Integration of sparse Bayesian learning and random subspace for fMRI Multivariate Pattern Analysis

Multivariate Pattern Analysis (MVPA) is frequently used to decode cognitive states from brain activities in fMRI study. Due to the discrepancy between sample and feature size, MVPA methods are suffered from the overfitting problem. This paper addresses this issue by introducing sparse modelling along with its advanced decoding method, Compressive Sensing (CS). As brain voxels have highly correlated in spatial domain, the prerequisite of CS methods are not well satisfied. We therefore propose a novel MVPA method to integrate linear Sparse Bayesian Learning (i.e. Bayesian Compressive Sensing) with random subspace method. Benefiting from the random subspace method, spatial correlation and feature-to-sample ratio are largely reduced. The experimental results from a real fMRI dataset demonstrate that our method has distinct prediction power compared to three other popular MVPA methods, and the detected relevant voxels are located in informative brain areas.