Directed graph neural networks with partial directed coherence for seizure prediction and epileptogenic network characterization.

While electroencephalography (EEG) analyses using undirected connectivity are well-established for seizure prediction and epileptogenic zone (EZ) network characterization, the complementary value of directed connectivity in graph neural networks (GNNs) remains unclear. We integrated partial directed coherence (PDC) graphs into a Digraph Inception Convolutional Network (DiGCN) to preserve directional edge information for seizure prediction, achieving accuracies of 96.11 ± 0.76%, 96.32 ± 0.70%, and 97.69 ± 0.67% on CHB-MIT, Siena, and the institutional cohort, respectively. In the institutional cohort, preictal modularity and small-world index were both significantly elevated (p < 0.001), reflecting increased modular segregation and topological reorganization. PDC in-degree was significantly elevated at EZ-concordant electrodes across five epilepsy subtypes (p < 0.05), whereas out-degree changes were not statistically significant. GNNExplainer identified compact, prediction-relevant subgraphs preserving these directed connectivity patterns. PDC-DiGCN provides an interpretable sensor-level framework for seizure prediction and EZ-concordant network characterization, though requiring larger prospective validation.