基于卷积神经网络的T波形态分类

T Wave Shape Classification Based on Convolutional Neural Network

T波形态分类有助于诊断心肌缺血、急性心包炎和心脏猝死等疾病,是心电图远程监控中一个重要的研究课题.传统的T波分类算法依赖于T波检测,在准确定位T波的关键点之后再提取T波特征,完成分类.但是由于T波位置可能发生一定程度偏移,T波的形态多变且受到多种噪声的干扰,T波检测是一个难题.为了解决上述问题,本文提出基于卷积神经网络的T波分类算法:首先根据QRS波群位置及医学统计规律确定一个T波候选段,然后采用卷积神经网络直接完成T波分类.由于卷积神经网络有稀疏连接、权值共享的特性,能够通过训练自动获取T波特征,并且其特征对微小平移具备不变性且对噪声不敏感,从而能够有效解决T波形态分类问题.最后在MIT-BIH QT心电数据库上对本文方法进行测试,实验结果表明,本文方法可以在T波起始点未确定的情况下,能够识别单峰直立、单峰倒置、低平、负正双向、正负双向五类T波形态,正确率达到了99.1%.

T wave shape classification which is helpful for the diagnosing of many cardiovascular diseases such as my- ocardial ischemia, acute pericarditis and sudden cardiac death, is an important research topic in electrocardiogram remote monitoring. The method of traditional T wave shape classification is based on the accurate detection of the T wave. It is implemented after the T wave delineation and feature extraction. However, T wave detection is difficult because of the position shift, morphologic variation and multi-noise. To resolve this problem, this paper proposes to classify T wave shape based on convolutional neural network. In the new method, firstly, a candidate data segment which contains the T wave is intercepted based on the location of the QRS wave and the medical statistical knowledge. Then the T wave is classified directly based on the convolutional neural network. Due to the advantages of sparse connection and weight share, the convolutional neural network can extract T wave feature by data training and it is robust to the poison shift and noise. So the convolutional neural network can resolve the T wave shape classification problem efficiently. The new method is tested on the MIT-BIH QT database; the experimental results show that the new method performs well in T wave shape classification without T wave delineation and the classification accuracy is 99.1%.