Multi-subject Independent Component Analysis
using a Maximum Likelihood Approach
Ying Guo
Department of Biostatistics
Emory University
Friday, November 2, 2007, 12:30–1:30 pm
GEMS classroom, 3rd Floor in
Shriner's Building
Coffee, tea, and cookies will be provided
Abstract
Independent component analysis (ICA) is becoming increasing popular for
analyzing functional magnetic resonance imaging (fMRI) data. ICA has been successfully applied for single-subject fMRI analysis. However, the extension of ICA for group inferences is not straightforward and remains an active research topic. Several methods, such as GIFT and tensor ICA, have been proposed for estimating group independent components based on multi-subject fMRI data. These group ICA models assume different underlying structures of group spatio-temporal processes when decomposing the high-dimensional imaging data. Currently, there are no methods for assessing the appropriateness of an assumed structure in a particular data set. Hence, the validity of the estimated group ICA components is often questionable. Another challenge in multi-subject ICA analysis is how to test group differences between subject subpopulations.
In this talk, I present a class of group ICA models which incorporate existing models as special cases. I present a maximum likelihood (ML) approach for estimating this class of ICA models. A modified EM algorithm is proposed to maximize the likelihood. The ML method offers a unified framework for estimating and comparing group ICA models with different underlying structures. Likelihood ratio tests (LRT) are derived to compare nested group ICA models. The LRT can be used to assess the goodness-of-fit of a model structure and to test the homogeneity of temporal responses between subgroups. The performance of the ML estimation and LRT is evaluated through simulation studies. I illustrate the proposed method using an fMRI study of neural activities between subjects who had long-term training in Zen meditation and those who didn't.