Self‐supervised data‐driven approach defines pathological high‐frequency oscillations in epilepsy

Abstract Objective Interictal high‐frequency oscillations (HFOs) are a promising neurophysiological biomarker of the epileptogenic zone (EZ). However, objective criteria for distinguishing pathological from physiological HFOs remain elusive, hindering clinical application. We investigated whether the distinct mechanisms underlying pathological and physiological HFOs are encapsulated in their signal morphology in intracranial electroencephalographic (iEEG) recordings and whether this distinction could be captured by a deep generative model. Methods In a retrospective cohort of 185 epilepsy patients who underwent iEEG monitoring, we analyzed 686 410 HFOs across 18 265 brain contacts. To learn morphological characteristics, each event was transformed into a time–frequency plot and input into a variational autoencoder. We characterized latent space clusters containing morphologically defined putative pathological HFOs (mpHFOs) using interpretability analysis, including latent space disentanglement and time‐domain perturbation. We built a predictive model to forecast postoperative seizure outcomes at 12 months based on the resection status of brain regions exhibiting mpHFOs. This model was compared to current clinical standards that evaluate outcomes based on the extent of seizure onset zone (SOZ) removal. Results mpHFOs showed strong associations with expert‐defined spikes and were predominantly located within the SOZ. The interpretability analysis discovered novel pathological features, including high power in the gamma (30–80 Hz) and ripple (>80 Hz) bands centered on the event with spike‐like activity. These characteristics were consistent across multiple variables, including institution, electrode type, patient demographics, and anatomical location. Predicting postoperative seizure outcomes using the resection ratio of mpHFOs outperformed unclassified HFOs (F1 = .72 vs. .68, p < .01) and matched current clinical standards using SOZ resection (F1 = .74, p = .76). Combining mpHFO data with demographic and SOZ resection status further improved prediction performance (F1 = .83, p < .01). Significance Our data‐driven approach using the generative artificial intelligence model yielded a novel, explainable definition of pathological HFOs, which has the potential to further enhance the clinical use of HFOs for EZ delineation.