基于深度学习的异常心拍识别方法研究_毕业论文.pdf

II

摘 要

心血管疾病的检测和预防是当下研究的热门。由于便利、经济等特点，心电信

号的检测成为了辅助医生诊断患者是否患有血管疾病的重要方法。传统的检测方

法需要临床医生根据自己的经验对心电信号进行分类，但医生对大量的数据进行

检测时容易产生疲劳，导致漏诊、误诊。针对以上问题，本文设计出了心电信号自

动检测方法，可以成为辅助临床医生检测的工具。随着对心电信号研究的深入，科

学人员发现，患者在发生睡眠呼吸暂停（Sleep apnea，SA）事件时，其心电信号会

(2) 为了将心电信号分成正常心拍或异常心拍，本文设计出了心电信号二分类

方法。现有深度学习方法在处理具有特异性的心电信号数据集时，所采用的交叉熵

损失寻找到分界面不会进一步优化，导致在心电信号这种具有特异性的数据集上

泛化能力较弱，针对这一不足，本文设计了一种基于结构风险最小化（Structural

Risk Minimization，SRM）的卷积神经网络（Convolutional Neural Network，CNN），

提升了模型对病人间数据集的泛化能力。相比于传统方法，所 设计的方法进一步提

升了分类器的性能。

(3) 在对心电信号进行二分类基础上，本文设计出了心电信号多分类方法。首

先，为了减轻传统网络进行多次卷积时导致的梯度消失问题，所设计的方法将每层

卷积的结果和上一层卷积的结果进行融合。接下来，针对使用的数据集类别严重不

以减轻数据不平衡的影响。实验结果表明，本文方法与其他方法相比具有更高的准

确率。同时，还有效解决了类别不平衡导致的敏感性过低问题。

关键词：心电信号；小波变换；卷积神经网络；多尺度残差网络；focal loss 函数

III

ABSTRACT

The detection and prevention of cardiovascular disease is a hot research topic. Because

of its convenience and economy, Electrocardiogram (ECG) detection has become an

important method to assist doctors in diagnosing patients with vascular diseases.

Traditional detection methods require clinicians to classify ECG signal according to their

own experience, but doctors are prone to fatigue when detecting a large number of data,

resulting in missed diagnosis and misdiagnosis. Aiming at the above problems, this thesis

clinicians in detection. With the in-depth study of ECG, scientists have found that the

ECG of patients will change regularly in the event of sleep apnea. According to this

change, we can judge whether the patient has sleep apnea. Therefore, this thesis proposes

a method to detect sleep apnea based on ECG signal.

The main work of this thesis is as follows:

(1) According to the sampling frequency and noise characteristics of ECG signal, this

learning methods are used to deal with the specific ECG data set, the cross entropy loss

adopted to find the interface will not be further optimized, resulting in the weak

generalization ability on the specific ECG data set. Aiming at this deficiency, this thesis

proposes a convolutional neural network (CNN) based on structural risk minimization

(SRM), which improves the generalization ability of the model to the data set between

patients. Compared with the traditional methods, the proposed method further improves

the performance of the classifier.

(3) Based on the two-classification of ECG signal, a multi-classification method of

ECG signal is proposed in this thesis. First, in order to alleviate the gradient disappearance

problem caused by multiple convolutions in traditional networks, the proposed method

IV

inputs. Then, the multi-scale residual network is used to extract the features of the derived

signal to obtain the sensitive features from different perspectives. Finally, the focal loss

function is introduced to replace the traditional cross entropy loss function, which makes

the model pay more attention to the learning of difficult samples in the training stage, so

as to reduce the impact of data imbalance. Experimental results show that this method has

higher accuracy than other methods. At the same time, it also effectively solves the

problem of low sensitivity caused by category imbalance.

KEYWORDS: Electrocardiogram; Wavelet Transform; Convolutional Neural Network;

Multi-Scale Residual Network; Focal Loss Function

V

目 录

1.3 论文主要研究内容与结构 ............................................................................... 6

第二章 心电信号及其预处理 ........................................................................................ 8

2.1 引言 ................................................................................................................... 8

2.2 心电信号的产生 ............................................................................................... 8

2.3 心电信号的去噪 ............................................................................................... 9

2.3.1 心电信号噪声分析 .............................................................................. 10

2.3.2 基于中值滤波去除基线漂移 .............................................................. 10

2.3.3 基于小波变换的去噪方法 ................................................................... 11

2.3.4 小波基数的选择 .................................................................................. 12

2.4 常见心律失常原因分析 ................................................................................. 14

3.2 数据集的处理 ................................................................................................. 17

3.3 基于传统机器学习的异常心拍分类 ............................................................. 18

3.3.1 支持向量机 .......................................................................................... 18

3.3.2 随机森林 .............................................................................................. 18

3.4 基于结构风险最小化的异常心拍分类 ......................................................... 19

3.4.1 卷积神经网络基本理论 ...................................................................... 19

3.4.2 基于结构风险最小化的损失函数 ...................................................... 20

VI

第四章 心电信号的多分类方法 .................................................................................. 26

4.1 引言 ................................................................................................................. 26

4.2 数据集预处理 ................................................................................................. 26

4.4 实验结果展示 ................................................................................................. 31

4.4.1 实验数据划分 ...................................................................................... 31

4.4.2 实验结果与分析 .................................................................................. 32

4.5 小结 ................................................................................................................. 36

第五章 睡眠呼吸暂停检测 .......................................................................................... 37

5.1 引言 ................................................................................................................. 37

5.2 数据集处理 ..................................................................................................... 38

5.2.1 信号去噪 .............................................................................................. 38

5.2.2 R 峰定位及派生信号提取 .................................................................. 38

5.2.3 数据不平衡处理 .................................................................................. 39

5.4.3 判断患者是否患有睡眠呼吸暂停 ...................................................... 45

5.4.4 在 UCD 数据集上测试模型 ................................................................ 45

5.4.5 同类研究结果对比 .............................................................................. 46

5.5 小结 ................................................................................................................. 47

第六章 总结与展望 ...................................................................................................... 48

6.1 全文的总结 ..................................................................................................... 48

6.2 对未来的展望 ................................................................................................. 49