Research On Abnormal Brain Topology In Mild Depression Based On EEG Signals

With the continuous development of computer and bio-information technology,many studies have attempted to use computers to aid in the diagnosis of psychiatric disorders,and in particular,electroencephalogram(EEG)has been widely used for the identification of depression due to its non-invasive nature,relatively low cost,and convenience.However,most current work on EEG for depression detection is based on supervised learning methods,which require specific labels to identify each instance in the EEG during training.Typically,binary supervised learning models based on depression detection require identifying the selected EEG band data and the labeling prior to model initialization.However,since EEG data in different frequency bands contain much physiological and disease information and dynamic changes,it is difficult for researchers to select the frequency band data that reflects the degree of depression.More importantly,in the current approach,the final prediction of the model is the score obtained by averaging all instances of the subject.However,the robust fitting capability of deep learning may lead to instances that do not exhibit significant symptoms negatively impacting the model performance.Therefore,to address the above issues,this paper first explores the abnormal brain topology of mild depression and then constructs a deep neural network-based model to assist in identifying mild depression based on the findings.The main work and innovations are as follows.1.To study functional brain network analysis in mild depression,this paper used the latest brain functional network analysis method and hierarchical clustering algorithm to explore the abnormal brain topology of mild depression patients based on the EEG data system in the visual search paradigm for the first time.The mild depression group exhibited a significantly longer reaction time compared to the normal control group,as revealed by the behavioral results.The brain functional network analysis revealed notable dissimilarities in functional connectivity between the two groups of participants,wherein the degree of long-range connectivity between hemispheres was considerably more pronounced than the short-range connectivity within hemispheres.Especially in the beta band,the local efficiency(LE)and clustering coefficient(CC)were significantly lower in patients with mild depression.The clustering structure of the frontal and parieto-occipital lobes was disrupted,with brain asymmetry in the frontal lobes.Moreover,there was a significant correlation between depressive symptoms and the mean functional connectivity of distant connections linking the left frontal and right parieto-occipital lobes.Our results suggest that patients with mild depression achieve long-distance connectivity between the two regions by sacrificing connections within the frontal and parieto-occipital regions,which may provide insight into abnormal cognitive processing mechanisms in depression.2.To investigate the multiple instance learning(MIL)assisted diagnosis models of abnormal brain topology of mild depression,this paper proposes a CAMMIL(MIL framework combining attention and max-pooling)model.The model framework uses the max-pooling layer at the instance level to capture information about depressive symptoms.It uses the wrapper layer’s attention weight to further integrate each instance’s contribution.In addition,this paper proposes a brain region features attention fused CNN net(BRFAFNet),which effectively enables whole brain features to be embedded in each brain region.The proposed method achieves 85% accuracy(Accuracy,ACC)and 84.1% area under the ROC curve(AUC),which is 13.1% more accurate than the current state-of-the-art depression detection methods.In addition,this paper also analyzed the band selection problem in the identification of mild depression,and it was found that there was a statistical difference in the band selection of beta band in mildly depressed patients.This may provide new insights for the study and detection of mild depression.The above study suggests an abnormal mechanism of sacrificing connections within frontal and parieto-occipital regions to achieve long-distance connections between the two regions in patients with mild depression,which may provide new insights into the abnormal cognitive processing mechanisms in depression.Secondly,the CAMMIL model,including BRFAFNet proposed in this paper,may provide a reliable methodological reference for mild depression detection under its active capture of frequency band data reflecting depressive symptoms,effectively embedding whole brain features into each brain region feature and higher model evaluation metrics.In addition,it is also significant that this paper found statistical differences in the frequency band selection of beta frequency band in mildly depressed patients.In summary,this paper has identified abnormal brain topology among patients with mild depression and showcased the ability of the CAMMIL model,built based on this structure,to enhance the detection rate of mild depression and offer a new viewpoint for detecting this condition.