Balancing the Stability and Predictive Performance for Multivariate Voxel Selection in fMRI Study

Recently, the Multivariate Pattern Analysis (MVPA) studies for fMRI not only focus on cognitive state prediction, but also explore the interpretations of brain activity using model predictors (selected voxels). A model is considered to be good for interpreting brain activity if the selected voxels are all relevant to the specific cognitive state. Classical MVPA methods select voxels based on their prediction power; the selected ones are those that provide the best prediction performance. This precision based voxel selection method can guarantee the prediction performance, but it cannot ensure that all the selected ones are relevant. The interpretation of brain activity is therefore not ideal. This paper addresses this issue by introducing the concept of stability to the MVPA studies. If only the stability is emphasized in the selection process, the probability of selecting irrelevant voxels is highly reduced with the sacrifice of the prediction precision. We, therefore, propose a method to combine the stability assessment with the prediction precision assessment. In this paper, the proposed voxel selection method is integrated into a linear sparse predictor, Random Subspace Sparse Bayesian Learning (RS-SBL). The experiment results of simulation datasets demonstrate that our method can simultaneously reduce false positive and false negative rates while maintaining the prediction performance.