A high-precision arrhythmia classification method based on dual fully connected neural network. (April 2020)

• Record Type: Journal Article
• Authors: Wang, Haoren; Shi, Haotian; Lin, Ke; Qin, Chengjin; Zhao, Liqun; Huang, Yixiang; Liu, Chengliang
• Is Part Of: Biomedical signal processing and control. Volume 58(2020)
• Publication Date: 2020-04
• Subjects: Electrocardiogram (ECG) -- Heartbeat classification -- Fully connected neural networks -- Inter-patient -- MIT database; 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.2020.101874
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: A novel method based on a two-layer architecture and fully connected neural networks was proposed for ECG heartbeat classification. In the second layer of the two-layer architecture, the threshold criterion was employed for N-S classification. Both the MIT arrhythmia database and the MIT supraventricular arrhythmia database were used for verification. A convolutional neural network was also employed to show the advantages of the proposed method over the classical deep learning method. Abstract: As an important arrhythmia detection method, the electrocardiogram (ECG) can directly reflect abnormalities in cardiac physiological activity. In view of the difficulty in the diagnosis of arrhythmia in different people, automatic arrhythmia detection methods have been studied in previous works. In this paper, we present a dual fully-connected neural network model for accurate classification of heartbeats. Our method is following the AAMI inter-patient standard, which includes normal beats (N), supraventricular ectopic beats (S), ventricular ectopic beats (V), fusion beats (F), and unknown beats (Q). Firstly, a total of 105 features are extracted from the preprocessed signals. Then, a two-layer classifier is introduced in the classification stage. Each layer contains two independent fully-connected neural networks, and the threshold criterion is also added in the second layer. For verification, both the MIT arrhythmia database (MITDB) and the MIT supraventricular arrhythmia database (SVDB) were adopted. The experiments demonstrate that the proposed method has high performance for arrhythmia detection. It also achieves high sensitivity for class S and V, which can easily detect potentially abnormal heartbeats. Furthermore, the proposed method can interfere with the classification effect for a certain disease and have more advantages in dataset size when comparing a convolutional neural network (CNN). Once properly trained, the proposed method can be employed as a tool to automatically detect arrhythmia from ECG.