类内类间距离在CNN识别心拍类中的应用研究

摘要

Heart disease is a serious threat to human health. Electrocardiogram(ECG)classification of heart beat plays an important role in clinical diagnosis and automatic diagnosis of heart disease. Although the existing ECG heart beat fea-tures generated by deep learning are superior to the heart beat features generated by traditional methods, there are serious data imbalance problems among various types of ECG, resulting in the poor performance of existing heart beat features generated based on deep learning methods. In order to solve this problem, based on the Convolutional Neural Network (CNN), this paper first constructs a common CNN model which can effectively express the common information of all kinds of heart beat on the basis of the equal data of all kinds of heart beat. Then based on the model of minimizing intra-class distances and maximizing inter-class distances, a category CNN model that can effectively reflect the propensity information of the corresponding heart beat categories is constructed on each type of heart beat data. Finally, the output of each heart beat category CNN model is integrated for identification and classification. The experimental results on the MIT-BIH database show that this method achieves 100% of the indicators of the classification of heart beat, solves the problem of automatic identification and classification of ECG four types of heart beats in the MIT-BIH database.%心脏疾病严重威胁人类身体健康,心电图(Electrocardiogram,ECG)心拍分类对心脏疾病的临床诊断和自动诊断具有重要意义.现有基于深度学习生成的ECG心拍特征虽然优于基于传统方法生成的心拍特征,但是因ECG中各类间存在着严重的数据不平衡问题,致使现有基于深度学习方法生成的心拍特征的性能仍不甚理想.针对这一问题,以卷积神经网络(Convolutional Neural Network,CNN)为基础,在各类心拍等量数据基础上构建能有效表达各类心拍共性信息的共性CNN模型,以共性CNN模型和最小化类内距离最大化类间距离模型为基础,分别在各类心拍数据上构建能有效反映相应心拍类别倾向性信息的类别CNN模型,综合各心拍类别CNN模型的输出进行识别与分类.在MIT-BIH数据库上的实验结果显示,该方法识别分类心拍的各项指标均达到100%,解决了MIT-BIH数据库中ECG四类心拍自动识别分类的问题.