癫痫脑电的分类识别及自动检测方法研究

发布时间：2018-01-26 02:01

本文关键词： 脑电信号癫痫发作分形特征微分盒维毯子覆盖技术稀疏表示核函数技术协作表示　出处：《山东大学》2014年博士论文　论文类型：学位论文

【摘要】：癫痫发作是脑内神经元阵发性异常超同步化电活动的临床表现,具有反复性、突发性和暂时性等特点。作为研究癫痫发作特征的重要工具,脑电图所反映的发作信息是其他生理学方法所不能提供的。利用信号处理技术和模式识别方法自动检测癫痫脑电信号,对于减轻医生负担并提高癫痫的诊断效率具有重要意义。目前,在脑电信号的分析研究中,非线性动力学的应用为癫痫脑电的识别提供了更加丰富的重要信息,但是多数非线性脑电特征具有较复杂的计算过程,无法保证检测算法的实时性。同时,传统的“脑电特征提取+分类器”的自动检测方法会提取多个脑电特征,然后组成特征向量或进行特征选择,这样进一步加剧了算法的计算复杂度,并且增加了特征选取的难题。本文立足于脑电信号的特征提取、分类识别和癫痫发作的自动检测的研究,围绕脑电信号的非线性特征提取、分形特性以及基于稀疏表示的脑电分类等内容展开以下研究：首先,本文将非线性动力学的重要分支——分形几何理论应用到脑电信号的分析与处理中。将常用于图像分形计算的微分盒维算法引入到一维脑电信号的分形分析中,计算了脑电信号的盒维数及其分形截距,并发现与盒维数相比,其分形截距能够更好的区分癫痫发作期和间歇期的脑电。之后,本文又通过改进毯子覆盖技术计算出脑电信号的多尺度毯子维及其分形截距,并发现在不同尺度上它们在临近癫痫发作前均会出现明显变化。其次,本文基于所提出的脑电分形特征进一步提出了癫痫发作检测与预测方法。将脑电信号的微分盒维的分形截距作为其非线性特征,然后结合极端学习机(ELM)分类器,提出了一种适于多导长程脑电的癫痫发作检测方法。采用BLDA算法对脑电的多尺度毯子维及其分形截距在发作前期的变化进行检测,从而实现了对癫痫发作的预报。实验验证的结果不仅说明了本文所提出的脑电分形特征的有效性,而且体现了所提出的检测和预测方法的良好性能。再次,本文依据稀疏表示分类方法,提出了一种基于Kernel稀疏表示的癫痫脑电识别算法。在该方法框架中,先通过求解最小l1范数优化问题求得待测脑电在脑电训练集上的稀疏表示系数,然后,分别计算发作期训练样本和间歇期训练样本对待测脑电的稀疏表示重构误差,通过比较误差的大小来确定待测脑电的类别。与常见的“脑电特征提取+分类器”的脑电分类方法不同,基于稀疏表示的脑电识别方法避免了脑电特征提取和选择的问题,更加完整地保留了脑电信号所携带的信息。为了进一步提高识别效果,本文将核函数技术与稀疏表示分类方法相结合,通过预先增强脑电样本的可分性来进一步提高对癫痫脑电的识别率。实验结果表明,基于Kernel稀疏表示的脑电分类方法取得了更加理想的分类性能。最后,在基于稀疏表示的癫痫脑电识别方法的基础上,进一步将计算待测脑电稀疏表示系数过程中所利用的最小l1范数优化问题替换为最小l2范数优化问题,从而可以通过正则化最小二乘算法(Regularized Least Square, RLS)解析地求得待测脑电的稀疏系数,避免了复杂的迭代运算,大大降低了算法的复杂性。由于改进后的方法强调来自所有类别的训练样本对测试样本的协作表示所起到的关键作用,因此称为协作表示分类方法。同样,本文将核函数技术与协作表示分类方法相结合,并且将两类脑电训练样本所对应的重构误差相减,所得的差值作为输出的决策变量,从而引入了平滑滤波等后处理环节,提出了较为完善的基于Kernel协作表示的癫痫发作检测方法。利用连续长程脑电数据对该方法的性能进行评价,实验发现,所提出的检测方法不但取得了较理想的检测结果,而且其较快的运算速度基本符合实时在线的发作检测的需求。本文的研究工作将有助于进一步推动癫痫自动检测在技术理论、算法和临床应用方面的研究,对于脑电信号的非线性特征提取、分形理论在脑电分析中的应用以及脑电信号的稀疏表示分类方法起到了积极的推进作用。由于实验所用脑电数据的局限性,本文所提出的几种癫痫脑电识别和自动发作检测方法还需要更大量的临床脑电数据来进一步验证它们的性能。
[Abstract]:A seizure is a clinical manifestation of brain neurons abnormal paroxysmal synchronized electrical activity has repeatedly, sudden and temporary. As an important tool of epilepsy, EEG information reflects the attack is methodology can provide other physiological. Automatic detection of epileptic EEG signal processing and utilization the pattern recognition method to reduce the burden on doctors and has important significance to improve the efficiency of diagnosis of epilepsy.
At present, the research on analysis of EEG signals, provides important information more abundant application of nonlinear dynamics for the identification of epileptic EEG, but most of the nonlinear characteristics of EEG with the calculation process is complicated, the real-time detection algorithm can not be guaranteed. At the same time, the traditional "EEG feature extraction + classifier" automatically detection method can extract multiple EEG features, then feature vector or feature selection, which further exacerbated the computational complexity of the algorithm, and increase the problem of feature selection and feature extraction. Based on the EEG signal, the automatic detection of the recognition and classification of seizures, nonlinear feature extraction on EEG signal, fractal characteristics and sparse representation based classification of EEG content following research:
First of all, this will be an important branch of nonlinear dynamics, fractal geometry theory is applied to the analysis and processing of EEG signals. The fractal analysis of differential box counting algorithm is commonly used in the calculation of fractal image into one-dimensional EEG, EEG signal box dimension and fractal intercept were calculated, and compared with the box the dimension of the EEG and intermittent period between the epilepsy better fractal intercept attack. Later, this paper improved blanket technology to calculate the multi-scale blanket dimension EEG and fractal intercept, and find the different scale of them near the seizure before there will be significant changes.
Secondly, the fractal characteristics of EEG based on the proposed detection and prediction of seizures. The fractal intercept differential box dimension of EEG signal as its nonlinear characteristics, and then combined with the extreme learning machine (ELM) classifier is proposed, which is suitable for the long time EEG seizure detection method. The BLDA algorithm is used for EEG multiscale blanket dimension and its fractal intercept were detected in the early attack changes, so as to achieve seizure prediction. Experimental results not only illustrate the effectiveness of the fractal characteristics of EEG in this paper, but also reflects the good performance of detection and prediction of the proposed method.
Again, based on the sparse representation classification method, this paper proposes a new Kernel based on sparse epileptic EEG recognition algorithm. In this method framework, first by solving the minimum L1 norm optimization problem to obtain the measured EEG EEG in the sparse representation coefficient, the training set is then calculated respectively the training sample and attack the intermittent period of training samples sparse EEG to said reconstruction error, by comparing the size of the error to determine the type of EEG measured. With the usual "EEG feature extraction + Classifier" EEG classification methods, EEG recognition method based on sparse representation avoids the problem of feature selection and extraction of brain power more, to retain the integrity of the EEG information carried by. In order to further improve the recognition effect, the kernel function and the sparse combination classification method, through the pre enhanced EEG samples can be divided into It further improves the recognition rate of epileptic EEG. Experimental results show that the EEG classification method based on Kernel sparse representation achieves a better classification performance.
Finally, on the basis of epileptic EEG recognition method based on sparse representation on further calculating EEG sparse representation of the minimum L1 norm optimization problem with minimum L2 norm optimization problem using coefficient process can be obtained by regularized least squares algorithm (Regularized Least Square, RLS) to obtain the sparse coefficient of EEG the measured analytically, avoid the complex iterative operation, greatly reduces the complexity of the algorithm. The improved method emphasizes collaboration from all categories of training samples of the test sample said the key role played by the so called collaborative representation classification method. Also, the kernel technology and collaboration said according to the classification method, and the corresponding two kinds of EEG training sample reconstruction error subtraction, the difference of output as the decision variables, then the smoothing filter at Physical link detection method is proposed based on Kernel collaboration said relatively perfect seizures. To evaluate the performance of the continuous long Cheng Nao electricity data of the experiment found that the method not only achieved the ideal results, but its fast basically meets the real-time online attack detection needs.
This research work will help to further promote the automatic detection of epilepsy in theory, algorithm research and clinical application of the nonlinear feature extraction for EEG signal, sparse fractal theory in application to the analysis of EEG and EEG signal classification method that has played a positive role in promoting. Due to the limitation of the brain the data used in the experiment, this paper proposed several kinds of epileptic EEG recognition and automatic seizure detection methods still need more clinical EEG data to verify their performance.

【学位授予单位】：山东大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：TN911.7

【参考文献】