Explainable Multimodal Graph Isomorphism Network for Interpreting Sex Differences in Adolescent Neurodevelopment

Abstract

Background: A fundamental grasp of the variability observed in healthy individuals holds paramount importance in the investigation of neuropsychiatric conditions characterized by sex-related phenotypic distinctions. Functional magnetic resonance imaging (fMRI) serves as a meaningful tool for discerning these differences. Among deep learning models, graph neural networks (GNNs) are particularly well-suited for analyzing brain networks derived from fMRI blood oxygen level-dependent (BOLD) signals, enabling the effective exploration of sex differences during adolescence. Method: In the present study, we introduce a multi-modal graph isomorphism network (MGIN) designed to elucidate sex-based disparities using fMRI task-related data. Our approach amalgamates brain networks obtained from multiple scans of the same individual, thereby enhancing predictive capabilities and feature identification. The MGIN model adeptly pinpoints crucial subnetworks both within and between multi-task fMRI datasets. Moreover, it offers interpretability through the utilization of GNNExplainer, which identifies pivotal sub-network graph structures contributing significantly to sex group classification. Results: Our findings indicate that the MGIN model outperforms competing models in terms of classification accuracy, underscoring the benefits of combining two fMRI paradigms. Additionally, our model discerns the most significant sex-related functional networks, encompassing the default mode network (DMN), visual (VIS) network, cognitive (CNG) network, frontal (FRNT) network, salience (SAL) network, subcortical (SUB) network, and sensorimotor (SM) network associated with hand and mouth movements. Remarkably, the MGIN model achieves superior sex classification accuracy when juxtaposed with other state-of-the-art algorithms, yielding a noteworthy 81.67% improvement in classification accuracy. Conclusion: Our model’s superiority emanates from its capacity to consolidate data from multiple scans of subjects within a proven interpretable framework. Beyond its classification prowess, our model guides our comprehension of neurodevelopment during adolescence by identifying critical subnetworks of functional connectivity.

Keywords

• deep learning
• graph neural network
• interpretability
• multi-modality
• multi-paradigm
• sex differences