| The booming development of neuroimaging has driven the direction of brain science to become increasingly prosperous,and exploring the pathological mechanisms of neurological diseases has become one of the current research hotspots in the field of brain science and biomedical engineering.Depression is a very common psychological disorder with high incidence and recurrence.Therefore,it is important to study the pathogenesis of depression and explore reliable physiological indicators for early diagnosis of depression.Both functional Near-Infrared Spectroscopy(f NIRS)and Electroencephalogram(EEG)are portable devices that have attracted much attention in the field of brain science for their high safety and other advantages.To address the current lack of objective and accurate criteria for depression diagnosis and the inability to identify depressed patients in a simple and effective way,this paper aims to improve the accuracy of identifying depression and provide a new method and basis for the clinical diagnosis of depression by using f NIRS and EEG signals as the research objects and feature layer fusion and decision layer fusion as the research tools.The details of the study are as follows:(1)Research on depression recognition based on feature layer fusion.In this paper,an emotional audio stimulation paradigm was designed,and information on prefrontal EEG and blood oxygen concentration changes of subjects were acquired by f NIRS synchronized with EEG in three emotional states.First,to explore generalizability and reproducible physiological indicators,a comparative analysis of single features of unimodal states was conducted in this paper.The results showed that the accuracy of changes in oxygen exchange was the highest with F1 score.Secondly,to explore the reliability and robustness of depression recognition in bimodality,this paper conducted a unimodal classification study of EEG and f NIRS,respectively,and then performed a splicing fusion of EEG and f NIRS at the feature layer.It was found that the average accuracy of bimodality was improved compared with both unimodality,and the F1 score under fearful emotion stimulus was improved by 0.12%,indicating that the bimodal fusion effectively improved the reliability and robustness of depression recognition.Finally,in order to explore simple and effective features in depression recognition,this paper utilized the maximum correlation maximum distance algorithm and Relief algorithm for feature selection.It was found that even using a small number of features could improve the accuracy rate,which provided some reference significance for feature selection algorithms in depression recognition.(2)Research on depression recognition based on decision layer fusion.Based on Relief feature selection,a decision layer fusion model was constructed to analyze depression representation information from multiple perspectives and levels with a modal dominance complementary strategy to achieve more robust and accurate depression recognition.To fully utilize the advantages of multiple classifiers,the Stacking integrated learning model was used in this paper.The results showed that the Stacking model improved the accuracy by 5.84% and the F1 score by 4.17% compared to the first-level model when pleased with audio stimuli;in addition,in order to take full advantage of the bimodal dataset,this paper also simplified the content-based decision layer fusion method and applied it to the data after Relief feature selection.It was found that the accuracy rate was improved by 3.34% and F1 score was improved by 2.07%.In conclusion,this study explores a depression recognition method with bimodal fusion of EEG and f NIRS.Based on the feature layer fusion,a suitable decision layer fusion method is selected to further improve the reliability and robustness in depression recognition.In addition,this paper also explores simple and effective methods of physiological indicators and feature selection,which are informative for depression recognition studies. |
PDF Full Text Request