DEEP LEARNING FOR ECG INTERPRETATION: A SYSTEMATIC REVIEW OF MODEL ARCHITECTURES, TRANSFER LEARNING, AND POPULATION BIAS IN AFRICAN CONTEXTS

Abstract

Automated ECG interpretation has been substantially advanced by deep learning, with several contemporary models now approaching the diagnostic accuracy of trained cardiologists. Despite this progress, a critical gap remains: virtually all existing models have been developed using data drawn from European, North American, or Asian patient populations. When these systems are deployed in African clinical settings, questions of generalizability and embedded population bias carry urgent clinical consequences. This systematic review pursued three aims: to compare the performance of major deep learning architectures on ECG interpretation tasks; to evaluate the practical utility of transfer learning in data-scarce environments; and to examine how thoroughly the existing literature accounts for population-level demographic differences, with particular attention to African patients. The review adhered to PRISMA 2020 guidelines. Searches were conducted across five databases: PubMed/MEDLINE, IEEE Xplore, Scopus, Web of Science, and Google Scholar, covering peer-reviewed work published between 2020 and 2025. From an initial pool of 1,847 records, 42 studies met all inclusion criteria following rigorous two-stage screening. Eligible studies were required to apply deep learning to ECG classification and report quantitative performance metrics. Extracted data encompassed model architecture, dataset origin and composition, preprocessing pipelines, validation methodology, and demographic reporting practices. Study quality was evaluated using a modified QUADAS-2 framework. Among the architectures examined, classical convolutional neural networks (CNNs) were most frequently used, followed by hybrid CNN-RNN models, ResNet variants, transfer learning pipelines, and attention-based designs. ResNet-based architectures consistently achieved the strongest performance on benchmark datasets such as MIT-BIH and PTB-XL. Transfer learning produced meaningful accuracy improvements, particularly where target datasets were small, suggesting it may offer a viable route forward for resource-limited settings. However, a number of methodological limitations were identified across the literature: limited external validation, inconsistent signal preprocessing, inadequate demographic reporting, and a complete absence of race- or ethnicity-stratified performance analyses. No included study used an exclusively African ECG dataset, and none evaluated model performance on African patients. This is especially concerning given well-documented differences in ECG parameters across ethnic groups. Black populations demonstrate higher rates of T-wave inversion and elevated Sokolow-Lyon voltages relative to standard reference ranges, differences that directly affect how AI models interpret cardiac tracings. In summary, while deep learning, particularly ResNet architectures augmented by transfer learning, delivers strong classification performance on benchmark data, the total exclusion of African populations raises genuine concerns about clinical fairness and real-world applicability. This review maps those gaps and offers a concrete methodological framework for developing ECG interpretation systems that are both population-aware and fit for use in African healthcare settings.