Swarm Decomposition Enhances the Discrimination of Cardiac Arrhythmias in Varied-lead ECG Using ResNet-BiLSTM Network Activations

Mohanad Alkhodari1, Georgios Apostolidis2, Charilaos Zisou2, Leontios Hadjileontiadis1, Ahsan Khandoker1
1Khalifa University, 2Department of Electrical and Computer Engineering, Aristotle University of Thessaloniki


The standard screening tool for cardiac arrhythmias remains to be the 12-lead electrocardiography (ECG). Despite carrying rich information about different types of arrhythmias, it is considered bulky, high-cost, and often hard to use. In this study, we sought to investigate the efficiency of using 6-lead, 4-lead, 3-lead, and 2-lead ECG in discriminating between 26 arrhythmia types and compare them with the standard 12-lead ECG. as part of PhysioNet/Computing in Cardiology 2021 Challenge. Our team, Care4MyHeart, developed a deep learning approach based on residual convolutional neural networks and Bi-directional long short term memory (ResNet-BiLSTM) to extract deep-activated features from ECG oscillatory components obtained using a novel swarm decomposition (SWD) algorithm. Alongside age and sex, these automated features were combined with hand-crafted features from heart rate variability and SWD components for training and classification. Our approach achieved a challenge score of 0.45, 0.43, 0.44, 0.43, and 0.42 using 10-fold cross-validation using the training set and 0.41, 0.39, 0.40, 0.39, and 0.37 using the hidden testing set for 12-lead, 6-lead, 4-lead, 3-lead, and 2-lead, respectively. Our final entry (12-lead) was ranked the 158th out of 264.