Incorporating Clinical and Heartbeat Level Features with Multichannel ECG for Cardiac Abnormality Detection Using Parallel CNN and GAP Network

Deepankar Nankani and Rashmi Dutta Baruah
Department of Computer Science and Engineering, Indian Institute of Technology Guwahati


Abstract

Background: Early and correct diagnosis of cardiac arrhythmias from Multichannel Electrocardiogram (MECG) is a challenging problem. We address this problem as part of the PhysioNet/Computing in Cardiology Challenge 2021 by classifying cardiac abnormalities from MECG.

Method: The proposed method incorporates clinical features including patient age, gender and heartbeat features with MECG to detect cardiac abnormalities. Initially, MECG is cleaned from noise, followed by resampling and segmentation. Then R-peaks are extracted from Lead II signal using Pan Tompkins detector to obtain heartbeat level features such as heart rate, RR Intervals, Mean QRS Amplitude, Hermite polynomial coefficients, statistical features, and Wave Amplitude based features. The feature vector consisting of clinical and heartbeat features is combined with MECG for classification using a Parallel Convolution Neural Network with Global Average Pooling (PCNN-GAP) network. The model extracts local features using smaller convolution kernels and global patterns using larger kernels. Lastly, the last layer sigmoid activation function classifies the rhythm into one or more cardiac abnormalities.

Results: Our team, skylark, achieved a score of 0.466, 0.426, 0.5, 0.491, and 0.514 (ranked 131, 137, 116, 118, and 107 out of 256 teams) for the 12-lead, 6-lead, 4-lead, 3-lead, and 2-lead versions of the hidden validation set with the Challenge evaluation metric.