文章摘要

文章信息/Info

Title:: Research on single-lead ECG beat classification based on semantic segmentation

作者:: 王豪1; 廖云朋2; 彭宽1; 黄忠朝1△; （1.中南大学基础医学院，长沙 410013；2.深圳市瑞康宏业科技开发有限公司，深圳 518000）

Author(s):: WANG Hao1; LIAO Yunpeng2; PENG Kuan1; HUANG Zhongchao1; （1. College of Basic Medical Sciences, Central South University, Changsha 410013, China;2. Shenzhen Rencare Co., Ltd, Shenzhen 518000, China）

Keywords:: ECG beat classification; U-Net; Semantic segmentation; Residual connection; Attention mechanism

摘要:: 为从心电图（electrocardiogram，ECG）中准确识别心拍，本研究提出了一种融合残差连接与注意力机制的改进一维U-Net语义分割模型,使用从上万名患者远程动态ECG记录中截取的148 340条单导联ECG数据，对正常窦性心律（Normal）、室性早搏（premature ventricular contraction，PVC）、房性早搏（atrial premature beat， APB）、左束支传导阻滞（left bundle branch block， LBBB）和右束支传导阻滞（right bundle branch block，RBBB）五种常见心拍进行分类。该模型以一定长度的ECG片段作为输入，通过添加背景标签，完成对所有采样点的语义分割，实现对各心拍进行定位的同时，完成类型识别。在测试集上的实验结果表明，该模型能够准确检测心拍位置，仅有0.04%的心拍被漏检；对Normal、PVC、APB、LBBB、RBBB心拍分类的F1分数分别为99.44%、99.03%、97.63%、95.25%和94.77%。该方法与传统U-Net模型相比，能够取得更好的心拍分类效果。

Abstract:: In order to accurately identify the beats from ECG signals, we proposed an improved 1D U-Net semantic segmentation model fusing residual connection and attention mechanism. 148 340 single-lead ECG datas intercepted from remote dynamic ECG records of tens of thousands of patients were used to classify five common beat types: normal sinus beats (Normal), premature ventricular contractions (PVC), atrial premature beat (APB), left bundle branch block (LBBB) and right bundle branch block (RBBB). The model took a certain length of ECG segments as input, and completed semantic segmentation of all sampling points by adding background labels, meanwhile completed type recognition while positioned each beat. The experimental results on the test set showed that the model could accurately detect the position of each beat, and only 0.04% of beats were missed, and the F1 scores of Normal, PVC, APB, LBBB and RBBB were 99.44%, 99.03%, 97.63%, 95.25% and 94.77%, respectively. Compared with the traditional U-Net model, the proposed model achieves better results of beat classification.