Graph diffusion convolution on multivariate weighted joint recurrence networks for seizure-state classification in intracranial EEG

Discriminating interictal and ictal brain states is important for both seizure-state classification and the characterization of seizure-related network reorganization in drug-resistant epilepsy. Intracranial electroencephalography (iEEG) provides high-resolution recordings of brain activity, but its sparse, irregular, and patient-specific electrode geometry requires models that go beyond Euclidean representations. In this work, we propose a multivariate weighted joint recurrence network coupled with a graph diffusion convolution network for binary interictal/ictal classification. Each iEEG segment is first embedded into a reconstructed phase space and transformed into a recurrence-based functional network, in which weighted edges quantify nonlinear dynamical coupling between implanted contacts. A graph diffusion convolution mechanism is then applied to capture higher-order dependencies across the implanted brain topology. On a nine-patient subset of the Hospital of the University of Pennsylvania iEEG Epilepsy Dataset, the proposed framework achieved an average accuracy of 93.4% and outperformed standard graph convolutional and Euclidean deep-learning baselines. In addition, recurrence networks derived from ictal and interictal periods showed state-dependent differences in weighted topological organization, suggesting seizure-related reconfiguration of nonlinear brain-network dynamics. These results indicate that recurrence-based graph diffusion provides an effective and interpretable nonlinear framework for seizure-state classification and for modeling ictal network reorganization in iEEG.