基于ResNet和Bi-LSTM模型融合的心电信号分类  

ECG Signal Classification Based on Fusion Model of Res Nte and Bi-LSTM Network

作　　者：叶兴扬 韦张跃昊 严卉[2] 周逸蒋[2] 朱建华[2] 洪慧[1] Ye Xingyang;Wei Zhangyuehao;Yan Hui;Zhou Yijiang;Zhu Jianhua;Hong Hui(不详;School of Electronic Information.Hangzhou Dianzi University,Hangzhou 310018,China)

机构地区：[1]杭州电子科技大学电子信息学院,浙江杭州310018 [2]浙江大学医学院附属第一医院,浙江杭州310006

出　　处：《航天医学与医学工程》2021年第3期244-251,共8页Space Medicine & Medical Engineering

基　　金：浙江省自然科学基金项目(LR19F020005)。

摘　　要：目的为提升临床心电图的分类效果,提出一种融合深度残差网络(Res Nte)和双向长短时记忆循环神经网络(Bi-LSTM)的心电信号分类算法。方法利用Res Nte模型提取和整合心电信号波形特征的能力,在Res Nte模型基础上融合Bi-LSTM模型,提取信号序列上下文信息。通过数据增强技术增加样本数量,解决训练集存在类别不平衡问题,进一步提升模型分类性能。结果文中提出算法的F1(查准率和召回率的调和平均)在The 2017 Physio Nte/Cin CChallen ge数据集上达到0.8571,优于近几年文献提出的其他几种模型。同时将该模型用于医院实际临床数据(92245条记录),模型F1指标达到0.8852;对比传统人工特征方法和单纯残差网络该模型具有更好的性能,尤其对于心房颤动的识别提升明显。结论基于Res Nte和Bi-LSTM模型融合的算法能有效提取心电信号特征,且分类效果较好,具有实际应用意义。Objective To improve the classification of clinical electrocardiogram(ECG),an ECG signal classification algorithm based on fusing deep Residual Network(ResNet)and Bi-directional Long Short-Term Memory(Bi-LSTM)neural network was proposed.Methods The algorithm not only made full use of the ability of the deep convolutional network to extract and integrate ECG signal features,but also added a Bi-LSTM neural network to further enhance the timing feature extraction capability.Results The FI(harmonic average of precision and recall)of the algorithm reached 0.8571 on2017 PhysioNet/CinC Challenge dataset and was superior to other algorithms proposed in recent years.Meanwhile,the algorithm was applied to a certain hospital clinical data(92,245 records),and the F1 reached 0.8852.Compared with the traditional artificial feature methods and the ResNet,this algorithm had better performance,especially in the recognition of Atrial Fibrillation.Conclusion The algorithm based on fusion model of ResNet and Bi-LSTM network can effectively extract ECG signal features,and the classification effect is better.It has practical application significance.

关 键 词：心电图 深度学习 残差网络 长短时记忆网络 信号处理 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]