Integrating uncertainty and practice variability for epilepsy prognostication after intracranial EEG : A meta‐analysis

OBJECTIVE: Intracranial electroencephalography (iEEG) practices and diagnostic testing vary among institutions, limiting standardization and optimization of surgical outcomes. This study estimates resection rates and surgical outcome effect sizes after iEEG and develops individual patient-level data (IPD) models using the missing-indicator method to approximate real-world clinical uncertainties and practice variability in iEEG and surgical decision-making. METHODS: Databases were queried for studies reporting ≥12-month outcomes after epilepsy surgery in patients with drug-resistant epilepsy (DRE) who underwent iEEG. Random-effects meta-analysis estimated pooled outcome proportions. IPD analysis was performed using multivariable generalized linear regression and logistic regression models with cross-validation and calibration assessment. Missing values for key predictors were retained as a "not specified" category to evaluate the impact of incorporating uncertainty in common clinical data on outcome prediction. RESULTS: Primary outcomes were seizure freedom (Engel class IA or International League Against Epilepsy [ILAE] 1) and favorable seizure outcomes (Engel I or ILAE 1-2). Of 4975 studies identified, 115 met the inclusion criteria. The pooled frequency of patients undergoing resection after iEEG was 67% (95% confidence interval [CI] 60%-74%), 52% achieved seizure freedom (95% CI 48%-56%), and 58% achieved a favorable outcome (95% CI 54%-62%). IPD analysis included 1305 patients from 29 studies and identified a lesional magnetic resonance imaging (MRI; odds ratio [OR] 1.94, CI 1.48-2.55, p < .0001) and a cortical malformation (OR 1.47, CI 1.08-2.00, p = .01) as independent predictors of a favorable outcome, whereas iEEG implantation strategy (bilateral implantation OR .61, 95% CI .42-.89, p = .01) was an independent predictor of a less favorable outcome. SIGNIFICANCE: About two-thirds of patients undergoing iEEG proceeded to resection, and about half achieved seizure freedom. Multivariable models incorporating clinical uncertainty and practice variability were well calibrated. However, modest discrimination suggests that routinely available clinical variables alone have limited ability to predict individual surgical outcomes. More standardized and granular clinical and practice-related data may improve individualized prognostication after iEEG.