Robust diagnostic performance of artificial intelligence–enabled electrocardiography across heterogeneous NSTEMI presentations: a prospective multicenter analysis
H S Lee, S Kang, J M Kwon, J M Son, K S Kim, K Kim, M S LeeAbstract
Background
Non–ST-segment elevation myocardial infarction (NSTEMI) encompasses a heterogeneous spectrum of clinical and electrocardiographic presentations. Diagnostic uncertainty is particularly pronounced in early presenters and in patients without overt ischemic ECG changes. Whether artificial intelligence–enabled electrocardiography (AI-ECG) maintains stable performance across this diversity remains unclear.
Purpose
To evaluate the robustness of AI-ECG across diverse NSTEMI phenotypes and to explore score distribution according to occlusion status.
Methods
In the prospective, multicenter ROMIAE cohort of 8,493 emergency department patients with suspected acute myocardial infarction, NSTEMI cases were analyzed. AI-ECG generated continuous probability scores from initial 12-lead ECGs using prespecified thresholds.1 Discrimination was assessed using AUROC. Prespecified subgroup analyses examined performance across age, comorbidities (hypertension, diabetes, coronary artery disease, prior myocardial infarction, heart failure), ischemic ECG features, rhythm abnormalities (atrial fibrillation, bundle branch block), and early presentation (≤3 hours). AI-ECG score distributions were compared between occlusive and non-occlusive NSTEMI.
Results
AI-ECG demonstrated stable discrimination for NSTEMI (AUROC ~0.80 range) without systematic degradation across clinical or ECG-defined subgroups. Performance remained consistent in early presenters, in patients without overt ischemic ECG features, and in those with rhythm abnormalities. (Figure 1) AI-ECG scores showed a graded distribution within NSTEMI phenotypes, with higher median values observed in occlusive compared with non-occlusive NSTEMI. (Figure 2) This gradient was evident despite absence of ST-segment elevation, suggesting that AI-ECG captures ischemic signal patterns beyond conventional criteria.
Conclusion
In a prospective multicenter cohort, AI-ECG maintained robust diagnostic performance across heterogeneous NSTEMI presentations and demonstrated graded signal distribution according to occlusion status. These findings support AI-ECG as a time-zero triage adjunct capable of reducing diagnostic uncertainty in NSTEMI while reflecting underlying ischemic burden.Figure 1Figure 2