基于深度学习的心电信号降噪和T波自动检测研究
发布时间：2018-12-15 16:47
【摘要】：心脏性猝死具有突发性、高发病率和高致死率的特点,是心血管领域关注的热点问题。在远程医疗背景下,动态心电监护是心脏性猝死预测的有效方法。心脏性猝死发生前,动态心电图的Q-T形态和间期时常伴有异常改变,从而使Q-T段的识别和分析成为了心脏性猝死早期预测的关键。其中T波信号微弱,易受噪声干扰,而且形态多变,因此在远程医疗背景下心脏性猝死预测的研究中,心电信号降噪和T波检测逐渐成为研究的重点和难点。本文考虑人体个体差异特征,以及远程医疗背景下信号噪声多、干扰大等因素,利用远程医疗背景下心电信号具有大数据特征的优势,引入深度学习,研究心电信号降噪和T波自动检测算法。主要工作如下:(1)心电信号中部分噪声的频谱和主要信号频谱存在重叠现象,传统的降噪方法很难将其滤除干净。为此,本文利用远程医疗背景下心电信号具有大数据特征的优势,提出了基于降噪自动编码器构建深度神经网络的心电信号降噪算法。通过堆叠多个降噪自动编码器,可以抽象输入信号的深层次特征。利用降噪自动编码器提取信号鲁棒性特征的能力,完成从含噪声信号中重构原始信号的得任务。基于心电信号之间的相似性构建训练数据,调整网络参数,使得构建的神经网络完成心电信号降噪。(2)针对基于降噪自动编码器的心电信号降噪算法中,部分降噪信中含有的锯齿状噪声残留情况,采用小波自适应阈值和压缩降噪自动编码器优化降噪模型。通过在损失函数中增加隐含层输出信号对输入信号雅可比矩阵的Frobenius范数平方项,来抑制网络隐含层过大变动对输出的影响,从而提高了网络降噪的性能。同时,小波自适应阈值法滤除部分噪声,使得较低的样本维度可以包含尽量全面的信号和噪声特征,进而降低网络各层的节点数,简化算法的计算复杂度。(3)现有T波检测算法中,T波形态检测和特征点检测之间是互相影响的,如果得知T波形态就可以提高T波的关键特征点检测精度,但是不知道T波关键特征点的信息又无法判断T波形态。为了解决T波形态检测和特征点检测之间的矛盾关系,本文考虑人体作为复杂的生物体具有个体差异的基本特征,提出了基于形态指导的T波自动检测算法,采用稀疏自动编码器提取T波形态特征,并将其分为单峰直立、单峰倒置、负正双向和正负双向四种类型。随后,根据每一类T波形态的特征,采用倾斜高斯函数进行数学建模。通过分析模板和T波段的相关性,实现T波的特征点检测。为了验证本文的研究成果,将所提的心电信号降噪算法和T波自动检测算法应用于所在科研团队自主研发的智慧心电监测平台中。经过实际采集的心电信号测试表明,本文提出的心电信号降噪算法可以在滤除复杂噪声的同时保持心电信号的主要特征波形。并且,本文提出的基于形态指导的倾斜高斯模板算法,实现了自动检测T波峰值和终点值。研究成果大大提高了远程医疗环境下心电监测系统的智能性。
[Abstract]:Sudden, high and high incidence of sudden cardiac death is a hot issue in the field of cardiovascular attention. In the remote medical background, dynamic ECG monitoring is an effective method of the prediction of sudden cardiac death. Before the occurrence of sudden cardiac death, the Q-T shape and interval of the dynamic electrocardiogram are often accompanied by abnormal changes, so that the identification and analysis of the Q-T segment is the key to the early prediction of sudden cardiac death. In the research of the prediction of sudden cardiac death in the remote medical background, the noise reduction and T-wave detection of the cardiac electrical signal are becoming the focus and the difficulty of the research. In this paper, the characteristics of human individual difference and the multiple factors of signal noise and large interference in the background of remote medical treatment are considered. In this paper, the advantage of the high data characteristic of the ECG signal under the background of remote medical care is used, and the depth study is introduced to study the noise reduction and T-wave automatic detection algorithm of the cardiac electrical signal. The main work is as follows: (1) The spectrum of some noise in the ECG signal and the main signal spectrum are overlapped, and the traditional noise reduction method is difficult to filter out. In this paper, based on the advantage of the high data characteristic of the ECG signal in the remote medical background, this paper puts forward the noise reduction algorithm of the ECG signal based on the noise reduction automatic encoder to construct the depth neural network. By stacking a plurality of noise reduction automatic encoders, the deep-level features of the input signal can be abstracted. The ability of the noise reduction automatic encoder to extract the robust feature of the signal is utilized to complete the task of reconstructing the original signal from the noise-containing signal. The training data is constructed based on the similarity between the electrocardiosignals and the network parameters are adjusted, so that the constructed neural network completes the noise reduction of the electrocardiosignal. (2) In the noise reduction algorithm of the ECG signal based on the noise reduction automatic encoder, the noise residual condition of the saw-tooth noise contained in the partial noise reduction signal is optimized, and a small-wave adaptive threshold and a compression noise reduction automatic coder are adopted to optimize the noise reduction model. By adding the implicit layer output signal to the Frobenius norm square of the input signal Jacobian matrix in the loss function, the influence of the over-large fluctuation on the output of the network hidden layer is suppressed, and the performance of the network noise reduction is improved. At the same time, the small-wave adaptive threshold method filters out some of the noise, so that the lower sample dimension can contain as much as possible signal and noise characteristics as much as possible, thus reducing the number of nodes of each layer of the network and simplifying the calculation complexity of the algorithm. (3) In the existing T-wave detection algorithm, the T-wave shape detection and the characteristic point detection are mutually affected, and if the T-wave form is known, the detection accuracy of the key characteristic points of the T-wave can be improved, but the T-wave shape cannot be judged by knowing the information of the key characteristic point of the T-wave. In order to solve the contradiction between T-wave shape detection and feature point detection, this paper takes into consideration the basic characteristics of human body as a complex organism, and puts forward a T-wave automatic detection algorithm based on morphological guidance. It can be divided into unimodal, unimodal, negative and positive two-way and positive and negative two-way. Then, according to the characteristics of each class of T-wave shape, the oblique Gaussian function is used for mathematical modeling. By analyzing the correlation between the template and the T-band, the feature point detection of the T-wave is realized. In order to verify the research results in this paper, the proposed method of ECG signal reduction and T-wave automatic detection is applied to the intelligent ECG monitoring platform developed by the research team. The results of the real-time ECG signal show that the noise reduction algorithm proposed in this paper can keep the main characteristic waveform of the ECG signal while the complex noise is filtered out. In addition, based on the form-based oblique Gaussian template algorithm, the T-wave peak value and the end-point value are automatically detected. The research results have greatly improved the intelligence of the ECG monitoring system in the remote medical environment.
【学位授予单位】：燕山大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：R540.4;TN911.4
[Abstract]:Sudden, high and high incidence of sudden cardiac death is a hot issue in the field of cardiovascular attention. In the remote medical background, dynamic ECG monitoring is an effective method of the prediction of sudden cardiac death. Before the occurrence of sudden cardiac death, the Q-T shape and interval of the dynamic electrocardiogram are often accompanied by abnormal changes, so that the identification and analysis of the Q-T segment is the key to the early prediction of sudden cardiac death. In the research of the prediction of sudden cardiac death in the remote medical background, the noise reduction and T-wave detection of the cardiac electrical signal are becoming the focus and the difficulty of the research. In this paper, the characteristics of human individual difference and the multiple factors of signal noise and large interference in the background of remote medical treatment are considered. In this paper, the advantage of the high data characteristic of the ECG signal under the background of remote medical care is used, and the depth study is introduced to study the noise reduction and T-wave automatic detection algorithm of the cardiac electrical signal. The main work is as follows: (1) The spectrum of some noise in the ECG signal and the main signal spectrum are overlapped, and the traditional noise reduction method is difficult to filter out. In this paper, based on the advantage of the high data characteristic of the ECG signal in the remote medical background, this paper puts forward the noise reduction algorithm of the ECG signal based on the noise reduction automatic encoder to construct the depth neural network. By stacking a plurality of noise reduction automatic encoders, the deep-level features of the input signal can be abstracted. The ability of the noise reduction automatic encoder to extract the robust feature of the signal is utilized to complete the task of reconstructing the original signal from the noise-containing signal. The training data is constructed based on the similarity between the electrocardiosignals and the network parameters are adjusted, so that the constructed neural network completes the noise reduction of the electrocardiosignal. (2) In the noise reduction algorithm of the ECG signal based on the noise reduction automatic encoder, the noise residual condition of the saw-tooth noise contained in the partial noise reduction signal is optimized, and a small-wave adaptive threshold and a compression noise reduction automatic coder are adopted to optimize the noise reduction model. By adding the implicit layer output signal to the Frobenius norm square of the input signal Jacobian matrix in the loss function, the influence of the over-large fluctuation on the output of the network hidden layer is suppressed, and the performance of the network noise reduction is improved. At the same time, the small-wave adaptive threshold method filters out some of the noise, so that the lower sample dimension can contain as much as possible signal and noise characteristics as much as possible, thus reducing the number of nodes of each layer of the network and simplifying the calculation complexity of the algorithm. (3) In the existing T-wave detection algorithm, the T-wave shape detection and the characteristic point detection are mutually affected, and if the T-wave form is known, the detection accuracy of the key characteristic points of the T-wave can be improved, but the T-wave shape cannot be judged by knowing the information of the key characteristic point of the T-wave. In order to solve the contradiction between T-wave shape detection and feature point detection, this paper takes into consideration the basic characteristics of human body as a complex organism, and puts forward a T-wave automatic detection algorithm based on morphological guidance. It can be divided into unimodal, unimodal, negative and positive two-way and positive and negative two-way. Then, according to the characteristics of each class of T-wave shape, the oblique Gaussian function is used for mathematical modeling. By analyzing the correlation between the template and the T-band, the feature point detection of the T-wave is realized. In order to verify the research results in this paper, the proposed method of ECG signal reduction and T-wave automatic detection is applied to the intelligent ECG monitoring platform developed by the research team. The results of the real-time ECG signal show that the noise reduction algorithm proposed in this paper can keep the main characteristic waveform of the ECG signal while the complex noise is filtered out. In addition, based on the form-based oblique Gaussian template algorithm, the T-wave peak value and the end-point value are automatically detected. The research results have greatly improved the intelligence of the ECG monitoring system in the remote medical environment.
【学位授予单位】：燕山大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：R540.4;TN911.4
【相似文献】
相关会议论文 前5条
1 王s，
本文编号：2380983
本文链接：https://www.wllwen.com/yixuelunwen/xxg/2380983.html
最近更新
教材专著
