A newly developed artificial intelligence (AI) model designed to identify blocked coronary arteries through electrocardiogram (ECG) readings has shown superior performance compared to expert clinicians and performed comparably to troponin T testing. These results were shared during the American College of Cardiology’s Annual Scientific Session (ACC.25). The open-source tool, freely available to clinicians, holds promise for accelerating the identification of patients needing urgent intervention for heart attacks.

While several AI solutions have previously targeted similar diagnostic goals, this particular model is the first crafted for widespread application across typical emergency department patients—not just high-risk, pre-selected groups. Researchers anticipate that its primary impact will be in diagnosing non-ST segment elevation myocardial infarction (NSTEMI), which remains the most challenging heart attack type to detect.

“ECG analysis in the emergency department often has high variability,” said Dr. Antonius Büscher, clinician scientist at University Hospital Münster in Germany and first author of the study.

“Whereas patients with ST segment elevation have changes in the ECG that are quite distinct, there is a large number of patients who need urgent coronary revascularization who do not have ST segment elevation; because the alterations on the ECG are not as clear-cut, those patients can experience treatment delays. Our goal was to accelerate this process to identify patients who might need revascularization earlier.”

The research focused on type 1 heart attacks—both STEMI and NSTEMI—which result from plaque buildup that ruptures within coronary arteries, impeding blood flow to the heart. The severity of damage is determined by the extent of oxygen deprivation and the speed of coronary revascularization, a procedure that reopens the blocked vessels. The two primary tools used to identify heart attacks include ECGs, which are quick and can detect electrical changes in the heart, and troponin T blood tests, which are more sensitive but may yield positive results from other medical conditions as well, leading to confusion in diagnosis. Emergency departments, facing high volumes of patients with chest pain or heart-related symptoms, need supplementary tools to better triage cases requiring immediate revascularization.

To address this, researchers created a deep learning algorithm capable of interpreting ECG patterns that signal an urgent need for revascularization. They trained it using data from nearly 145,000 emergency department visits in a U.S.-based center and evaluated its accuracy on an additional 35,000 visits from the same location. The model then underwent external validation on more than 18,000 patient records from a German emergency department. Validation covered both the initial training target—coronary revascularization—and the more clinically relevant outcome: diagnosis of a type 1 heart attack. The AI categorized patients into low, intermediate, or high risk groups based solely on their ECG results.

Performance was measured using the area under the receiver operating characteristic curve (AUC), a standard metric where 1.0 indicates perfect discrimination and 0.5 indicates no better than chance. The AI model reached an AUC of 0.91 in internal testing, outperforming human clinicians (0.65) and conventional troponin testing (0.71). In the external cohort, the model achieved an AUC of 0.85 for identifying type 1 heart attacks and 0.81 for predicting revascularization. Its overall diagnostic accuracy exceeded that of clinicians (0.74 and 0.70), and it matched the accuracy of high-sensitivity troponin T (0.87), which had greater sensitivity but lower specificity.

“We saw that performance of the model was better than human interpretation and nearly matched the accuracy of high-sensitivity troponin testing,” Büscher noted. “While the model is not intended to serve as a standalone diagnostic test, it could greatly enhance ECG interpretation in the emergency department and reduce uncertainty during the emergency department workup.”

During validation, the team found several cases in which heart attacks had initially gone undiagnosed. Büscher pointed out that the model would have flagged these cases hours earlier, indicating its potential to serve as an early alert system for clinicians or to automatically prompt troponin testing.

One distinguishing feature of this model is its transparency—unlike many AI systems, it offers self-explainability. Researchers could trace which ECG features the model relied on, and connect them with established clinical markers. This positions the model as part of a larger movement toward AI-supported decision-making in clinical practice.

“A common goal with these types of models is to make most use of the data that we have, and to complement physicians’ clinical reasoning by picking up some nuances in the data that might be obscure to the human eye,” Büscher said. “I think we are still at the beginning in this field, but in the not too far future, these types of models will be used as routine diagnostic tests.”

Although performance may decline slightly when applied in populations different from those it was trained on, Büscher expressed confidence that the tool will be suitable in most developed healthcare settings. Since the model is openly accessible, institutions can adapt and optimize it using their own patient data for better local performance.

A prospective study is currently being planned to evaluate the model's effectiveness and practical benefits in real-world clinical settings.

The study was supported by the Interdisciplinary Centre for Clinical Research at University Münster and published online in the European Heart Journal at the time of the conference.

Innovations In Cardiology

Discover groundbreaking advancements that are transforming the future of heart care.