Event Related Potential Extraction And Effective Connectivity Analysis Methods

Event related potential(ERP)evoked by external stimulus or the onset of a certain cognitive task can be recorded by electroencephalography.Since it can reflect the mechanism of human neural system,it has been widely used by the investigators in cognitive neuroscience,psychology,etc.As the development of neural science,brain connectivity is becoming a key issue in explaining brain‘s complex function.So the investigators are paying more and more attention on connectivity based on ERP besides ERP amplitude difference,latency shift,etc.We proposed a series of ERP extraction and effective connectivity analysis methods in this thesis: single experimental condition ERP model and its extraction algorithm,multi-experimental conditions ERPs joint model and their extraction algorithm,effective connectivity analysis method based on time-varying multivariate autoregressive model.Single experimental condition ERP model uses the phase-locking property of ERP.Trial-by-trial concatenated ERP can be represented by a linear combination of a series of sine bases which are independent of the ERP waveform.As so,given the sine bases,the property of multi-channel concatenated ERP can be expressed by a spatio-spectral matrix.We proposed an algorithm to estimate the parameters in the model,in which,the spatiospectral matrix and the noise covariance matrix are updated iteratively to maximize the posterior.The spatio-spectral matrix is decomposed to maximize the signal-to-noise ratio of ERP.Then,ERP spatial pattern,spectral pattern and spatial filter can be estimated.This method can determine the number of ERP components automatically which helps to prevent over-fitting.Multi-experimantal conditions ERP joint model assume that an ERP component shares the identical spatial pattern across conditions while its amplitude and latency differs.An algorithm,named Bayesian estimation of ERP component,is proposed to estimate the posterior of the parameters in the model.The algorithm updates the posterior of ERP spatial pattern,temporal wave,amplitude factor,latency shift and noise covariance matrix iteratively using variational Bayesian.This method does not require certain pattern of noise covariance matrix and can determine the number of ERP components.Estimating time varying effective connectivity depends on the correct estimation of the corresponding time-varying multivariate autoregressive model.A time-varying multivariate autoregressive model is proposed for ERP effective connectivity analysis,which assumes the one-order differences of model coefficients are sparse.An algorithm based on variational Bayesian principle is developed to estimate the posterior of time-varying multivariate autoregressive model coefficients.The order of model can also be determined using the variational lower bound.Since the key parameters in the model can be automatically determined,the method can be easily used on real ERP data.The proposed models and algorithms can extract ERP and estimate the effective connectivity correctly.They are new tools for ERP analysis in cognitive neuroscience,psychology,etc.