Research On Optimization Methods Of Channel Space And Intrinsic Feature Of EEG Signals For Depression Recognition

In recent years,the rapid development of the economy and society has improved people's material life level,but also brought the accelerated pace of life and the intensification of social competition,increasing the physical and mental pressure of the human race,the prevalence rate of depression continues to rise,seriously endangering human life and health.Depression is a mental illness with affective and cognitive dysfunction,which noticeably affects patients' thoughts,behaviors,feelings,and sense of well-being,and also a leading cause of suicide.Current clinical diagnosis of depression is mainly based on self-report questionnaires and interviews,while diagnoses by psychiatrists vary depending on their clinical profession and the diagnostic approaches.However,in clinical practice,the number of experienced psychiatrists and depressed patients is extremely unbalanced,besides,clinical practice takes a lot of time.Therefore,relying on the booming Artificial Intelligence(AI)and Affective Computing technology to develop an objective and effective depression recognition method is crucial.Electroencephalogram(EEG)signals are one of the physiological data that can objectively reflect the inner working status of the human brain and are regarded as a tool to enable efficient and effective clinical depression diagnosis and detection.In current research of depression recognition,the insufficient optimization and selection of the multichannel space and the limited channel subspace of EEG signals can limit the effect of depression recognition.Nevertheless,the traditional Empirical Mode Decomposition(EMD)intrinsic feature extraction method can not extract effective features from highly complex and non-stationary EEG signals because the inverse for the matrix product of the Intrinsic Mode Functions(IMFs)does not exist.Based on these problems,the paper starts with the optimization of EEG channel space to explore the channel weighting optimization and the optimal channel subspace selection methods for EEG signals.On the other hand,starts with the optimization of intrinsic features of EEG signals to extract effective features that can characterize EEG intrinsic characteristics.The main work and innovations of this paper are as follows:1.To solve the problem of insufficiently optimized channel weight setting of EEG in depression recognition,we propose the loss minimization and the adaptive channel weighting method based on the common and individual information of EEG channels,which fully explores the EEG spatial information from the multichannel space of EEG.The loss minimization channel weighting method starts with the common information of EEG channels to fully explore the EEG spatial information that best reflects the characteristics of depression recognition by learning the weights that minimize EEG channel space and class loss.The adaptive channel weighting method first constructs the optimal kernel matrix for the EEG features of each channel and calculates the similarity between each EEG segment and each class.Then the adaptive weight of each channel of each EEG segment is calculated using the similarity.Learning adaptive weights can dynamically explore the optimal spatial information of each EEG segment.The adaptive channel weighting method addresses individual information for each EEG segment,the weights are dynamically set by taking into account the different contributions of each channel of each EEG segment to the depression recognition task.Thus,each segment of EEG can be spatially optimized.Experimental results on two EEG datasets demonstrate that the proposed channel weighting methods can effectively improve the effect of depression recognition compared to some other state-of-the-art methods.2.Aiming at the problem of insufficient selection of the limited channel subspace of EEG signals,we propose an optimal channel selection for EEG-based depression recognition via Kernel-Target Alignment(KTA),which is oriented to the multichannel space of EEG,and explores the channel subspace most relevant to depression recognition tasks to optimize the EEG channel space.This method first considers a modified KTA to measure the alignment between the EEG channel subspace and the target space.Afterwards the method proposes an optimal channel selection strategy to optimize the channel subspace,so as to find the channel subspace with the highest alignment between EEG spatial information and the target for depression recognition.This channel selection method can dynamically select the channel subspace that most suitable for depression recognition based on the spatial information of EEG,and it can also explore the correlation between the brain area corresponding to the channel and depression recognition.Moreover,this method can also effectively reduce the computational complexity,space complexity,and experimental setup complexity,in addition it can efficiently use EEG spatial information with high rationality and interpretability.Experimental results on EEG datasets show the effectiveness of the proposed channel selection method.3.To solve the problem that the traditional EMD intrinsic feature extraction method can not extract effective features,we propose Singular Value Decomposition(SVD)-based and regularization parameter-based intrinsic feature extraction methods by introducing SVD and regularization parameter,respectively.The proposed methods are oriented to the intrinsic characteristics of EEG and extract effective EEG features that can fully reflect the characteristics of depression recognition.The SVD-based intrinsic feature extraction method takes advantage of SVD to decompose the matrix product of the IMFs,getting one rectangular diagonal matrix with non-negative real numbers on the diagonal and computes the pseudo-inverse of matrix product of the IMFs by replacing every non-zero diagonal entry with its reciprocal in the diagonal matrix to achieve the purpose of optimizing intrinsic features of EEG signals.The regularization parameter-based intrinsic feature extraction method leverages the addition of the regularization parameter diagonal matrix and the matrix product of the IMFs to ensure the invertibility of the sum matrix and takes advantage of the inverse of the sum matrix to represent the inverse of the matrix product of the IMFs to extract the intrinsic features of EEG signals effectively in an optimized way.Experimental results on several EEG datasets demonstrate that the effective features extracted by the improved intrinsic feature extraction methods can fully reflect the intrinsic characteristics of EEG and effectively improve the effect of depression recognition.By optimizing the channel space and intrinsic features of EEG,channel weighting,channel selection,and feature extraction are used as entry points to improve the effect of depression recognition.The proposed channel weighting methods and channel selection method can explore the optimized channel weights and the optimal channel subspace of EEG from the multi-channel space of EEG to optimize spatial information of EEG signals;the proposed feature extraction methods start with the optimization of intrinsic features of EEG signals to explore the intrinsic differences of different types of EEG signals in depression recognition.The purpose of this study is to solve the problems existing in the existing depression recognition research framework.We have obtained gratifying results both in the theory and the application.The research in this paper will further promote the development of intelligent applications of depression recognition.