Brain Network Research Of Emotion Based On Dynamic Mode Decomposition

Emotion is the product of the combined action of physiology and psychology.Emo-tion plays a very important role in thinking and decision-making,social cognition and social interaction.The production of emotion is closely related to the brain.The syner-gistic effect of cerebral cortex and subcortical neurons promotes the production of emo-tion.EEG signals are widely used in emotion recognition due to their advantages of high temporal resolution,good rhythm,not easy to camouflage,and contain important physio-logical information of the human body.Nowadays,more and more scholars use complex networks to construct EEG brain networks for emotion recognition.The connectivity be-tween brain regions provides the basis for the construction of complex networks.Using graph theory to study the functional connectivity of the brain can well reflect the coupling relationship between brain regions.However,the analysis of brain networks requires new approaches.Therefore,Thesis conducts research on the construction of emotion-based brain network.The details are as follows:(1)In Thesis,dynamic mode decomposition is applied to EEG data to convert high-dimensional dynamic data into a simple representation based on spatiotemporal coherent structures,namely DMs.Each DM corresponds to an eigenvalue,and its phase and am-plitude determine the oscillation frequency and stability of the decomposed DM,respec-tively.The stability is divided according to the size of the eigenvalue?_i,and the part with|?_i|<0.95 is called relatively stable DMs,and the part with 0.95<|?_i|<1 is called very stable DMs.Under the four bands of alpha,beta,delta,and theta,the eigenvalues,eigen-vectors,and initial amplitudes in each mode are combined to generate relatively stable components and very stable components.(2)Using relatively stable components extracted from DMD,very stable components extracted from DMD and raw EEG data,Thesis constructs an EEG brain network based on Pearson correlation,phase-locking and partial correlation.The brain network constructed by very stable components is Brain Network 1(BN1),the brain network constructed by relatively stable components is Brain Network 2(BN2),and the brain network constructed by raw EEG data is Traditional Brain Network(TBN).After thresholding,under the four bands of alpha,beta,delta,and theta,the network topology features are extracted for the three brain networks of BN1,BN2 and TBN.The extracted network topology features are:feature path length,global efficiency,local efficiency,clustering coefficient and The rich club coefficient.Statistical analysis of the network topological characteristics under the four bands using a two-sample T test showed that most of the topological properties of BN1,BN2 and TBN were significantly different between high Valence/high Arousal and low Valence/low Arousal(p value<0.05).And in each band,compared with high Valence/high Arousal,the clustering coefficients of most nodes and the local efficiency of the network in low Valence/low Arousal have increased,and the characteristic path length has decreased.It is speculated that when an individual is under negative emotions,the information interaction of the brain is more complex than under positive emotions,the functional connections of the brain network are more closely,the network contains more stable structures,and the information transmission speed is faster,which can achieve Efficient information exchange.(3)Four classifiers are used for classification based on the network topology features of the full frequency band.Since the classification performance of BN2 is not significantly higher than that of TBN,only the classification effects of BN1 and TBN are compared.The classification effect is generally about 3%-8%higher than that of TBN.The biggest difference in the classification results is the classification of high Valence and low Va-lence based on the topological features of BN1 and TBN constructed by PLV.The average accuracy of BN1 and TBN is 66.01%and 62.73%,the average precision is 65.94%and61.81%,the average recall is 71.46%and 68.25%,and the average F1-Score is 65.14%and 62.76%.Most of the accuracy,precision,recall and F1-Score of BN1 are significantly higher than those of TBN,which indicates that the very stable components extracted from DMD are helpful for EEG emotion classification.Thesis proposes a new method to use DMD to extract very stable components and relatively stable components from EEG emotional signals to construct an EEG brain net-work,which provides new ideas for EEG emotion recognition,and is useful for better understanding emotional mechanisms,identifying emotional states.