Bayesian Computational Neuroscience
In this Section:
Schizophrenia
is an illness with enormous public health significance,
affecting over 30 million people and inflicting immense personal and
economic cost worldwide. Treatments for this terrible disease are
still limited, at least in part because of poor understanding its
anatomical, neural, cognitive, and genetic substrates. Modern
morphometric and functional neuroimaging technologies can provide us
noninvasive views of the behavior of the schizophrenic brain, but
identifying significant circuits (and, worse, misconnections or
breakdowns in circuits) in such rich data sources is a challenging
problem.
In this work, my research group and our colleagues at
The MIND Research Network
are developing advanced dynamic
Bayesian network
(DBN) analysis methods to infer and model functional activity networks
from neuroimaging data. Given functional neuroimaging data
(fMRI,
MEG, and/or
EEG)
gathered from both healthy and schizophrenic
patients, we attempt to find DBNs that characterize each of the
populations and/or the differences between the populations. The
conditional independence relations encoded in these models provide
semantic indications of brain activity networks. The content of the
conditional probability tables descibe the form of the functional
relationships among related regions.
This approach provides three distinct approaches over existing modeling mechanisms for functional neuroimaging data:
- Data driven
- By employing Bayesian Network structure search mechanisms, we can identify independence relations directly via their posterior probabilities. Thus, our methods do not require strong initial anatomical connectivity or hypothesis-driven model assumptions.
- Nonlinear
- Unlike common linear Gaussian dynamical models, we take our conditional probability distribution family to be discrete multinomials. Multinomials are a universal function familiy over the finite-dimensional simplex, so we have the capacity to represent any, arbitrarily nonlinear functional relationship among random variables.
- Multivariate
- DBNs capture the statistical dependence relationships among sets of variables. Thus, rather than characterising the relationship of only pairs of regions or of all regions simultaneously (e.g., under a linear combination model), we can identify small, discrete groups of interacting brain regions.
