Authors :
Presenting Author: Victor Karpychev, PhD – Emory University
William Yun, MPH – Emory University
Rebecca Roth, BS – Emory University
Daniel Drane, PhD – Emory University School of Medicine
Patricia Dugan, MD – NYU Langone Health
Heath Pardoe, PhD – NYU Langone
Anto Bagic, MD, PhD – University of Pittsburgh Department of Neurology
Alexandra Parashos, MD – Medical University of South Carolina
Kathryn Davis, MD – Center for Neuroengineering and Therapeutics and Penn Epilepsy Center, Department of Neurology, University of Pennsylvania
Ruben Kuzniecky, MD – Hofstra/Northwell University
Greydon Gilmore, PhD – Emory University School of Medicine
Nealen Laxpati, MD, PhD – Emory University School of Medicine
Leonardo Bonilha, M.D., Ph.D. – University of South Carolina
Ezequiel Gleichgerrcht, M.D., Ph.D. – Emory University
Rationale:
While 60–70% of patients with drug-resistant temporal lobe epilepsy (TLE) achieve seizure-free (SF) status in the first postoperative year, long-term outcomes often deteriorate (Lamberink et al., 2020), with many patients eventually becoming non-seizure-free (NSF). Given that TLE disrupts functional and structural brain hubs (Gleichgerrcht et al., 2020; Royer et al., 2022), in this study, we investigated whether preoperative hub metrics derived from functional (FC) and structural (SC) connectivity could predict long-term surgical outcomes.
Methods:
We prospectively collected resting-state fMRI from 102 patients (52% of SF) and diffusion-weighted MRI (dMRI) data from 167 patients (56% of SF) prior to surgery for TLE. For 94 patients (48% of SF), we acquired fMRI and dMRI. Seizure outcomes were followed longitudinally, at least 2 years after surgery (post-surgical duration: M = 65.2, SD = 38.9 months). Based on FC and SC, we computed betweenness centrality (BC) as a measure of hubness for 374 ROIs (Glasser and Desikan-Killiany atlases). To classify SF vs. NSF, we trained Elastic Net Logistic Regression models within a nested 4-fold cross-validation framework. FC, SC, and combined FC-SC models also included age of onset, epilepsy duration, post-surgical duration, lesions, and were each evaluated using 100 iterations, with AUC as the performance metric. To avoid overfitting, feature selection involved F-tests (α < 0.05) and LASSO regression on training folds. We compared AUC based on FS and SC with FS-SC using 5000-permutation paired t-tests. Finally, we aligned the spatial pattern of the difference in BC between NSF and SF with normative cortical maps from neuromaps (Markello et al., 2022) using spin-based Spearman correlations.
Results:
For FC, BC predicted long-term seizure outcome with a mean AUC of 0.75 (SD = 0.08), compared to 0.71 (SD = 0.06) for SC, and 0.76 (SD = 0.09) for FC-SC. The FC-SC model significantly outperformed the SC [t(198) = -5.73, pperm < 0.001], but not the FC model [t(198) = -1.15, pperm < 0.25]. Fig.1 shows the importance of the features in the FC-SC model. Positive SHAP values indicate the stronger capability of the features to predict SF; negative SHAP values indicate the stronger capability of the features to predict NSF. For SC, there was a trend that greater BC differences between SF and NSF (Fig.2A) were positively correlated to higher rates of myelination (rho = 0.11, pspin = 0.06). For FC, regions with greater BC differences (Fig.2B) were spatially aligned with cortical areas showing higher rates of myelination (rho = 0.16, pspin = 0.01).
Conclusions:
Preoperative betweenness centrality derived from functional connectivity predicts long-term post-surgical seizure outcomes in TLE with higher accuracy than structural hub metrics alone. The spatial correspondence between betweenness centrality differences and cortical myelination suggests that network vulnerability in highly myelinated regions may underlie poor long-term outcomes.
Funding:
EG was supported by awards UL1TR002378 and KL2TR002381 by the National Center for Advancing Translational Sciences (NCATS) of the National Institutes of Health (NIH). Imaging data were obtained through NINDS R01NS110347.