Recently, the problem of estimating functional connectivity in the context of fMRI has been addressed using two different approaches. One approach treats the activity of regions of interest (ROI) as unobserved, but uses restrictive bilinear model to describe interactions between ROIs [4]. The other approach applies a general dynamic Bayesian netwok (DBN) model with multinomial conditional probability densities, but treats the state of ROIs as completely observed and uses quantized fMRI data [2]. Our goal is to employ DBN framework to allow modeling of general multiway nonlinear interactions and integration of varios sources of data, while at the same time avoiding loosing information at data quantization phase. A very general and a powerful framework that allows inference in such models is particle filtering (PF) [3]. Unfortunately this approach is known to be computationally expensive when the dimensionality of the problem grows. In this work we show how to improve efficiency of PF by introducing additional data sources (MEG and fMRI joint analysis) and how to exploit the structure of the observations in the DBN to further reduce the curse of dimensionality. We have conducted a series of sim-
[1]
Karl J. Friston,et al.
Dynamic causal modelling
,
2003,
NeuroImage.
[2]
Sergey M. Plis,et al.
Parallel Subspace Sampling for Particle Filtering in Dynamic Bayesian Networks
,
2009,
ECML/PKDD.
[3]
Timothy J. Robinson,et al.
Sequential Monte Carlo Methods in Practice
,
2003
.
[4]
Hamilton E. Link,et al.
Discrete dynamic Bayesian network analysis of fMRI data
,
2009,
Human brain mapping.
[5]
Volker Tresp,et al.
Discovering Structure in Continuous Variables Using Bayesian Networks
,
1995,
NIPS.