Modeling Dynamics of Large-Scale Effective Connectivity in fMRI Data

• Activity Number: 89
• Type: Invited
• Date/Time: Sunday, August 9, 2015 : 8:30 PM to 9:15 PM
• Sponsor: Korean International Statistical Society
• Author(s): Hernando Ombao* and Chee-Ming Ting and Sh-Hussain Salleh
• Companies: UC Irvine and Universiti Teknologi Malaysia and Universiti Teknologi Malaysia
• Keywords: Autoregressive ; EM Algorithm ; State-space models ; Subsapce VAR ; Effective Connectivity ; fMRI

Abstract

Most analyses of effective brain connectivity employ stationary vector autoregressive (VAR) models which assume a static inter-dependence between distinct brain regions. Emerging evidence suggests the presence of temporal variations in connectivity during rest or active conditions. A possible way of quantifying these changes is to use a time-varying VAR (TVAR) state-space model. However, estimation is unreliable for networks with large number of brain regions. We present our newly developed subspace VAR (SVAR) model based on a latent factor model, where the directed interactions in the large fMRI data are driven by a few common factors following a VAR dynamic. To capture the evolving causal interactions, we propose a time-dependent variant of the SVAR model, by combining the factor model with a TVAR factor process. We propose a robust two-step estimation procedure by first estimating the factor model using the principal component (PC) methods, followed by the expectation maximization (EM) algorithm. The proposed method is evaluated on a resting-state fMRI dataset of 25 healthy subjects.