A Study Of Mild Depression Classification Based On EEG And EMs Features

Depression is a kind of mental illness,which widely distributed all over the world and seriously disturbs the physical and mental health of modern people.The diagnosis of depression mainly relied on a combination of questionnaires and doctor's consul-tation.Due to the uncontrollable factors and the subjectivity of doctors,this method may cause misjudgment.In recent years,with the maturity of machine learning tech-nology,more and more research combined depression-related physiological indicators with machine learning,and achieved better depression diagnosis results.Among them,Electroencephalography(EEG)is an electrophysiological signal generated by brain ac-tivity,taking the advantages of being non-invasive and not easy to disguise,while Eye Movements(EMs)are relatively intuitive reflection of a person's mental state,so these two kinds of physiological indicators are often used in the study of depression.There is a certain correlation between the EEG modal and EMs modal,It is essential to fully ex-plore the correlation between the two modals' data to obtain a perfect feature represen-tation for training an accurate model.Autoencoder is a commonly used unsupervised feature learning tool.Compared with autoencoders based on back-propagation neural networks,the training of Extreme Learning Machine Sparse Auto Encoder(ELM-SAE)directly determines the network parameters through linear solving methods without tun-ing,so It has a faster training speed and has unique advantages in engineering applica-tions.Based on ELM-SAE,the main works of this paper includes:(1)Using emotional face free viewing paradigm,the EEG data and EMs data of 18 mild depressed subjects and 21 normal controls were collected from students.For EEG data,this paper extract 17 linear features and 9 non-linear features on 16 electrodes from ?,?,?,slow_?,? and the full-band 6 frequency bands respectively.For EMs data,baseline correction are performed on pupil waveforms,after that,we innovatively ex-tract 17 baseline pupil characteristics reflecting the instantaneous change characteristics of the pupil,and 16 features derived from the original eye tracker are spliced together for classification research.The features of the two modalities are preprocessed respec-tively to obtain the EMs characteristics and EEG characteristics under the conditions of positive stimulation and negative stimulation.(2)Using EEG and EMs single-modal features,7 classifiers were trained to clas-sify depression under two conditions.On the ? band features of EEG single-model,using SVM with Gaussian kernel we achieved the highest classification accuracy rate of 74.56%under positive stimulation conditions.On the single modal features of EMs,the decision tree are used to obtain the highest classification accuracy rate of 62.93%under the condition of positive stimulation.Using ELM-SAE to obtain the hidden layer features of the two modalities,on the EEG single modal full-band features,using Gaus-sian kernel SVM under the condition of positive stimulation achieved the highest clas-sification accuracy rate of 75.32%.In terms of EMs single modal features,using KNN to achieve the highest classification accuracy rate of 64%under negative stimulus con-ditions.The hidden layer feature representation obtained by ELM-SAE has improved the classification effect under multiple classifiers,which proves its effectiveness.(3)In order to fully explore the correlation between the two kinds of features of EEG and EMs,this paper proposes two feature layer fusion strategies,hidden layer splicing and hidden layer mixing,based on ELM-SAE.Using the hidden layer splicing strategy,under the condition of positive stimulation,the full-band EEG and EMs hidden layer feature splicing achieved the highest classification accuracy rate of 78.53%on the linear SVM classifier.Using the hidden layer hybrid strategy,the full-band EEG and EMs hidden layer hybrid features achieved the highest classification accuracy rate of 78.42%on the linear SVM classifier under the condition of positive stimulation.Compared with the highest results of EEG single-mode hidden layer,the results were increased by 3.21%and 3.1%respectively.By analyzing the results,it is found that the classification effect under the positive stimulus is better than the classification result under the negative stimulus.The results of the paired-sample t-test show that most of the results under ? and the full-band are significantly higher than the results of ?,? and slow_? bands.(4)The feature-level fusion method has different characteristics in different wave-bands.In order to improve the stability of the model and make full use of dominant bands,this paper proposes a decision-level fusion algorithm based on Bayesian the-ory.On the basis of the feature representation learned by the feature-level fusion,under the positive stimulus conditions,the three modes of ?,? and the full-band are used to perform the decision-level fusion under four decision vectors,and finally The high-est classification accuracy rate of 80.44%was achieved under Gaussian kernel SVM.Compared with the highest result of feature-level fusion,it is increased by 1.91%.In summary,the proposed fusion method can effectively utilize the advantages of multiple modalities,improve the recognition rate of mild depression,and has certain reference value for the applications of mild depression recognition.