Comparing Clustered and Distributed Sampling for Random Subspace Brain Mapping

Random subspace sampling provides a simple and intuitive solution to the problem of the large dimensionality of functional magnetic resonance imaging (fMRI) data for brain state decoding. Recently proposed methods for estimating voxelwise contributions to the ensemble decoding task yield biologically plausible and interpretable maps, but their sensitivity may suffer since random sub sampling ignores the spatial smoothness typical of brain activity. We therefore compare random subspace maps with those generated with explicit modeling of spatial relationships through regionally clustered sampling. On realistic simulated data, clustered sampling dramatically boosted voxel detection sensitivities and specificities from an area under the receiver operating characteristic curve in the range of 0.6 to 0.9. These results suggest that future efforts in classifier ensemble approaches are devoted to accommodate the spatial organization of brain activation.