A Contrastive Learning Framework for ECG Anomaly Detection

Abstract

ECG is important for the recognition and diagnosis of cardiac arrhythmias as a physiological signal characterizing the condition of the heart. A lot of studies have started to experiment with statistical and traditional machine learning methods to analyze and detect ECG data, thus to the heart and other organs of intelligent auxiliary treatment. Although a lot of work has been done in ECG signal processing, the existing work still suffers from the following deficiencies:i) Since the number of various types of signals is unbalanced when classifying ECG signals, and end-to-end deep learning models are very sensitive to unbalanced data, which can affect the automatic detection and classification tasks. ii) The models lack robustness due to inconsistent ECG data representation. For this reason, in this paper, we first design a data augmentation-based contrast learning module to alleviate the data imbalance and robustness problems of the model. Thus, a new contrast learning ECG abnormality detection framework is designed by capturing the underlying patterns of ECG signals. Many experiments show that our abnormality detection framework outperforms the baseline methods, which provides a new view for cardiovascular disease prevention and automatic diagnosis.