Deep learning for ECG-based cardiovascular disease diagnosis: a systematic review and meta-analysis

Abstract Deep learning (DL) has emerged as an approach to identify complex trends in electrocardiogram data. However, there are inconsistencies in the methods used and results. We aimed to assess the efficacy of the state-of-the-art DL algorithms on ECG data to classify cardiovascular diseases (CVDs). We performed a systematic review and meta-analysis (PROSPERO, reference: CRD42023478008) using eight online databases. We identified 1068 articles, of which 85 were included in this review, covering a range of DL approaches. Results highlight the applicability of DL models in detecting arrhythmias and myocardial infarctions while highlighting issues such as noise and interference, imbalanced datasets, and the need for rigorous standardization methods. Meta-analysis of 27 articles provides quantitative insights into the overall efficacy of DL methods, where the summary operating point is located at 96% specificity and 97% sensitivity. We found limited focus on encryption (4.7%), uncertainty (0%), and explainability (9.4%) in DL models, which suggests redesign of early CVD identification by integrating explainable, uncertainty-aware, and robust DL techniques with remote monitoring technology. DL approaches show high diagnostic accuracy for ECG-based CVD classification; however, limited external validation, inconsistent reporting, and sparse explainability temper immediate clinical adoption. With rigorous standardisation and prospective validation, these methods may enhance early detection and workflow efficiency.