Classification Study Of Action Intention Understanding Based On EEG Signals

Action intention understanding refers to a kind of psychological thinking activity in which people spontaneously understand the intention behind others’ actions when observing others’ actions.It is of great significance to human mind,language,and social development.In recent years,a variety of advanced brain imaging techniques have been used to collect brain signals of action intention understanding so as to decode the neural mechanism of action intention understanding and carry out the classification of brain signals.This paper carries out research that mainly focuses on the shortcomings of the previous classification of action intention understanding brain signals.The classification of action intention understanding brain signals is a kind of very challenging work.Previous studies in this field generally have the defect of low classification accuracy.At present,it is still difficult to find an ideal solution to this problem.Based on the EEG signals,this research tries to effectively solve the difficulty in the classification of action intention understanding brain signals from the perspective of different feature extraction methods.Our specific research contents mainly include the following aspects:1)The EEG signals of action intention understanding are classified by brain network metric features.Three sub-studies are carried out in this study.In sub-study 1,based on three kinds of action intentions,this paper first constructs functional connectivity matrices in multiple frequency bands by using synchronization likelihood(SL)algorithm,then calculates 11 weighted brain network metrics in these matrices,and then uses a statistical threshold to select the most useful metric features,finally,uses the selected metric features as classification features to carry out the binary classification task of action intention understanding.In experimental results,8 metrics from δ band and 5 metrics from θ band show statistical significances(p<0.05).Almost all the classification accuracies of single significant metrics are higher than the random level,and the classification accuracies of these significant metrics fusion are better,some even close to80%.In addition,p-values of permutation test on the real classification accuracies of support vector machine(SVM)classifier are all less than 0.05.The experimental results show that the new feature extraction method is very effective for the classification of action intention understanding EEG signals,and the combination method used in this study is extremely useful for the classification task.In sub-study 2,this paper first uses phase lag index(PLI)and weighted phase lag index(WPLI)to construct functional connectivity matrices in 63 time windows and 5 frequency bands after EEG source trace,then calculates 9 weighted brain network metrics on these functional connectivity matrices as classification features,finally uses the SVM classifier to perform the binary classification task of action intention understanding.In experimental results,the classification accuracy of PLI combined with WPLI is better than that of PLI and WPLI alone.Most of the classification accuracies are more than 70%,some even close to 80%.In the statistical test of dynamic brain networks,many important nodes appear in the prefrontal lobe,posterior occipital lobe,parietal lobe and temporal lobe.The experimental results show that the weighted brain network can effectively retain the data information,and the feature integration method proposed in this study is extremely effective for the study of action intention understanding.Mirror neuron and mental system,as cooperators,participate in the process of action intention understanding.In sub-study 3,PLI and WPLI are first used to construct functional connectivity matrices on 5 frequency bands and 63 time windows of source space,then nine kinds of brain network metrics are calculated as initial classification features in these matrices,and then variance analysis and cross validation are used to calculate the variances of all initial classification features under three kinds of action intention stimuli,and the features corresponding to the top variances are selected as the final classification features according to the set threshold,finally,the SVM,linear discriminant analysis(LDA)and multilayer perceptron(MLP)classifiers are used to implement the multi-class classification task of action intention understanding.In experimental results,α and full bands get better classification results(the average classification accuracy is higher than the random level and the highest average classification accuracy is even more than 70%).Additionally,the MLP classifier performs the best.Compared with other classification methods,our new method has some advantages.The experimental results suggest that the brain activity of action intention understanding is closely related to the α band,the new feature extraction process is an effective measure for the multi-class classification of action intention understanding,and neural network classifer is extremely useful for the multi-class classification of action intention understanding brain signals.2)The EEG signals of action intention understanding are classified by the features of phase synchronization indices.In this study,the functional connection matrices are first established by three phase synchronization indices(phase locking value,PLV;PLI;WPLI)in multiple microstate time windows,then the sum of the statistically significant difference edges is calculated in each micro state time window as the classification feature,finally,the SVM classifier is used to perform the binary classification task of action intention understanding.The experimental results show that the classification accuracy under all conditions is more than 65%,and the classification accuracy in α band is higher than that in β band;The classification accuracy of the α and β bands fusion data set is higher than that of single α or β band.The highest classification accuracy is more than 95%,and the lowest is 80%.In the experimental results of brain network statistical analysis,many differential connection edges appear in the α band,but few in the β band.In addition,many important vertexes appear in the temporal lobe,prefrontal lobe and posterior occipital lobe.The experimental results show that the new feature extraction method is very effective for the classification of action intention understanding,and the action intention understanding is closely related to the temporal lobe,prefrontal lobe and occipital lobe.3)The improved discriminative spatial patterns(DSP)algorithm is used to extract features for the classification of action intention understanding EEG signals.This paper first extracts several specific components(N70,P120,N170-P200,P300,P400-700)from the preprocessed EEG time series to form a matrix and reshape the matrix into a vector,then uses the modified DSP algorithm to calculate the projection vector in the stretched vector,then restores the calculated projection vector into a projection matrix,and then transforms the preprocessed EEG data into new data space and use the new data as classification features by the projection matrix,finally uses k-nearest neighbor(KNN)classifier to implement the binary classification of action intention understanding.In experimental results,the full band and fusion band get better classification accuracies.Compared with the previous methods,the new method has higher classification accuracy under some conditions.The experimental results show that the new feature extraction method not only effectively avoids the complex value problem of traditional DSP algorithm in solving generalized eigenvectors,but also completes the classification task of action intention understanding well.In addition,the experimental results also show that the fusion of features in different frequency domains can improve the classification accuracy in a certain extent.4)The power spectral density(PSD)features are extracted and selected by statistical method to classify action intention understanding EEG signals.This study first calculates the power spectral density of different frequency bands in the preprocessed EEG data,then calculates the p-values of all electrodes under different action intention understanding comparison conditions by rank sum test,then arranges these p-values in ascending order,selects the corresponding electrodes with the highest p-values,and uses the power spectral densities of these significantly different electrodes as the feature data,and then uses rank sum test and cross validation to filter the former power spectral density features,finally uses the KNN classifier to carry out binary classification on the selected features.In experimental results,significant difference electrodes appear in the prefrontal,temporal,parietal and occipital regions after strict false discovery rate(FDR)correction.In terms of classification accuracies,most of average classification accuracies under different conditions are higher than 70%,some even close to 80%.The experimental results show that the new feature extraction and selection method not only effectively analyzes the neural mechanism of action intention understanding,but also performs the classification task of action intention understanding well.In a word,this paper reasonably uses various feature extraction methods to carry out comprehensive classification research on action intention understanding.From the perspective of classification samples,it involves not only the classification of action intention understanding on group level,but also the classification of action intention understanding on trial level;From the perspective of classification types,it includes both binary classification and multi-class classification.In addition to the classification study,this paper also implements some researches about the neural mechanism of action intention understanding.