Interictal EEG source connectivity to localize the epileptogenic zone in patients with drug‐resistant epilepsy: A machine learning approach

Abstract

ObjectiveTo deconstruct the epileptogenic networks of patients with drug‐resistant epilepsy (DRE) using source functional connectivity (FC) analysis; unveil the FC biomarkers of the epileptogenic zone (EZ); and develop machine learning (ML) models to estimate the EZ using brief interictal electroencephalography (EEG) data.MethodsWe analyzed scalp EEG from 50 patients with DRE who had surgery. We reconstructed the activity (electrical source imaging [ESI]) of virtual sensors (VSs) across the whole cortex and computed FC separately for epileptiform and non‐epileptiform EEG epochs (with or without spikes). In patients with good outcome (Engel 1a), four cortical regions were defined: EZ (resection) and three non‐epileptogenic zones (NEZs) in the same and opposite hemispheres. Region‐specific FC features in six frequency bands and three spatial ranges (long, short, inner) were compared between regions (Wilcoxon sign‐rank). We developed ML classifiers to identify the VSs in the EZ using VS‐specific FC features. Cross‐validation was performed using good outcome data. Performance was compared with poor outcomes and interictal spike localization.ResultsFC differed between EZ and NEZs (p < .05) during non‐epileptiform and epileptiform epochs, showing higher FC in the EZ than its homotopic contralateral NEZ. During epileptiform epochs, the NEZ in the epileptogenic hemisphere showed higher FC than its contralateral NEZ. In good outcome patients, the ML classifiers reached 75% accuracy to the resection (91% sensitivity; 74% specificity; distance from EZ: 38 mm) using epileptiform epochs (gamma and beta frequency bands) and 62% accuracy using broadband non‐epileptiform epochs, both outperforming spike localization (accuracy = 47%; p < .05; distance from EZ: 57 mm). Lower performance was seen in poor outcomes.SignificanceWe present an FC approach to extract EZ biomarkers from brief EEG data. Increased FC in various frequencies characterized the EZ during epileptiform and non‐epileptiform epochs. FC‐based ML models identified the resection better in good than poor outcome patients, demonstrating their potential for presurgical use in pediatric DRE.

Recent Advances in Wearable Healthcare Devices: From Material to Application

“…Thanks to the boom in deep learning, many deep learning models are developed and combined with EEG to detect mental fatigue [151], Parkinson's disease [152], depression [153], schizophrenia [154,155], epilepsy [156][157][158], and neurocognitive disorders [159,160]. In addition to symptom detection, EEG is reported to localize the epileptogenic zone [161,162] and it has the potential to guide the surgery. Kucikiene et al [163] found different EEG features of focal epilepsy and psychogenic nonepileptic seizures.…”