Predictive Modeling of fMRI Brain States Using Functional Canonical Correlation Analysis

We present a novel method for predictive modeling of human brain states from functional neuroimaging (fMRI) data. Extending the traditional canonical correlation analysis of discrete data to the domain of stochastic functional measurements, the method explores the functional canonical correlation between stimuli and fMRI training data. Via an incrementally steered pattern searching technique, subspaces of voxel time courses are explored to arrive at (spatially distributed) voxel clusters that optimize the relationship between stimuli and fMRI in terms of redundancy. Application of the method for prediction of naturalistic stimuli from unknown fMRI data shows that the method finds highly predictive brain areas, i.e. brain areas relevant in processing the stimuli.