【作者】 楼恩平；

【导师】 张胜；

【作者基本信息】 浙江师范大学 ， 计算机软件与理论， 2009， 硕士

【摘要】 精神抑郁症是一种常见的、慢性复发性疾病,表现为心境显著和持久的低落,伴有相应的思维和行为的改变。研究表明,抑郁症患者脑电信号(EEG)在节律、波形幅度和功率谱等参数中存在着不同于健康人的特征。自发脑电信号含有丰富的频率成分,不同的生理状态和病因下某些频段的能量在头皮不同区域的分布会发生变化,因此可以提取不同频段上的能量作为分类器的特征参数实现抑郁症自发脑电信号的分类。本文使用小波分析和特征向量法对抑郁症患者及健康人脑电信号进行特性分析和特征提取,结合支持向量机分类器实现两类信号的准确分类。具体内容如下:1.通过小波变换和小波包变换对原始采样信号进行多尺度小波分解,得到的不同尺度的频带分量,提取EEG在不同频段上的能量特征。小波变换是一种多尺度信号分析方法,具有良好的时频局部化特性,非常适合于分析像EEG这样非平稳信号的瞬态特性和时变特性。小波包分解具有任意多尺度特点,避免了小波变换固定时频分解的缺陷(如高频段频率分辨率低),为时频分析提供了极大的选择余地,更能反映信号的本质和特征。2.将特征向量功率谱估计法应用到抑郁症脑电信号的特征提取中,我们对试验对象的脑电信号功率谱幅度进行统计分析,选取功率谱幅度的最大值、最小值、平均值和标准偏差作为信号的分类特征参数。特征向量法功率谱估计是基于矩阵特征分解的一种非参数建模谱估计方法,它主要适用于混有白噪声的正弦信号的频率估计及功率谱估计,甚至对于信噪比很低的信号,也能取得很高的谱分辨率。3.在完成抑郁症患者和健康人脑电信号特征提取的基础上,我们使用支持向量机对这两类信号进行分类研究。支持向量机是根据统计学理论提出的一种机器学习方法,它集成了最大间隔超平面、Mercer核、凸二次规划和松弛变量等多项技术。支持向量机的方法根据结构风险最小化原则,提高学习机的泛化能力,它将优化问题转化为求解一个凸二次规划的问题,二次规划所得的解是唯一的且为全局最优解,这样就不存在一般神经网络的局部极值问题。支持向量机由于较好地解决了小样本、非线性、高维数、局部极小点等实际问题,在若干具有挑战性的应用中,获得了目前为止最好的性能。支持向量机已经逐渐成为解决模式分类问题的首选工具。实验结果表明,采用以上三种特征提取方法提取的特征参数作为分类器的输入向量均可以取得理想的效果,抑郁症脑电信号分类准确率达87%以上。该研究成果将为精神抑郁症的病理临床诊断提供一种新的途径,为下一步基于自发脑电的抑郁症疾病诊断实用系统开发奠定基础。更多还原

【Abstract】 Melancholia is a kind of common dysfunction disease characterized by an obvious reduction in intellectual and physiological vigor. Some related studies found that electroencephalogram (EEG) signals of melancholic differ from that of healthy persons in rhythm, wave amplitude and power spectrum amplitude. A number of frequency components are included in spontaneous EEG and the energy corresponding to different frequency bands, which is detected in different physiological states and pathogeny, changes with the scalp area. Thus the energy corresponding to a certain frequency sub-band can be taken as a feature parameter of the classifier to realize the classification of melancholia EEG. In this paper, the use of wavelet analysis and eigenvector estimation was proposed for the extraction of discriminating features from melancholia and nomal people’s EEG. Subsequently, we achieved the classification combining with SVM. The main work done in this dissertation is as follows:1. The EEG recordings were decomposed into various frequency bands through multiscale decomposition by the method of wavelet transaction (WT) and wavelet package transaction (WPT) respectively. And then, we extracted the energy feature using wavelet coefficients. WT is a multiscale signal analysis method which key feature is the time -freqency localsation, and it is suitable for capturing transient nature of nonstationary EEG signals. With the trait of arbitrary multiscale decomposition, WPT cover the shortage of fixed time-frequency decomposition in WT (i.e. poor frequency resolution for high frequency component). Therefore, WPT has better time-frequency charactristic and provides more choice in time-frequency signal analysis.2. By applying the method of eigenvector estimation to the feature extraction of EEG signal, we carry a statistical analysis on the EEG power spectrum amplitude. And then, we take the maximum, minimum, mean and standard deviation of EEG power spectrum amplitude as characteristic parameters. Eigenvector estimation is a non-parametric method based on an eigen-decomposition of the correlation matrix of the noise-corrupted signals. It is best suited to the signals assumed to be composed of several specific sinusoids buried in noise. Even when the signal-to-noise ratio (SNR) is low, the eigenvector estimation can still obtain a high resolution of frequency spectra.3. Having finished the feature extraction of the melancholic and healthy persons’ EEG, we achieved the classification of these two kinds of EEG by Support Vector Machine (SVM). SVM is a machine learning method based on statistics theory, which includes a number of techniques such as the largest interval hyper plane, Mercer kernel, convex quadratic programming and relaxation variables etc. According to the principle of minimizing structural risk, SVM enhances the generalization capability of learning machine and converts the optimization problem into a convex quadratic programming problem. Since the solution of this convex quadratic programming problem is unique and global, local extremum problem existing in general neural networks doesn’t occur. Since practical problems such as nonlinear problem, high dimension problem and local extremum problem have been resolved, SVM obtains the best performance in a variety of practical applications with much challenge. Consequently, SVM has gradually become a superior tool to solve the problem of pattern classification.Experiments demonstrate that taking the feature parameter, which is extracted by the above three feature extraction methods, as the input eigenvector can achieve ideal clsssification accuracy which arrives 87%. This paper presented a new method for melancholia diagnose. The present research provides a basis for the ongoing study "research of melancholic diagnose based on spontaneous EEG".更多还原