Analysis Of EEG Features And Study Of Automatic Classification In Patients With First-Episode Depression

Background and PurposeDepression is a common psychiatric disorder characterized by high incidence,high relapse rate,and high suicide rate,which may impose a heavy burden on individuals and their families and even society.Currently,the etiology and pathogenesis of depression are not clear.Clinical diagnosis is mostly obtained through psychiatric examination of patients by psychiatric clinicians,which limits the objectivity and reliability of diagnosis.Therefore,it is important to find objective electrophysiological indicators of depression to improve the diagnosis of depression.Recently,resting-state EEG studies have provided valuable information for the study of mechanisms in patients with depression.However,previous studies have tended to focus on the analysis of a particular EEG feature.Some of them are lack of integrated analysis of multiple features.In view of the abnormalities of whole-brain neural activity and functional connectivity in patients with depression,our present study will analyze multi-modal EEG features ranging from single-band relative power to multi-band cross-coupling,from the functional connectivity in any two scalp brain regions to temporal dynamic changes of the whole-brain neural networks.The purpose of the study would explore specific electrophysiological indicators in patients with first-episode depression.Further,we will explore the role of these electrophysiological indicators in clinical diagnosis of depression using machine learning.MethodsForty patients with first-episode depression and forty healthy controls were recruited for this study,and EEG data were collected from all subjects in the resting state with eyes closed for 10 minutes.The EEG data were preprocessed and the EEG characteristics of both groups were extracted,and the relative power of each frequency band in each scalp brain region of interest,the functional connectivity of any two scalp brain regions in each frequency band,the relevant parameters(Duration,Occurrence,Coverage)of the four typical topographic topographies(A,B,C,D)in the EEG microstates and phase amplitude coupling value of the two coupling frequency bands were calculated.Finally,the obtained related indicators of the two sets of EEG features with differences were used as feature sets.Three methods of feature selection including SVM-based recursive feature elimination(Support Vector Machines-Recursive Feature Elimination,SVM-RFE),L1 discipline based logistic regression(Least Absolute Shrinkage and Selection Operator-Logistic Regression,LASSO-LR),and principal component analysis(Principal Component Analysis,PCA)were used.Five classifier algorithms including decision tree(Decision Tree,DT),support vector machines(Support Vector Machines,SVM),gradient boosting decision tree(Gradient Boosting Decision Tree,GBDT),K-nearest neighbor(k-Nearest Neighbor,KNN),and Naive Bayes(Na?ve Bayesian,NB)were utilized to classify depressed patients.Result1.After the analysis of the resting-state EEG data from the patients with depression and the healthy controls,the differences of the EEG features between the two groups were compared as follows:1)In the left occipital region,the relative power in the alpha band was significantly lower in the depression group than that in the normal control group(t=4.829,FDRp =0.01),but the relative power in the delta band was significantly higher in the depression group than that in healthy controls in this region(t=-3.357,FDRp=0.01).In the right occipital region,the relative power in the delta(t=-4.934,FDRp=0.01)and theta(t=-4.007,FDRp=0.01)bands was significantly higher in the depression group than that in the healthy controls.The relative power of the alpha band in the depression group was significantly lower than that in the healthy controls(t=6.142,FDRp=0.01).In the right parietal region,the relative power of the alpha band was significantly lower in the depression group than that in the healthy controls(t=2.886,FDRp=0.03).2)The weighted phase lag index(w PLI)in the alpha band were significantly lower in the depression group than that in the healthy controls,but the w PLI in the gamma band in depression group were significantly enhanced compared to that in the healthy controls.No significant differences were found between the two groups in other bands.3)In the analysis of microstates,no significant differences were observed between the depression group and the healthy controls in the parameters related to the four types of topographic maps(duration,occurrence and coverage).4)In the calculation of the phase-amplitude coupling,we found that there existed coupling in the delta-gamma band in both groups,but the statistical results revealed no significant difference in the coupling value-modulation index(MI)of the whole-brain average of delta-gamma between the two groups.5)There was no significant correlation between each EEG feature and the HAMD-17 and HAMA scales.2.Based on the above comparison of resting EEG data between the two groups,EEG feature metrics that differed between the two groups were selected,and machine learning algorithms were further used to show that 1)in feature set 1(relative power of ROI in each band),the highest accuracy of 88.2% was achieved using the KNN classifier when using PCA feature selection.2)In the explanatory model using SHAP values,the influence ranked in the top three features are: relative power of alpha band in left parietal region,relative power of alpha band in left occipital region,and relative power of alpha band in right parietal region.ConclusionIn this study,abnormal EEG neural oscillations were found in patients with first-episode depression,which may reflect an imbalance of excitation,inhibition and hyperactivity in the cerebral cortex.Meanwhile,the machine learning in combination with the feature selection algorithm revealed that the relative power of the alpha band in the left parietal region is expected to be an objective electrophysiological indicator of depression.