Abstract: Cardiac auscultation is the basic way for primary diagnosis and screening of Congenital Heart Disease（CHD）. In this paper, 1, 800 heart sound signals are used to train the mainstream deep learning networkssuch as RNN（Recurrent Neural Network）, LSTM （Long Short-Term Memory network）, and CNN（Convolutional Neural Networks）. The result shows that the RNN is prone to over-fitting. The LSTM network losses value of 0.257 and the accuracy of 0.872. The CNN losses value of 0.25 and the accuracy of 0.896. The CNN has greater potential than other two networks. The classification algorithm of CHD based on CNN has the universal applicability due to big training samples. Compared with ordinary networks the CNN has an advantage that is it can extract features automatically. This work is hopeful to be used in machine assisted auscultation of CHD.

Key words: Congenital Heart Disease, deep learning, machine-assisted auscultation, convolutional neural network

摘要： 心脏听诊是先心病初诊和筛查的主要手段。传统心音分类算法普适性差，过程复杂，不利于将来实时化决策。采用1 800个心音信号对几种时间序列分类的主流深度学习网络进行训练，结果显示循环神经网络易出现过拟合；长短时记忆网络分类损失值0.257，准确率0.872；卷积神经网络损失值0.25，准确率0.896。实验表明卷积神经网络相比较其他两种网络具备更大的潜力。基于卷积神经网络的先心病分类算法，因训练样本量大，使网络普适性得到了保证。与其他分类器相比，CNN的另一个优势是其可自动提取特征。该研究有望用于机器辅助听诊。

关键词: 先心病, 深度学习, 机器辅助听诊, 卷积神经网络