A High-Dimensional State-Space Model for the Joint Analysis of EEG and MEG Data

• Activity Number: 623
• Type: Invited
• Date/Time: Thursday, August 13, 2015 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistical Graphics
• Author(s): Farouk Salim Nathoo*
• Companies: University of Victoria
• Keywords: Variational Bayes ; Sequential Monte Carlo ; Multimodal Neuroimaging ; Joint Model ; State-Space Model ; High-Dimensional Data

Abstract

I present a state-space model for solving the neuroelectromagnetic inverse problem based on a joint model combining EEG and MEG data. This inverse problem is ill-posed and leads to a bivariate system of underdetermined dynamic models parameterized over a high-dimensional state-space. We assume linear non-Gaussian dynamics for the neural activity that incorporate spatial dependence across neighboring brain locations as well as bilateral dependence corresponding to symmetric locations on opposite hemispheres of the brain. Modeling neural activity over a large number of spatial locations and time points leads to challenges for Bayesian computation. I will discuss solutions based on mean-field approximations, hybrid variational Bayes/MCMC algorithms, and sequential Monte Carlo, along with parallel computing implementations. Results from simulation studies and an application examining the neural response to face perception will be presented.