Identifying Networks within an fMRI Multivariate Searchlight Analysis

There is great interest in understanding how different brain regions represent information across space and time. Information-based searchlight analyses systematically examine the information encoded within clusters of functional magnetic resonance imaging (fMRI) voxels across the brain. Significant searchlights contain information that can be used to decode conditions of interest, but significant discriminability can be achieved in a variety of ways. We have developed and report on a new analysis method that can identify sub-networks of searchlights. Notably, unlike methods that collapse trials by condition, such as Representational Similarity Analysis, our method groups searchlights based on them having similar temporal changes in information. We present this method and apply it to fMRI data collected as participants viewed words, faces, shapes, and numbers. After running a searchlight analysis with a 4-way Gaussian Naive Bayes (GNB) classifier, the accuracy vector was submitted to a multi-subject Independent Component Analysis (ICA) to group searchlights based on their decoding timeseries. The ICA identified seven components (sub-networks) of searchlights. These networks identified sets of brain areas that have been commonly associated with the processing of faces, words, shapes and numbers. For instance, two of the components drew strongly on the face-processing network, including fusiform cortex. Switching the classification scheme to faces versus non-faces reconfigured the observed network to reflected face-related systems. These results demonstrate that this method can divide searchlight maps into meaningful components.