An Interpretable Classification of ECGs with Varying Dimensions using Graph Convolutional Network and Deep Wavelet Decomposition Network

Xiang Wang, Zehao Lei, Xiao Han, Jie Yang
Institute of Image Processing and Pattern Recognition, Shanghai Jiao Tong University.


Introduction: A deep neural network architecture was described to classify multiple cardiac abnormalities using ECG data with varying leads. The model was created by the team “SJTU_PAMI_Lab” for the PhysioNet/Computing in Cardiology Challenge 2021.

Methods: In this work, ECGs were preprocessed using high-pass filters and set to a consistent duration with 4096 samples by randomly clipping or zero-padding the signal. An interpretable model combining graph convolutional network (GCN) and deep wavelet decomposition network (DWDN) was proposed to learn effective features. 3-layers GCN was to extract the spatial features between different leads. Moreover, it learned a weights to interpret the dependences of leads. 10-layers DWDN was used to build frequency-aware deep learning models and to learn the temporal features for ECG sequences. The modified ResNet was embedded into DWDN to improve the efficiency of the model. Also, to handle class imbalance, a simple constrained grid-search was applied. Four models with similar architecture were built for twelve-leads ECGs, six-leads ECGs, three-leads ECGs and two-leads ECGs, respectively.

Results: based on the weighted accuracy metric provided by the Challenge, four models (for twelve-leads ECGs, six-leads ECGs, three-leads ECGs and two-leads ECGs) achieved a 5-fold cross-validation score of (0.706, 0.710, 0.691, 0.660), sensitivity and specificity of (0.910, 0.918, 0.892, 0.888) and (0.969, 0.970, 0.953, 0.946), respectively on the training data.

Conclusion: The proposed classification model performed well on the training data with varying leads. Such models potentially interpreted the relationships between different ECG leads, which was beneficial for ECG research and diagnosis.