Late and Early Fusion Graph Neural Network Architectures for Integrative Modeling of Multimodal Brain Connectivity Graphs

The integration of structural and functional brain connectivity provides a holistic view of the brain’s organization, but its application in Graph Neural Network (GNN) models for predicting “brain age” is understudied, and a systematic benchmark of optimal data fusion strategies is currently lacking. We systematically benchmark the performances of early and late fusion multimodal architectures against single-modality models for brain age prediction using structural and functional connectomes, using five different GNN backbones on 747 healthy participants (median age, 16.3 years; IQR 13.5-18.5 years) obtained from the Philadelphia Neurodevelopmental Cohort. The late fusion architecture improved performance over the structural-only baseline in three of five models, with the GCN model achieving the highest overall score in cross-validation (R2=0.639±0.05). The early fusion architecture showed inconsistent results and did not offer a reliable improvement over the single-modality baseline. Finally, it is observed that optimal model architecture depends on the data type: structural brain graphs favor deep, narrow models to capture their hierarchy, whereas functional brain graphs require wider, shallower models.