A spatio-temporal support vector machine searchlight for fMRI analysis

We apply support vector machines (SVMs) in the context of fMRI analysis, in order to identify brain regions that are predictive of the experimental conditions. For the spatial SVM, we use the data within local 3D windows, called a searchlight, to train an SVM classifier to distinguish different experimental protocol conditions. Brain regions with high classification accuracy are identified as being implicated in the experimental task. Similarly for the temporal SVM, we use temporal sequences for every voxel to train a classifier. A major technical challenge is the higher computational overhead associated with SVMs. We overcome this by using parallel programming techniques based on MPI (message passing interface) that achieve load balancing. We report results on two separate datasets used previously in the literature. The SVM searchlight produces results comparable to the GLM for the spatial domain. In the temporal domain, the SVM searchlight was applied to a publicly avail able dementia dataset, and identified prominent novel regions such as the frontal cortex and pre-motor cortex which did not appear in the earlier study.