基于卷积自编码器网络的脉搏波分类模型-《郑州大学学报(工学版)》

文章信息/Info

Title:: Pulse Wave Classification Model ba＜x＞sed on Convolutional Auto-Encoder

作者:: 逯鹏; 王汉章; 毛晓波; 赵宇平; 刘超; 尚莉伽; 孙智霞; 郑州大学电气工程学院;中医药智能科学与工程技术研究中心;中国中医科学院;北京市东城区中小学卫生保健所;郑州大学第五附属医院;

Author(s):: Ji Peng; Wang Hanzhang; Mao Xiaobo; Zhao Yuping; Liu Chao; Shanglija; Sun Zhixia;; School of Electrical Engineering, Zhengzhou University; Research Center of Intelligence Science and Engineering Technology of Chinese Medicine; Chinese Academy of Chinese Medicine Sciences; Institute of Healthcare, Eastern Primary and Middle Schools in Dongcheng District, Beijing; Fifth Affiliated Hospital of Zhengzhou University;

Keywords:: Cardiovascular disease; pulse wave; convolutional neural network; self -encoder

摘要:: 基于深度学习的脉搏波分类依赖大量标记数据，临床数据有限、标注成本高影响了脉搏波分类识别效果。设计一种基于卷积自编码器网络(Convolutional Autoencoder Networks，CAE-Net)的脉搏波分类模型，首先，利用卷积神经网络(Convolutional Neural Network，CNN)局部特征提取能力和自编码器(Autoencoder，AE)的压缩重构及降维特性构建卷积自编码器(CAE)，结合脉搏波波形特点，在平均绝对误差损失函数中引入脉搏波时域特征约束，提升CAE对脉搏波低维特征自学习能力；其次，重用预训练CAE编码层网络和权重，构建卷积自编码器网络(CAE-Net)，并利用有标记脉搏波数据进行网络微调。通过在心血管疾病数据集上实验表明，CAE-Net的分类准确率为98.00%，F1值达到94.40%，相较于其它脉搏波分类模型，设计网络所提取特征的区分度较高，同时减少了对标注脉搏波数据依赖，在小样本脉搏波数据具有较好分类效果。

Abstract:: Classification of pulse wave ba＜x＞sed on deep learning relies on a large number of labeled data, however, limited clinical data and expensive labeling costs hinder the pulse wave classification and recognition. A pulse wave classification model ba＜x＞sed on convolutional autoencoder networks (CAE-net) is designed in this paper, firstly, the convolutional autoencoder (CAE) is constructed, which combines the local feature extraction ability of convolutional neural network (CNN) and the compression reconstruction and dimension reduction characteristics of autoencoder (AE), and considering the characteristics of pulse wave, the time domain feature constraint of pulse wave is introduced into the mean absolute error loss function to improve the self-learning ability of CAE for low dimensional features Secondly, the CAE-Net is constructed by reusing the coding la＜x＞yer network and weights of the pre-training CAE, then the network is fine tuned by using labeled pulse waves. Experiments on cardiovascular disease dataset show that the classification accuracy of CAE-Net is 98%, and the F1 value is 94.40%. Compared with other classification models, the designed network can extract features with high discrimination, reduces the dependence on the labeled pulse waves, and performs well in the classification of small sample pulse wave data.