Advancing diagnosis of bipolar disorder using brain morphometric similarity networks in a graph AI framework

Brain similarity networks (BSNs), extracted from structural magnetic resonance imaging, provide a validated framework for studying brain network organization and encode neurodevelopmental information relevant for psychiatric disorders. Recently, a neurodevelopmental hypothesis has been proposed for bipolar disorder (BD), where evidence demonstrates neuroprogression phenotypes differing from controls. BSNs offer a promising framework for investigating BD's neural correlates but remain largely underexplored. Parallelly, graph neural networks (GNNs) have emerged as suitable deep learning models for exploiting network-level information. This study aimed to investigate BSNs for discriminating subjects with BD from controls within a GNN framework using the multi-site StratiBip network, composed of 605 controls and 501 subjects with BD. Leveraging advanced analysis tools, we developed a multi-site classification framework including: i) the state-of-the-art MIND algorithm for computing morphometric similarity (MS) networks based on gray matter volumes (GMV), ii) MS integration with age, sex, and GMV, iii) a leave-one-site-out cross-validation for multi-site model generalizability evaluation. The best model achieved a mean multi-site accuracy of 68%. Explainability analyses revealed meaningful MS patterns in the basal ganglia, frontal and temporal lobes, and a particularly relevant integration with age. This study provides interpretable insights into the role of MS in BD and unveils evidence supporting ageing-related processes as a significant component of BD pathophysiology.