Two‐stage spatiotemporal graph neural network framework for seizure onset zone localization

Objective Accurate localization of the seizure onset zone (SOZ) is critical for the presurgical evaluation of drug‐resistant epilepsy. Intracranial electroencephalography captures high‐resolution seizure dynamics, but interpreting complex spatiotemporal patterns across implanted electrodes is clinically challenging. We propose a two‐stage, uncertainty‐aware framework for SOZ localization that integrates patient‐specific seizure modeling with cross‐patient probabilistic ranking to estimate the likelihood of each electrode belonging to the SOZ. Methods The framework was evaluated on a bi‐institutional cohort of 36 patients (20 with good surgical outcome, 16 with poor surgical outcome), comprising 184 seizures. In the first stage, a patient‐specific spatiotemporal graph neural network modeled seizure onset as a network‐level transition using directed functional connectivity graphs derived from burst‐integrated electrophysiological features. Trained solely to detect preictal‐to‐ictal transitions, this model generated electrode probability trajectories without SOZ supervision. In the second stage, these trajectories were summarized into interpretable descriptors and used to train a patient‐agnostic classifier assigning SOZ likelihood scores to electrodes. Evaluation and hyperparameter optimization were performed using nested leave‐one‐patient‐out cross‐validation restricted to good outcome patients, ensuring strict patient‐level separation between training, validation, and test data. Results Seizure onset detection was stable across patients, with comparable binary cross‐entropy values between outcome groups (median = .495 vs. .438, p = .108). In good outcome patients, SOZ electrodes were prioritized among top‐ranked contacts (area under the curve = .86 ± .06), with all patients exhibiting at least one SOZ electrode within the top positions. In poor outcome patients, predictions exhibited reduced spatial separability, lower likelihood magnitudes, and near‐zero confidence gaps, reflecting attenuated model confidence. Significance This two‐stage framework integrates spatial interactions, temporal dynamics, and probabilistic ranking within a pipeline. By providing graded SOZ likelihood estimates and uncertainty‐aware prioritization, the approach is a decision‐support tool to assist hypothesis refinement during presurgical evaluation rather than as a standalone diagnostic system.