Causal Brain Connectivity: Integrating Granger Directed Graphs in fMRI Analysis

Functional Magnetic Resonance Imaging (fMRI) provides rich, time-resolved signals of neural activity, yet standard correlation-based analyses obscure the direction and temporal order of interregional influences. We present a unified framework that embeds Granger-causal inference within Graph Convolutional Networks (GCNs) to capture effective connectivity. We compare three models: (1) an MLP on flattened BOLD time series, (2) a GCN over undirected Pearson-correlation graphs, and (3) a directed GCN incorporating Granger-causal edges. Lag orders are selected via Akaike and Bayesian criteria. On large-scale fMRI cohorts, our directed GCN matches undirected performance in both classification and regression tasks while uncovering biologically plausible, asymmetric information flows that remain stable across hyperparameter settings. These results demonstrate that Granger-causality-informed graph models can enrich interpretability without sacrificing predictive power, marking a step toward causality-aware, graph-based analysis of brain networks.