EEG-Based Spatio-Temporal Filters Modeling And Its Application In Depression Recognition

With the increasing pressure of modern society and heavy work,the incidence of mental disorders,represented by depression,continues to rise,which has brought a heavy burden to both society and families.However,due to the subtle symptoms of depressive disorders and the cross-symptoms with other diseases,and the fact that existing diagnostic methods are significantly affected by the subjectivity and variability of doctors’ experience,it is difficult to objectively and effectively identify and detect depression patients,which hinders early intervention of the disease.Electroencephalogram(EEG)could intuitively reflect the functional structure of the human brain and the activity state of the central nervous system.Event-related potentials(ERP),in particular,are electric signals induced by stimuli that can reflect specific cognitive activities of the brain.Therefore,ERP can be used as an objective biomarker to distinguish patients with depression from healthy individuals.The time and space domains of ERP signals contain rich information,but most of the current methods only analyze ERP from a single dimension,lacking comprehensive consideration of the information of two dimensions.To solve this problem,this thesis introduces a spatio-temporal filtering method to analyze the ERP signals and constructs a classification recognition model based on spatio-temporal linear discriminant analysis(STLDA)according to the spatio-temporal characteristics of the signals for effective detection of depression.The main work of this thesis includes:(1)In-depth analysis of the temporal and spatial domain characteristics of ERP signals in patients with depressive disorders.To explore the spatial distribution characteristics at the scalp electrode positions as well as the time-locked and phase-locked characteristics in the temporal domain of ERP signals of patients with depressive disorders,in this thesis,the averaged signals of several trials are used to draw the EEG waveforms of different electrodes for the entire time period,and the signals of a specific time period are averaged in the temporal dimension to draw the scalp topography,thus the spatial distribution characteristics of ERP signals at different electrode positions are qualitatively analyzed.Further,repeated measures analysis of variance is used to quantitatively evaluate the distribution positions of the scalp electrodes which have significant differences in the ERP signals between patients with depressive disorders and healthy people.At the same time,this thesis uses the Neuro RA Decoding method to explore the differences in the coding patterns of the brain of two groups of people after receiving the same stimulation from the temporal dimension,so as to determine the more separable regions of the ERP signals in the temporal dimension.This thesis conducts analysis experiments based on two public ERP datasets related to depression,namely: 1)Depression RL dataset,which induces the Rew-P component of EEG through a probabilistic learning task;2)Malaysia dataset,which induces the P300 component of EEG through the visual Oddball stimulus.The results show that:on the two datasets,there are significant differences in ERP signals in the prefrontal region between depressive disorder patients and healthy people,and polarity inverted ERP components appear around 300 ms after stimulation.In terms of temporal dimension,in addition to the 300 ms region considered in previous studies,the ERP signals of the Depression RL dataset also contain information that is helpful for classification at around 450 ms and 750 ms after stimulation.The Malaysia dataset also has better separability in the ERP signals around 450 ms,650ms,and 850 ms after stimulation.Therefore,this thesis demonstrates that ERP contains rich information in both spatial and temporal dimensions,which lays the foundation for further construction of a depressive disorder recognition model based on the spatio-temporal characteristics of signals.(2)Aiming at the analysis conclusion that there are significant differences in the spatio-temporal domain of ERP signals between patients with depressive disorders and healthy people,a depressive disorder recognition model based on the Spatio-temporal Linear Discriminant Analysis method is constructed.Based on the original matrix form of EEG data,the projection matrix is constructed separately in the spatial dimension and temporal dimension to learn the spatio-temporal pattern of the ERP signals,and the optimization of the spatial projection matrix and the optimization of the temporal projection matrix are combined in the same optimization goal.The optimal spatial projection matrix and time projection matrix are obtained through joint iterative optimization of the two projection matrices,so that the data after dimensionality reduction has better separability.Finally,the category center of the training data after dimensionality reduction is obtained,and the minimum distance method(Minimum Distance to Means,MDM)is used to discriminate new data to effectively identify patients with depressive disorders.(3)The results of the experiments on the classification and recognition of depressive disorders show that compared with LDA,SKLDA,and SWLDA,the STLDA method has better performance.On the Depression RL dataset,the balanced accuracy and F1 score of STLDA are 58.15% and 50.05%,which are increased by 2.0%-3.8%and 5.2%-10.8%,respectively.On the Malaysia dataset,the accuracy and F1 score of STLDA are 81.30% and 84.55%,which are improved by 1.6%-3.4% and 1.8%-4.0%,respectively.Compared with the experimental results that use data obtained by averaging several trials,STLDA has less performance degradation when using single-trial data.The results of the ablation experiment also show that the STLDA method which combines temporal domain and spatial domain is better than the matrix-based LDA applied only from spatial or temporal dimension,which proves that STLDA can comprehensively consider the relationship between the two dimensions.Through the display and analysis of the projection matrix,the correlation between the weight assigned to each channel or time point and the class separability information contained is further proved.The results show that there are significant differences in the ERP signals in the prefrontal region between patients with depression and healthy people,and they also have better separability in time regions other than 300 ms.In summary,the STLDA method can comprehensively consider the information of the ERP signals in the temporal dimension and spatial dimension.The spatial and temporal projection matrices could be jointly optimized and then the optimal results could be obtained,which makes the data after dimension reduction have better separability.Finally,a more interpretable and stable recognition model for the depressive disorder could be obtained.