Loading ...

CHICAGO — A new artificial intelligence (AI) model performed about as well as high-sensitivity troponin for the detection of non–ST-elevated myocardial infarction (NSTEMI) on an ECG, but better than clinicians.

“Essentially, the model has rediscovered features on the ECG that we, as clinicians, already recognize as ischemic, but it qualifies these features more precisely and correlates them more effectively with clinical outcomes,” said Antonius Büscher, MD, from the Institute of Medical Informatics at the University of Münster in Münster, Germany.

If the performance is validated in randomized trials, AI should be able to select candidates for revascularization more quickly, substantially reducing the door-to-needle time, Büscher explained during his presentation at a late-breaker session at the American College of Cardiology (ACC) Scientific Session 2025. The study was published simultaneously in the European Heart Journal.

AI Is 50% More Accurate Than Clinicians

For the study, accuracy was evaluated with a receiver-operating characteristic curve on two separate sets of data after training. In the 35,995-patient test cohort from the Beth Israel Deaconess Medical Center in Boston, the receiver-operating characteristic for predicting revascularization during the index admission were 0.91 for AI, 0.71 for conventional troponin, and 0.65 for clinicians.

In an external validation conducted with an 18,673-patient test cohort from Büscher’s center, the receiver-operating characteristics were 0.85 for high-sensitivity troponin, 0.81 for AI, and 0.74 for clinicians.

The lower AI performance in the external cohort “was normal and expected,” according to Büscher, given that the European data involved a different population managed in a different healthcare system with varying clinical practices, and the ECG machines were not the same as the ones at Beth Israel Deaconess where the AI was trained.

Although high-sensitivity troponin was superior to AI in the European cohort, there is an inevitable loss of accuracy when using AI to evaluate a population different than the one it was trained on.

When the ability to identify type 1 MI on an ECG was tested in the external cohort, the AI model also outperformed clinicians, suggesting it “genuinely learned to recognize the features that are associated with MI,” Büscher said.

AI Appears to Read ECG Better for Ischemia

The exact criteria used by the AI neural network to make decisions is generally considered a black box, so Büscher and his coinvestigators used a heatmap to evaluate the relative importance of each of the 12 ECG leads.

“We confirmed that it relied on well-known ischemic ECG changes, like ST-segment and T-wave alterations,” Büscher explained. In other words, the AI model appears to evaluate the same types of changes as clinicians, but in far greater detail.

This confirmation is reassuring, said Thomas Maddox, MD, professor of medicine at the Washington University School of Medicine in St. Louis, who was the ACC-invited discussant for the session.

“One thing that you have done very well — and that we should pay attention to and probably demand from other analyses of AI approaches like this — is explainability,” Maddox said. “We know that machine learning and deep neural networks, like the one you deployed here,” get into details so minute — in this case on an ECG — that they cannot be appreciated with the human eye.

Because of the black-box nature of AI, “some degree of faith” has been needed in the past to believe that “what the machine is telling us represents the biologic reality,” Maddox said. The evidence that AI is reading the ECG in more detail than a clinician provides “comfort” in the validity, “relative to what we understand clinically.”

If this AI model is validated in the randomized trials that are now expected, it will meet a major unmet need, according to Büscher and Maddox. Unlike the diagnosis of ST-elevated MI, which is “straightforward,” the diagnosis of NSTEMI continues to be a realm of “uncertainty,” Büscher said.

To augment an ECG finding, troponin and other biomarkers are often used, but these require time to process and involve the risk for false-positive findings, he explained. AI will likely increase confidence and shorten the diagnostic process, getting patients with NSTEMI to the cath lab more quickly.

The training of AI was based on historical data, and included diagnostic decisions that are “potentially biased and might not necessarily reflect clinical outcomes in all patients,” Büscher acknowledged. Although this might have been mitigated by the large dataset, there are plans to refine the model and perform further validation.

The AI used in this study is an open-source system, which “allows the scientific community to adapt it and refine it for specific patient populations and clinical settings,” Büscher said.

Büscher and Maddox reported no conflicts of interest.