A Study Of Depression Detection And Data Mining Algorithm Based On EEG Signal

At present,depression has become a major health and economic burden worldwide.However,many disadvantages exist in current diagnosis,such as patient denial,subjective biases and inaccuracy.Hence,effective detection of depression needs a reliable and objective evaluation method.Electroencephalography(EEG)with high temporal resolution,easy to record and non-invasive advantages becomes our optimal choice.According to the morbidity degree,depression can be divided into mild,moderate and severe depression,and depression is a changing process.So accurate detection,prevention and avoid state of it getting worse will be an urgent problem to be solved.In recent years,although many studies have used classifiers and feature selection methods based on EEG signals to detect depression,there are few studies aiming at mild depression,and the recognition accuracy also needs to be improved.So in this paper we process and analyze EEG signals of 10 mild depressive patients and 10 normal controls under emotional facial expression pictures task,and hope to find prominent frequency band and brain region that significantly related to mild depression,as well as an optimal combination of classification algorithms and feature selection methods which can be used in future mild depression detection.Meanwhile,for the purpose of simplifying experiment task and achieving the pervasive application,this paper also collects 5min resting state EEG data for 30 moderate to severe depressive patients and 17 normal controls.The aim is to investigate whether some features of resting state EEG can be applied as a biomarker to distinguish depression patients from normal controls.Our main work and contributions will be described as below:(1)8 linear and 9 non-linear features were extracted from EEG data of mild depression patients.Because of the higher features dimension,in order to remove the redundant and low discriminative performance features,BestFirst(BF),GreedyStepwise(GSW),GeneticSearch(GS),LinearForwardSelection(LFS)and RankSearch(RS)based on Correlation Features Selection(CFS)were applied for features selection,then 5 typical classifiers: BayesNet(BN),Support Vector Machine(SVM),Logistic Regression(LR),k-nearest neighbor(KNN)and RandomForest(RF)were used to classify data.Experiment results indicated that the combination of GSW+KNN could achieve the optimal performance.Beta frequency band played a more prominent role than alpha and theta frequency bands in depression identification.Accuracies achieved 92.00% and 98.00% for Emo_block and Neu_block beta band data respectively,and in this case,the average classification accuracy of GSW+KNN was respectively increased 4.17% and 9.25% than GSW+ other classifiers(BN,SVM,LR and RF).T-test results validated the effectiveness of this method.Analysis of features distribution in the brain regions which could achieve the highest accuracy was done,and results showed that left parietotemporal lobe and linear features had greater effect on mild depression detection.Simplified EEG system(FP1,FP2,F3,O2 and T3)combined with linear features yielded accuracies above 91.00% for both the two block tasks,which laid a solid foundation of the implementation of portable system.(2)According to the conclusions drawn from above research,and combined with relevant research,we can know that most of non-linear features exist some problems,such as large amount of calculation,hard to real-time monitoring and poor repeatability,so as to better realize the real-time monitoring of depression,this paper only extracted 8 linear features for moderate to severe depressive patients,and GSW+KNN was used to classify single-channel electrode of 128 channel electrodes with linear features,accuracy reached 80.85%.Brain topography analysis method was used to analyze the features that had significant differences(activity and complexity calculated based on Hjorth parameter),results showed that both of the values of activity and complexity in alpha,beta and theta frequency band for depressive patients are higher than normal controls.And differences are mainly in parietal-occipital brain region and left temporal lobe.The correlations between features(activity and complexity)and PHQ-9 scale scores indicated that both the activity value of alpha,beta and theta frequency band in parietal brain region(E51,E64,E89 and E92),and complexity value of alpha frequency band in occipital lobe(E73,E75 and E76)are positive related to PHQ-9 scores.Hence we concluded that these two features,especially the activity feature can be used as a sensitive index of depression.Finally,according to the aforementioned two parts research results,we can obtain more meaningful conclusions.First,in this paper,GSW+KNN is the optimal combination method that can achieve more accurate depression identification,and the short running time makes it more suitable to be applied to real-time systems.Second,linear features especially activity feature play a very important role in depression recognition,and the increased activity value of alpha,beta and theta frequency band in parietal lobe(E51,E64,E89 and E92)may be used as an important biological indicator for early-stage defense and auxiliary diagnosis of depression.