2024年4月20日发(作者：u盘启动盘无法进入引导安装系统)

摘要

心血管疾病的患病率和死亡率逐年上升，是当今世界头号致死疾病。心音信号可以

反应心脏的结构和状态信息，在心血管疾病的诊断上有着重要的意义。在心音智能诊断

方面，有少量产品已经问世，但是由于受到算法性能的限制，这些产品只能分辨心音正

常与否，需要医生来确诊异常心音的类别，存在着诊断滞后、成本高的问题。本文针对

以上问题，根据心音的特点，提出了基于深度学习与集成学习的心音分类算法，并研发

了智能听诊系统，具体工作如下：

（1）构建了基于深度学习的心音分类网络（Heart Sound Net，HSNet）：HSNet为

本文从经典卷积神经网络结构AlexNet和VGGNet出发，根据心音时域和频域的特点构

建的一种心音分类网络。本文利用公开数据集进行实验，并与相关文献中的分类结果进

行对比分析。HSNet在正常心音和异常心音的二分类任务中达到了97.32%的准确率，较

改进的AlexNet高0.27%，且HSNet预处理过程简单，易于工程实现。验证HSNet的有

效性后，本文将HSNet用于本实验室自主研发的电子听诊器所采集的心音数据集，在正

常心音、房性早搏、室间隔缺损和其他类别的四分类任务中达到98.72%的准确率。

（2）提出了基于深度学习和集成学习的心音网络提升算法（Heart Sound Net

Boosting，HSNBoost）：针对HSNet调优后性能达到瓶颈的问题，HSNBoost使用深度

学习模型HSNet作为特征提取器，并使用集成学习模型XGBoost（eXtreme Gradient

Boosting）对提取特征后的心音进行分类，在公开数据集和自建数据集上的分类准确率

较HSNet分别提高了1.68%和2.33%。

（3）研发了智能听诊系统：本文研发了智能听诊系统的用户端及服务端。在用户端

实现了用户注册和登录、用户个人信息管理、心音录制和用户问诊的功能。服务端实现

了用户信息管理、HSNBoost预测与结果整合、半监督心音标注和HSNBoost模型更新

的功能。

关键词：心音分类；深度学习；集成学习；智能听诊系统

I

Abstract

The prevalence and mortality of cardiovascular diseases are increasing year by year,

making it the number one deadly disease in the world today. Heart sound signals can reflect the

structure and state of the heart and have important implications for the diagnosis of

cardiovascular disease. In the intelligent diagnosis of heart sounds, some products have been

introduced, but due to the limitations of algorithm performance, these products can only

distinguish whether the heart sounds are normal or not. They need doctors to diagnose the

abnormal heart sounds precisely, and there are problems of diagnosis delay and high cost.

Based on the above characteristics, this paper presents a heart sound classification algorithm

based on deep learning and ensemble learning, and builds an intelligent auscultation system.

The specific work is as follows:

(1)Heart Sound Network (HSNet) based on convolutional neural network was constructed.

Based on the classical convolutional neural network structure AlexNet and VGGNet, this paper

constructs a heart sound classification network based on the characteristics of heart sound time

domain and frequency domain. In this paper, the public data set is used for experiments, and

the classification results in the relevant literature are compared and analyzed. HSNet achieved

an accuracy of 97.32% in the normal classification of heart sounds and abnormal heart sounds,

0.27% higher than the modified Alex Net, and the HSNet preprocessing process is simple and

easy to implement. After verifying the effectiveness of HSNet, this paper train HSNet with

self-built datasets, achieving 98.72% accuracy in normal heart sounds, atrial premature beats,

ventricular septal defects, and other categories.

(2)Heart Sound Net Boosting (HSNBoost) based on deep learning and ensemble learning

was proposed. HSNBoost uses the deep learning model HSNet as the feature extractor, and uses

the ensemble learning model XGBoost(eXtreme Gradient Boosting) to classify the heart

sounds after extracting features. The classification accuracy on the public dataset and self-built

dataset is 1.68% and 2.33% higher than HSNet respectively.

(3)Intelligent auscultation system was developed. This paper develops the user client and

server of the intelligent auscultation system. The functions of user registration and login, user

personal information management, heart sound recording and user inquiry are implemented on

the user client. The server implements the functions of user information management,

HSNBoost prediction and result integration, semi-supervised heart sound label and HSNBoost

model update.

II