Modeling Spatiotemporal Structure in fMRI Brain Decoding Using Generalized Sparse Classifiers

The curse of dimensionality constitutes a major challenge to functional magnetic resonance imaging (fMRI) classification. Coupled with the typically strong noise in fMRI data, prediction accuracy is often limited. In this paper, we propose exploiting the inherent spatiotemporal structure of brain activity to regularize the typically ill-conditioned fMRI classification problem. To impose a spatiotemporal prior, we employ a recent classifier learning formulation for building Generalized Sparse Classifiers (GSC). This formulation combines a generalized ridge term with the LASSO penalty, which is integrated into classifier learning to permit various general properties, such as spatial smoothness, to be modeled. Here, we exploit this flexibility of GSC to build a spatio temporally-regularized sparse linear discriminant classifier, and contrast its performance on real fMRI data against a number of state-of-the-art classification techniques. Our results show that incorporating a spatiotemporal prior jointly improves prediction accuracy and result interpretation, which demonstrate the added value of such prior information in fMRI spatiotemporal classification.