The Classification Of The Phases Of Schizophrenia Based On EEG Data

Electroencephalogram is noninvasive technology that records biological brain electrical activity and has potential applications in various fields such as human-computer interaction,neuroscience and clinical medicine.For example,various types of EEG data have been collected from different subjects for analysis of human behavior controlling ways,the principle of human thinking and the related causes of mental illness.Among them,the application of EEG data for the analysis of the risk of schizophrenia is a new field of research.Schizophrenia is a type of mental illness,which has the characteristics of high recurrence rate and high morbidity,bringing a serious disease burden to the family and society.At present,the clinical diagnosis of schizophrenia in the medical sector mostly comes from expert interrogation,and further pathology research and disease treatment are limited to the results of interrogation.For the development of schizophrenia pathology research and the improvement of diagnosis and treatment programs supported by the information technology,we need a method of schizophrenia analysis and classification based on data.The electroencephalogram(EEG)has been focused on the characteristics of non-invasive and multi-dimensional collection of brain activity.Based on the EEG data of manually calibrated health control(HC),first episode of schizophrenia(FES)or clinical high-risk(CHR),we hope to learn the automatic classification method based on EEG data and provide effective technical support for diagnosis.needs more effective methods beside of traditional data analysis methods.From the perspective of high-dimensional statistics,we could find that each sample of the EEG data contains the changes of EEG signals over a period of time.Therefore,the EEG of the subjects is analyzed through a combination of preprocessing,feature extraction and machine learning classification.Based on the classification of subjects,the results that can be applied to the study of pathology are given from the perspective of medical research.From the perspective of deep learning,more accurate EEG classification results can be obtained by using a combination of preprocessing,data morphological transformation and deep neural network.This study explores the algorithms for the analysis of brain wave data from two aspects including high-dimensional statistics and deep learning.The main work of this thesis is divided into the following three parts:(1)This study analyzes the EEG data in frequency domain from the aspect of high-dimensional statistics and proposes a classification algorithm including three stages: data preprocessing,feature extraction and machine learning classification.In the data preprocessing stage,the key band of the frequency domain data are obtained.In the feature extraction stage,the linear eigenvalue statistics(LES)based on Random Matrix Theory(RMT)is used as the classification feature.In the machine learning classification stage,the classifier based on SVM is applied.The frequency band analysis method is studied based on the results of classification.The original EEG data is divided into several bands to obtain features and classified.After that the classification results are used to construct the voter.Then the feature weights of the frequency bands are fitted to obtain the significance of each frequency band in distinguishing three stages of schizophrenia,and the probable factors causing schizophrenia are analyzed.The results help to study the pathology of psychosis,meanwhile the results of weight distribution can be applied to the machine learning classifier to achieve higher classification accuracy.(2)From the aspect of deep learning,this study analyzes the EEG time-domain data by generalizing the individual features into group features by dividing the single sample of data into pieces and increasing the amount of samples to satisfy the requirements for training the deep neural network model.With a large number of EEG time-domain data,this study shows a combination algorithm of convolutional neural network and support vector machine which is suitable for classification of EEG time-domain data to improve the classification accuracy and contribute to the research of auxiliary diagnosis of schizophrenia.(3)In order to make better use of spatial information and high temporal resolution of EEG temporal data,the EEG signals at each moment are converted from vectors to micro-state information matrix with positional relationship by data preprocessing.Based on time series data,this thesis proposes a cascade network architecture that combines spatial-information-oriented convolutional neural network and recurrent neural network oriented to time series features to improve the classification accuracy of the three-stages of schizophrenia.In the three experiments,the EEG data of three-phase samples from Shanghai Mental Health Center are analyzed using the above classification algorithms,which verifies their effects on the pathology research and the auxiliary diagnosis of schizophrenia.