What Leads to Arrhythmia: Active Causal Representation Learning of ECG Classification

Shaofei Shen, Weitong Chen, Miao Xu

Abstract: The electrocardiogram (ECG) classification has attracted great attention as a crucial tool to detect arrhythmia which can be an early sign of heart disease. However, the key challenge of the current ECG classification methods is the lack of annotated data when applied to new patients. On the one hand, enormous ECG data are produced and they require a high labelling cost for supervised classification. On the other hand, the morphological and temporal features of ECG in individual patient can vary significantly. Therefore, the heartbeat classification models cannot be trained on adequate data and usually faced a huge performance degradation when tested on new patients without enough annotated data. Although the current works have worked on reducing labelling costs through active learning, these methods do not focus on patient differences and cannot guarantee performance when patient differences increase. Other works that aim to solve the patient differences only focus on the correlations but not the causal relations behind the data. In this paper, we firstly analyse the patient differences in ECG heartbeat in a causal view and propose Active Causal Representation learning of ECG heartbeat Classification (ACREC) to learn the stable features that have a direct causal effect on the outcome variable. The experiment results show our method can outperform other methods when handling patient differences. After active learning, our model can select the most informative data to annotate and achieve reliable performances. Moreover, we also conduct the ablation study to validate the effect of each part in our model.