The SEEG brain network predicts epileptic surgical outcomes of radiofrequency thermocoagulation.

Predicting the postoperative outcome of stereoelectroencephalography-guided radiofrequency thermocoagulation (SEEG-guided RF-TC) remains challenging despite its increasing use in epilepsy treatment. Although SEEG-guided RF-TC has attracted extensive clinical interest, reliable biomarkers for treatment efficacy are still lacking. This study aims to address this gap by analyzing the altered brain network to predict postoperative outcome. Thirty-one focal cortical dysplasia epileptic patients who underwent RF-TC based on SEEG were enrolled in this study. They were included in the favorable outcome and poor outcome groups according to the follow-up. Partial Directed Coherence and Directed Transfer Function were applied to construct SEEG brain networks, and then brain network features were extracted. Subsequently, the differences in the presurgical and postsurgical brain network features were compared using the Wilcoxon test in the favorable and poor outcome groups, respectively. Finally, four machine learning models were applied to predict the outcome of RF-TC. After RF-TC surgery, the Characteristic Path Length (L) and average Betweenness Centrality (BC) increased while the average Clustering Coefficient (C) and Assortativity Coefficient (R1, R2) decreased in the favorable outcomes group. In contrast, there were no significant changes in the patient group with poor outcomes. The Support Vector Machine (SVM) model achieved the highest performance, with accuracy, sensitivity, specificity, and ROC values of 0.887, 0.821, 0.920, and 0.879, respectively. This study sheds light on the mechanisms of epilepsy from the perspective of brain networks and introduces a novel therapeutic strategy by altering network features. These feature alterations can also support machine learning models in effectively distinguishing favorable from poor outcomes.