Brain networks have been extensively studied in neuroscience, to better understand human behavior, and to identify and characterize distributed brain abnormalities in neurological and psychiatric conditions. Several deep graph learning models have been proposed for brain network analysis, yet most current models lack interpretability, which makes it hard to gain any heuristic biological insights into the results. In this paper, we propose a new explainable graph learning model, named hierarchical brain embedding (HBE), to extract brain network representations based on the network community structure, yielding interpretable hierarchical patterns. We apply our new method to predict aggressivity, rule-breaking, and other standardized behavioral scores from functional brain networks derived using ICA from 1,000 young healthy subjects scanned by the Human Connectome Project. Our results show that the proposed HBE outperforms several state-of-the-art graph learning methods in predicting behavioral measures, and demonstrates similar hierarchical brain network patterns associated with clinical symptoms.
Keywords: HCP; brain functional connectome; explainable AI; graph learning; regression.