Explainable deep learning based techniques for ECG-Based heart disease classification: A systematic literature review and future direction

OBJECTIVE: This study aims to improve the understanding of explainability in deep learning (DL) architectures utilized for heart disease (HD) classification through electrocardiogram (ECG) data. It offers a systematic review of methodological choices, analyses their impacts on model interpretability, and highlights significant challenges in this domain while suggesting opportunities for future research. METHODS: A systematic literature review (SLR) was conducted based on Kitchenham and Charters guidelines. To address the proposed research questions, this article provides an SLR of academic articles on Explainable AI (XAI)-based DL for the classification of HDs using ECGs dated from January 2018 to September 2024. Results were synthesized by identifying the valuable insights into the datasets, preprocessing methods, and techniques employed in the development of XAI-based DL models for the classification of ECG-based HDs. RESULTS: This study identified 6448 primary studies that utilized machine learning and DL techniques for the classification of HD based on ECG data, of which 51 used an XAI-based DL architecture. Our finding indicates that there are 25 different datasets that were used, 16 different DL architectures were introduced, and eight novel XAI techniques were introduced, while most of the selected studies used the conventional XAI approaches such as SHAP, Saliency Maps, Grad-Cam, and LIME. CONCLUSIONS: A considerable amount of XAI-based DL architectures were identified for the classification of ECG-based HD. The techniques frequently selected by these studies have potential limitations such as data standardization, inconsistent explainability, temporal dependency visualization, lack of XAI benchmarking, lack of standardized metrics, and many more, which are presented with proposed future research directions.