Research On Effective Brain Connectivity Methods And Applications Based On Granger Causal Mode

Neuromanagement is an international emerging frontier field that uses neuroscience and other life science technologies to study management problems.It mainly examines human cognition and decision-making behavior from a new perspective by studying brain activities and thinking processes.With the continuous development of brain imaging technology,especially functional brain imaging,we can record the information of brain neural activity from different aspects and angles.The methodology of neuroscience provides an important tool for the realization of neural information management,which not only promotes the understanding of the neurobiological basis of cognitive behavior,but also facilitates the diagnosis and treatment of brain diseases.The brain is a complex dynamic information processing system.The analysis of statistical dependence and directional information flow between different brain regions in the brain network from the perspective of functional integration based on functional brain imaging data has become one of the research hotspots in recent years.Effective connectivity describes the intensity and direction of information transmission between brain regions,which can effectively explore the functional logic of neurocognition and the abnormal state of physiological mechanism of diseases.Among them,Granger causality analysis(GCA)is used as a powerful tool to evaluate directional effects among multivariate time series.It has been widely used in functional magnetic resonance imaging(fMRI)data.Based on the GCA model,this paper studies the existing problems in the construction of brain directed network and dynamic modeling,and explores the information processing of visual system and intelligent diagnosis management of central neurodegenerative diseases.Firstly,a dynamic brain network analysis method based on sliding time window(STW)was investigated to address the problem that the traditional static connectivity assumption cannot reveal the time-varying information interaction among neurons in the brain.From the perspectives of functional connectivity and effective connectivity,we quantified the changes of undirected and directed networks of the brain over time in resting-state and task-based state,and further explored the relationship between static and dynamic measures of connectivity,as well as the consistency of the trend in dynamic connectivity.The specific study of network dynamics was divided into three steps:(1)The STW method was used to construct dynamic networks of visual cortex subregions,and the time-varying Pearson correlation coefficients and Granger causal values were used to quantify the connection strength and information flow of each window over time,respectively,to explore the time-varying coupling and causal interactions information of human visual cortex subregions(V1 ～ V5)in resting state and task modulation effects;(2)Extract features of dynamic connectivity such as variance,dispersion index and amplitude of low-frequency fluctuation,and investigate the relationship with static connection measures;(3)Estimate the consistency of changing trends in dynamic functional connectivity and dynamic effective connectivity,and observe whether the fluctuations of different connections are similar from a global perspective.The results showed that there are indeed extensive connections between subregions within the visual system,and these connections are time-varying in both the resting and task-related states.The strength of static functional connections was negatively correlated with the variance of dynamic functional connections,and the strength of static effective connections was positively correlated with the variance of dynamic effective connections.Compared to the task-related state,dynamic functional connectivity within the visual cortex in resting state showed stronger consistency,while dynamic effective connectivity showed weaker consistency.Therefore,dynamic brain connectivity analysis is expected to more accurately characterize brain functional networks and help us understand the neural mechanisms of specific cognitive functions.Further,a novel data-driven Change Points Detection(CPD)model was proposed to address the shortcomings of commonly used dynamic network modeling methods such as STW.The brain network was partitioned according to the sparsity of connections between brain regions and the smoothness of changes,and the directed graph was estimated for time between adjacent change points,which was described as a time-varying effective connection network,reflecting the dynamic process of information transfer in neural activity.Specifically,the connectivity metrics at each time point were estimated from the fMRI data of each subject,and the number and location of significant changes in the multi-dimensional connectivity time series were automatically estimated by fused lasso regression,i.e.,the moments when part or all of the one-dimensional connectivity time series changed together in some specific way.Then GCA was used to estimate the time-varying brain directed connectivity network at the level of individual subjects.The performance of the proposed CPD model was evaluated by extensive simulation studies and compared with existing dynamic connectivity detection(DCD)algorithms.Then the dynamic Granger causal network was applied to the resting state fMRI data of patients with Alzheimer’s disease(AD)and healthy control group as the original feature set.The results showed that the CPD model can achieve higher classification accuracy compared with static and STW-based dynamic connectivity methods,which indicated that the CPD algorithm-based dynamic effective connectivity networks can be used as classification features to provide help for the intelligent diagnosis of brain diseases.Finally,to address the shortcomings of GCA in brain directed network detection,such as not being applicable to systems with too many variables and insufficient time series length,and not incorporating more useful information as a pure data-driven approach,a regularized Granger Causality model integrating prior information(RPGC)was proposed.For multivariate dataset,considering the sparsity of the brain network and the delayed correlation among variables,the regression coefficients were regularized with ridge regression to alleviate the overfitting problem.Meanwhile,the maximum delayed cross-correlation was added as a priori knowledge to put certain constraints on the strength and direction of causal relationships among variables,so as to obtain a more accurate and reliable effective connection network.Simulated datasets containing different number of variables and time series lengths validated the effectiveness of the method which was compared with different methods.The results showed that the RPGC method provides better causality recovery ability.In real fMRI data,higher classification accuracy and sensitivity were also obtained by exploring the directional information transfer of abnormal brain regions between AD patients and healthy controls.Meanwhile,abnormal causal relationship related to temporal lobe was found to be an important feature of AD,further demonstrating the importance of time delay information for predictive Granger causality analysis model.It also provides ideas for disease information management and diagnosis and treatment decision.