Research On MDD Recognition Using Machine Learning Based On Behavior And ERP Data

Depression is one of the most common mental diseases in the world.According to the statistics of the World Health Organization(WHO),350 million people around the world have suffered from depression to varying degrees,accounting for 4.4% of the world's total population,and the incidence continues to rise year by year.Patients with depression show persistent depression,loss of pleasure,severe cognitive impairment and even suicidal attempts.It brings heavy psychological and economic burden to its individual,family and society.Among them,major depression disorder(MDD)has a greater threat than other degrees of depression.Statistics show that China's annual economic losses caused by depression as high as 51.37 billion yuan,has become China's second largest burden of disease.Based on this,the Office of the National Health Commission released Exploring the work plan of characteristic service for depression prevention in September 2020,providing a comprehensive set of solutions for screening,prevention and treatment of depression.Proving the diagnosis and treatment of depression has become the focus of current work.However,the traditional MDD diagnostic methods have the problems of low diagnostic efficiency,strong subjectivity and high misdiagnosis rate.And lack of objective and reliable diagnostic indicators,according to statistics,general practitioners can only correctly identify 47.3 % of depression.Therefore,it is of great practical significance to explore the potential neurophysiological basis of MDD,find more effective and objective auxiliary diagnostic physiological indexes and more effective auxiliary diagnostic methods for MDD for its early diagnosis and prevention.With the development of computer-aided diagnosis technology,a large number of studies have found that patients' behavior information and EEG information can effectively assist in the diagnosis of depression.Electroencephalogram(EEG)has been widely used in the identification of depression,but it also faces the challenges of noise signal complexity,large number of leads and high data dimension.As a special task evoked potential,event-related potential(ERP)can explore the neuroelectrophysiological changes in the cognitive process of patients by giving task stimulation with special psychological significance.It has the characteristics of significance,controllability and low data dimension,which can effectively compensate for the shortcomings of EEG.However,ERP data and behavioral data are seldom used in machine learning assisted diagnosis of depression.Therefore,based on the emotional face-word Stroop task,through the behavioral data and ERP data of the subjects,this paper used machine learning method to analyze 31 MDD patients and 31 healthy controls,in order to seek a more accurate,effective and fast physiological indexes and computer-aided diagnosis method for MDD.The main contents and results of this paper are as follows :(1)The behavioral data collected through E-prime were preprocessed,and statistical analysis was carried out in the two dimensions of reaction time and accuracy.The results showed that there were significant differences in the two dimensions between MDD patients and healthy controls.The response time and accuracy of MDD patients were significantly slower or lower than those of healthy controls.Then,in order to explore the effectiveness of behavior data as a computer-aided diagnosis index of MDD,traditional machine learning,ensemble learning and deep learning methods were used to identify MDD based on the reaction time data of implicit correct rate information of MDD patients and healthy controls.The results show that compared with other machine learning models,convolutional neural network(CNN)has the highest recognition accuracy of 89.76 % ± 19.18 %,which can effectively identify MDD.(2)ERP was used to explore the mechanism of emotional inhibition control ability in MDD patients.The results showed that there were differences in the classical components of P1,N170,N2,P2,EPN,N300,P3 b,P450,CSP and LPP between MDD patients and healthy controls.In this study,a new ERP component was found in the parietal lobe,and it was named as Difference Wave of Depression(DWD).It was found that there was polarity reversal in the evoked potential of DWD components induced by emotional face stimulation in MDD patients and healthy controls.Specifically,positive waves were induced in MDD patients and negative waves were induced in healthy controls.All of the above results show that there are some related disease characteristics in ERP data.(3)In order to explore whether ERP components can be effectively used as a physiological index for computer-aided diagnosis of MDD,Support Vector Machine,e Xtreme Gradient Boosting,Light Gradient Boosting Machine,Convolutional Neural Network,Long Short-Term Memory and Temporal Convolutional Network are also applied to explore the recognition effect of ERP data in MDD.The results show that compared with other machine learning models,convolutional neural network has the highest recognition accuracy of 90.71 % ± 14.17 %,which can effectively identify MDD patients.At the same time,ERP data can greatly reduce the data dimension,so as to greatly improve the efficiency of model operation,and can be effectively used for computer-aided diagnosis of MDD.(4)This study proposes a multi-modal deep learning neural network based on behavior data and ERP data,which is named behavior-ERP parallel sequential convolutional neural network(BEPTCNN).The model uses Convolutional Neural Network and Temporal Convolutional Network parallel processing model architecture,which integrates individual behavior information,spatial information and time information of ERP data,and can effectively take into account the relevant characteristics of MDD in behavior and EEG signals.It has been proved that the recognition accuracy and F1 scores of BEPTCNN for MDD are significantly higher than those of other machine learning or deep learning models,and the recognition accuracy can reach 95.48 % ± 7.31 % and F1 score can reach 0.95 ± 0.08.The above studies show that MDD have a large number of disease-related characteristics in behavioral information and EEG signals.ERP data can be applied to computer-aided diagnosis of MDD.However,ERP data are still less applied to the research of deep learning,which may become one of the research directions of deep learning in the future.Secondly,the BEPTCNN model proposed in this paper may become a new deep learning model for computer-aided diagnosis of MDD due to its more comprehensive data information and higher recognition accuracy.However,due to the limited number of data,it can be verified on more data sets in the future to determine the generalization ability of BEPTCNN.