Noninvasive Preoperative Evaluation in Drug-Resistant Epilepsy Based on the Brain Network Topological Dynamics

Drug-resistant epilepsy (DRE) affects over 50 million individuals worldwide, yet surgical resection, the most effective treatment, achieves seizure freedom in only approximately 50% of cases. A reliable, noninvasive method for preoperative surgical outcome prediction is therefore critically needed to avoid the risks of invasive intracranial monitoring. This retrospective study developed a fully noninvasive, personalized framework that integrates scalp EEG source imaging, brain network topological features, and machine learning. In 43 DRE patients (110 seizures), Global Diffusion Efficiency (GEDIFF) and Network Density (KDEN) emerged as the most robust prognostic markers, both showing ${p} \lt 0.001$ , while transitivity ( $T$ ), betweenness centrality (BC), and node strength (STR) in the epileptogenic zone provided complementary discriminative information. In leave-one-out cross-validation, KNN achieved 86.0% accuracy and 88.0% precision, while in five-fold cross-validation, CatBoost reached 86.4% accuracy and 90.1% precision, with ExtraTrees attaining an AUC of 94.6%. This noninvasive performance is comparable to that of invasive iEEG-based models, yet it was achieved using only 16-channel scalp EEG, the most widely available clinical montage. These findings suggest that favorable surgical outcomes are associated with more isolated epileptogenic zone networks, supporting the clinical potential of noninvasive brain network analysis for personalized preoperative evaluation. Source code is available at: https://github.com/wyl1994/Source-code-for-our-TNSRE-paper.