A deep spatiotemporal graph learning architecture for brain connectivity analysis.

Abstract

In recent years, the conceptualisation of the brain as a "connectome" as summary measures derived from graph theory analyses, has become increasingly popular. Still, such approaches are inherently limited by the need to condense and simplify temporal fMRI dynamics and architecture into a purely spatial representation. We formulate a novel architecture based on Geometric Deep Learning which is specifically tailored to the one-step integration of spatial relationship between nodes and single-node temporal dynamics. We compare different spatiotemporal modelling mechanisms and demonstrate the effectiveness of our architecture in a binary prediction task based on a large homogeneous fMRI dataset made publicly available by the Human Connectome Project (HCP). As the idea of e.g. a dynamical network connectivity is beginning to make its way into the more mainstream toolset which neuroscientists commonly employ with neuroimaging data, our model can contribute to laying the groundwork for explicitly incorporating spatiotemporal information into every association and prediction problem in neuroscience.