Research On Group Sparse Canonical Correlation Analysis Fusion Method For Anxiety Disorder Detection

With the rapid development of society,people bear high pressure,and the number of people suffering from anxiety disorders is increasing yearly.In order to formulate a treatment plan in time,a precise detection of anxiety disorder is very essential.But now most of the detection of anxiety disorder only depend on the doctors' clinical experience and related self-assessment scales.Because of the stigmatization of mental illness,many patients will deliberately conceal their true mental state and provide doctors with more subjective information,then lead to misjudgments and misdiagnosis.In recent years,with the rapid development of cognitive science and behavior,EEG and eye movement have been widely used in the detection of anxiety disorder,because they can reflect the functional activities of the brain and human attentional bias.However,unimodal is easily affected by noise interference or data loss.Therefore,from the perspective of multimodal fusion,we make full use of the correlation between EEG and eye movement and obtain coordinated representation to achieve more accurate detection of anxiety disorder.The main research work and results of this paper are as follows:(1)Proposed a multimodal fusion method based on group sparse canonical correlation analysis: Because the structure and function of each brain region are heterogeneous,EEG with similar functions in the same brain region will undergo similar changes during stimulation,and the EEG features of the same brain region have a strong correlation;At the same time,the correlation between eye movement features of the same category is higher than that of other categories.In order to fully model the correlation between this two signals and obtain a more specific coordinated fusion representation,we propose a new multimodal fusion method based on Group Sparse Canonical Correlation Analysis(GSCCA).It first divides the EEG and eye movement features into 13 groups and 4 groups,respectively,and then uses GSCCA to model the correlation between this two signals.It can not only avoid the over-fitting problem caused by Canonical Correlation Analysis(CCA)method,but also make full use of the group structure information between EEG and eye movement features to obtain accurate detection of anxiety disorder.(2)Proposed a multimodal fusion method based on kernel group sparse canonical correlation analysis: Although the multimodal fusion method based on GSCCA can make good use of the correlation between EEG and eye movement and their own group structure information,it mainly models the linear correlation between this two signals.In order to further investigate the nonlinear complex correlation between EEG and eye movement,we further propose a multimodal fusion method based on Kernel Group Sparse Canonical Correlation Analysis(K-GSCCA).It uses Gaussian kernel function to transform the EEG and eye movement features into the kernel space first,and effectively fuse the EEG and eye movement features to obtain a more effective nonlinear coordinated fusion representation,which can achieve a better detection of anxiety disorder.(3)Selected 92 subjects(45 anxiety subjects and 47 normal subjects)from the HBN database to verify the performance of the above two fusion models.The experimental results show that the multimodal fusion methods based on GSCCA and K-GSCCA all have certain advantages,and the highest accuracy rate based on SVM classification is 77.93 %.In addition,we have completed related experiments on eye movement features and EEG features of different frequency bands as input.The result shows that,by optimizing the gamma band EEG features and effectively fusing them with the eye movement features can greatly improve the detection accuracy with the SVM classification algorithm.Especially,the highest accuracy rate can reach 87.47 % in the fusion method based on KGSCCA.The experimental results not only verified the ability of the gamma band EEG feature to distinguish anxiety subjects,but also verified the advantages of the fusion methods based on GSCCA and K-GSCCA.In summary,multimodal fusion methods based on GSCCA and K-GSCCA first effectively utilize the group structure information of EEG and eye movement,and model the linear and nonlinear correlations respectively to obtain more effective coordinated representation,then further achieve more efficient detection of anxiety disorder.