Detecting Cardiac Abnormalities with Multi-Lead ECG Signals: A Modular Network Approach

Ryan Clark1, Mohammadreza Heydarian2, Mohammad Siddiqui1, Sajjad Rashidiani3, Md Asif Khan3, Thomas Doyle3
1School of Biomedical Engineering, McMaster University, 2Department of Electrical and Computer Engineering, McMaster University, Hamilton, ON, Canada, 3Department of Electrical and Computer Engineering, McMaster University


Abstract

Introduction: Globally, heart disease has been the leading cause of death for more than two decades. The 2021 PhysioNet challenge offers a unique opportunity to develop intelligent architectures to handle a variety of real life clinical scenarios when a 12 lead ECG is not a viable option. We describe four deep CNN architectures to classify cardiac abnormalities from twelve-lead, six-lead, three-lead, and two-lead ECGs. These four networks were created for the PhysioNet/CinC Challenge 2021, by the Biomedic2ai team. 

Methods: Each ECG signal was down-sampled to 100 Hz, and segmented to 16-second length by zero-padding shorter signals and windowing longer signals with a 50% overlap sliding window. A one-dimensional deep CNN (1D-dCNN) model was used to preserve sequentially related features embedded in the signals. Extracted statistical features were also used to assist the deep model for classification. This feature extraction module was added to the 1D-dCNN, creating a ‘wide and deep modular network’. This framework allows for addition or removal of modules to optimize classification models for twelve-lead, six-lead, three-lead, and two-lead ECGs. Each model was trained until stable training and validation accuracies were achieved over 5 epochs. 

Results: Our method achieved a cross-validated scoring metric of 0.67, 0.21, 0.45 and 0.46 on training data, and the challenge validation score provided by the CinC challenge submission system on a hidden dataset were 0.52, 0.13, 0.47, and 0.47 for 12, 6, 3, and 2 leads, respectively.

Conclusion: Our current model offers modularity and demonstrates potential with the wide modular network. The 6-lead model’s lower performance offers opportunity for development to reach parity with 12, 3, and 2 lead models. The framework also provides the ability to integrate clinical knowledge in future modules to improve the overall performance.