Deep Learning for Myocardial Infarction Detection Using Electrocardiogram Images: A Systematic Review

Myocardial infarction is a leading cause of global mortality. Electrocardiograms (ECG) are the standard diagnostic tool; however, even among medical experts, the accuracy of ECG-supported diagnoses varies. To improve diagnostic accuracy, deep learning has emerged as a promising approach. Although numerous studies demonstrate its potential, the field lacks a unified characterization of state-of-the-art methods using ECG images. This systematic review, following PRISMA guidelines, addresses this gap by analyzing studies presenting deep learning models trained on ECG images for myocardial infarction detection. Searches across six scientific databases yielded 361 records, which were filtered to 47 articles using inclusion, exclusion, and quality criteria. Guided by 12 research questions, this review contributes (i) a characterization of deep learning architectures for myocardial infarction detection using ECG images; (ii) an assessment of training and evaluation practices of deep learning models; (iii) a description of state-of-the-art results in terms of machine learning metrics; (iv) a pioneering exploration of the use of vision transformers for myocardial infarction detection; (v) a compilation of ECG databases; and (vi) future research directions aimed at advancing deep learning approaches for myocardial infarction detection, e.g., involving domain experts in evaluating deep learning models to guarantee their safe deployment in clinical settings.