DWTFrTV-DEAL: Denoising and deep ensemble learning for arrhythmia classification using ECG signals

Abstract

Electrocardiogram (ECG) signals adeptly capture the intricate electrical dynamics of the cardiac system, offering a means to assess its physiological well-being. To address the rising cardiovascular diseases (CVDs), various Machine Learning (ML)/Deep Learning (DL) techniques are employed to improve the accuracy, speed, and robustness of classifying Arrhythmia. While considerable attention has been given to designing architectures and selecting data sets, the significance of pre-processing data and its classification must be underscored. The presence of diverse disturbances during the acquisition of ECG signals influences the accurate feature extraction, which leads to lower classification accuracy. Hence, the paper introduces an efficient method of denoising and classification of ECG signals to substantially enhance the accuracy of using DL models for arrhythmia classification. The proposed work in the paper delineates two distinct methodologies, one for denoising ECG signals and another for classification. The initial approach involves mitigating noise in the ECG pattern through Discrete Wavelet Transform using Fractional Total Variation (DWTFrTV), while the subsequent step entails classifying the denoised signals utilizing a combination of two stacked learner classifiers. The base learner (set of optimized classifiers) has three models and the meta learner is the regression model. The proposed model is trained and tested over the MIT-BIH Arrhythmia dataset where the results and comparative analysis with MIT-BIH Normal Sinus Rhythm Database and St. Petersburg INCART Database demonstrates that the proposed model yields a notable improvement with an average accuracy of 99.9% and MAE: 0.05 (for Training) and average accuracy 99.7% and MAE: 0.09 (for Testing) over recent existing models.