融合单导联心电图传统与深度特征的常见心律失常识别方法研究

Research on Intelligent Recognition Method of Common Arrhythmia Combining Traditional and Deep Features of Single-Lead ECG

心律失常是心血管疾病中最为常见的疾病类型之一。基于便携式设备,对少数导联的心电图进行长期智能监护,有利于提高心律失常的检出率,但产生的海量长程ECG数据会给临床医生带来极大的工作负担,也容易导致漏检和误判。为此,提出一种融合单导联心电图传统与深度特征的常见心律失常自动识别方法。首先针对常见心律失常,提取频域、时域和形态上的传统特征;然后搭建残差块深度卷积神经网络和双向长短时记忆网络,用于提取深度特征;接着在通过深度网络融合3种不同尺度的特征后,对常见的心律失常节律和正常节律进行分类识别;最后采用2018年中国生理信号挑战赛和2017年全球房颤挑战大赛分别提供的6877组静态和8528组动态心电图数据来验证所提出的研究方法。在只采用一个导联的静态心电图信号情况下,在分类识别6种心律失常节律和1种正常节律上获得0.855的平均F1分数,优于现有的相关方法;在单导联动态心电图上,新的研究方法在分类识别房颤、其他心律失常和正常节律的平均F1分数为0.827,与2017年全球房颤挑战大赛中两个并列第一的方法性能相当,优于其他相关方法。因此,所提出的研究方法在常见心律失常的辅助诊断和穿戴式远程监护中具有较好的应用前景。

Cardiac arrhythmia is one of the most common types of cardiovascular diseases.Long-term monitoring of few-leads ECGs based on portable devices is helpful to improve the detection rate of arrhythmias.But the large amounts of long-range ECG data generated impose a great burden on clinicians,which leads to missed detection and misjudgments.Therefore,this paper developed an automatic identification method for common arrhythmias by combining features of the traditional single-lead ECG with deep network features.The new method first extracted the traditional features in frequency domain,time domain,and morphology of common arrhythmias.Then a residual block deep convolutional neural network and a bidirectional long-short memory network were built to extract the deep network features.These three types of the features were fused in one deep network to classify heart rhythms including normal and arrhythmias.Finally,6877 sets of static and 8528 sets of Holter data provided by the 2018 Chinese Physiological Signals Challenge and the 2017 Global Atrial Fibrillation Challenge were used to verify the method in this paper.With single-lead of static ECG signal,the method achieved an average F1 score of 0.855 for categorizing six arrhythmic rhythms and one normal rhythm,which is better than the existing relevant methods.As for single-lead dynamic ECG,the method achieved an average F1 score of 0.827 for categorizing AF,other arrhythmias,and normal rhythms,which is comparable to two methods tied for first in 2017 Global AF Challenge and superior to other related methods.Thus,this method has a good prospect of application in wearable remote monitoring and the auxiliary diagnosis of common arrhythmias.