Authors :
Presenting Author: Keisuke Hatano, MD, PhD – Wayne State University
Naoto Kuroda, MD, PhD – Wayne State University
Hiroshi Uda, MD, PhD – Wayne State University
Michael Cools, MD – Children's Hospital of Michigan
Aimee Luat, MD – Children’s Hospital of Michigan
Eishi Asano, MD, PhD – Wayne State University
Rationale:
Our fundamental hypotheses are that epileptogenicity in drug-resistant focal epilepsy exhibits a spatial gradient and that maximizing resection of highly epileptogenic areas while preserving least epileptogenic cortex optimizes seizure control. The rate of high-frequency oscillations (HFOs) is proposed as a biomarker to quantify epileptogenicity, predict the seizure onset zone (SOZ), and forecast postoperative seizure outcomes. Investigators suggest that morphological features of HFOs can enhance localization accuracy. However, spatial and age-related variability in HFO rates has also been reported within nonepileptic areas. Thus, we aimed to determine if integrating HFO rate, morphology, anatomical location, and patient age into a comprehensive model improves prediction accuracy for SOZ and postoperative seizure outcomes compared to models based solely on HFO rate or on standard-care evaluation.Methods:
We analyzed 20-minute intracranial EEG recordings obtained during stage-2 or slow-wave sleep from 177 patients who underwent focal resection and had at least one year of postoperative follow-up. HFO events were detected using four algorithms implemented in RIPPLELAB. HFO morphological parameters included power, duration, maximum peak spectral frequency, and the number of spectral peaks within the 80–300 Hz range. Cortical lobes generating HFOs, patient age, and electrode type (subdural or depth) were incorporated as predictors. Using an extreme gradient boosting algorithm trained on data from patients recruited in or before 2015, we constructed two predictive models: an HFO-rate-only model and a comprehensive HFO model also incorporating morphology, anatomical location, patient age, and electrode type. Accordingly, each HFO model assigned a SOZ probability to each electrode site. We performed receiver operating characteristic (ROC) analyses using an independent test dataset comprising patients recruited in 2016 or later and assessed how accurately each HFO model discriminated SOZ from non-epileptic sites. For each patient, we calculated the difference between the mean SOZ probabilities of resected and preserved sites; this summary measure represents the extent to which areas of higher epileptogenicity were resected compared to preserved areas with lower epileptogenicity. We evaluated whether this summary measure derived from the comprehensive HFO model predicted ILAE class 1 outcomes more accurately than that from the HFO-rate-only model, and also determined if it improved prediction relative to standard-care evaluation.Results:
The comprehensive HFO model classified SOZ sites more accurately than the HFO-rate-only model (area under the curve [AUC]: 0.980–0.996 vs. 0.74–0.85; p< 0.00001). The comprehensive HFO model more accurately classified postoperative seizure outcomes (AUC: 0.72–0.76 vs. 0.51–0.61; p≤0.00026). Incorporating the summary measure from the comprehensive HFO model enhanced standard-care outcome prediction (AUC increased from 0.72 up to 0.83; p=0.0009).Conclusions:
Incorporating HFO rate, morphology, anatomical location, and patient age improves localization of the epileptogenic zone.Funding:
The Uehara Memorial Foundation, 202441017 (K.H.), NIH NS064033 (E.A.).