Learning Functional Brain Connectivity with Time-Series Bayesian Network from fMRI Data

• Activity Number: 466
• Type: Contributed
• Date/Time: Wednesday, August 4, 2010 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): Xiangxiang Meng*+ and Xiaodong Lin
• Companies: University of Cincinnati and Rutgers University
• Keywords: time-series Bayesian network ; functional magnetic resonance imaging ; Bayesian model averaging

Abstract

We propose a novel continuous time-series Bayesian network (TBN) approach to learn the connectivity network among active brain regions in a functional magnetic resonance imaging (MRI) experiment. Unlike ordinary Bayesian networks or Dynamic Bayesian Networks, TBN captures the temporal hemodynamic dependence of the entire fMRI series between brain regions, by defining a network score metric directly using the spectral density matrices in the frequency domain. A heuristic Bayesian model averaging method is further applied to estimate the network structure from a pool of candidates with close network scores when a limited length of fMRI scans is obtained from children. We demonstrate the proposed network model with multivariate time series simulated from vector regressive models, and fMRI data collected from 313 children in a story comprehension task.