癫痫脑电的分形分析及自动检测方法研究

发布时间：2018-04-27 03:03

  本文选题：癫痫脑电 + 发作检测　； 参考：《山东大学》2016年博士论文

【摘要】：癫痫是一种常见的慢性脑部疾病,影响全世界近1%的人口。长期反复突然的癫痫发作,给患者带来极大的痛苦和严重的身心伤害。癫痫发作是多种病因引起的大脑神经元群突发性异常超同步化放电的结果,约80%的癫痫患者存在脑电图异常现象。因此,脑电图检查与分析是癫痫疾病诊断、病灶定位和发作类型判断的重要手段。而借助计算机技术,研究癫痫脑电信号的自动分析与检测方法,对提高癫痫诊断的效率和研制闭环癫痫刺激器,具有重要意义。脑电图(Electroencephalogram, EEG)信号作为大脑神经元电活动在头皮表面或大脑皮层的总体反应,具有复杂的非线性特性。虽然EEG信号的非线性分析得到了癫痫自动检测研究人员的重视,但是EEG信号的分形特性研究较少。分形理论是现代非线性科学的一个重要分支,研究EEG信号的分形特性,有助于进一步了解癫痫发作过程中大脑混沌动力活动的内在本质。同时,由于癫痫发作的机理非常复杂,发作类型和过程多种多样,不同的癫痫患者,甚至是同一患者的不同次发作,其发作过程都不相同。因此,目前的癫痫发作自动检测技术还难以满足临床应用所提出的准确性、实时性和鲁棒性要求。针对癫痫脑电分析和发作检测领域存在的上述问题,本文对癫痫脑电信号的分形及多重分形特性进行系统、深入的研究,并将机器学习和模式识别领域的前沿算法或分类器模型引入到癫痫发作检测领域,研究准确度高、实时性好的癫痫发作检测方法。具体研究内容包括以下几方面。首先,研究EEG信号的Higuch分形维数在癫痫发作前期的演化规律,并从发作机理的角度进行分析解释；将发作前期EEG信号Higuchi分形维数的变化,作为癫痫发作的先兆特征,结合贝叶斯线性判别分析器,提出一种癫痫发作预测算法,对发作前期脑电进行检测识别。该算法在Freiburg癫痫脑电数据集上,达到较高的预测灵敏度和较低的误报率,同时具有较低的计算成本。然后,对比研究发作期与间歇期脑电的K近邻分形维数,发现两类脑电信号的K近邻分形维数具有显著的统计差异性。于是引入梯度Boosting集成学习算法,提出一种基于K近邻分形维数和梯度Boosting的癫痫发作检测方法。在Freiburg长程脑电数据集上,不但取得了较高的检测灵敏度和较低的误检率,而且对发作期起始时刻(Onset)的检测延时小,21例癫痫患者的平均检测时延仅为2.46秒。接着,本文从对癫痫EEG进行单一分形维数的算法研究和特性分析,进一步扩展和深入到研究EEG信号的多重分形特性,用多重分形谱深层次地刻画癫痫脑电的局部奇异性和分形特性的不均匀性。在证明癫痫脑电信号具有多重分形特性的基础上,对EEG信号多重分形谱参数的物理意义进行解释,并通过对比研究,发现发作期与间歇期EEG的多重分形特性和谱参数(α0、αmin、αmax、Δα、f(αmin)、 f(αmax)、Δf, R)都具有显著的统计差异性。最后,将癫痫患者EEG信号的多重分形谱特征与相关向量机相结合,提出一种融合多导联判决结果的癫痫发作检测系统。在对相关向量机输出的类概率进行后处理的过程中,将多导联的判决结果进行融合,使其更符合临床医生的诊断过程。该癫痫发作检测系统在Freiburg癫痫脑电数据集上进行性能测试,取得了较高的检测灵敏度和识别率。同时该检测系统具有较低的计算复杂度,对一小时三导联EEG进行处理大约只需要1.2分钟,表现出很好的检测实时性。本文在对脑电信号单一分形维数的计算中,所采用的Higuchi算法和K近邻算法,都是直接从信号时域进行,不需要重构相空间,算法简单,计算复杂度低；而对EEG进行多重分形分析所采用的Moment方法,相对于其他研究领域中常用的多重分形去趋势波动分析法,也具有物理意义简单明确,计算量小等优点。因此,本文基于EEG的各分形特征建立的癫痫发作检测算法,大大降低了EEG分析和特征提取所需的时间,保证检测算法具有较好的实时性。另外,本文所提出的几种癫痫发作检测算法中,分别采用了贝叶斯线性判别分析、基于集成学习思想的梯度Boosting和基于贝叶斯稀疏学习理论的相关向量机等前沿的学习算法和分类器模型,对脑电模式进行分类识别,从而保证检测算法具有较高的检测准确度。因此,本文的研究工作进一步推进了癫痫脑电的非线性特性研究,并且为研究检测准确度高、实时性好的自动检测方法,提供了新的思路。本文所提出的癫痫发作自动检测算法将在临床大量癫痫脑电数据上,进行性能验证与完善。
[Abstract]:Epilepsy is a common chronic brain disease that affects nearly 1% of the world's population. A long and recurrent seizure has caused great pain and serious physical and mental injury to the patient. Epileptic seizures are the result of a sudden abnormal hyper synchrotron discharge of a group of brain neurons caused by a variety of causes, and about 80% of the epileptic patients have electroencephalogram. Therefore, the examination and analysis of electroencephalogram (EEG) is an important means for the diagnosis of epilepsy, the location of the focus and the type of seizure, and the method of automatic analysis and detection of epileptic EEG with the help of computer technology is of great significance to improve the efficiency of epileptic diagnosis and develop a closed loop epilepticus. (Electroencephalog Ram, EEG) signals have complex nonlinear characteristics as the electrical activity of the brain neurons in the scalp surface or the cerebral cortex. Although the nonlinear analysis of EEG signals has been paid attention to by the researchers of the automatic detection of epilepsy, the fractal characteristics of the EEG signal are seldom studied. Fractal theory is an important part of the modern nonlinear science. The study of the fractal characteristics of EEG signals helps to further understand the intrinsic nature of the chaotic dynamic activity of the brain during epileptic seizures. At the same time, due to the complex mechanism of epileptic seizures, the types and processes of seizures are varied, different epileptic patients, even the different episodes of the same patient, have different episodes. Therefore, the current automatic detection techniques for epileptic seizures are difficult to meet the accuracy, real-time and robustness requirements of the clinical application. In view of the above problems in the field of epileptic EEG analysis and seizure detection, the fractal and multifractal characteristics of epileptic EEG signals are introduced in this paper, and the machine learning and modeling are studied in depth. The forward algorithm or classifier model in the field of pattern recognition is introduced into the field of epileptic seizure detection, and the methods of epileptic seizure detection with high accuracy and good real-time are studied. The specific research contents include the following aspects. First, we study the evolution of the Higuch fractal dimension of the EEG signal in the pre epileptic seizure, and divide it from the point of view of the attack mechanism. The change of the fractal dimension of the pre paroxysmal EEG signal Higuchi, as the precursor of the epileptic seizures, combines with the Bias linear discriminant analyzer, and proposes an epileptic seizure prediction algorithm, which can detect and recognize the EEG in the preparoxysmal EEG. The algorithm achieves high predictive sensitivity and low in the Freiburg epileptic EEG data set. By comparing the K nearest neighbor fractal dimensions of the brain electricity in the attack and the intermittent period, we find that the K nearest neighbor fractal dimension of the two kinds of EEG signals has significant statistical difference. Then the gradient Boosting integrated learning algorithm is introduced to propose a epilepsy based on the near neighbor fractal dimension of K and the gradient Boosting. The detection method of epileptic seizure. On the Freiburg long range EEG data set, not only high detection sensitivity and low misdetection rate have been obtained, but also the detection delay of onset time (Onset) is small. The average detection delay of 21 epileptic patients is only 2.46 seconds. Then, this paper studies the single fractal dimension algorithm of epileptic EEG The multifractal characteristics of EEG signal are further expanded and studied, and the local singularity and fractal characteristics of epileptic EEG are deeply depicted with multifractal spectrum. On the basis of the multifractal characteristics of epileptic EEG, the physical meaning of the EEG signal multifractal parameters is explained. By contrast, the multifractal characteristics and spectral parameters (alpha 0, alpha min, alpha max, delta alpha, f (alpha min), f (alpha max), delta f, R) were found to have significant statistical differences. Finally, the multiple fractal spectrum characteristics of EEG signals in epileptic patients were combined with the correlation vector machines, and a kind of epilepsy was proposed to fuse the multiple lead decision results. In the process of post processing of the class probability of the output of the related vector machine, the decision results of the multi lead are fused to make it more consistent with the diagnosis process of the clinician. The epileptic seizure detection system performs the performance test on the Freiburg epileptic EEG data set, and has obtained high detection sensitivity and recognition rate. The detection system has a low computational complexity, and it takes only 1.2 minutes to deal with the one hour three lead EEG. In the calculation of the single fractal dimension of the brain electrical signal, the Higuchi algorithm and the K nearest neighbor algorithm are all directly from the signal time domain and do not need to reconstruct the phase space. The algorithm is simple and the computational complexity is low, and the Moment method used in multifractal analysis for EEG is also characterized by simple physical meaning and small calculation, compared with the common multifractal detrending wave analysis method in other research fields. Therefore, the epileptic seizure detection algorithm based on the fractal characteristics of EEG is established in this paper. The time required for EEG analysis and feature extraction is greatly reduced to ensure that the detection algorithm has good real-time performance. In addition, the Bayesian linear discriminant analysis, the gradient Boosting based on the integrated learning idea and the correlation vector machine based on Bayesian sparse learning theory are used respectively in several epileptic seizure detection algorithms proposed in this paper. The advanced learning algorithm and classifier model are used to classify the EEG pattern, thus ensuring the high detection accuracy of the detection algorithm. Therefore, the research work of this paper further advances the study of the nonlinear characteristics of epileptic EEG, and provides a new thought for the study of the automatic detection method with high detection accuracy and good real-time performance. The automatic epileptic seizure detection algorithm proposed in this paper will be validated and perfected on clinical epileptic EEG data.

【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;TN911.6
，

本文编号：1808873