Towards Depression Recognition Using EEG And Eye Tracking: A Content Based Ensemble Classification Model CBEM

Depression,threatening the well-being of millions,has become one of the major diseases in the past decade.Major depression significantly affects a person's family,personal relationships and other general health aspects.However,the assessment methods of diagnosing depression rely almost exclusively on patient-reported or clinical judgments of symptom severity.Current diagnostic techniques of depression have obvious disadvantages,which are associated with patient denial,poor sensitivity,subjective biases and inaccuracy.All these disadvantages make depression diagnosis a labor intensive work.Recently,some studies try to employ some bio-signals such as Electroencephalography(EEG),Eye Movements(EMs)combining with data mining methods to build an automatic and accurate method to detect depression.These biosignals are objective,harmless and easy to collect.In this paper we propose a content based ensemble classification model(CBEM)to improve the accuracy of depression detection.CBEM divides the dataset into different data subsets by the content of the data and then training classification models based on these data subsets.Finally,using these models to classify one subject and combining these results as the final diagnose for the subject.This paper uses 3 datasets from different human recognition experiments to test the effect of CBEM including eye movements,resting-state EEG and Stroop EEG experiments.And these 3 experiments contain 36,34,40 subjects respectively.The following part will elaborate the main work of this paper.(1)In this paper,we used some classic methods to denoise,fill empty values,remove outliers and perform data normalization.Based on previous studies we extract features from EEG datasets.Then we use BayesNet,Logistic,RandomForest,J48,NaiveBayes,SVM,KNN algorithms to classify test subjects.In human recognition experiments,one subject usually has more than one data tuple and we call this ‘non-integrated' data in this paper,while in CBEM we average all data tuples into one data tuple for every subject and we call it ‘integrated' data.We used same classification algorithms on both integrated data and non-integrated data to see if there is a significant difference between them.According to the results,the traditional classification on nonintegrated data achieves 73.89%,75.02%,78.19% respectfully on three datasets while on integrated data the traditional classification achieves 70.28%,78.24%,77.32% respectfully.The result of Wilcoxon signed-rank test shows that there is no significant difference between results of the integrated data and results of the non-integrated data.(2)In the dynamic CBEM,dataset will be divided into several data subsets according to the experiment paradigms.Take the eye tracking dataset for instance,in the eye tracking experiment,there are five kinds of face,divide the dataset by the face types so there will be five data subsets corresponding to the face types.Then perform the data integration,employ several classification algorithm on these data subsets,based on the results of data subsets,the CBEM will dynamically choose some data subsets to train classifiers.The test data will test one these classifiers,the majority vote will be the final diagnosis for a certain subject.The dynamic CBEM achieves accuracies of 78.50%,85.00%,89.50% respectfully in three datasets.The result of Wilcoxon signed-rank test shows the results of dynamic CBEM are significant different with traditional classification.(3)According to the results of data subsets in the dynamic CBEM,we found that some data subsets and classification algorithms are more frequently chosen by CBEM,so we improve the dynamic by using some fixed data subsets and classification algorithms to train classifiers and we call such model ‘Static Model'.The static CBEM achieves accuracies of 82.50%,92.65%,92.73%,the result of Wilcoxon signed-rank test shows that the results of static CBEM are significant different with dynamic CBEM.Compared with dynamic CBEM,the static CBEM has higher accuracies and lower deviation,the calculation time of static CBEM is less than dynamic CBEM.These advantages make static CBCM more suitable than actual application requirements.According to all the results in this paper,we make such a conclusion that CBEM can truly improve the depression detection accuracies by dividing dataset into different subsets is meaningful and cooperating with vote strategy.