From features to functional maps: An attention based framework for explainable memory-related brain network analysis

Memory encoding, during naturalistic narrative-experiences, engages distributed brain networks that extend beyond commonly reported brain regions of interest (ROIs). Although, several recent studies have reported the role of white-matter (WM) and cerebellum in memory recall, most studies overlook noncanonical regions outside the traditional default mode network comprising of hippocampal-medial temporal lobe system. In this paper, for establishing memory related networks, we suggest the involvement of entire brain with memory recall. For this purpose, brain ROIs are identified through a Graph Attention Network (GAT) model developed using whole-brain functional connectivity (FC) so as to capture inter-region dependencies and highlight connections most relevant to recall performance. The GAT model is trained using fMRI data of 180 subjects listening to four audio narratives. Our proposed model achieves 84% accuracy in classifying high versus low memory recall using four-fold cross-validation. In addition, most memory recall studies lack interpretability of brain networks. To bridge this gap we developed GATxp, an attention-based explainer, that converts the GAT model's learned attention weights into interpretable sub-network ROI maps. Our analysis reinforces the initial proposition that memory-related networks span default mode, frontoparietal control, visual, and critically-cerebellar and WM regions. Finally, we validate these findings three ways: using Network-Based Statistics tool, Coordinate-based Meta-analytic decoding (NeuroQuery), and literature review, with 84 of 90 identified ROIs showing significant memory recall related associations. This work demonstrates how GAT with intrinsic attention mechanisms can identify distributed, whole-brain sub-networks supporting ecologically valid memory encoding, providing both accurate classification and neuro-scientific interpretability.