基于互信息的多导联心电图排序方法

doi:10.16180/j.cnki.issn1007-7820.2024.02.008

摘要/Abstract

摘要：

基于卷积神经网络的心电图(Electrocardiograph, ECG)自动分类研究从默认12导联顺序的心电图中提取特征,未考虑导联顺序对卷积网络特征提取的影响。为解决该问题,文中提出了一种基于互信息的两端递增排序方法,使用互信息衡量导联之间的相关性,并根据导联之间的相关性以及二维卷积的特点将关系密切的导联相邻排序。实验结果表明,多导联心电图排序方法在3个数据库和3个卷积网络分类模型上取得了显著效果,F1、正确率、召回率、精确率以及杰卡德系数数值分别提升了0.011、0.009、0.007、0.014和0.013,汉明损失值减低了0.002。

关键词: 心电图, 心率不齐, 卷积神经网络, 互信息, 多导联, 排序, 分类, 相关性

Abstract:

The studies of automatic Electrocardiograph(ECG) classification based on convolutional neural networks all extract features from the ECG with the default 12-lead sequence, ignore the influence of lead sequence on feature extraction of convolutional network. To solve the problem, this study proposes a 2-end increasing sorting method based on mutual information, which uses mutual information to measure the correlation between leads. According to the correlation between leads and the characteristics of two-dimensional convolution, the adjacent connections of closely related leads are sorted.The experimental results show that the multi-lead ECG sorting method has achieved remarkable results on three databases and three convolutional network classification models.F1, accuracy, recall, accuracy, and Jacquard's coefficient of the proposed method increases by 0.011,0.009,0.007,0.014, and 0.013, while Hamming's loss decreases by 0.002.

Key words: electrocardiogram, arrhythmia, convolutional neural network, mutual information, multi lead, sorting, classification, correlation

中图分类号:

TN183

南娇,孙占全. 基于互信息的多导联心电图排序方法[J]. 电子科技, 2024, 37(2): 55-60.

NAN Jiao,SUN Zhanquan. Sorting Method of Multi Leads ECG Based on Mutual Information[J]. Electronic Science and Technology, 2024, 37(2): 55-60.

图/表 8

图1

图2

表1

表2

表3

表4

表5

表6

参考文献 18

[1]	周明成. 心血管疾病心脏康复现状与发展及思考[J]. 实用心脑肺血管病杂志, 2021, 29(9):6-9.
	Zhou Mingcheng. Current situation, development and thi-nking of cardiac rehabilitation for cardiovascular diseases[J]. Practical Journal of Cardiac Cerebral Pneumal and Vascular Disease, 2021, 29(9):6-9.
[2]	Gordon P, Kerton M. Cardiac arrhythmias in the critically ill,anaesthesia and intensive care medicine[J]. Circulation, 2021, 22(5):301-312.
[3]	Sharma A N, Baranchuk A. Ambulatory external electr-ocardiography monitoring:Holter,extended holter,mobilecardiac telemetry monitoring[J]. Cardiac Electrophysiology Clinics, 2021, 13(3):427-438.
[4]	Ponnusamy S S, Vijayaraman P. Electrocardiography guided left bundle branch pacing[J]. Journal of Electrocardiology, 2021, 68(9):11-13.
[5]	徐利, 徐久强, 冯家乐. 结合CEEMDAN与改进区间阈值的ECG降噪研究[J]. 小型微型计算机系统, 2020, 41(8):1576-1579.
	Xu Li, Xu Jiuqiang, Feng Jiale. Research on ECG denoising combined with CEEMDAN and improved interval threshold[J]. Journal of Chinese Computer Systems, 2020, 41(8):1576-1579.
[6]	刘姝, 邵杰, 张颐婷, 等. 基于双谱和谱特征的心电信号分类方[J]. 电子科技, 2021, 34(2):21-26.
	Liu Shu, Shao Jie, Zhang Yiting, et al. ECG classificationbased on bispectrum and spectral features[J]. ElectronicScience and Technology, 2021, 34(2):21-26.
[7]	张浪飞, 李诗楠, 梁竹关, 等. 基于自适应多元多尺度熵的心电信号分类研究[J]. 计算机应用研究, 2022, 39(5):1505-1509.
	Zhang Langfei, Li Shinan, Liang Zhuguan, et al. Classific-ation of ECG signal based on adaptive multivariate multiscale entropy[J]. Application Research of Computers, 2022, 39(5):1505-1509.
[8]	万静, 张晓瑞, 何云斌, 等. 12导联高频心电信号的特征提取及聚类[J]. 计算机应用研究, 2015, 32(10):2934-2939.
	Wan Jing, Zhang Xiaorui, He Yunbin, et al. Feature ext-raction and clustering of 12 lead high frequency electrocardiogram signals[J]. Application Research of Computers, 2015, 32(10):2934-2939.
[9]	Honarvar H, Agarwal C, Somani S, et al. Enhancing convolutional neural network predictions of electrocardiograms with left ventricular dysfunction using a novels ubwaveform representation[J]. Cardiovascular Digital Health Journal, 2022, 3(5):220-231.
[10]	Acharya U R, Fujita H, Lih O S, et al. Automated detection of coronary artery disease using different durations of ECG segments with convolutional neural network[J]. Knowledge-Based Systems, 2017, 132(15):62-71.
[11]	Wang C, Yang S, Tang X, et al. A 12-lead ECG arrhyth-mia classification method based on 1D densely connected CNN[C]. Shenzhen: Proceedings of the First Machine Learning and Medical Engineering for Cardiovascular Health and Intravascular Imaging and Computer Assisted Stenting, 2019:1125-1133.
[12]	葛靖, 刘子龙. 基于CNN和LSTM的睡眠呼吸暂停检测算法[J]. 电子科技, 2021, 34(2):21-26.
	Ge Jing, Liu Zilong. The alorithm based on CNN and LSTM for sleep apnea syndrome detection[J]. Electronic Science and Technology, 2021, 34(2):21-26.
[13]	陈鹏, 刘子龙. 基于GAN-CNN的心律失常识别[J]. 电子科技, 2022, 35(3):45-50.
	Chen Peng, Liu Zilong. Arrhythmia recognition based on GAN-CNN[J]. Electronic Science and Technology, 2022, 35(3):45-50.
[14]	Yildirim O. A novel wavelet sequence based on deep bidirectional LSTM network model for ECG signal classification[J]. Comuters in Biology and Medicine, 2018, 96(2):189-202.
[15]	Karpuz D, Hallioglu O, Toros F, et al. The effect of metilpheniy date, risperidone and combination therapy on ECG in children with attention-deficit hyperactivity disorder[J]. Journal of Electrocardiology, 2017, 50(4):410-415.
[16]	Ratnaparkhi A, Bormane D, Ghongade R. Performance analysis of fuzzy rough assisted classification and segmentation of paper ECG using mutual information and dependency metric[C]. Palladam: International Conference on ISMAC in Computational Vision and Bio-Engineering, 2018:778-786.
[17]	Ghosh S M, Bandyopadhyay S, Mitra D. Nonlinear classification of emotion from EEG signal based on maximized mutual information[J]. Expert Systems with Application, 2021, 185(10):114-120.
[18]	Hannun A Y, Rajpurkar P, Haghpanahi M, et al. Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network[J]. Nature Medicine, 2019, 25(1):65-69.

模型	排序方法	A_p	H_p	J_p	R_p	P_p	F1
VGGNet	默认排序	0.727	0.043	0.673	0.853	0.761	0.804
	两端递增排序	0.727	0.042	0.681	0.853	0.771	0.810
	两端递减排序	0.699	0.045	0.659	0.841	0.752	0.794
	类最大树排序	0.724	0.042	0.680	0.854	0.769	0.809
GoogLeNet	默认排序	0.749	0.044	0.683	0.841	0.781	0.809
	两端递增排序	0.758	0.041	0.696	0.848	0.795	0.820
	两端递减排序	0.717	0.044	0.669	0.853	0.755	0.802
	类最大树排序	0.749	0.043	0.682	0.829	0.794	0.811
ResNet	默认排序	0.729	0.042	0.678	0.854	0.767	0.808
	两端递增排序	0.728	0.042	0.681	0.859	0.766	0.810
	两端递减排序	0.742	0.044	0.673	0.831	0.779	0.804
	类最大树排序	0.728	0.042	0.680	0.866	0.759	0.809

模型	排序方法	A_p	H_p	J_p	R_p	P_p	F1
VGGNet	默认排序	0.709	0.012	0.743	0.876	0.831	0.853
	两端递增排序	0.741	0.011	0.756	0.903	0.823	0.861
	两端递减排序	0.710	0.012	0.738	0.881	0.819	0.849
	类最大树排序	0.718	0.012	0.748	0.880	0.832	0.855
GoogLeNet	默认排序	0.725	0.011	0.759	0.880	0.847	0.863
	两端递增排序	0.738	0.011	0.767	0.888	0.849	0.868
	两端递减排序	0.724	0.012	0.755	0.874	0.847	0.860
	类最大树排序	0.726	0.011	0.759	0.876	0.850	0.863
ResNet	默认排序	0.727	0.011	0.750	0.887	0.832	0.859
	两端递增排序	0.735	0.011	0.761	0.894	0.836	0.864
	两端递减排序	0.719	0.012	0.748	0.883	0.830	0.855
	类最大树排序	0.724	0.011	0.758	0.877	0.847	0.862

模型	排序方法	A_p	H_p	J_p	R_p	P_p	F1
VGGNet	默认排序	0.325	0.024	0.496	0.724	0.612	0.663
	两端递增排序	0.325	0.024	0.501	0.704	0.635	0.668
	两端递减排序	0.333	0.024	0.499	0.702	0.633	0.665
	类最大树排序	0.316	0.024	0.493	0.717	0.613	0.661
GoogLeNet	默认排序	0.345	0.023	0.516	0.715	0.650	0.681
	两端递增排序	0.357	0.022	0.522	0.745	0.636	0.686
	两端递减排序	0.336	0.022	0.513	0.757	0.614	0.678
	类最大树排序	0.344	0.023	0.515	0.735	0.632	0.680
ResNet	默认排序	0.352	0.023	0.514	0.740	0.628	0.679
	两端递增排序	0.354	0.023	0.517	0.737	0.634	0.682
	两端递减排序	0.331	0.023	0.509	0.725	0.631	0.675
	类最大树排序	0.347	0.023	0.515	0.726	0.639	0.680

模型	排序方法	A_p	H_p	J_p	R_p	P_p	F1
VGGNet	互信息	0.727	0.042	0.681	0.853	0.771	0.810
	欧式距离	0.717	0.044	0.667	0.854	0.752	0.800
	余弦相似度	0.711	0.044	0.665	0.848	0.755	0.799
	相关系数	0.721	0.043	0.672	0.851	0.762	0.804
GoogLeNet	互信息	0.758	0.041	0.696	0.848	0.795	0.820
	欧式距离	0.744	0.043	0.681	0.839	0.783	0.810
	余弦相似度	0.728	0.042	0.681	0.859	0.766	0.810
	相关系数	0.729	0.044	0.670	0.835	0.773	0.803
ResNet	互信息	0.728	0.042	0.681	0.859	0.766	0.810
	欧式距离	0.742	0.044	0.673	0.831	0.779	0.804
	余弦相似度	0.735	0.043	0.677	0.851	0.768	0.807
	相关系数	0.716	0.044	0.669	0.836	0.769	0.801

模型	排序方法	A_p	H_p	J_p	R_p	P_p	F1
VGGNet	互信息	0.741	0.011	0.756	0.903	0.823	0.861
	欧式距离	0.679	0.013	0.735	0.846	0.848	0.847
	余弦相似度	0.713	0.012	0.747	0.883	0.828	0.855
	相关系数	0.714	0.012	0.744	0.874	0.832	0.853
GoogLeNet	互信息	0.738	0.011	0.767	0.888	0.849	0.868
	欧式距离	0.724	0.011	0.758	0.881	0.844	0.862
	余弦相似度	0.708	0.012	0.747	0.846	0.865	0.855
	相关系数	0.724	0.012	0.755	0.874	0.850	0.861
ResNet	互信息	0.735	0.011	0.761	0.894	0.836	0.864
	欧式距离	0.724	0.011	0.752	0.886	0.832	0.859
	余弦相似度	0.714	0.012	0.743	0.900	0.809	0.852
	相关系数	0.730	0.011	0.755	0.886	0.836	0.861