Investigation On Epileptic Seizure Electroencephalogram Classification Technology Based On Singular Value Decomposition And Riemannian Geometry

Epilepsy is a disease of the brain and nervous system,which is mainly caused by paroxysmal hypersynchronous abnormal discharge of central nerve cells of the brain,which has the characteristics of mutability,temporality and repeatability.According to the World Health Organization in 2019 estimates that nearly 50 million people in the world are subjected to epilepsy,which is approximately 1% to 2% of the world’s population.EEG has been used a quatitative assessment of brian function,directly reflects most of the pathological information of spontaneous electrophysiological activities in the brain and provides an effective basis for analyzing the activity state of epilepsy.Most clinicians make judgements about seizures by visual inspection of patients’ 24-hour long-term recording EEG signals.However,visual inspection is time-consuming,subjective and empirical.Therefore,designing a convenient real-time epilepsy monitoring system has prefound significance for reducing the workload of medical staff.Since EEG signals are one-dimensional signal sequences,existing signal processing and deep learning algorithms have problems of low computational efficiency and too high complexity when extracting high-dimensional spatial feature information through EEG.In order to solve the above problems,this thesis uses covariance features to characterize the dynamic information of the brain system in high-dimensional space.Then,the covariance features are deeply analyzed by singular value decomposition and Riemannian geometry tools.The purpose is to extract more comprehensive,deeper and more discriminating feature imformation for the covariance features of EEG signals.The robustness and effectiveness of the proposed algorithm are verified in two public single channel epileptic EEG datasets.The major research work and creative points of this thesis are as follows:（1）Aiming at the problem of insufficient representation of dynamic features in highdimensional space by EEG signals,the epilepsy EEG classification technology based on logarithmic normalized singular value is proposed.The main idea can be summarized as establishing the joint distribution of time domain and singular spectrum domain based on singular value decomposition,so as to achieve the specific representation of epileptic seizure state characteristics.High dimensional phase space is obtained by phase space reconstruction of one dimensional EEG,and convariance characteristic information is constructed for high dimensional phase space.Singular value decomposition is used to decompose the covariance matrix to obtain singular value eigenvetor.Singular spectral value feature vector is obtained by logarithmic normalized of singular value feature vector.The computer results shows that singular spectral value eigenvector is better to mine the intrinsic pathological information of EEG signals.The performance of accuracy of 96% and 91.3%are obtained for the three class problems in the University of Bonn dataset and India Neurology and Sleep Centre dataset,respectively.（2）Aiming at the problem that the dimension of covariance feature of EEG signal is too high in the Riemann tangent plane,the epileptic EEG classification technology based on Riemann tangent plane feature reduction is proposed.The covariance matrix is constructed based on the phase space reconstruction technique as the data points of the Riemannian manifold are embedded into the tangent space with the Riemannian mean point as the tangent point by logarithmic operator.Then,vectorization is performed on the SPD data points of tangent space to obtain the manifold features of EEG.T-SNE algorithm is used to reduce the dimension of manifold vector features,and the classification of epileptic states is realized by combining with traditional classification algorithms.Experimental results show that the t-SNE algorithm can better reduce the feature redundancy problem,significantly enhance the separability of sample points in low-dimensional space,and improve the recognition performance in multiple classification problems of epileptic seizure recognition.The identification accuracy,sentivity and specificity in the Bonn University are above 98%,96% and 99%,respectively.The identification accuracy and sentivity in the INSC University are above 91% and 96%,respectively.（3）Aiming at the problem that the covariance features of EEG signal are not separable in Riemannian space,the epileptics EEG recognition technology based on Riemannian geodesic discriminant is proposed.Specifically,the phase space reconstruction technique is used to construct a symmetric positive definite matrix,which is located at Riemannian manifold space.Then,we project all sample points of the Riemannian space to the Riemannian tangent plane space by logarithmic operator,and use Fisher discriminant analysis logarithmic points to perform geodesic discriminant,and then the SPD sample points after geodesic discriminant are mapped to the Riemannian manifold space by exponential operator.Finally,the minimum distance to Riemannian mean discriminant algorithm is used to realize automatic recognition in normal state,interictal and ictal state.The experimental results show that geodesic discriminant with Fisher discriminant analysis algorithm can increase the separability between samples,thus improving the result of epilepsy recognition and classification.The identification accuracy,sentivity and specificity in the Bonn University are above 97.5%.The identification accuracy and sentivity of the INSC University are more than 93%.The research work of this thesis based on singular value decomposition and Riemannian geometry revolves around the classification technology of EEG signals in epileptic seizure state,which has played an active and effective role in promoting practical process of the home-type and convenient of the epilepsy recoginition equipment.