Prognostic method of MRI-negative epilepsy surgery based on brain network characteristics and machine learning

Objective: This study aims to establish an individualized prognosis prediction model for MRI-negative epilepsy surgery using the network-based statistics predict (NBS-Predict) method, by combining brain network characteristics and machine learning techniques. The goal is to accurately predict patients' prognosis and provide a basis for enhancing the surgical decision - making process and personalized diagnosis and treatment. Methods: This study comprised 31 MRI-negative epilepsy patients, classified by Engel criteria into 17 seizure-free and 14 seizure-recurrence patients, and their T1-weighted MRI data were obtained. By using NBS-Predict, we first identified the key brain regions and extracted five graph characteristics of the these regions. Then, the cmcRVFL+ machine learning model was employed to classify these features, enabling the prediction of surgical outcomes. Results: Based on the cmcRVFL+ model, this study extracted five graph theory indicators (betweenness centrality, degree centrality, nodal clustering coefficient, nodal) from key networks. The accuracy of the model based on all five indicators exceeded 80%. Among them, degree centrality showed the best classification performance with the highest accuracy. Its AUC, sensitivity, specificity, precision and F1-Score were 0.875, 0.956, 0.967, 0.970 and 0.964, respectively. Conclusion: Our results demonstrate that this approach can effectively capture critical network associated with prognosis and predict the prognosis of epilepsy patients, achieving a maximum accuracy of 96%. It provides a new tool for the personalized diagnosis and treatment of MRI-negative epilepsy patients.