|M.Sc Student||Sumszyk Mikael|
|Subject||Multichannel Electroencephalogram Signal Analysis for|
Brain-Computer Interface and Functional Tomography
|Department||Department of Electrical Engineering||Supervisor||Professor Emeritus Gideon Inbar (Deceased)|
|Full Thesis text|
In this research, a new dynamical model for event-related brain activities is proposed. It is based on a decisive map of brain physical connections released recently (Sporns 2008) and coined the Connectivity Backbone (CB). In the following, the physical connections in the brain should be differentiated from its effective connections. Indeed, two regions are said to be effectively connected only if there is some causal relationship in their activity. Sommer (2000), pointed out that physical connections are useful predictors of activation patterns in the cortex. In this paper, we use the CB for the first time as a prior to constrain the estimation of the brain effective connections (Paus 2007). Following Penny (2002), we propose to model the brain effective connections using a Bayesian multivariate autoregressive (MVAR) model where the prior on the parameters is modeled by a centered multivariate Gaussian distribution. However unlike Penny (2002), we propose a covariance matrix that integrates the information encoded in the CB. Several possible designs for the covariance matrix are proposed and compared. In order to ensure fast computations, we use the Variational Bayes method to find a satisfactory parametric approximation to the posterior distributions. Since the analysis of EEG-based event-related brain activity necessarily requires to deal with short data segments, and ideally also calls for the simultaneous use of tens of channels, traditional approaches will not perform well. On the contrary, by constraining the solution space, the use of the CB as prior for high dimensional systems may succeed in providing reliable estimation. Extensive and realistic simulations on a synthetic CB - where several levels of noise are considered - show the potential of our approach which may have broad applications in neuroscience.