Discrimination of Cardiac Abnormalities Based on Multifractal Analysis in Reservoir Computing Framework

• Published in: IEEE open journal of instrumentation and measurement 2023, Vol.2, p.1-11
• Main Authors: Purkayastha, Basab Bijoy, Barma, Shovan
• Format: Article
• Language: English
• Subjects: Arrhythmia; Arrhythmia classification; Complexity theory; echo state network (ESN); Electrocardiography; Fractals; Heart; Heart beat; multifractal analysis; multivariate multifractal singularity spectrum; nonlinear dynamics; Reservoirs
• ISSN: 2768-7236; 2768-7236
• DOI: 10.1109/OJIM.2023.3332344

This study proposes a multiclass classification technique based on multifractal spectra for different types of cardiac arrhythmias which are associated with irregularity and/or complex dynamics of the heart. Indeed, the degree of complexity of such dynamics is diverse for different states of cardiac condition. Certainly, such physiological responses of the heart dynamics can be discriminated by analyzing electrocardiogram (ECG) signals through different channels. Earlier, ECG-based works for discriminating cardiac arrhythmias consider the heart as a black box system and the analysis is mostly surrounded with time domain statistical averages or spectral analysis. The works ignore one of the key parameters, i.e., the presence of time-localized irregularities which are strongly associated with different kinds of arrhythmias and contribute to subtle variations in the amplitude and shape of the signal dynamical system while analyzing the signal. Therefore, in this work, we proposed a new method based on multifractal analysis to classify different kinds of cardiac conditions. Here, we followed the dynamical systems approach and computed the multifractal spectrum of the embedded phase space structure of the ECG signal. We performed the classification task by an echo state network to reduce the computational burden. For validation, three well-known datasets (Shaoxing Peoples' Hospital dataset, PTB diagnostic ECG database v1.0.0, and 2017 PhysioNet/CinC Challenge dataset) have been considered. The results and analysis show that the proposed method can achieve a maximum accuracy of up to 96%, which is significantly high. Further, an optimum number of channels/leads has also been evaluated in multichannel ECG analysis. The result and analysis reveal that the effectiveness of the model in classifying various categories of cardiac disorders from ECG.