Research On Individual Emotional State Recognition Based On EEG Brain Networks

In recent years,with the development of theories and technologies related to cognitive neuroscience,researchers have proposed the concept of Affective BrainComputer Interface(a BCI)based on traditional Brain-Computer Interface(BCI)technology.The a BCI aims to enhance the ability of interactive devices to perceive,generate,and express human emotions through the effective recognition of individual emotional states,thereby improving intelligent emotional human-computer interaction.Additionally,the a BCI intends to provide a potentially effective intelligent assistant for the personalized closed-loop diagnosis and treatment of clinical mental disorders.However,due to the complex cognitive nature of emotion processing in the brain,which involves information interaction and collaborative processing across multiple functional brain regions,there remains a dearth of emotional EEG analysis methods capable of effectively elucidating the cognitive mechanisms of emotion and efficiently decoding emotional states.Thus,to further explore the interaction mechanisms of various functional brain regions during individual emotional cognitive processes and achieve robust emotional states recognition,this work develops a series of EEG emotion recognition models based on individual emotional EEG brain networks.This is accomplished through the development of emotional EEG network feature mining and graph representation learning algorithms,which provide support for advancing research and the application of affective brain-computer interface technology.The specific contributions of this work are as follows:1.To reveal the inherent spatial topological differences among individual emotional EEG brain networks,we propose an individual emotional EEG brain network decoding model based on a supervised learning strategy named MESNP(Multiple Emotion-related Spatial Network topology Patterns,MESNPs),which can effectively mine the discriminant spatial topological patterns of different emotional EEG brain networks.Specifically,the MESNP model learns spatial topological patterns of key activation brain areas and EEG brain networks relevant to individual emotional cognitive states,enabling the mining of discriminative graph spatial topological patterns among individual emotional state EEG networks to further enhance the distinguishability of emotional EEG networks and achieve robust decoding of individual emotional states.To validate the feasibility and stability of the proposed MESNP model in individual emotional EEG brain network recognition,both offline and online individual emotional EEG decoding experiments are conducted.Experimental results indicate that compared to traditional EEG feature extraction algorithms,the proposed MESNP model achieves higher robustness in individual emotional EEG identification,obtaining an online real-time decoding accuracy of 84.56% for individual emotional states(positive/negative,2classes),thus providing an effective solution for real-time decoding and dynamic monitoring of individual emotional states.2.To mitigate the impact of outlier samples(e.g.,EEG artifacts,false labeled EEG samples)on the robustness of emotional EEG decoding,we propose a discriminative EEG brain network manifold learning model named L1-SGL(Supervised Graph Learning model in the L1-norm space)based on the MESNP model.The L1-SGL model integrates a supervised graph learning strategy with K-nearest neighbor sparsity,constructing sparse homogeneous and heterogeneous emotional EEG network sample adjacency subgraphs to suppress the influence of artifacts and mislabeled samples on model robustness.Additionally,to achieve robust graph representation of emotional EEG brain networks,we define an optimal discriminative EEG brain network graph embedding subspace in the L1-norm space,enabling low-dimensional discriminative manifold graph representation and robust decoding of emotional EEG brain networks.Experimental results demonstrate that the L1-SGL model has achieved offline emotional decoding accuracies of 95.29%(SEED,3 classes),85.55%(MAHNOB,3 classes),84.20%(DEAP,4 classes),and 84.23%(MAHNOB,4 classes),as well as an online real-time emotional decoding accuracy of 86.30%(positive,neutral,negative,3 classes).Consistent experimental results indicate that the L1-SGL model can effectively suppress the impact of outlier samples on model robustness,enabling efficient graph representation and robust decoding of emotional EEG brain networks.The L1-SGL may provide a potential solution for the application of affective brain-computer interface technology in real-world scenarios.3.To effectively characterize the cognitive structural similarity of the emotional EEG brain network,we further propose an emotional EEG brain network embedding model based on the L1-SGL with cognitive-inspired learning strategy named L1-CGE(Cognition-inspired Graph Embedding model in the L1-norm space),which aims to achieve cognitive graph representation of individual emotional EEG brain networks and robust assessment of individual emotional states.Specifically,in the L1-CGE model,we introduce a cognitive-inspired metric of EEG network structural similarity based on EEG network topology to capture the node affinity of emotional EEG brain networks,constructing a potential high-dimensional manifold topology graph of individual emotional EEG networks.Moreover,by utilizing the objective function of the optimal graph embedding subspace defined in the L1 norm space,we learn the optimal cognitive graph representation of emotional EEG brain networks,extracting low-dimensional cognitive graph features of emotional EEG brain networks to achieve effective assessment of individual emotional states.Experimental results demonstrate that the L1-CGE model has achieved the offline decoding accuracies of 95.01%(SEED)and 85.17%(MAHNOB),as well as an online real-time decoding accuracy of 84.38% in the 3-class individual emotional EEG decoding tasks.The experimental results effectively validate the robustness and effectiveness of the L1-CGE model in decoding individual emotional states,which provides an effective support for the development of online affective braincomputer interface systems.4.To settle the cognitive interpretability issues of deep learning strategies in emotional EEG recognition applications,we utilize emotional EEG brain networks as the cognitive prior graph to characterize individual emotional cognitions and propose an emotional EEG graph learning system named BF-GCN(Graph Convolutional Network framework with Brain network initial inspiration and Fused attention mechanism),in which the cognitive-inspired and data-driven learning strategies are fused to learn the potential graph features of emotional EEG signals.Specifically,in the BF-GCN,the emotional EEG brain networks are first used as cognitive prior-inspired graph features,and the attention learning mechanism is used to fuse the data-driven and cognitiveinspired learning strategies to further settle the limitations of single data-driven learning strategies in the cognitive interpretability graph representation of emotional EEG signals.Experimental results based on public emotional EEG datasets demonstrate that the BFGCN model can effectively learn the potential graph pattern features of the emotional EEG signals,and achieve the average decoding accuracies of 97.44%(SEED,3 classes)and 89.55%(SEED-IV,4 classes)for subject-dependent emotional EEG decoding tasks,as well as the recognition accuracies of 92.72%(SEED,3 classes)and 82.03%(SEEDIV,4 classes)for the subject-independent emotional EEG decoding experiments.The experimental results have validated the effectiveness and superiority of the proposed BFGCN learning system in exploring emotional EEG deep graph features,which may provide a promising research direction for constructing robust,efficient,and generalizable emotional EEG decoding models.