Training and validation of a 12-lead ECG-based deep-learning model for myocardial infarction subtypes

Abstract A convolutional neural network was developed to detect acute myocardial infarction (AMI) subtypes from digital 12-lead ECGs. Trained on 173,396 hospitalized patients’ ECGs, the model underwent fine-tuning and internal validation in 7591 patients and external validation in 4370 patients with suspected AMI. Both prospective validation studies employed central diagnostic adjudication. For non-ST-segment elevation MI (NSTEMI), the model achieved an area under the receiver operating characteristic curve (AUROC) of 0.81 [95%-confidence interval (CI) 0.79–0.83], with NSTEMI type 1 at 0.82 [0.80–0.84] versus type 2 at 0.75 [0.71–0.79]. Performance was superior in younger patients without prior heart disease. For ST-segment elevation MI (STEMI), the model outperformed physician interpretation (AUROC 0.96 [95%-CI 0.94–0.98] vs. 0.89 [0.85–0.93], p < 0.001). Occlusion MI detection reached AUROC 0.91 [95%-CI 0.89–0.93], though NSTEMI-OMI cases were frequently missed. Overall calibration was good for all MI subtypes. This ECG-based deep learning model demonstrates good discrimination and calibration across AMI subtypes, indicating potential clinical utility for rapid risk stratification and early cardiology intervention.