摘 要：心音为疾病的诊断提供了初步的线索，有助于医生对疾病的评估，但传统的心音诊断训练费用昂贵，难以推广应用。针对以上问题，本文提出了一种基于CNN-LSTM的心音自动诊断分类方法，并给出了该方法的体系结构。网络结构由两个局部特征学习块和一个长短期记忆层组成，局部特征学习块主要包括一个卷积层和一个池化层。CNN利用卷积层和池化层来学习局部相关性，同时提取层次相关性。LSTM层用于从学习到的局部特征中学习长期相关性。文章中设计的网络可以充分利用这两种网络的优点，克服它们各自的缺点。实验采用了著名的Peter Bentley心音数据集，以梅尔频率倒谱系数作为心音特征，实验结果表明，设计的CNN-LSTM在心音识别中具有较好的效果，准确率约提高5%。所设计的网络结构在Peter Bentley数据集上的识别率达到85.4%，远高于LSTM和CNN分别在同一数据集上获得的准确率75.6%和80.5%。
关键词：特征提取；深度学习；CNN；LSTM；CNN-LSTM；心音分类
中图分类号：TN912.3 文献标识码：A 文章编号：2096-4706（2019）22-0079-05
Research on Heart Sound Classification Based on CNN-LSTM
YU Qiankun，DANG Xin，CHEN Jianfei
（School of Computer Science and Technology，Tianjin Polytechnic University，Tianjin 300387，China）
Abstract：Heart sound provides preliminary clues in the diagnosis of the disease and helps doctors to assess the disease. However，traditional trainning of this skill are expensive and hardly to be widely used. Based on the above problems，a new classification method based CNN-LSTM is proposed for the automatic diagnosis of the heart sound. The network structure consists of two local feature learning blocks and a long-term and short-term memory layer. The local feature learning block mainly consists of a convolution layer and a pooling layer. CNN uses convolution layer and pooling layer to learn local correlation and extract hierarchical correlation at the same time. LSTM layer is used to learn long-term dependencies from the learned local features. The designed network in this paper can take advantage of the strengths of both networks and overcome the shortcomings of them. The widely known Peter Bentley heart sound dataset is used in this paper. The Mel-Frequency Cepstral Coefficients used as heart sound feature in this work. The experimental results show that the designed CNN-LSTM has a good effect in heart sound recognition，and the accuracy is improved by about 5%. The recognition rate of the designed network structure on Peter Bentley data set is 85.4%，which is much higher than the accuracy of 75.6% and 80.5% obtained by LSTM and CNN on the same data set respectively.
Keywords：feature extraction；deep learning；CNN；LSTM；CNN-LSTM；heart sound classification
参考文献：
[1] BALILI C C，SOBREPENA M C C，NAVAL P C. Classification of heart sounds using discrete and continuous wavelet transform and random forests [C]// 2015 3rd IAPR Asian Conference on Pattern Recognition （ACPR），November 3-6，2015. Kuala Lumpur，Malaysia：IEEE，2016.
[2] LEUNG T，WHITE P，COLLIS W，et al. Classification of heart sounds using time-frequency method and artificial neural networks [C]//Proceedings of the 22nd Annual International Conference of the IEEE Engineering in Medicine and Biology Society （Cat. No.00CH37143），July 23-28，2000. Chicago，USA：IEEE，2000.
[3] NILANON T，PURUSHOTHAM S，LIU Y. Normal/Abnormal Heart Sound Recordings Classification Using Convolutional Neural Network [C]//2016 Computing in Cardiology Conference （CinC），September 11-14，2016. Vancouver，Canada：IEEE，2016.
[4] BENTLEY P，NORDEHN G，COIMBRA M. Classifying Heart Sounds Challenge [EB/OL].（2012-03-30）. http://www.peterjbentley.com/heartchallenge/index.html.
[5] CHAKRABORTY K，TALELE A，UPADHYA S. Voice recognition using MFCC algorithm [J]. International Journal of Innovative Research in Advanced Engineering （IJIRAE），2014，1（10）：158-161.
[6] MOHAN B J，RAMESH B N. Speech recognition using MFCC and DTW [C]//2 2014 International Conference on Advances in Electrical Engineering （ICAEE），January 9-11，2014. Vellore，India：IEEE，2014：1-4.
[7] SUJADEVI V G，SOMAN K P，VINAYAKUMAR R，ET AL. DEEP MODELS FOR PHONOCARDIOGRAPHY （PCG） CLASSIFICATION [C]// 2017 International Conference on Intelligent Communication and Computational Techniques （ICCT），December 22-23，2017. Jaipur，India：IEEE 2017.
作者简介：
于乾坤（1992-），男，汉族，河南周口人，硕士研究生，研究方向：自然语言处理和音频信号处理；
党鑫（1983-），男，汉族，天津人，副教授，工学博士，研究方向：音频信号处理、计算医学应用和嵌入式系统开发。