DSNet: A Decoupled Siamese Network for ECG Classification

Mengyu Sun, Pengyao Xu, Xiaoyun Xie, Yinglong Wang

Abstract: The scarcity of rare classes in ECG data has traditionally hindered the widespread application of ECG analysis. We propose an innovative few-shot ECG classification method that addresses this challenge using a Decoupled Siamese Network (DSNet). The method consists of two main components: a weight-sharing feature embedding module and a relation assessment module. The weight-sharing embedding module incorporates multiple CNN blocks and a Brownian distance covariance matrix to facilitate robust feature representation. The similarity degree of input pairs is determined by the relation module, which meticulously assesses whether the inputs belong to the same class. Notably, this work is the first to classify few-shot ECG data using Brownian distance covariance (BDC). Furthermore, our study does not use any pooling layers to avoid losing important data. The CPSC2018 dataset was used for the experiments, which show that our technique performs exceptionally well for a range of values of K (the number of samples per class). In the 50-shot scenario, the identification accuracy of DSNet is 85.25%, representing an improvement of 2.19% to 34.25% over comparison methods. Through this proposed method, accurate detection of arrhythmias can be achieved, effectively reducing the burden on cardiovascular physicians.

External IDs:doi:10.1007/978-981-96-6585-3_22