癫痫发作自动检测算法研究

发布时间：2017-12-28 09:23

  本文关键词：癫痫发作自动检测算法研究　出处：《山东大学》2016年博士论文　论文类型：学位论文

  更多相关文章： 脑电信号 癫痫检测 小波变换 扩散距离 协作表示 核方法 对称正定矩阵 稀疏表示

【摘要】：癫痫是一种由脑部神经元群阵发性异常电活动导致的慢性神经系统疾病,其发作具有突发性、反复性特点,并伴随意识丧失、昏厥、四肢抽搐等临床表现,严重危及人们的身心健康与生命安全。据统计,全球有超过1%的人口遭受该疾病的困扰。癫痫发作的病因复杂多样,发病机制迄今尚未完全明确。作为研究癫痫疾病的一种重要手段,脑电图利用电极记录脑部神经细胞的电活动,包含了大量的生理与病理信息,在癫痫的临床诊断、病灶定位与治疗等方面都发挥着极其重要的作用。目前,脑电图分析和癫痫检测主要依靠医务人员根据临床经验视觉观察来完成。然而,庞大的数据量使得该项工作非常枯燥与耗时,且医务人员的主观性对癫痫发作的判断也会造成影响。因此,利用计算机自动分析脑电信号并有效检测癫痫发作是十分迫切和必要的,它不仅可以减轻医生的负担,提高癫痫的诊断效率,而且在有效治疗癫痫疾病,改善患者生活质量,并深入揭示癫痫发病机制等方面具有重大意义。本文立足于癫痫自动检测这一课题进行相关研究,围绕不同时期脑电信号非相似度距离特征提取,基于稀疏表示和协作表示理论的脑电信号分类等内容展开研究,并提出了几种有效的癫痫自动检测算法。本论文的研究内容与创新点主要包括以下几点：(1)将距离测度应用到脑电信号分析中,提出了基于扩散距离与贝叶斯线性判别分析(BLDA)的癫痫发作检测算法。该算法应用小波变换对脑电信号进行时频分析,并将三个频段的脑电信号组合构成脑电分布。然后,计算癫痫发作期与间歇期脑电分布之间的扩散距离,定量描述不同时期脑电信号之间的差异性。根据同类别脑电信号之间的差异性低于不同类别脑电信号的原则,将扩散距离作为脑电特征与BLDA分类器相结合,实现癫痫发作的识别与检测。与推土机距离(EMD)相比,扩散距离不仅能有效地区分发作期脑电信号和间歇期脑电信号,而且具有更强的抗噪性与更低的计算复杂度。BLDA通过正则化方法避免了在含噪声的数据集上出现过拟合问题,可以得到更好的分类效果。在Freiburg长程脑电数据库上的实验结果验证了脑电信号扩散距离特征的有效性,而且表明该癫痫检测算法具有良好的分类识别性能。(2)以稀疏表示理论为基础,提出了一种基于多层核协作表示分类方法的癫痫自动检测算法。该算法将脑电信号进行多层小波分解,然后在各层上结合核方法与协作表示对子频带脑电信号进行分类,并提出一种新颖的判决决策有效融合各层与各导联的判断结果,最终构建出多层核协作表示分类与检测系统。在核协作表示分类框架中,求解最小72范数优化问题得到测试样本在训练字典上的稀疏向量并计算两类训练样本集对测试样本的重构误差,将其差值作为判决变量进行分类识别,避免了传统检测算法中的特征选取与分类器设计难题。核函数的应用增强了脑电信号的可分性,有利于提高算法的分类性能。协作表示使用l2范数代替稀疏表示的l1范数进行求解,在保证分类性能的同时大大降低了算法复杂度。多层的系统设计与判决准则将脑电信号的频域和空间信息有机结合,进一步提高了检测准确性。将该算法在长程脑电数据库上进行测试评估,取得了较理想的检测灵敏度与较低的误检率。实验结果表明该算法对癫痫脑电信号具有较好的检测性能且实时性较高。(3)研究了对称正定(SPD)矩阵的稀疏表示算法并提出一种Log-Euclidean高斯核稀疏表示分类方法进行癫痫发作自动检测。该算法使用协方差描述子对多导联脑电信号建模,并修正协方差矩阵使之成为SPD矩阵。脑电信号协方差矩阵形成的空间是一个非线性黎曼流形,应用Log-Euclidean高斯核函数将其嵌入到线性可再生核希尔伯特空间(RKHS)中进行稀疏表示,并计算重构误差进行分类识别。协方差描述子结合了脑电信号在时域、频域及空间域上的统计特性并能够有效抑制噪声。不同于欧氏空间中向量的稀疏表示,Log-Euclidean高斯核函数考虑了流形数据的几何结构,使得SPD矩阵的稀疏表示合理有效。而且,传统稀疏表示检测算法需要对每个导联的脑电信号迭代处理,具有重复运算且系统设计复杂的缺点。该算法应用SPD矩阵的稀疏表示同时处理多导联脑电数据,成功解决了传统算法的缺陷,有效降低了算法的复杂度。在长程脑电数据库的实验结果表明,该算法不仅具有更加理想的检测性能,鲁棒性较强,且运行速度更快,基本满足在线检测系统对检测准确性与实时性的要求。本论文的工作有助于促进脑电信号分析和癫痫自动检测算法在理论和临床实际应用方面的研究,积极有效地推动了癫痫自动检测技术的发展。由于实验数据的局限性,本文提出的癫痫自动检测算法的有效性和鲁棒性还有待进一步验证与提高。
[Abstract]:Epilepsy is a chronic neurological disorder caused by brain neuron group paroxysmal abnormal electrical activity, its onset is sudden and repeated characteristics, accompanied by loss of consciousness, fainting, twitching limbs and other clinical manifestations, seriously endanger people's health and life safety. According to statistics, more than 1% of the population in the world is suffering from the disease. The causes of epileptic seizures are complex and diverse, and the pathogenesis has not yet been fully defined. As an important means to study epilepsy, EEG uses electrodes to record electrical activity of brain neurons, which contains a lot of physiological and pathological information. It plays an extremely important role in clinical diagnosis, location and treatment of epilepsy. At present, electroencephalogram analysis and epileptic detection are mainly done by medical staff according to clinical experience visual observation. However, the large amount of data makes the work very boring and time-consuming, and the subjectivity of the medical staff will also affect the judgment of epileptic seizures. Therefore, the use of computer automatic EEG signal analysis and detection of seizures is very urgent and necessary, it can not only reduce the burden on doctors, improve the diagnostic efficiency and effectiveness in the treatment of epilepsy, seizure disorders, improve the quality of life of patients, and reveal the pathogenesis of epilepsy is of great significance. This paper is based on the automatic detection of epilepsy related research on this topic, on different stages of extraction of similarity distance characteristics of non electric signal, sparse representation and cooperation study theory of EEG signal classification based on content, and puts forward several effective method for automatic detection of epilepsy. The research contents and innovations of this paper include the following points: (1) applying distance measure to EEG analysis, a epileptic seizure detection algorithm based on diffusion distance and Bayesian linear discriminant analysis (BLDA) is proposed. The algorithm uses wavelet transform to analyze the time frequency of EEG and combine the EEG signals of three bands to form the EEG distribution. Then, the diffusion distance between epileptic and intermittent EEG distribution is calculated, and the difference between EEG signals in different periods is quantitatively described. According to the principle that the difference of EEG signals between the same kind is lower than the different kinds of EEG signals, the diffusion distance is used as the combination of EEG features and BLDA classifier to realize the recognition and detection of epileptic seizures. Compared with bulldozer distance (EMD), diffusion distance can not only effectively distribute EEG and intermittent EEG, but also has stronger noise immunity and lower computational complexity. Through the regularization method, BLDA avoids the over fitting problem in the noisy data set, and can get better classification effect. Experimental results on Freiburg long range electroencephalogram database verify the validity of EEG diffusion distance characteristics, and show that the epileptic detection algorithm has good classification and recognition performance. (2) based on the sparse representation theory, an automatic detection algorithm for epilepsy based on multi-layer kernel cooperative representation classification is proposed. The algorithm divides the EEG wavelet decomposition, and then combined with the kernel method and cooperation on band EEG signal classification in each layer, and proposes a novel effective decision making fusion judgment results of each layer and each lead, and ultimately build a multi nuclear cooperation in classification and detection system. In the nuclear cooperation classification framework, minimum 72 norm optimization problem to get the test samples in the training dictionary sparse vectors and calculate the two kinds of training sample set the reconstruction error of the test sample, the difference as the decision variables for classification, to avoid the traditional detection algorithm of feature selection and classifier design problem. The application of kernel function enhances the separability of EEG signal, which is helpful to improve the classification performance of the algorithm. The cooperative representation uses the L2 norm instead of the L1 norm of sparse representation to solve the problem, which greatly reduces the complexity of the algorithm while ensuring the performance of the classification. Multi-layer system design and decision criteria combine the frequency domain and spatial information of EEG signal, and further improve the detection accuracy. The algorithm is tested and evaluated on the long range EEG database, and the ideal detection sensitivity and low false detection rate are obtained. The experimental results show that the algorithm has good detection performance and high real-time performance for epileptic EEG. (3) the sparse representation algorithm of symmetric positive definite (SPD) matrix is studied, and a Log-Euclidean Gauss kernel sparse representation classification method is proposed for automatic detection of epileptic seizures. The covariance descriptor is used to model the multi lead EEG, and the covariance matrix is modified to make it a SPD matrix. The space formed by the covariance matrix of EEG is a nonlinear Riemann manifold. The Log-Euclidean Gauss kernel function is applied to the sparse representation in the linear regenerative kernel Hilbert space (RKHS), and the reconstruction error is calculated for classification and recognition. Covariance descriptors combine the statistical characteristics of EEG in time domain, frequency domain and space domain, and can effectively suppress noise. Unlike the sparse representation of vectors in Euclidean space, the Log-Euclidean Gauss kernel function takes into account the geometric structure of manifold data, making the sparse representation of SPD matrix reasonable and effective. Moreover, the traditional sparse representation detection algorithm needs to iterate over the EEG signals of each lead, which has the disadvantages of repeated operation and complex design of the system. The algorithm uses sparse representation of SPD matrix to deal with multi lead EEG data simultaneously. The algorithm has successfully solved the defects of the traditional algorithm and effectively reduces the complexity of the algorithm. Experimental results in long range EEG database show that the algorithm has better detection performance, robustness and faster operation, and basically meets the accuracy and practicality of online detection system.
【学位授予单位】：山东大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：R742.1;TN911.7
，

本文编号：1345476