A Research On EEG Microstate Recognition Model For Depression Based On Convolutional Neural Network

Background Depression has become a worldwide medical problem,and will further aggravate the social medical burden,prevention and control of depression is imperative.The diagnosis of depression is usually made by clinical professionals,according to the Diagnostic and Statistical Manual of Mental Disorders,DSM and International Classification of Diseases（ICD）are used for diagnosis,but misdiagnosis and missed diagnosis still occur due to subjective evaluation.How to establish an objective and effective method for clinical diagnosis for depression has become the spotlight research.Electroencephalogram（EEG）microstate is a neural activity index based on instantaneous global potential.It has the same physiological basis as the functional connection network of functional magnetic resonance imaging（f MRI）and the fixed state in magnetoencephalogram（MEG）,and can make up for the disadvantage of low spatial resolution of EEG signals to a certain extent.It can be used as a more effective signal index to provide reference basis for the identification,diagnosis,treatment and prognosis of mental and neurological diseases and emotional disorders.Today,with the rapid development of artificial intelligence technology,artificial intelligence technology is gradually applied in the medical field.Intelligent algorithms can achieve the purpose of accurate diagnosis and treatment by analyzing the information of patients.As one of the common deep learning methods,convolutional neural network（CNN）can automatically extract physiological signal characteristics of neuropsychiatric diseases,and then classify and diagnose diseases according to these signal characteristics,which is helpful to explore objective biomarkers of various neuropsychiatric diseases.At present,models with high identification and classification accuracy have been obtained in schizophrenia,epilepsy,Alzheimer’s disease and other neuropsychiatric diseases,providing a new idea for the clinical diagnosis of depression.Therefore,on the problems of diagnosis of depression,to get a breakthrough in objective biomarkers,our research focused on different depressive state individuals’ resting state and task-based characteristics of EEG microstate to build a depression identification model based on CNN,which will provide a new perspective to a more accurate and applicable depression diagnosis scheme.Objects and methods 1.Subjects were recruited from the community and hospitals,included in the normal individual group（26）,the depressive state group（26）,and the depressive patient group（26）.2.Research methods（1）General information collection and grouping.Healthy individuals were recruited in Chongqing,and the collection tools included general demographic information tool（selfmade）and Hamilton Depression Scale.Those with Hamilton Depression Scale score < 7 were included in the normal group,and those with Hamilton Depression Scale score < 17 were included in the depressive state group.Patients diagnosed with depression by clinicians were included in the depressive patient group.（2）EEG acquisition and pretreatment.Neuroscan 64 channel EEG acquisition system and Curry 7.0 recording system were used to collect EEG signals,including 8min resting EEG data with eye-open and eye-closed,and the EEG signals within free viewing task.Pretreatment includes electrode positioning,filtering,bad channel elimination,independent component analysis（ICA）,segmentation,etc.,to obtain pure and homogeneous EEG data.（3）Paradigm construction of free viewing task.Using E-Prime 2.0 software,100 images of positive,negative and neutral emotional faces in CFAPS（Chinese facial affective picture system）were randomly selected as stimulus materials,and each type of stimulus was marked.The images appeared for 3000 ms with an interval of 1000 ms.（4）Analysis of resting EEG microstate characteristics.Matlab was used for resting state EEG microstate analysis,and AAHC algorithm was used to obtain the optimal clustering number of 4.The parameter values of the four kinds of microstate in the eye-open and eye-closed state were statistically analyzed to obtain the resting state EEG microstate features of individuals in different depressive states.（5）Analysis of characteristics of task-based EEG microstates.EEG signals were recorded during the free viewing task.EEG data were extracted when positive,negative and neutral stimuli appeared in the range of 0¹000ms.After superimposing and averaging datasets,Cartool was used for cluster analysis to obtain the optimal EEG microstate cluster number and parameters.Various parameters were analyzed statistically in order to obtain the features of EEG microstates of individuals in task-based state.（6）Construction of CNN model.According to the CNN structure,The basic framework of Res Net was built according to the structure of CNN,including convolutional layer and residual structure,input layer,batch normalization layer,activation function layer,full connection layer and output layer.On this basis,a complete network model is constructed based on Maxpooling,LSTM,1Dconv and LSTM-1DCONV,respectively.Two-dimensional images of EEG microstates in resting state and task-based state were drawn,and 13-fold cross validation was used.The CNN model for depression recognition based on EEG microstates was preliminarily constructed by debugging the model and comparing the training results.3.Statistical methods SPSS23.0 was used for statistical analysis,mainly including descriptive statistical analysis,paired sample t-test,mixed analysis of variance,etc.Inspection level: α=0.05.Results 1.The optimal clustering number of resting EEG microstates was 4（ABCD）.Under the condition of eye-closed,there was no significant difference in global explanation variance（GEV）of EEG microstates in individuals with different depressive states（P>0.05）.The difference of GEV was statistically significant in the three groups under eye-open state,and the GEV in the depressive patient group was lower than that in the normal individuals and the depressive state group（P<0.05）.The duration of microstate A in normal individuals and depressive state group was higher than that in depressive patient group（P<0.05）.The duration of microstate D in depressive state group was higher than that in depressive patient group（P<0.05）.The frequency of microstate D in normal individuals and depressive state group was lower than that in depressive patient group（P < 0.05）,and the coverage rate of microstate A in depressive state group was higher than that in depressive state group（P<0.05）.The duration of microstate A,microstate B and microstate D was longer in the eye-closed state than in the eye-open state（P<0.05）.There was significant difference in the transition rate of EEG microstates among individuals with different depressive states（P<0.05）.2.Under the stimulation of different emotional faces,the optimal clustering number of EEG microstates of individuals in different depressive states was 4,which was different from the resting state category,and the relatively high potentials were located in different brain regions,named as 1⁴ categories.The global explanation variance（GEV）of task-based EEG microstates ranged from 49.61% to 54.12%,and there was no significant difference among individuals with different depressive states（P>0.05）.There were statistically significant differences in the EEG microstate parameters of individuals with different depressive states under different stimulus conditions（P<0.05）,and there was an interaction between the dispersion and global explanation variance of different depressive states and different stimulus conditions（P<0.05）.The global field power（GFP）,dispersion（Dis）and global explanation variance（GEV）of EEG microstates induced by neutral emotion face stimulation were the highest,and the differences were statistically significant（P<0.05）.The global field power（GFP）and dispersion of EEG microstates were the lowest in the depressive patient group,and the global explanation variance（GEV）was the highest in the depressive state group,followed by the depressive patient group and the healthy group（P<0.05）,while the other differences were not statistically significant（P>0.05）.Microstate 3 under the condition of different depressive state individual stimulus has the higher global field power and the higher global explained variance with low dispersion（p<0.05）,Microstate 2 only appearing in depressive patient group in 3 different emotional face stimulus conditions,has the higher global field power and the higher global explained variance with low dispersion（p<0.05）.3.The accuracy of convolution neural network model verification set and test set constructed by resting EEG microstate images is higher than that constructed by task-based images（P<0.05）.Different models have certain interaction with resting and task-based acquisition conditions（P<0.05）.ConvNet_maxpool and ConvNet₁DConv had higher efficiency,while ConvNet_LSTM had lower efficiency（P < 0.05）.The efficacy of resting state EEG data model and task-based EEG data model was significantly different among individuals with different depressive states,and the classification efficacy of depressive state group was lower（P < 0.05）.Conclusion 1.The global explanation variance of resting EEG microstates is different in depressive patient group in eye-open and eye-closed condition.In the resting state,different types of EEG microstates occur and change frequently in the eye-open state,but maintain a stable state while in eye-closed condition.The decreased duration of microstate A,the increased appearance of microstate B and the transition rate of microstate D may be closely related to individual depressive state.2.The clustering of EEG microstates in task-based is different from that in resting state,and the global explanation variance is lower than that in resting state.The parameters of task-based microstate were affected by 3 different depressive state groups and 3 different stimulus conditions.Microstate 2 may be the unique type of microstates in depressive patient group,and microstate 3 was stable in the three individual groups.3.The CNN recognition model of depression based on EEG microstates can obtain rather higher accuracy.The model constructed by resting EEG microstates images has better classification efficiency.The ConvNet_maxpool and ConvNet₁DConv models had higher efficiency,while ConvNet_LSTM model had lower efficiency（P < 0.05）.In the future,there is still space for improvement in the classification model construction of depressive state individuals and the classification model construction based on task-based EEG microstate images.