An explainable class-balanced dilated temporal convolutional network framework for multi-class ECG arrhythmia classification

Electrocardiogram (ECG) signal categorization is important for the identification of cardiac arrhythmia and diagnosis. In addition, ECG signals have been compromised by various forms of noise, and the imbalance in the dataset, which is caused by class imbalance, poses a significant challenge to the deep learning algorithms. This paper introduces a novel hybrid approach to ECG signal classification, which combines the discrete wavelet transform (DWT) of a signal to remove noise, R-peak localization with the Pan-Tompkins algorithm of data to extract features, and a dilated temporal convolutional network (dilated TCN) to classify the data into multiple classes using the MIT-BIH arrhythmia database. The cardiac signal helps identify characteristics for the detection of the R-peak with the Pan-Tompkins algorithm. ADASYN and Tomek link have been employed to equalize all classes and to reduce the imbalance in the five classes. The balanced arrhythmia classification has been trained with dilated TCN for the prediction of cardiac arrhythmia-affected signals. The overall classification accuracy of 99.70% and a macro-average F1-score, recall, and precision of 99.70%, and statistical significance analysis on the test dataset have been achieved by the proposed methodology. Class-specific feature attributes and physiologically relevant features were integral and consistent with clinical interpretability using the Grad-CAM and SHAP methods. The proposed Hybrid Wavelet-driven Pan-Tompkins features and Dilated TCN (HWPDTCN) method establishes a robust and accurate framework for automated ECG arrhythmia analysis.