Prediction of Long-Term Postsurgical Seizure Recurrence From MRI Brain Hub Disruption in Patients With Temporal Lobe Epilepsy.
TL;DR
The model achieved high specificity in predicting long-term seizure recurrence, which supports its potential clinical utility for postoperative risk stratification and counseling rather than surgical exclusion and underscores the translational potential of network-level biomarkers to complement conventional predictors.
Abstract
Background
AND
Objectives
Patients with temporal lobe epilepsy (TLE) can achieve seizure freedom in the early period after surgery, yet up to half experience seizure recurrence in the following years (i.e., long-term). TLE is associated with disruption of highly connected brain regions (hubs), which may reduce the likelihood of long-term surgical success. We tested whether disruption of physiologic (normative) hubs predicts long-term seizure outcomes.
Methods
In a prospective, multimodal cohort of patients with drug-resistant TLE from 6 centers who underwent resective or laser ablative surgery and had more than 2 years of follow-up (mean = 5.4 years, SD = 3.2 years), we derived structural and functional connectomes from preoperative diffusion-weighted MRI and resting-state fMRI. Using a large multicenter healthy-control cohort, we identified normative connector hubs and quantified patient-specific disruption within these hubs using the graph-theory measure-participation coefficient. To classify seizure-free (positive class) and non-seizure-free (negative class) outcomes, we trained machine learning models using patient-specific disruption of the participation coefficient derived from structural and functional connectomes and their combination (multimodal approach). We evaluated model performance in an independent cohort. Models further incorporated clinical and demographic variables, as well as gray and white matter volumes.
Results
In our cohort of 175 patients, the multimodal approach outperformed a model based on clinical and demographic variables only and unimodal approaches, achieving high specificity (mean = 80.0%, SD = 9.9%) and moderate-to-high negative predictive value (mean = 63.9%, SD = 3.6%). Using 362 healthy controls to define normative connector hubs, Shapley Additive Explanation analyses identified disruption of the participation coefficient in the hippocampi and connector hubs of the dorsal attention network as predictive of long-term seizure recurrence, which includes areas not typically targeted in TLE surgery.
Discussion
Disruption of normative hub architecture provides biologically interpretable biomarkers of long-term seizure outcomes in TLE. Contrary to previous studies, the model achieved high specificity in predicting long-term seizure recurrence, which supports its potential clinical utility for postoperative risk stratification and counseling rather than surgical exclusion. By validating performance in an independent cohort under conservative evaluation, our study underscores the translational potential of network-level biomarkers to complement conventional predictors.
