Comparative Diagnostic Accuracy of AI-Enhanced Mobile Limb-Lead ECG versus Cardiologist-Interpreted Standard 12-Lead ECG for Acute Myocardial Infarction

Abstract Background Deep learning models developed on standard 12-lead electrocardiograms (ECGs) can detect acute myocardial infarction (AMI), but their performance on limb-lead ECGs from mobile devices is uncertain. We evaluated an artificial intelligence (AI) model applied to mobile limb-lead ECGs for AMI detection, using cardiologist interpretation of truncated limb-lead and 12-lead ECGs as comparators. Methods We performed an analysis of a prospective observational cohort of 153 patients referred for coronary angiography at a Korean tertiary center (September 2022–October 2023). Each patient underwent mobile limb-lead and standard 12-lead ECG acquisition before revascularization. The model was applied without retraining to the mobile ECGs at two thresholds prespecified during model development (high sensitivity, high specificity), and three cardiologists independently interpreted truncated limb-lead and 12-lead ECGs. AMI was diagnosed according to the Fourth Universal Definition of Myocardial Infarction, with coronary angiography in all patients. Results Twenty-five patients (16.3%) had AMI, including 3 ST-elevation and 22 non-ST-elevation myocardial infarctions. The model achieved an area under the receiver operating characteristic curve of 0.843 (95% confidence interval, 0.758–0.928). At the high-sensitivity threshold, the model showed higher sensitivity than cardiologists interpreting truncated limb-lead and standard 12-lead ECGs (88.0% vs. 20.0% and 40.0%; both p < 0.001) and numerically higher negative predictive value (NPV; 95.9% vs. 85.5% and 88.6%), but lower specificity (55.5% vs. 92.2% and 91.4%; both p < 0.001) and positive predictive value (PPV; 27.8% vs. 33.3% and 47.6%). At the high-specificity threshold, specificity was similar to that of cardiologists interpreting standard 12-lead ECGs (93.8% vs. 91.4%; p = 0.58), and sensitivity did not differ significantly (48.0% vs. 40.0%; p = 0.77). Conclusions In this cohort, an AI-ECG model developed on standard ECGs discriminated AMI when applied to mobile limb-lead ECGs without retraining. At the high-sensitivity threshold, it showed higher sensitivity and numerically higher NPV but lower specificity and PPV than cardiologists’ ECG interpretation. These findings support its potential as an adjunctive screening aid where 12-lead ECG acquisition or expert interpretation is delayed or unavailable.