Classification of 12-lead ECG Signals with Adversarial Multi-Source Domain Generalization

Abstract

The ECG classification is a critical task in the early and correct diagnosis of cardiovascular diseases. Although various models have been developed to tackle the heartbeat classification problem, their performance degrades on ECG signals recorded in varied testing conditions due to the distribution discrepancy among different sources of data. In this work, we have developed a multi-source domain generalization model to address the distribution discrepancy problem that occurred when the collection of the data is from multiple sources with various acquisition conditions. We have employed a combination of convolutional neural network (CNN) and long short term memory (LSTM) for feature extraction. Further, we exploit the adversarial domain generalization method to overcome probable heterogeneity between the train and test datasets. To increase generalization, we also utilized different augmentation techniques including random ECG pad and crop, adding low-frequency artifacts, and lead dropout. We evaluate our proposed model on cardiac abnormality classification based on 12-lead ECG signals associated with' Classification of 12-lead ECGs for the Phy-sioNetlComputing in Cardiology Challenge 2020'. Our method, achieved a challenge validation score of 0.609, and full test score of 0.437 placing us (Sharif AI Team) 5th out of 41 teams in the final official ranking.