CA-GCN: A Causality-Aware Graph Coupling Network for Multimodal Learning of Brain Connectivity

Understanding the coupling relationship between structural connectivity (SC) and functional connectivity (FC) is essential for advancing our knowledge of individual brain function and organization. Recent deep learning techniques, particularly graph neural networks (GNNs), have shown strong potential in modeling SC-FC coupling relationships. However, existing multimodal graph-learning methods often overlook intrinsic causal relationships and task-specific SC-FC coupling patterns, thereby limiting both predictive performance and interpretability. To address these issues, we propose a Causality-Aware Graph Coupling Network (CA-GCN) that disentangles causal representations and models causal SC-FC coupling patterns in a task-relevant manner. Specifically, we first develop a Multimodal Causal Disentanglement (MCD) module that isolates SC and FC causal features related to the task label via Granger causality-inspired learning. Then, we design a Multimodal Causal Coupling (MCC) module that identifies SC-FC coupling patterns through contrastive learning of cross-modal attention. Evaluated on HCPYA for fluid cognition prediction and CamCAN for brain age prediction, CA-GCN consistently outperforms single-modal and state-of-the-art multimodal graph methods. Moreover, the interpretable saliency maps of CA-GCN reveal the key causal brain regions and SC-FC coupling patterns.