AI in cardiovascular disease classification—a systematic literature review

Despite advances in integrating Artificial Intelligence (AI) with Electrocardiogram (ECG) analysis, several limitations remain, including limited dataset diversity, class imbalance, underrepresentation of rare conditions, and lack of external validation, which affect generalizability and minority class performance. Many approaches rely on “black-box” models with limited interpretability, while deep and hybrid architectures can be computationally expensive, and multimodal integration remains constrained. Following the PRISMA methodology, 490 records were retrieved, 223 studies were selected after screening, and four additional studies were included through snowballing, resulting in a total of 227 studies. This review addresses whether multimodal AI improves prediction over single-modality approaches, how Explainable AI (XAI) supports interpretability, and how self-supervised, transformer, and transfer learning models compare with traditional methods. Findings show that multimodal approaches frequently show improved performance, with many studies reporting gains of 2–6% even though results vary across datasets and tasks. Advanced architectures can extract complex features, but their effectiveness depends on data quality and evaluation metrics. XAI methods give context-dependent support for interpretability, while self-supervised and transfer learning approaches show potential in low-label settings. Future research should focus on lightweight models, standardized evaluation, and improved clinical validation. This review highlights emerging trends in multimodal learning, XAI, and advanced architectures in ECG-based CVD analysis.