Research On Nonlinear Feature Extraction And Classification Method For Epileptic EEG Based On Complex Network Topological Structure Statistical Properties

Epilepsy is a kind of ancient nervous system disease, which impacts a wide range of people and seriously affects the patients’ physical and mental health. Moreover, there is a large treatment gap in the society, which further intensifies the difficulties both in life and work of the patients and their families. The Electroencephalography(EEG) is a tool which reflects the discharge phenomenon of the cerebral cortex. As a time series, it contains a large amount of physiological information about brain working. At present, the diagnoses of epileptic seizure mainly rely on that the medical professionals identify the patients’ EEG with their naked eye. However, this method has some shortcomings, such as costing a lot of time, having high level of subjectivity, and being inefficient. Therefore, the studies of high-performance epileptic EEG automatic classification algorithm, with the aid of computers, have greatly clinical application significance. The EEG signals exhibit nonlinear dynamic characteristics because of the complex structure and irregularly discharge of the cortical neurons. Thus analyzing EEG through nonlinear time series analysis method recently becomes the first choice of the academics. Whereas, the classification performances of the frequently-used nonlinear features are relatively poor, such as approximate entropy and sample entropy of EEG. This directly affects the performance of the epileptic EEG classification algorithm and seriously hinders applying it to the clinical diagnoses.Recently, the application of utilizing the topological statistical properties of complex networks to study the deep nonlinear dynamical characteristic of time series gets increased attention. It also gives a new perspective of the nonlinear time series analysis. Thanks to the emergence of construction algorithms, which transform the time series into complex network, and coupled with rich and widely studied topological statistical properties of complex network, the complex network analysis method of time series began to be applied to analyze real-life nonlinear data, such as financial stocks time series, traffic flow series, climate change recording, river flow, and so on.Based on the researches on the analysis method of time series complex network, this thesis mainly studies the nonlinear feature extraction method for epileptic EEG on the base of the topological statistical properties of complex network. Combined the extracted features respectively with support vector machine classifier, the epileptic EEG automatic classification algorithms are constructed, which are used to distinguish the two types of epileptic EEG(interictal and ictal EEG).Firstly, the node set of time series complex network is constructed based on the node set construction algorithm; Then the standard Euclidean distance is utilized to measure the similarities between any two nodes in the node set; According to the conversion parameters(θ), the edge set of the time series complex network is established. Through the above process, the time series complex network is constructed.Secondly, the time series complex network is analyzed by applying the topological statistical properties of the complex networks. Based on the analysis results, four corresponding features are extracted as a classification features to distinguish the different types of time series in this thesis. The four nonlinear classification features are the entropy of degree distribution(NDDE), the partial sum of clustering coefficient distribution(Pclu), the weighted mean value of vertex strength distribution(wmean), and the small value sum of weight differences distribution(WDrα). Through applying the extracted classification feature to distinguish interictal and ictal EEGs in clinical epileptic EEG database, the results show that the classification accuracy and feature extraction time by using the extracted features to classify the two kinds EEG signals both have great improvement compared with using other frequently-used features, such as approximate entropy and sample entropy.Finally, the extracted features are applied into the epileptic EEG automatic classification algorithms. Four automatic epileptic EEG classification algorithms are constituted by feeding the four two-dimensional classification feature vectors, which are constructed by combing the four extracted nonlinear features and the fluctuation index(FI), respectively, into the support vector machine classifier. The performances of the proposed automatic epileptic EEG classification algorithms are verified by the experiments on classifying standard epileptic EEG database. Compared with the existing classification algorithms, the overall classification accuracy all have been improved.As the directly reflection carrier of brain activity, the EEG signal contains physiological dynamical information of the brain. By applying the construction algorithm of time series complex network, the EEG signal is converted into a complex network. Through applying the statistical properties to analyze the topological structure of EEG signal complex network, we can extract deep level information contained in the EEG signals. By extracting classification features with excellent performance, the high-performance epileptic EEG automatic classification algorithms can be constructed. The research idea of this article and the classification algorithm itself both provide great help for clinical research and application. It may constantly improve the medical assistance level.