癫痫发作自动检测方法的研究

发布时间：2018-05-30 03:36

本文选题：癫痫发作检测 + 短时傅里叶变换　；参考：《中国科学院研究生院(长春光学精密机械与物理研究所)》2014年硕士论文

【摘要】：癫痫(Epilepsy)是一种常见的脑功能障碍性疾病,其基本病理学表现为大量神经元集群异常同步化放电,而头皮脑电和皮层脑电是脑神经电活动的反映,是研究癫痫发作的主要手段。基于脑电信号的癫痫发作自动检测可以实现脑电分类和病灶定位,并能够提高检测效率,对癫痫的临床治疗具有重要意义。针对累积能量特征不稳定的问题,本文结合时频分析方法和能量特征定义了时频能量特征,并针对大数据量、高特征值空间的快速和准确分类问题,提出一种基于最大相关最小冗余准则和带参数的极限学习机的癫痫发作自动检测方法。首先,对皮层脑电和头皮脑电分别使用短时傅里叶变换和希尔伯特变换,提取时频分布的能量块,并结合空间信息,得到时频能量特征集。然后,利用序列前向选择搜索策略生成不同大小的特征子集,其中的评价准则采用基于最大相关最小冗余的信息准则,以特征子集作为评价单位,使用基于分类准确率的特征选择方法选择最优特征子集。最后,使用基于粒子群算法的支持向量机、基于粒子群算法的BP神经网络和带参数的极限学习机对癫痫不同状态进行分类和判别。实验结果： 1)在皮层脑电中,使用短时傅里叶变换提取的时频能量特征的分类准确率为0.97,优于使用经验模态分解提取的时频能量特征。 2)使用基于粒子群算法的支持向量机、基于粒子群算法的BP神经网络和带参数的极限学习机三种分类器进行10折交叉验证,头皮脑电的分类准确率为0.85左右,皮层脑电中包含发作期和亚临床发作期的分类组的分类准确率为0.82左右,而发作间期和发作期的分类准确率达到0.97以上。 3)在皮层脑电的发作间期和发作期分类组,基于粒子群算法的支持向量机的分类准确率为0.98,训练时间为28.1s；基于粒子群算法的BP神经网络的分类准确率最高可达0.99,但是随特征子集大小变化而有明显的起伏,分类器性能不稳定；带参数的极限学习机的分类准确率为0.97,但训练时间仅为0.8s。带参数的极限学习机的分类准确率和训练速度两方面的综合性能优于基于粒子群算法的支持向量机和基于粒子群算法的BP神经网络。 4)利用带参数的极限学习机对72小时的皮层脑电中数据进行连续检测,误检率为1次/24小时,平均发作开始时刻检测延迟为0.1s。结果表明,在皮层脑电的发作期和发作间期分类中,基于短时傅里叶变换的时频能量特征集是有效的。基于最大相关最小冗余准则的序列前向选择方法和带参数的极限学习机的方法能够实时准确地检测癫痫发作。
[Abstract]:Epilepsys is a common disorder of brain function. Its basic pathological manifestation is the abnormal synchronous discharge of a large number of neurons. The scalp EEG and cortical EEG are the reflection of EEG activity and the main means to study epileptic seizures. The automatic detection of epileptic seizures based on EEG signals can achieve EEG classification and focus location, and can improve the detection efficiency. It is of great significance for the clinical treatment of epilepsy. In this paper, the time-frequency energy features are defined by time-frequency analysis method and energy feature, and the fast and accurate classification problem of large data volume and high eigenvalue space is discussed. An automatic epileptic seizure detection method based on maximum correlation minimum redundancy criterion and parameter based extreme learning machine (LLM) is proposed. Firstly, the cortical EEG and scalp EEG are extracted from the energy blocks of time-frequency distribution by using short-time Fourier transform and Hilbert transform, respectively, and the time-frequency energy feature set is obtained by combining spatial information. Then, different size feature subsets are generated by the sequence forward selection search strategy. The evaluation criteria are based on the information criterion of maximum correlation and minimum redundancy, and the feature subset is used as the evaluation unit. A feature selection method based on classification accuracy is used to select the optimal feature subset. Finally, support vector machine based on particle swarm optimization, BP neural network based on particle swarm optimization and extreme learning machine with parameters are used to classify and distinguish different states of epilepsy. Experimental results: 1) in cortical EEG, the classification accuracy of time-frequency energy features extracted by short-time Fourier transform is 0.97, which is better than that extracted by empirical mode decomposition. 2) support vector machine based on particle swarm optimization, BP neural network based on particle swarm optimization and extreme learning machine with parameters are used for 10 fold cross validation. The accuracy of scalp EEG classification is about 0.85. The classification accuracy was about 0.82 in the cortical EEG group which included the attack period and subclinical attack stage, but the classification accuracy rate of the interictal phase and the attack stage was more than 0.97. 3) in the interictal and interictal groups of cortical EEG, The classification accuracy of support vector machine based on particle swarm optimization is 0.98 and the training time is 28.1 s. The classification accuracy of BP neural network based on particle swarm optimization is up to 0.99, but it fluctuates obviously with the change of feature subset size. The performance of the classifier is unstable, and the classification accuracy of the LLMs with parameters is 0.97, but the training time is only 0.8 s. The classification accuracy and training speed of LLMs with parameters are superior to those of SVM based on PSO and BP neural network based on PSO. 4) the data of 72 hours cortical electroencephalogram were continuously detected by the extreme learning machine with parameters. The false detection rate was 1 / 24 hours, and the detection delay was 0.1 s at the average onset time. The results show that the time-frequency energy feature set based on short-time Fourier transform is effective in the classification of cortical EEG seizures and interictal phases. The method of sequence forward selection based on maximum correlation minimum redundancy criterion and the method of extreme learning machine with parameters can detect seizures accurately and in real time.
【学位授予单位】：中国科学院研究生院(长春光学精密机械与物理研究所)
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：R742.1;TP18

【参考文献】