An AI model using 12-lead ECG detected STEMI with AUC 0.97, NSTEMI with AUC 0.80, and OMI with AUC 0.90, showing higher accuracy for NSTEMI type 1 than type 2.
Does an AI-powered 12-lead ECG-only model accurately diagnose acute myocardial infarction and its subtypes in patients presenting to the emergency department?
A deep-learning model based solely on 12-lead ECGs demonstrated very high diagnostic accuracy for STEMI and OMI, and high accuracy for NSTEMI, potentially aiding early triage of chest pain patients.
Abstract Background Accurate and timely diagnosis of acute myocardial infarction (AMI) remains a challenge in clinical practice. While the 12-lead electrocardiogram (ECG) is an essential tool for identifying AMI, manual interpretation by healthcare professionals is skill-dependent and only identifies a minority of patients with clear signs of acute ischemia requiring urgent intervention. Automated ECG analysis using artificial intelligence has the potential to overcome these limitations and enhance patient care. Purpose To train and validate an AI-powered 12-lead ECG-only model for the detection of AMI and its subtypes on three large and high-quality datasets. Methods A convolutional neural network was trained on digital 12-lead ECG data of hospitalized patients (n=178’682, from 10/2021 to 09/2024, Figure 1) with discharge diagnoses as labels. Utilizing transfer-learning, the model was fine-tuned and internally validated on a 80% (n=6’721) / 20% (n=1’645) split of a prospective single-center cohort of adult chest-pain patients presenting to the Emergency Department (ED) (01/2019-01/2022). External validation was performed in a large prospective international multicenter study (04/2006-06/2018) of patients presenting to the ED with suspected AMI (n=3’839). Central adjudication of the final diagnosis (including AMI subtypes) was performed by two independent cardiologists using all clinical information and serial cardiac troponin concentrations according to the Fourth Universal Definition of Myocardial Infarction. The primary outcome was a diagnosis of ST-segment Elevation Myocardial Infarction (STEMI) and Non-STEMI (NSTEMI). Secondary outcomes included the differentiation between NSTEMI type 1 (due to atherothrombosis) and type 2 (due to oxygen supply-demand mismatch) and a diagnosis of Occlusion Myocardial Infarction (OMI). Results Internal validation showed good performance with an AUC of 0.96 (95%-CI 0.94-0.97) for STEMI and 0.82 (95%-CI 0.79-0.85) for NSTEMI. Results were similar in the external validation cohort, with an AUC of 0.97 (95%-CI 0.96-0.98) for STEMI and 0.80 (95%-CI 0.78-0.82) for NSTEMI (Figure 2). The model showed higher discrimination for NSTEMI type 1 than type 2 in both internal (type 1: AUC 0.82, 95%-CI 0.79-0.85, type 2: AUC 0.77, 95%-CI 0.70-0.83) and external validation (type 1: AUC 0.80, 95%-CI 0.78-0.83, type 2: AUC 0.69, 95%-CI 0.64-0.74). Secondary analysis revealed a very high diagnostic accuracy for OMI with an AUC of 0.90 (95%-CI 0.88-0.92). Overall calibration was good for STEMI (intercept -1.21, slope 1.1), NSTEMI (intercept 0.43, slope 0.85) and OMI (intercept 0.02, slope 0.85). Conclusion Our model showed very high diagnostic accuracy for STEMI and OMI and high accuracy for NSTEMI. Based on 12-lead ECG data only, the model more accurately identified NSTEMI type 1 compared to NSTEMI type 2. Whether care guided by our model can improve the early diagnosis of AMI requires prospective evaluation.
Zimmermann et al. (2025) studied this question. An AI model using 12-lead ECG detected STEMI with AUC 0.97, NSTEMI with AUC 0.80, and OMI with AUC 0.90, showing higher accuracy for NSTEMI type 1 than type 2.