Prince Isaac R. Pantino, Fitzsixto Angelo L. Singh, Christine F. Pena
Neuro-RAGX is a reproducible framework for explainable analysis of resting-state functional connectivity that combines graph neural modeling with retrieval-augmented grounding. Using precomputed region-to-region connectivity matrices from the Human Connectome Project Young Adult cohort, the framework constructs multi-view brain graphs and trains graph attention autoencoders to learn subject-level representations. The models achieve substantially lower reconstruction error than a population-mean baseline and exhibit consistent connectivity motifs across multiple atlas and connectivity configurations. We also compare attention-selected edges with rankings based on raw edge strength and find that attention highlights complementary connectivity structure rather than simply selecting the strongest connections. A hybrid retrieval pipeline that integrates dense retrieval, BM25, and late-interaction ranking then links learned motifs to supporting textual evidence, providing citation-grounded explanations. These results position NeuroRAGX as an auditable workflow for hypothesis generation in connectomics while still requiring expert interpretation to assess biological plausibility and avoid overclaiming causal structure. © 2026 IEEE.
University of San Carlos, Department of Computer, Information Sciences and Mathematics, Cebu, Philippines
