Automated arrhythmia classification using depthwise separable convolutional neural network with focal loss. (August 2021)

• Record Type: Journal Article
• Authors: Lu, Yi; Jiang, Mingfeng; Wei, Liying; Zhang, Jucheng; Wang, Zhikang; Wei, Bo; Xia, Ling
• Is Part Of: Biomedical signal processing and control. Volume 69(2021)
• Publication Date: 2021-08
• Subjects: Arrhythmia classification -- Convolutional neural network -- Depthwise separable convolution -- Focal loss; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2021.102843
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: We propose a novel DSC-FL-CNN method for Arrhythmia Classification with imbalanced 17 types ECG datasets. The Focal loss is designed to solve the class imbalance problem in ECG classification. The depthwise separable convolutional layers is used to reduce the number of parameter selection. Abstract: Arrhythmia was one of the primary causes of morbidity and mortality among cardiac patients. Early diagnosis was essential in providing intervention for patients suffering from cardiac arrhythmia. Convolution neural network (CNN) was widely used for electrocardiogram (ECG) classification. However, the conventional CNN method only worked well for balanced dataset. Therefore, a depthwise separable convolutional neural network with focal loss (DSC-FL-CNN) method was proposed for automated arrhythmia classification with imbalance ECG dataset. The focal loss contributed to improving the arrhythmia classification performances with imbalance dataset, especially for those arrhythmias with small samples. Meanwhile, the DSC-FL-CNN could reduce the number of parameters. The model was trained on the MIT-BIH arrhythmia database and it evaluated the performance of 17 categories of arrhythmia classification. Comparing with state-of-the-art methods, the experimental results showed that the proposed model reached an overall macro average F1-score with 0.79, which achieved an improvement for arrhythmia classification.