Arrhythmia classification algorithm based on multi-head self-attention mechanism. (January 2023)

• Record Type: Journal Article
• Authors: Wang, Yue; Yang, Guanci; Li, Shaobo; Li, Yang; He, Ling; Liu, Dan
• Is Part Of: Biomedical signal processing and control. Volume 79(2023)Part 2
• Publication Date: 2023-01
• Subjects: Arrhythmia classification -- Electrocardiogram (ECG) -- Attention mechanism -- Feature extraction; 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.2022.104206
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: ECG signal preprocessing method based on wavelet transform to reduce noise. Linear projection layer is designed to acquire semantic features of ECG signal. Novel position encoding is proposed to obtain time and voltage series information. Novel arrhythmia classification algorithm based on multihead self-attention mechanism. The proposed ACA-MA outperforms other state-of-the-art methods. Abstract: Cardiovascular disease is a major illness that causes human death, especially in the elderly. Timely and accurate diagnosis of arrhythmia types is the key to early prevention and diagnosis of cardiovascular diseases. This paper proposed an arrhythmia classification algorithm based on multi-head self-attention mechanism (ACA-MA). First, an ECG signal preprocessing algorithm based on wavelet transform is put forward and implemented using db6 wavelet transform to focus on improving the data quality of ECG signals and reduce the noise of ECG signals. Second, a linear projection layer for acquiring semantic features of ECG signals is designed using the matching relationship between ECG tag and segmented ECG signals. Third, a position encoding-based spatiotemporal characterization method of ECG signal sequences is designed to integrate time series information into a matrix operation. Fourth, a multi-head self-attentive mechanism capable of capturing global contextual information is proposed to extract relationships and semantic features between ECG segments and achieve semantic association and information stitching of nonadjacent ECG signals. Finally, experimental results on the arrhythmia dataset MIT/BIH show that ACA-MA outperforms other state-of-the-art methods with an overall classification accuracy of 99.4%, a specific rate of 99.41%, and a sensitivity of 97.36%.