Artificial Intelligence for Myocardial Infarction Detection via Electrocardiogram: A Scoping Review

Background/Objectives: Acute myocardial infarction (MI) is a major cause of death worldwide, and it imposes a heavy burden on health care systems. Although diagnostic methods have improved, detecting the disease early and accurately is still difficult. Recently, AI has demonstrated increasing capability in improving ECG-based MI detection. From this perspective, this scoping review aimed to systematically map and evaluate AI applications for detecting MI through ECG data. Methods: A systematic search was performed in Ovid MEDLINE, Ovid Embase, Web of Science Core Collection, and Cochrane Central. The search covered publications from 2015 to 9 October 2024; non-English articles were included if a reliable translation was available. Studies that used AI to diagnose MI via ECG were eligible, and studies that used other diagnostic modalities were excluded. The review was performed per the PRISMA extension for scoping reviews (PRISMA-ScR) to ensure transparent and methodological reporting. Of a total of 7189 articles, 220 were selected for inclusion. Data extraction included parameters such as first author, year, country, AI model type, algorithm, ECG data type, accuracy, and AUC to ensure all relevant information was captured. Results: Publications began in 2015 with a peak in 2022. Most studies used 12-lead ECGs; the Physikalisch-Technische Bundesanstalt database and other public and single-center datasets were the most common sources. Convolutional neural networks and support vector machines predominated. While many reports described high apparent performance, these estimates frequently came from relatively small, single-source datasets and validation strategies prone to optimism. Cross-validation was reported in 57% of studies, whereas 36% did not specify their split method, and several noted that accuracy declined under inter-patient or external validation, indicating limited generalizability. Accordingly, headline figures (sometimes ≥99% for accuracy, sensitivity, or specificity) should be interpreted in light of dataset size, case mix, and validation design, with risks of spectrum/selection bias, overfitting, and potential data leakage when patient-level independence is not enforced. Conclusions: AI-based approaches for MI detection using ECGs have grown quickly. Diagnostic performance is limited by dataset and validation issues. Variability in reporting, datasets, and validation strategies have been noted, and standardization is needed. Future work should address clinical integration, explainability, and algorithmic fairness for safe and equitable deployment.