Artificial intelligence-enhanced electrocardiography for acute myocardial infarction detection: a systematic review

Artificial intelligence (AI) shows promise for improving electrocardiogram (ECG)-based acute myocardial infarction (AMI) detection, but clinical readiness remains uncertain. We aimed to conduct a systematic synthesis informed by a structured search of AI-enhanced ECG systems to evaluate their performance, validation quality, and implementation readiness.

We conducted a structured search of PubMed and Embase (publication date limits: January 1, 2017, to August 31, 2025; last searched: February 18, 2026) for English-language human studies that developed or validated AI models for ECG-based AMI diagnosis, extracting data on architectures, clinical applications, validation approaches, and performance metrics; we excluded non-original publications and studies that were non-English or non-human studies or did not use ECG input or did not address AMI diagnosis.

We included 88 studies; the total number of participants was not estimable because sample sizes were inconsistently reported across studies. Among 88 identified studies, convolutional neural networks predominated (51/88, 58%). Most studies were retrospective (80/88, 91%) and used 12-lead ECG (64/88, 73%). Reported performance varied widely [area under the receiver operating characteristic curve (AUROC), 0.700-0.999; sensitivity, 67.7-100.0%; specificity, 73.3-100.0%], with promising results for detecting subtle ischemic patterns in non-ST-elevation and occlusion myocardial infarction (OMI). However, only 33/88 (37.5%) performed external validation. Public datasets were used in 50/88 (57%) and institutional patient cohorts in 43/88 (49%); several studies used both sources. More complex architectures did not consistently demonstrate superior accuracy, though heterogeneity in study designs limits definitive conclusions.

AI demonstrates substantial technical potential for AMI detection, particularly for subtle ischemic patterns. However, critical gaps impede clinical deployment: insufficient external validation, reliance on curated datasets with limited generalizability, absence of standardized evaluation frameworks, and insufficient evidence on patient-centered outcomes. Future research must prioritize prospective multicenter validation, standardized benchmarks with rigorous reference standards, and real-world implementation studies examining clinical outcomes and workflow integration. Technical feasibility is established; clinical impact now depends on validation rigor and pragmatic deployment.