A Study Of Depression Recognition And Source Localization Based On Resting State EEG Signal

Depression has become a global mental disease that endangers millions of people and is characterized with high morbidity and mortality.Depression has brought heavy health and economic burden to society.Recently,with the increasing number of patients with depression,more and more attention has been paid into depression.However,the traditional diagnosis mainly relies on patients' questionnaires self-test and doctor's inquiry.These subjective methods are easily influenced by the patients' compliance and the doctors' experience.Therefore,we need to find an objective,effective and convenient assessment method to improve the accuracy of diagnosis of depression.Electroencephalography(EEG)is a method which uses electrophysiological signal to record brain activity and can give researchers and doctors an overall reflation of the electrophysiological activity of brain neurons on the cerebral cortex.EEG has several advantages such as safe,easily to operate and non-intrusive which make EEG become a promising method for the researchers of the human recognition area.Many researchers try to use EEG data combined with data mining classification algorithms to automatically detect depression,most of their works are based on task-state EEG signals,the resting-state EEG is relative less involved and the accuracies of previous studies need further improvement.In this paper,128 channels of resting-state EEG data of 17 depressed subjects and 17 normal subjects were selected for analysis and the data of alpha band,beta band and theta band were extracted and processed separately.By using feature selection and classification algorithms to get depression detection accuracies then using a voting strategy among three bands EEG data to improve the accuracies.The results shows that compared with the traditional classification,the voting strategy can significantly improve the classification accuracies.In order to discover some brain differences between depressive patients and normal people,we perform single electrode analysis and source localization study on the EEG data to find some electrodes and brain regions that can distinguish depressive subjects from normal subjects.The main works and contributions of this paper:(1)In this paper,10 linear and 8 non-linear features are extracted from the alpha band,beta band and theta band respectively for the analysis.To avoid the dimension curse,Greedy Stepwise(GSW),a feature search method based on correlation feature selection(CFS),is used to select features.As for the classification,five typical classification algorithms are employed to classify the depressed and the normal subjects including Bayesian belief network(BN),support vector machine(SVM),nearest neighbor classification(KNN),decision tree(C4.5)and random forest(RF).The highest classification accuracies for the alpha band data,beta band data and theta band data are 68.92%±3.61%,76.67%±1.72% and 76.76%±2.36% respectively.To improve the classification accuracies,the voting strategy is employed.The highest classification accuracies obtained by voting strategy in three bands are 77.65%±2.84%,83.24%±2.42% and 81.47%±4.71% respectively.Compared with the traditional classification results,the voting strategy has significantly improved the classification results.Among the three bands of resting-state EEG data,the beta band data achieves the highest accuracy in the traditional classification situation and voting situation.The results indicates that beta band may have a greater impact on identifying depressed subjects and normal subjects and the traditional classification algorithm combined with voting strategy can significantly improve the accuracies of depression detection,which can provide an accurate and automatic solution for the depression recognition,and can be used as an assistant tool for the depression diagnosis.(2)To discover the brain differences between depressed subjects and normal subjects in the resting-state EEG experiment,we firstly made single electrode analysis based on beta band EEG data.The results of T test show that FP1,F3,C3 and O2 electrodes can significantly distinguish the depressed and the normal.The results of single electrode classification also demonstrates that the FP1,F3,C3 and O2 electrodes are better than other electrodes.As for the reason of these phenomena,the FP1,F3 and C3 electrodes are located in the left prefrontal lobe,frontal lobe and central sulcus of the frontal lobe,and according to the results of sLORETA analysis that the brain of the depressed subjects were less active in the left frontal lobe and the left lower frontal cortex than normal subjects.The results of single electrode analysis also show that there were significant differences between depressed patients and normal people in the left prefrontal lobe.These results indicating the left prefrontal lobe and depression are correlated.According to the two parts work in this paper,we can conclude that using voting strategy can significantly improve the accuracies of the traditional classification,and the single electrode analysis and source localization demonstrate that there is a significant difference in the left frontal lobe of the brain between depressed subjects and normal subjects in the resting-state EEG,and the brain activity of depressed subjects is weaker than the normal subjects.These conclusions can provide us more reliable classification accuracies and the more special brain location for depression analysis for the future studies.