Research On EEG Depression Identification Based On Feature Selection And Ensemble Classification

Since the 21 st century,with the rapid development of social economy,people’s life and work pressure has gradually increased,thus greatly increasing the incidence of anxiety,schizophrenia,epilepsy and other mental diseases.Depression,as one of them,has had a serious impact on the lives of patients and their families,so we urgently need to find efficient diagnosis and treatment for the disease.At present,most diagnostic methods for depression are to ask patients to make scales and ask some patients at the same time,so as to make a general judgment of the patient’s condition,and then give more accurate results based with the diagnostic criteria of ICD-10 depression.The method relies heavily on the professional level of expertise,is subjective and does not quantify depression by specific physiological index values and medical images.And EEG as an electrophysiological technology has been widely used in the diagnosis of psychological diseases,to assist in the clinical diagnosis of doctors.In this thesis,we use 128 EEG data from a group of patients with major depressive disorder published by the UAIS laboratory in 2020 to make a diccategorization of depressed patients and normal people.First,we preprocessed the original EEG signals containing the impurities and removed the contained impurities to obtain the pure EEG.At the same time,in order to reduce the complexity of the experiment,we only selected the pure EEG signal,the forehead 7,which is closely related to depression,for the subsequent analysis.Then we extracted 11 features for each guide including linear and nonlinear features separately,yielding a feature matrix consisting of 77 features.Next,three feature selection methods for this feature matrix,namely K-S test,feature selection based on random forest and feature selection based on genetic algorithm,are used to perform preliminary feature selection,and select the final features by feature voting.Since the sample sizes of the two categories in the final feature matrix are different,we also resampled the feature matrix using the SMOTE oversampling technology,so that the number of samples of the two classes is equal.Finally,we input the oversampled feature matrix into KNN,SVM,CART decision tree classifier as well as the Soft Voting-based integrated classifier for classification.According to the experimental results,the Soft Voting-based integrated classifier is significantly better than the single classifier,and the accuracy,recall,f1,and g-mean values are 0.8216,0.8247,0.8258,0.8261,respectively.