【作者】 袁琦；

【导师】 周卫东；

【作者基本信息】 山东大学 ， 信号与信息处理， 2014， 博士

【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。更多还原

【Abstract】 Epilepsy is a chronic neurological disorder characterized by an ongoing liability to recurrent epileptic seizures that result from abnormal, excessive or hypersynchronous neuronal activity in the brain. As an important tool for research of epileptic seizures, electroencephalogram (EEG) contains a mass of physiological and pathological information which cannot be offered by other physiology methods. The automatic seizure detection based on signal processing and pattern recognition is significant in both relieving heavy workload of doctors and improving the diagnosis efficiency for epilepsy.At present, the application of the nonlinear dynamics in the analysis of EEG signals brings more rich information to the identification of ictal EEGs. But the high computational cost of most of nonlinear EEG features blocked the real-time seizure detection. Meanwhile the conventional methods of "EEG feature extraction+classifier" usually compute several kinds of EEG features, and then organize them as feature vectors or carry out feature selection. Such ways not only increase computation complexity further, but also add the problem of EEG feature selection. This thesis is based on the research of EEG feature extraction, epileptic EEG recognition and automatic seizure detection, and conducts the following studies in the aspects of EEG nonlinear feature extraction, EEG fractal characteristics, and EEG classification via the theory of sparse representation.Firstly, as an important branch of the nonlinear dynamics, the fractal geometry theory is introduced to the analysis and process of EEG signals. The differential box-counting approach for estimating the fractal dimension of an image is improved to fit for one-dimensional EEG signals. Compared to the box-counting dimension, the fractal intercept has stronger discriminatory power for the classification of ictal and interictal EEGs. Afterwards, the blanket technique is introduced to calculate the multi-scale blanket dimension and fractal intercept of EEG signals. It is found that there are obvious changes of their values at different scales before the occurrence of the epileptic seizure.Secondly, the novel seizure detection and prediction methods are proposed based on these EEG fractal features. The EEG fractal intercept via differential box-counting approach is combined with extreme learning machine (ELM) to build a detection method which is suited for multi-channel continuous EEG signals. And then the BLDA classifier is trained to discriminate preictal from interictal patterns of the multi-scale blanket dimension and fractal intercept, which enable to predict seizure events. The experimental results not only proved the validity of the presented EEG fractal features, but also showed the good performance of the proposed seizure detection and prediction methods.Thirdly, a novel EEG classification method based on kernel sparse representation is presented. In the scheme of sparse representation based classification (SRC), a test EEG sample is sparsely represented on the training set by solving l1-minimization problem, and the represented residuals associated with ictal and interictal training samples are computed. The test EEG sample is categorized through the comparison of the residuals. Unlike the conventional EEG classification methods, the calculation and choice of EEG features are avoided in this framework. Moreover, the kernel trick is employed to generate a kernel version of the SRC method for improving the separability between ictal and interictal classes. The kernel SRC method shows superior performance in the experiments of EEG classification.Finally, following the SRC scheme, the l1-minimization problem is replaced with l2-minimization to estimate the sparse coefficients of test EEG sample analytically so that the time-consuming iterative operations are avoided. This improved method emphasizes the key role played by using all training samples to collaboratively represent the test EEG sample, so it is named collaborative representation based classification (CRC) method. Besides, the kernel trick is combined with the CRC method. The decision variable is defined as the residual associated with the non-seizure training samples minus the residual associated with the seizure training samples. And then the post-processing including the smoothing technique is added. Thus a novel seizure detection method for long-term EEG recordings is proposed based on the kernel collaborative representation method. The experiments on the long-term EEG dataset indicate that this detection method not only achieves satisfactory results but also has the potential for the real-time seizure detection.The research work in this thesis contributes to the development of the study of automatic seizure detection in the aspects of technique theories, algorithms and clinical application. This thesis also promotes the research on nonlinear EEG feature extraction, the application of fractal theory in analysis of EEG signals, and EEG classification via sparse representation. Due to the limitation of the used EEG database size, the proposed ictal EEG identification and seizure detection methods are need to be evaluated further on a much larger range of EEG data.更多还原