# Locating the epileptogenic zone for drug-resistant epilepsy through neuroelectrophysiological brain network topology

## TL;DR

This study introduces a new method to locate the brain area causing drug-resistant epilepsy using brain network analysis and machine learning, improving surgical outcomes.

## Contribution

A hierarchical framework combining scalp EEG and SEEG with machine learning for accurate epileptogenic zone localization.

## Key findings

- The epileptogenic zone shows reduced network metrics like node degree and local efficiency in SEEG data.
- Scalp EEG-based SVM models achieved high diagnostic accuracy (AUC = 0.927) for EZ localization.
- SEEG models with normalization techniques improved performance (AUC = 0.872) for precise surgical planning.

## Abstract

Drug-resistant epilepsy (DRE) constitutes approximately one-third of the epilepsy population, posing a significant challenge due to low seizure freedom rates. Accurate localization of the epileptogenic zone (EZ) is the prerequisite for successful surgery. However, the limitations of conventional visual inspection underscore an urgent need for novel localization strategies based on quantitative brain network topology.

This study established a hierarchical analytical framework to independently analyze neuroelectrophysiological signals from scalp EEG (used for macroscopic hypothesis formulation) and stereoelectroencephalography (SEEG, used for mesoscopic confirmation). We included 25 patients with favorable surgical outcomes and constructed brain networks from ictal and interictal recordings. Subsequently, we evaluated the diagnostic value of these network features using machine learning classifiers [including Support Vector Machine (SVM), Random Forest, etc.].

In SEEG, the EZ exhibited significantly reduced topological metrics (specifically node degree, clustering coefficient, and local efficiency) compared to non-EZ regions (P < 0.001), indicating that the epileptogenic focus is a functionally isolated node. The SVM model based on interictal scalp EEG features achieved superior diagnostic performance (AUC = 0.927, Accuracy = 85.7%, Sensitivity = 85.7%, Specificity = 85.7%). In the SEEG modality, we applied Log-transformation and Z-score normalization to overcome individual variations in implantation schemes. This processing significantly boosted the performance of the interictal SEEG model (SVM) (AUC = 0.872, Accuracy = 81.9%, Sensitivity = 83.1%, Specificity = 80.7%).

These findings confirm the stability of the EZ's topological signature in the resting state and demonstrate a stepwise workflow: scalp EEG provides coarse localization of the potential EZ to guide SEEG implantation, while SEEG offers more precise surgical recommendations for EZ localization.