Research On Construction Methods Optimization And Application Of Functional Brain Network For Depression Recognition

With the development of brain science and technology,using various neuroimaging techniques to explore the pathological mechanism of brain diseases has become a current research hotspot in the fields of brain science and medicine.Depression is a globally prevalent psychiatric disorder,which causes serious harm to individuals,families and society.Therefore,it is of great significance to study the abnormal neurophysiological bases of depression,and explore its objective and reliable physiological quantitative biomarkers for clinical auxiliary diagnosis,early prevention and intervention.Existing studies have shown that patients with depression have abnormal functional connectivity patterns.And electroencephalography(EEG)with its non-invasive,high time resolution,portability and practicability has attracted much attention in the study of brain functional connectivity.However,the current analysis methods of functional brain network based on EEG are affected by various subjective and objective factors,there is no recognized“gold standard” for the construction methods of functional brain network;and there is no unified conclusion on EEG biomarkers system for depression recognition.Therefore,based on the above two key problems,the present work takes EEG as the research object,uses complex brain network analysis methods as the research method,and explores the specific changes of brain network topology structure in patients with depression as the research purpose,aiming to provide reliable technical method and reproducible EEG physiological indicators for accurate screening and the auxiliary diagnosis of depression.The main contributions and innovations are as follows:1.Due to the coupling methods are influenced by artefacts of volume conduction,and binarization techniques are often subjective and arbitrary in the functional brain network analysis based on EEG signals.In this work,Imaginary Part of Coherency +Cluster Span Threshold(ICoh + CST)was constructed,which could effectively avoid the influence of volume conduction.On the other hand,by ensuring a trade-off of sparsity and density of network structure,it was easier to capture the subtle differences of topology changes.The performance of different coupling methods(Coherence,ICoh,Pearson Correlation Coefficient,Phase Lag Index and Phase Lock Value)and binarization techniques(CST,Efficiency Cost Optimization Threshold,Minimum Spanning Tree and Density)were systematically evaluated and compared.The results showed that ICoh + CST method was significantly better than other combination methods.Based on this method,right hemisphere function deficiency,symmetry breaking and randomized network structure were found in patients with depression.In addition,in order to explore the robustness and generalization of ICoh+ CST method,this paper analyzed the EEG signals of sensor layer and source layer under different data sets.Based on the change of time window length,the influence on the constructed network attributes was evaluated.It was found that ICoh + CST had the robustness,and the network attributes on the source layer could reach a stable state in a shorter time window.Moreover,this work also explored the network topology differences between depressive patients and normal subjects at sensor layer and source layer on different data sets.The experiments results showed that EEG signal based on source layer were more likely to obtain consistency results.In summary,this work optimizes the existing functional brain network analysis methods,and provides a reliable reference method for discovering EEG physiological indicators of depression with generalization and consistency.2.Aiming at the uncertainty of effective indicators for depression recognition in EEG research,this work studied the effective physiological indicators for classifying depression based on multi-types of EEG features(such as linear,nonlinear,functional connectivity and network attribute features).The results showed that the combination of multi-type EEG feature set with a robust classifier could obtain better and more stable depression recognition effect.By analyzing the distribution characteristics with discriminant performance EEG features,it was found that compared with other types of features,functional connectivity features occupied the largest proportion,of which functional connections in intra-hemisphere played an important role in depression recognition.The recognition accuracy of high frequency bands(alpha and beta)was higher than that of low frequency bands(delta and theta),and the dysfunction of parietal occipital lobe was more obvious in high frequency bands.In addition,the statistical results showed that there were significant differences in each type of EEG features between the depressive and normal groups.This study is of great significance for the extraction of heterogeneous EEG biomarkers,and the comprehensive interpretation and quantification of brain function in patients with depression.3.In view of the characteristics of small sample size and individual differences in the data of patients with depression in clinical research,this paper proposed a clustering-fusion based feature selection method(HCSNF),which aimed to mine the representative brain network atlas that could characterize different populations,and achieve the acquisition of reproducible EEG physiological indicators.Firstly,hierarchical clustering algorithm was used to cluster the functional connectivity data of patients with depression and normal subjects respectively,and the sub-population clusters representing the two groups were obtained.Secondly,different fusion strategies were used to fuse the sub-population clusters of each group,and then the representative brain network atlas that could characterize each group was obtained.Finally,the absolute difference of representative brain network atlas between the two groups was calculated to screen out the connectivity features with discriminant performance.The results showed that the HCSNF feature selection method was significantly better than the traditional feature selection methods.Especially in the beta frequency band,this method could improve the accuracy by more than 6% when using support vector machine classifier to classify EEG data in sensor layer.In addition,it was found that long-distance connections between parietal occipital lobe and other brain regions played a crucial role in depression recognition.Therefore,this study may provide methodological guidance for the discovery of reproducible EEG physiological indicators,and provide new insights into the common neuropathological mechanisms of heterogeneous depression diseases.