计算机与现代化 ›› 2021, Vol. 0 ›› Issue (08): 52-57.

• 图像处理 • 上一篇    下一篇

基于深度学习的临床心电图分类算法

  

  1. (1.中国科学院大学,北京100049;2.中国科学院微电子研究所,北京100029; 
     3.新一代通信射频芯片技术北京市重点实验室,北京100029)
  • 出版日期:2021-08-19 发布日期:2021-08-19
  • 作者简介:刘守华(1992—),男,安徽六安人,硕士研究生,研究方向:数据挖掘,E-mail: liushouhua@ime.ac.cn; 王小松(1981—),男,副研究员,博士,研究方向:高性能混合信号集成电路,生物电传感技术,物联网技术; 刘昱(1975—),男,研究员,博士,研究方向:高性能模拟,射频CMOS集成电路,硅基毫米波集成电路,超低功耗短距离无线通信系统,物联网相关技术及医疗电子系统集成技术。
  • 基金资助:
    天津市院市合作专项(18YFYSZC00130)

Clinical Electrocardiogram Classification Algorithm Based on Deep Learning

  1. (1. University of Chinese Academy of Sciences, Beijing 100049, China;  
     2. Institute of Microelectronics of Chinese Academy of Sciences, Beijing 100029, China; 
     3. Beijing Key Laboratory of Radio Frequency IC Technology for Next Generation Communications, Beijing 100029, China) 
  • Online:2021-08-19 Published:2021-08-19

摘要: 心电图反映了人体心脏健康状况,是临床诊断心血管类疾病的重要依据。随着心电图数量的快速增长,计算机辅助心电图分析的需求愈加迫切,心电图自动分类作为实现计算机辅助心电图分析不可或缺的技术手段,具有重要的医学价值。由于心电信号非常微弱、抗干扰性差,传统心电图分类算法存在测试集上效果好,实际临床应用效果欠佳的问题。为此,本文研究一种基于多导联二维结构的一维卷积ResNet网络结构,通过平移起始点、“加噪”等数据增强手段增 加训练样本多样性,并采用Focal Loss损失函数优化病人个体的心电图分类模型。该模型利用2万条完整的8导联心电图数据,共计34类心电异常事件进行分类实验,取得了0.91的F1值、93.96%的准确率和87.89%的召回率的分类性能。实验结果表明,该心电图分类算法模型具有较优的深层特征挖掘与分类能力,验证了其在心电异常自动分类上的有效性。

关键词: 深度学习, 残差网络, 卷积神经网络, 心电图, 数据分布, 损失函数

Abstract: Electrocardiogram (ECG) which can reflect the health state of human heart is widely used in clinical examination on heart diseases as an important basis. With the increasing number of ECG data, the demand of  the computer-assisted electrocardiogram analysis has become urgent. Electrocardiogram automatic classification as an indispensable technical means of computer aided electrocardiogram analysis has important medical value. However, because of the weakness and low anti-interference of ECG signal, the traditional ECG classification algorithms have the problems of good effect on test set and poor effect in clinical application. So, this paper introduces a ResNet network structure of one-dimensional convolution based on multi-lead two-dimensional structure, increases the diversity of training samples by means of data enhancement such as translation starting point and adding noise, and uses Focal Loss function to optimize the ECG classification model of individual patients. The model uses 20000 complete 8-lead ECG data and a total of 34 types of abnormal ECG events for classification experiments. The results obtained are: F1 score 0.91, accuracy 93.96%, recall rate 87.89%. Experiment results show the proposed algorithm has better ability of deep feature mining and classification, which verifies its effectiveness in arrhythmia classification.

Key words: deep learning, residual network, convolutional neural network(CNN), ECG, data distribution, loss function