Deep learning (DL) has been introduced in automatic heart-abnormality classification using ECG signals, while its application in practical medical procedures is limited. A systematic review is performed from perspectives of the ECG database, preprocessing, DL methodology, evaluation paradigm, performance metric, and code availability to identify research trends, challenges, and opportunities for DL-based ECG arrhythmia classification. Specifically, 368 studies meeting the eligibility criteria are included. A total of 223 (61%) studies use MIT-BIH Arrhythmia Database to design DL models. A total of 138 (38%) studies considered removing noise or artifacts in ECG signals, and 102 (28%) studies performed data augmentation to extend the minority arrhythmia categories. Convolutional neural networks are the dominant models (58.7%, 216) used in the reviewed studies while growing studies have integrated multiple DL structures in recent years. A total of 319 (86.7%) and 38 (10.3%) studies explicitly mention their evaluation paradigms, i.e., intra- and inter-patient paradigms, respectively, where notable performance degradation is observed in the inter-patient paradigm. Compared to the overall accuracy, the average $F_1$ score, sensitivity, and precision are significantly lower in the selected studies. To implement the DL-based ECG classification in real clinical scenarios, leveraging diverse ECG databases, designing advanced denoising and data augmentation techniques, integrating novel DL models, and deeper investigation in the inter-patient paradigm could be future research opportunities.

Deep learning (DL) has shown great promise in electrocardiogram (ECG) analysis, revolutionizing cardiovascular medicine by enabling precise and efficient diagnosis. This study explores the integration of expert knowledge into DL models, creating a flexible structure that adapts during training to refine this knowledge and guides feature separation in higher-dimensional space, thereby improving multi-label ECG classification. By leveraging domain expertise on the relationships between specific ECG abnormalities and their corresponding changes across lead dimensions, a lead-wise prior knowledge framework (LPKF) was introduced to enhance the learning efficiency of DL models. The effectiveness of this framework was validated by the higher classification performance of LPKF-enhanced models compared to their original versions. In addition, the LPKF-enhanced Inception model outperforms recent state-of-the-art DL methods, underscoring the benefits of integrating learnable expert knowledge. The study also demonstrated the interpretability of the LPKF-enhanced Inception model using gradient-weighted class activation mapping, revealing its capability to identify crucial diagnostic symptoms from ECG signals that align with clinical criteria.