基于脑电多域特征融合的跨任务认知负荷研究

Cross-task cognitive load based on EEG multidomain feature fusion

为探究不同认知负荷下的人脑活动,设计了三种实验范式(N-Back、心算和Sternberg),采集被试者在三类认知负荷下的脑电图(electroencephalogram,EEG)信号,并对EEG信号进行预处理、特征提取和特征分类。模型通过相位锁相值(phase locking value,PLV)计算了EEG的功能连接特征,将PLV功能连接矩阵作为脑功能网络的边;以微分熵特征作为网络的节点信息,实现EEG频域与空间域特征的融合,利用图注意力神经网络完成了跨任务的认知负荷分类。模型在跨任务三分类认知负荷识别中取得了57.12%的平均分类准确率。基于复杂网络理论,从全局和局部两个层次分析了不同负荷状态下大脑网络结构的变化,随着认知负荷程度增加,theta与alpha频段的全局聚类系数逐渐减小,delta与theta频段的全局效率则有所提高;theta频段下的额叶、顶叶与颞叶脑区电极的局部效率呈上升趋势。网络全局与局部的度量变化表明随着人脑认知负荷程度的提高,功能脑网络的拓扑结构在发生改变。

[Objective]At present,most experimental results in cognitive load research focus on a single psychological task,with limited exploration and expansion of cross-task cognitive load identification.To explore human brain activity under different cognitive loads,electroencephalogram(EEG)signals were collected from various tasks.Cross-task cognitive load recognition was achieved by fusing frequency domain and spatial domain features,and the functional brain network of EEG signals was visualized and analyzed based on graph theory.[Methods]Three experimental paradigms(N-Back,mental arithmetic,and Sternberg)were designed to collect EEG signals from participants under three types of cognitive loads.The original EEG signals were preprocessed through operations such as downsampling,filtering,rereferencing,and independent component analysis to remove noise from the original data.The feature extraction method was applied to extract features,including differential entropy(DE),power spectral density(PSD),and the phase locking value(PLV),in terms of frequency and spatial domains.The degree to which each feature represents cognitive load was evaluated.Moreover,a functional brain network architecture was constructed based on the graph theory and phase lag values.The frequency domain features and functional brain network were fused through graph attention networks(GATs),and the brain network topology was efficiently learned through attention modules.The network parameters of the functional brain network were calculated using the graph theory,including global clustering coefficients,global efficiency,and local efficiency.Differences in functional brain networks were analyzed under different loads based on statistical changes in network parameters.[Results]The results of the cross-task cognitive load show that:①DE has a stronger classification performance than PSD,and PLV has a better classification performance than traditional frequency domain features.The classification accuracy based on a single feature can reach up to 49.6%.②The classification results of integrating the frequency domain and spatial domain features through GATs are better than those of a single feature,and the cross-validation accuracy is 57.12%.③The global parameter analysis results of graph theory-based functional brain networks show that as the load level increases,the global clustering coefficients in the theta and alpha frequency bands gradually decrease under different tasks,while the global efficiency of the delta and theta frequency band networks improves.④The local parameters of the functional brain network were divided into different brain regions based on the distribution of electrode nodes.The statistical results showed that as the load increased,the local efficiency of electrodes in the frontal,parietal,and temporal brain regions increased under different tasks.[Conclusions]Compared with traditional frequency domain features,functional connectivity effectively measures the interactions between different regions of the brain,explaining human brain activity from a spatial perspective.GATs can also fuse frequency and spatial domain features,effectively using the topological structure between EEG channels to learn more discriminative EEG cognitive load representations.However,changes in functional brain network parameters indicate that the functional brain network structure reorganizes to varying degrees with increasing cognitive load.