign up for Health Lab newsletters today. Get medical tips from top experts and learn about new scientific discoveries every week.

Sign up for the Health Lab Podcast. Add us wherever you listen to your favorite shows.