Localization and anticipation of seizure onset and propagation using network architecture of pathological high-frequency oscillations.

OBJECTIVE: In focal epilepsies, seizures originate from a presumed seizure onset zone (SOZ) and propagate through a network of dynamically interconnected brain regions. Accurate identification of the SOZ and propagation pathways is critical for surgical planning in patients with drug-resistant epilepsy. We tested the hypothesis that pathological high-frequency oscillation (pHFO) network properties can characterize seizure propagation at the individual-patient level. METHODS: We measured pHFOs (80-500 Hz) in >9000 channel-hours of intracranial electroencephalographic data recorded during early epilepsy monitoring unit recordings and across 46 seizures in 20 patients implanted with subdural (n = 11) or depth electrodes (n = 9). An additional prospective validation patient with two seizures was included for independent testing. We used graph theory-based features to map pHFO networks, measured propagation latency, and applied machine learning frameworks to determine how network metrics inform seizure propagation. We then evaluated the predictive utility of these network features for identifying both the origin and propagation pathways of pHFOs and examined the spatiotemporal dynamics of the pHFO network prior to seizure onset. RESULTS: Spatial and temporal patterns of pHFO generation and propagation during interictal and peri-ictal periods closely align with seizure onset and spread. Machine learning models achieved robust performance, with within-subject clustering identifying the SOZ at 95% accuracy. Across patients, elastic-net models achieved robust generalization (SOZ vs. all channels: accuracy = 93.6%, area under the receiver operating characteristic curve = .90, positive predictive value = .54, Matthews correlation coefficient = .53). Patients with favorable surgical outcomes showed significantly greater concordance between pHFO network predictions and clinically identified epileptogenic regions. Notably, pHFO network structure underwent progressive reorganization beginning up to 30 min before clinical seizure onset. SIGNIFICANCE: Network-based pHFO analysis provides a quantitative, patient-specific approach to localizing seizure sources, mapping propagation pathways, and characterizing preictal network reorganization. Temporally resolved mapping of pHFO network architecture may improve surgical targeting and provide a framework for identifying seizure-related network transitions.