EEG-based Large-Scale Networks Modeling

Brain function does not depend on isolated brain regions but rather arises from the efficient functioning of large-scale networks that characterize hierarchical complex interactions in the brain.Understanding the brain largely involves comprehending largescale brain networks.Electroencephalography(EEG)-based modeling of large-scale networks holds promise in providing a more detailed understanding of the dynamic activity of the brain more safely and conveniently.However,due to the lack of reliable network constructing methods adapted to the complex physiological characteristics of EEG,there is still a lack of systematic understanding regarding large-scale EEG networks.Therefore,this dissertation aims to develop robust methods for constructing large-scale networks,specifically,1)Traditional methods for constructing time-varying directed networks are often affected by poor model fitting as well as discrepancies between prior assumptions and actual observations,which can lead to unreliable network estimations that deviating from the true brain interactions.This work first focuses on scalp EEG and develops a novel method for constructing time-varying directed networks-Nonparametric Dynamic Granger Causality based on Multi-Space Spectrum Fusion(nd GCMSF).It integrates cross-space complementary information to generate robust spectral representations and estimates the whole brain’s large-scale dynamic causal interactions based on nonparametric spectral matrix decomposition.The proposed method avoids limitations imposed by model-driven approaches while enhancing the robustness of non-parametric estimation approaches,provides a way to explore essential dynamic causal interaction without constraints,and is more suitable for the complex physiological characteristics of EEG.Systematic experiments have verified the superiority of nd GCMSF over traditional methods in terms of noise resistance and characterization of real brain dynamic interactions.In addition,research based on nd GCMSF reveals the evolution mechanism of lateralization in the large-scale networks during motor execution of hemiplegic patients.The network-derived features also are reliable indicators for diagnosing types of hemiplegia and assessing patients’ motor function.2)The low-level scalp EEG network construction is constrained by volume effects and spatial information ambiguity,which does not facilitate a deep understanding of large-scale complex activities.Functional network connectivity(FNC)characterizes the information exchange between functional subnetworks within the brain,which is more context-sensitive,macroscopic,and higher hierarchical.Therefore,this work proposes to combine EEG source imaging techniques and uses Bayesian nonnegative matrix factorization(BNMF)to extract spatiotemporal information of cortical functional subnetworks in the source space to construct EEG large-scale cortical FNC in a datadriven way.BNMF relies on the brain’s non-negative activation characteristics,and combined with prior information introducing,it achieves robust extracting of spatiotemporal information of non-negatively activated subnetworks.Experiments on two independent real EEG datasets,i.e.,P300 and Ultimate Gambling(UG),demonstrate the superiority of BNMF in identifying the spatiotemporal information of subnetworks with non-negative activation,clearer spatial distribution,and physiological reality over traditional methods.Moreover,FNC analysis based on BNMF reveals hierarchical interactions during the UG decision-making process,involving decision-making workspaces dominated by the default mode network and specific large-scale network interaction patterns corresponding to different decision behaviors.3)Another branch of FNC analysis based on prior brain atlases excels in experimental reproducibility and physiological interpretability.However,as different functional subnetworks contain different functional regions with diverse spatial distributions,the dynamic activities of these subnetworks defined based on the atlas are represented by a “multivariate” time series.The traditional “univariate” measurement methods cannot effectively measure the coupling relationship between subnetworks.This work further proposes to reconstruct cortical activity based on EEG source imaging,extract multidimensional temporal information from it using prior brain atlases,estimate the interaction between subnetworks by employing multivariate correlation analysis method S estimator,and subsequently construct large-scale cortical EEG FNC in a template-driven approach.Systematic experiments have shown that S estimator exhibits strong robustness in low signal-to-noise ratio and short data length scenarios,making it more suitable for measuring the interaction between subnetworks and constructing largescale cortical FNC in a general context of EEG.Additionally,we further reveal the significant role of cerebellar networks in evoking P300-a kind of event-related potential that is closely related to cognitive processes such as attention and memory.4)Brain activity is essentially dynamic.Therefore,this work further combines wavelet coherence that can measure the brain’s dynamic interactions with the aforementioned S estimator to develop a novel time-varying multivariate correlation analysis method called wavelet coherence S estimator(WTCS).Then,using WTCS to measure the dynamic coupling between subnetworks treating each subnetwork as a functional entity,we achieve constructing large-scale EEG cortical dynamic functional network connectivity(d FNC).This extended the EEG cortical FNC research based on prior brain atlases to dynamic scenarios.Simulation studies demonstrate the robustness and reliability of WTCS in estimating d FNC.Furthermore,the cortical d FNC analysis reveals both “primary” and “P3-like” peaks in brain large-scale organization,as well as evolving network topology during P300 evoking processes.In summary,aiming at challenges such as unconstrained robust scalp time-varying directed network estimation,reliable extraction of cortical functional subnetworks,and robust measurement of relationships among functional subnetworks that contain multivariate signals,this dissertation develops methods for constructing large-scale networks from various perspectives,ranging from scalp EEG to cortical activities,from low-level to high-level node definitions,from data-driven to atlas-driven,and from static to dynamic interactions.This dissertation presents a series of reliable methods and new perspectives for the characterization of large-scale EEG networks and reveals the underlying mechanisms of large-scale networks in various higher cognitive processes,including movement,decision-making,and attention.