Language and Memory Network Alterations in Temporal Lobe Epilepsy: A Functional and Structural Connectivity Study.

Abstract

Background and purpose: This study evaluated preoperative alterations and postoperative reorganization of the joint language-memory network (LMN) from the perspective of resting-state functional and structural connectivity in temporal lobe epilepsy (TLE). Graph theory and machine learning approaches were used to explore automatic lateralization.

Materials and methods: Resting-state fMRI and DTI data were obtained from 20 healthy subjects and 35 patients with TLE. Functional and structural connectivity were calculated within the LMN before and after temporal lobectomy. ANOVA was performed to identify significant connectivity differences between groups. Four local graph measures were extracted from functional and structural connectivity matrices. Standard feature selection techniques and genetic algorithm methods were applied to select the optimal features. Subsequently, K-nearest neighbor, support vector machine, Naive Bayes, and logistic regression classification methods were used to classify healthy controls (HC) and presurgical TLE groups, as well as presurgical left TLE (L-TLE) and right TLE (R-TLE) groups. Also, relationships between psychological scores and the selected features were evaluated using a linear regression method.

Results: The results demonstrated increased functional and decreased structural connectivity in patients with TLE before surgery. After surgery, significant connections revealed reduced functional connectivity and increased structural connectivity in patients with TLE. Functional analysis identified the left parahippocampal region in L-TLE and the right temporal regions in R-TLE as key areas. Structural connectivity analysis showed that memory-related areas in the bilateral occipital region and the left language-related area were the origins of alterations. The genetic algorithm method achieved the highest classification performance using a support vector machine for fMRI and DTI graph measures, with accuracy rates of 97% and 88% for distinguishing L-TLE from R-TLE, and 93% and 87% for distinguishing patients with TLE from HC, respectively. Moreover, a significant relationship was observed between the best-selected features and memory-assisted cognitive tests.

Conclusions: Presurgical functional hyperconnectivity, postsurgical hypoconnectivity, and newly observed bilateral structural connectivity after surgery highlight both functional and structural alterations in the LMN network. Additionally, the study underscores the potential for machine learning for TLE diagnosis and lateralization. A limited sample size, particularly in the postsurgical group, was one of the constraints of this study.