基于多特征分支卷积神经网络的心电图分类算法  

Electrocardiogram Classification Algorithm Based on Multi Feature Branch Convolutional Neural Network

作　　者：王建荣[1,2] 程伟[2] 邓黎明 李国翚 WANG Jianrong;CHENG Wei;DENG Liming;LI Guohui(College of Intelligence and Computing,Tianjin University,Tianjin 300000,China;School of Automation and Software,Shanxi University,Taiyuan 030000,China;Department of Product R&D,Tianjin Development Zone Orking High Tech.Co.,Ltd,Tianjin 300000,China)

机构地区：[1]天津大学智能与计算学部,天津300000 [2]山西大学自动化与软件学院,山西太原030000 [3]天津开发区奥金高新技术有限公司产品研发部,天津300000

出　　处：《测试技术学报》2024年第2期161-169,共9页Journal of Test and Measurement Technology

基　　金：国家重点研发计划资助项目(2018YFC2000701);中国博士后科学基金资助项目(2021M692400);山西省基础研究计划资助项目(202203021221017)。

摘　　要：我国心血管疾病发病率、病死率呈逐年上升趋势。但由于心电图数据规模大且繁杂,临床医护人员在心电图筛查时,工作负担大且容易出现误诊或者漏诊的情况。基于此,利用CPSC-201812导联数据,提出了一种基于多特征分支卷积神经网络的多导联心电信号的智能分类与分析。首先,将CPSC-201812导联数据分为9个类别,基于12导联推导出8导联心电信号并分别提取局部特征。然后,通过双向GRU编码和注意力机制计算出不同类别的注意力权重向量,并将特征信息串联融合成特征向量,从而实现多导联心电图分类。实验结果表明:在验证集上取得了较好的分类效果,正常类别的F1值达到81.2%,平均F1值达到84.2%。特别地,在识别房颤(AF)和右束支传导阻滞(RBBB)这两类别心律失常时F1值分别达到95.1%和93.1%。The incidence rate and mortality of cardiovascular diseases in China are increasing year by year.However,due to the large scale and complexity of electrocardiogram data,clinical medical staff have a heavy workload and are prone to misdiagnosis or missed diagnosis during electrocardiogram screening.Based on this,in this paper we proposes an intelligent classification and analysis of multi-lead electrocardiogram signals based on multi feature branch convolutional neural networks using CPSC-2018 twelve lead data.Firstly,divide the CPSC-201812-lead data into 9 categories,derive 8-lead electrocardiogram signals based on the 12 leads,and extract local features separately.Then,the attention weight vectors of different categories are calculated through bidirectional GRU encoding and attention mechanism,and the feature information is concatenated and fused into feature vectors to achieve multi-lead electrocardiogram classification.The experimental results showed that good classification performance was achieved on the validation set,with an F1 value of 81.2% for normal categories and an average F1 value of 84.2%.Especially,when identifying two types of arrhythmia,atrial fibrillation(AF)and right bundle branch block(RBBB),F1 values reached 95.1% and 93.1%,respectively.

关 键 词：心律失常 心电图 卷积神经网络 GRU网络 注意力机制 

分 类 号：R540.4[医药卫生—心血管疾病] TP183[医药卫生—内科学]