Mixed Membership Stochastic Blockmodels for the Human Connectome

Alzheimer’s disease and other neurological diseases are often characterized by brain atrophy. It is hypothesized that such degradation directly affects connectivity as measured by whole brain tractographies and their derived connectivity networks. It is unclear, however, that current network construction methods provide either the most useful or efficient representation of the underlying connectivity structure. In the present work, we study the applications of a generative network model that can be used for automated cortical parcellation as well as network summary. We evaluate its performance through an independent classification task. In particular, we study whole brain tractographies from 96 subjects from the Alzheimer’s Disease Neuroimaging Inititive (ADNI). We fit a Mixed Membership Stochastic Blockmodel (MMSB) to both an anatomically generated connectome as well as a larger, finely resolved connectome. We reduce each network to a much smaller block connectivity representation, and then use a generic Support Vector Machine to classify the resulting matrices by disease category. Our results suggest that mixed membership blockmodels produce parsimonious representations of existing anatomic connectomes, as well as useful parcellations of higher resolution networks.