AI enables rapid detection of coronary microvascular dysfunction from standard ECGs

Using an AI model developed at the University of Michigan, doctors may soon be able to diagnose elusive types of heart disease within seconds, according to a recent study.

The researchers trained a model to use common electrocardiograms to detect coronary microvascular dysfunction, a complex condition that requires advanced imaging techniques to diagnose.

Their predictive tool significantly outperformed previous AI models in nearly every diagnostic task, including predicting myocardial flow reserve, the gold standard for CMVD diagnosis.

The result is NEJM AIThis is a monthly magazine. New England Medical Journal family.

Our model creates a way for clinicians to use 10-second ECG strips to accurately identify symptoms that are notoriously difficult to diagnose and often overlooked in the emergency department. ”

Venkatesh L. Murthy, MD, Ph.D., senior author, Associate Chief of Cardiology for Translational Research and Innovation at UM Health Frankel Cardiovascular Center and Melvin Rubenfire Professor of Preventive Cardiology at UM Medical School

Approximately 14 million people visit the ER or outpatient clinic each year for chest pain.

Unlike coronary artery disease, which occurs due to blockage of large blood vessels in the heart, CMVD affects smaller blood 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 AI models work

These high-value scans are expensive and rarely accessible outside of specialized centers.

When Murthy and his research team looked for data to train their AI models, the limited number of scans available was a challenge.

They solved this problem using self-supervised learning (SSL).

They started by pretraining a deep learning model called Vision Transformer on more than 800,000 unlabeled electrocardiogram waveforms and fine-tuned it on a small labeled PET scan dataset.

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

After training on the basics, researchers taught the model to accurately analyze advanced PET data using 12 different demographic and clinical prediction tasks, including tasks not possible with current ECG AI models.

The model not only successfully predicted CMVD across a variety of databases, but also consistently improved diagnostic accuracy for more common heart disease prediction tasks compared to previous state-of-the-art models.

The four diagnostic tasks used by the model often include an electrocardiogram obtained during an exercise stress test.

However, the results showed only minimal improvement in performance using stress ECG compared to resting ECG.

The future of cardiac AI

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

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

Murthy's team believes that by using hard-to-access “gold standard” data from PET MPI scans to train models, they can extend the electrocardiogram's ability to predict hard-to-detect microvascular diseases like CMVD.

“People who come to the ER with chest pain may have CMVD, but it shows up as 'clear' on an angiogram,” said co-author Sasha N. Goonwardena, MD, associate professor of internal medicine and cardiology at the UM School of Medicine.

“For hospitals with limited resources or non-specialty centers, using our ECG AI model to predict myocardial perfusion reserve and CMVD is an easy, cost-effective, non-invasive way to identify when patients would benefit from advanced testing for critical conditions.”