没有合适的资源?快使用搜索试试~ 我知道了~
资源推荐
资源详情
资源评论
III
Abstract
The cardiovascular disease is the leading cause of death currently and has aroused
health risks worldwide. Arrhythmia detection has a high medical and social value in the
field of cardiovascular disease diagnosis and prevention. The electrocardiogram(ECG)
signal, the most sensitive signal for detecting cardiac arrhythmias, is the most
convenient and widespread of the many available methods for detecting cardiovascular
disease. However, the lack of medical resources and the imbalance in the number of
doctors and patients result in a large number of patients not being detected in time.
Therefore, It is important to develop a computer-based algorithm that can automatically
detect a wide range of arrhythmic disorders to assist healthcare professionals in
furthering a rapid understanding of a patient's physical condition.
This research belongs to the emerging field of medical-industrial interaction, with
the starting point of achieving automated detection of the multi-lead electrocardiogram
signal analysis, hoping to deepen the technical performance of multi-class ECG health
monitoring instruments and provide an efficient and accurate algorithmic solution to
play an active role in the prevention and treatment of cardiovascular diseases in China.
Therefore, this paper provided a novel hybrid network model based on a deep learning
network and a traditional morphological feature selection network, which used the
CPSC2018 database published in the Biomedical Signals Challenge release, for the
automatic classification and detection of multiple arrhythmia diseases with the
multi-lead ECG signal. The main work accomplished were as follows:
(1) This paper designed a pre-processing algorithm that is fully driven by the raw
ECG signals from the 12-lead dataset to enhance the universality and clinical
applicability of the study. Taking into account the variable length of each lead signal and
its susceptibility to interference, the various types of noise were removed to ensure that
a ECG signal with high quality would be used for subsequent training. At the same time,
the raw signals of different lengths were sliced, thus ensuring compatibility in signal
timing length and robustness in the detection of sudden heart rhythms.
(2) In the arrhythmia multi-classification task, this paper proposed a novel hybrid
IV
network model as the main study of the ECG signal classification algorithm. This
hybrid network model was formed by fusing a deep learning network and a traditional
morphological feature selection network. Firstly, the ATI-CDBiLSTM deep learning
network was used to learn high-dimensional depth features at different scales. Among
them, the CDBiLSTM module was used to learn spatial and temporal fusion
information on the ECG signal, and combined with the Attention module to adjust the
weight distribution among different features, which solved the inadequate use of the
ECG information and difficult-interpretation of features by current LSTMs models.
Secondly, inspired by the ability of traditional morphological features to reflect actual
diagnostic information, this paper introduced a feature selection network using the
random forests. Based on the different importance of lead II and other leads in the
detection, two feature selection strategies were employed to obtain the best set of
traditional morphological features, which gained the best classification performance
with the least amount of features. Finally, the features extracted from the above two
models are combined and trained by the XGBoost classifier to achieve more accurate
classification. The above models are also analyzed and the superiority of this new
hybrid network is verified by relevant comparisons.
The proposed ATI-CDBiLSTM network achieved a macro F1 score of 0.802 for
this classification task, while the macro F1 obtained using the traditional morphological
feature selection network is 0.793 and the macro F1 of the new hybrid network model is
0.824. The results showed that the combination of CNNs and LSTMs were highly
superior in processing temporal signals, and the Attention mechanism substantially
improved the ST segment anomalies where the classification results have always been
poor. The traditional morphological features introduced also increased the ability to
classify certain classes, complementing the comprehensiveness of the classification. The
advantage of this novel hybrid network algorithm solution was its ability to handle
temporal signals of different lengths and achieve excellent automatic detection and
classification capabilities, which has significant clinical implications and provided a
portable solution in the construction of wearable, home-based ECG monitors.
Keywords: 12-lead Electrocardiogram, Arrhythmia Detection, Bidirectional Long
Short-term Memory(BiLSTM), Attention Module, Model Fusion
目 录
摘 要
....................................................................................................... I
Abstract................................................................................................. III
第 1 章 绪论...........................................................................................1
1.1
课题的研究背景及意义
............................................................................ 1
1.2 心电图异常分类算法研究现状................................................................ 2
1.3 研究重难点及内容安排............................................................................ 5
1.3.1
研究重难点
..................................................................................... 5
1.3.2 内容安排......................................................................................... 6
第 2 章 课题研究理论基础.................................................................. 9
2.1
心电信号基础原理
.................................................................................... 9
2.1.1 心电图基础理论............................................................................. 9
2.1.2 多导联心电信号的形态特征....................................................... 10
2.1.3
常见心律失常
ECG
图谱
............................................................. 12
2.1.4 常见心电干扰............................................................................... 13
2.2 深度学习的基础理论的基础理论.......................................................... 15
2.2.1
卷积神经网络
............................................................................... 15
2.2.2 双向长短时记忆网络................................................................... 15
2.2.3 注意力机制................................................................................... 17
2.3
本文使用的
HRV
特征参数
.................................................................... 19
2.3.1 HRV 的时域特征........................................................................... 20
2.3.2 HRV 的频域特征........................................................................... 20
2.3.3 HRV
的统计学特征
....................................................................... 21
2.3.4 HRV 的非线性特征....................................................................... 22
2.4 特征选择.................................................................................................. 24
2.4.1
决策树
........................................................................................... 25
2.4.2 随机森林....................................................................................... 25
2.5 XGBoost 模型融合................................................................................... 27
2.6 本章小结.................................................................................................. 30
第
3
章 心律失常分类研究框架及预处理
........................................ 31
3.1 总体流程图.............................................................................................. 31
3.2 实验数据.................................................................................................. 32
3.2.1
实验环境
....................................................................................... 32
3.2.2 CPSC2018 数据集介绍................................................................. 33
3.2.3 数据集的训练与测试................................................................... 34
3.3
预处理方案设计
...................................................................................... 34
3.4 预处理算法方案...................................................................................... 35
3.4.1 小波变换....................................................................................... 35
3.4.2
平稳小波变换去除运动伪影
....................................................... 37
3.4.3 基于经验模态分解的 QRS 定位算法......................................... 39
3.4.4 基于小波阈值去噪....................................................................... 41
3.5
数据均衡处理及归一化
.......................................................................... 42
3.6 分类评价指标.......................................................................................... 44
3.7 本章小结.................................................................................................. 46
第
4
章 基于
CPSC2018
数据集的心律失常分类研究
.................... 47
4.1 问题描述与模型分析.............................................................................. 47
4.2 基于 ATI-CDBiLSTM 模型实验与分析.................................................47
4.2.1
深度学习模型
............................................................................... 47
4.2.2 实验结果及分析........................................................................... 51
4.3 传统形态学特征提取模型实验与分析.................................................. 54
4.3.1
两种特征选择策略
....................................................................... 54
4.3.2 形态学特征实验分析................................................................... 58
4.4 混合网络实验结果分析.......................................................................... 61
4.4.1
结合形态学特征后的模型性能分析
........................................... 62
4.4.2 与其他识别算法的对比分析....................................................... 64
4.5 本章小结.................................................................................................. 65
第 5 章 总结与展望.............................................................................67
5.1
总结
.......................................................................................................... 67
5.2 展望.......................................................................................................... 68
参考文献...............................................................................................69
致谢
....................................................................................................... 77
攻读硕士学位期间科研成果.............................................................. 79
剩余80页未读,继续阅读
资源评论
2201_75761617
- 粉丝: 20
- 资源: 7339
上传资源 快速赚钱
- 我的内容管理 展开
- 我的资源 快来上传第一个资源
- 我的收益 登录查看自己的收益
- 我的积分 登录查看自己的积分
- 我的C币 登录后查看C币余额
- 我的收藏
- 我的下载
- 下载帮助
安全验证
文档复制为VIP权益,开通VIP直接复制
信息提交成功