Quantifying the Information Content of Brain Voxels Using Target Information, Gaussian Processes and Recursive Feature Elimination

Multivariate pattern classification is emerging as a powerful tool for analysis of fMRI group studies and has the advantage that it utilizes spatial correlation between brain voxels. However, this makes quantifying the information content of brain voxels and localizing informative brain regions difficult. In this paper we a probabilistic Gaussian process classifiers to compute a sensitive measure of the information content of brain voxels (`target information´/TI) which we combine with a recursive feature elimination strategy. We apply this approach to a pharmacological fMRI study investigating rewarded working memory and compare it to sparse logistic regression. We show our approach is better suited to fMRI group studies, yielding more accurate classifiers and a sparse representation of informative brain regions. We also show that TI furnishes better estimates of voxel information content than existing approaches.