An end-to-end model for ECG signals classification based on residual attention network. (February 2023)

• Record Type: Journal Article
• Authors: Lu, Xiang; Wang, Xingrui; Zhang, Wanying; Wen, Anhao; Ren, Yande
• Is Part Of: Biomedical signal processing and control. Volume 80:Part 2(2023)
• Publication Date: 2023-02
• Subjects: Arrhythmia -- Atrial fibrillation -- Residual attention network -- Electrocardiogram; 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.2022.104369
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: A RA-NET model for ECG classification combining residual structure and attention mechanism is proposed. An algorithm to fill other recordings of the same category is presented. The model is quite competitive compared to the models proposed by other scholars utilizing the same dataset. Abstract: Arrhythmia is the most threatening disease among cardiovascular diseases, and in the last few years, the automatic detection of arrhythmia using neural networks have been intensely focused by physicians. In our work, we propose an effective method to automatically classify electrocardiogram (ECG) signals utilizing residual attention network (RA-NET). RA-NET combines the residual structure and attention mechanism, which can not only generate the attention weight of atrial fibrillation (AF) category to enhance the effective information, but also avoid the network degradation problem in the deep network. Besides, a novel filling algorithm for filling sample values of other recordings with the same category is presented, which is combined with RA-NET to validate model on the PhysioNet Challenge 2017 dataset. According to the comparison with other relevant classification models and filling methods, the experimental results demonstrate that the model we proposed achieves an excellent classification performance, the average of F1 -score and sensitivity reach 0.8289 and 0.8955, respectively. For AF category, the precision, F1 -score and specificity achieve 0.8763, 0.8835 and 0.9858, separately. With its preeminent performance, the proposed model is capable to play an important auxiliary role in single-lead AF detection.