GSMD-SRST: Group Sparse Mode Decomposition and Superlet-Transform-Based Technique for Multilevel Classification of Cardiac Arrhythmia

Cardiac arrhythmia is caused due to the irregularity of the heartbeat and heart rhythm, which increases the complications leading to the risk of heart strokes. Atrial fibrillation (AF) and ventricular fibrillation (VF) are the most prevalent cardiac arrhythmia. Timely and precise detection of arrhythmia reduces mortality rates, which are essential to prevent various heart diseases. The proposed framework utilizes the group sparse mode decomposition (GSMD) technique for decomposing the nonstationary electrocardiogram (ECG) signals into the intrinsic mode functions (IMFs). Then, the super-resolution technique is applied to obtain frequency-domain spectrograms corresponding to IMFs of the input ECG record. The superlet transform (SLT) is a high-resolution technique used to reshape the 1-D ECG signal into a 2-D ECG signal. The spectrograms obtained through SLT are fed to the deep neural networks for arrhythmia classification. ECG records are fetched from different databases like MIT-BIH arrhythmia, MIT-BIH malignant ventricular ectopy, MIT-BIH CU ventricular tachyarrhythmia, MIT-BIH AF, and Mendeley-II dataset. Different deep neural models, such as VGG19, RESNET-18, GoogleNet, and AlexNet, are used for the classification of arrhythmia from the ECG signals. The proposed method with VGG19 shows the best results with an overall accuracy of 98.2% on Physionet dataset, whereas, for the combined dataset (Physionet + Mendely-II), the proposed model achieved an accuracy of 99.2% using GoogleNet. The classification is done between healthy heart, AF, and VF. The proposed system minimizes the overall intricacy as it reduces the requirement of an extra step for noise removal and peak detection.