By Michael H. Crawford, MD, Editor
Synopsis: A large Korean study of an artificial intelligence electrocardiogram (ECG) interpretation algorithm for identifying patients with acute myocardial infarction showed a high degree of accuracy for diagnosing acute myocardial infarction and identifying patients at risk for 30-day major adverse cardiac events in an emergency department setting, which was similar or superior to standard risk stratification methods.
Source: Lee MS, Shin TG, Lee Y, et al. Artificial intelligence applied to electrocardiogram to rule out acute myocardial infarction: The ROMIAE multicentre study. Eur Heart J. 2025;46(20):1917-1929.
The electrocardiogram (ECG) is a critical tool deployed early after the arrival of patients with symptoms and signs suggestive of acute myocardial infarction (AMI), but accurate interpretation of these ECGs can be a challenge, even with well-trained physicians. Thus, these investigators from South Korea conducted a multicenter study to validate the predictive performance of an artificial intelligence (AI)-enhanced 12-lead ECG analysis tool (AI-ECG) compared to two existing AMI risk stratification models: the History, ECG, Age, Risk factors and Troponin (HEART) and Global Registry of Acute Coronary Events 2.0 (GRACE-2) scores.
In the initial retrospective validation study, the AI-ECG had an area under the receiver operating characteristic (AUROC) curve was 0.95 for ST-elevation MI (STEMI) and 0.90 for non-STEMI. For the prospective observational study, the emergency department (ED) physicians were blinded to the AI-ECG results. Adult patients presenting to the 18 study EDs < 24 hours after the onset of chest pain were included in ROMIAE. Excluded were patients with out-of-hospital cardiac arrest, traumatic chest pain, and those with chest pain clearly distinct from AMI, such as pneumothorax.
The primary outcome was type 1 or 2 AMI diagnosed during the index hospitalization. The secondary outcome was major adverse cardiovascular events (MACE) defined as death, MI, stroke, target vessel revascularization, or stent thrombosis within 30 days. Between 2022 and 2023, a total of 8,493 patients were enrolled in the study (median age 62 years, 65% men); 19% were diagnosed with AMI (95% type 1), of whom 40% had a STEMI and 60% had a non-STEMI. The AUC for the diagnosis of AMI by AI-ECG was 0.88, which was the same as the HEART score but superior to the GRACE-2 score of 0.71. The AUC for 30-day MACE was 0.87 for AI-ECG, 0.86 for the HEART score, and 0.71 for the GRACE-2 score.
When AI-ECG was added to the HEART score, there was a net reclassification improvement of 20% compared to the HEART score alone. The authors concluded that the diagnostic accuracy for AMI and predictive power for 30-day MACE of AI-ECG was similar or superior to traditional risk stratification methods, thus supporting the potential for the use of AI-ECG for early AMI detection.
Commentary
The missed diagnosis of AMI is an important ED issue that occurs approximately 1% to 2% of the time. An acceptable level is considered 1% or less. At this time, AI-ECG is the only diagnostic model that meets this threshold (sensitivity 99.6% and negative predictive value 99.1%). Considering that the HEART score includes troponin, this is remarkable. Also, waiting for troponin values to come back from the laboratory can take up to an hour. AI-ECG has the potential for earlier diagnosis, even in the pre-hospital phase, since many ambulances have ECG capabilities. In addition, in settings where physicians experienced in ECG reading are not available, AI-ECG would be of particular value. Not only would AMI be diagnosed early, but low-risk patients who could be discharged would be identified.
Prior studies of AI-ECG have focused on comparing it to ECG device diagnostics and physician interpretations retrospectively, whereas ROMIAE evaluated AI-ECG as a diagnostic tool in a clinical setting and as a prognosticator of future outcomes. Although powerful overall in detecting AMI, subgroup analyses identified groups in which it was less accurate: patients older than 65 years of age, obese patients, and patients with chronic illnesses. Also, it was less accurate if left bundle branch block, left ventricular hypertrophy, atrial fibrillation, or paced rhythm was present on the ECG.
ROMIAE has limitations to consider beyond these subgroups. The study population was Korean, so the results might not apply to other groups. Only short-term outcomes were evaluated. There was no assessment of the clinical effect of deploying AI-ECG, especially regarding ED workflow and efficiency. Comparing AI-ECG to the GRACE-2 score is akin to setting up a “straw man,” since it was initially designed to estimate in-hospital and six-month mortality for hospitalized acute coronary syndrome patients, not for diagnosing AMI. Finally, some may find it difficult to accept the decisions of a black box when we do not fully understand how it works.
Michael H. Crawford, MD, is Professor Emeritus of Medicine and Consulting Cardiologist, UCSF Health, San Francisco.
A large Korean study of an artificial intelligence electrocardiogram (ECG) interpretation algorithm for identifying patients with acute myocardial infarction showed a high degree of accuracy for diagnosing acute myocardial infarction and identifying patients at risk for 30-day major adverse cardiac events in an emergency department setting, which was similar or superior to standard risk stratification methods.
You have reached your article limit for the month. Subscribe now to access this article plus other member-only content.
- Award-winning Medical Content
- Latest Advances & Development in Medicine
- Unbiased Content