EEG-based Diagnosis Of Schizophrenia Using Double-level Analysis And Feature Selection Methods

Electroencephalogram(EEG)signals are originate from postsynaptic potentials generated by a large number of neurons,It contains a large number of physiological and pathological information,which can reflect the brain state of brain activity.Numerous studies have been shown that schizophrenia patients and normal controls can be distinguished clearly with EEG signal characteristics.Therefore,the diagnosis of schizophrenia based on EEG features has become a research topic for many experts and scholars.Traditional diagnosis of schizophrenia is based on clinical diagnostic scale and subjective judgment of the psychologists,the subjectivity and misjudgment rate of the results are still relatively high.With the rise of machine-learning technology,this technique has been applied to the diagnosis of this disease by more and more researchers,and many researchers have proved the feasibility and significance of this research.In recent years,with the development of BCI technology,EEG physiological characteristics have become an important indicator for visual,motor,imagination and pathology.Most studies of schizophrenia diagnosis based on EEG features such as ERP peak amplitude,power spectrum,etc,have achieved good results.Due to the number of EEG features is much larger than samples,the problem of inaccuarcy and over-fitting are often found in the classification model.Therefore,we propose to use feature selection algorithms for reducing feature dimensions before constructing diagnostic models.This method not only reduces feature redundancy,but also greatly improves the accuracy and robustness of the diagnostic model,and it also reduces the complexity of diagnostic system.This paper mainly uses four kinds of filter feature-selection algorithms,such as Relieff,MI,SD,and m RMR.These algorithms are introduced in detail.We proposes a double-level feature extraction method of EEG signals from sensor level and source level,in addition,our diagnosis system based on Support Vector Machine(SVM)and K Nearest Neighbor(KNN)classifier for schizophrenia.The system consists of five parts: EEG signal preprocessing,feature extraction,feature selection,model construction,and model evaluation.Our EEG data were recorded from first-episode schizophrenia,which eliminates the influence of drugs and other factors during experimental,besides,it can also increase the reliability of the results.In our system,a total of 720 features were extracted at the sensor level,428 features from source level,and 1128 features at the double level were used as the input of the classifier respectively.The results of the system evaluation showed :(1)EEG features can be used as schizophrenia Physiological indicators of diagnosis;(2)we get the optimal system when both the source-level and the sensor-level features were used;(3)When using the Relieff feature selection algorithm,and double level features(7 features at the source level and 13 features at the sensor level)the best diagnostic model can be obtained.The optimal scheme are,SVMRelieff-double:CCR=95.2%�5.85,AUC=0.97�0.03;KNN-Relieff-double:CCR=96.5%�4.79,AUC=0.97�0.03.