The Spatial Divergence of Hypometabolism and Atrophy in Mesial Temporal Lobe Epilepsy: Combining 18F‐FDG PET and Structural MRI for Predicting Postoperative Seizure Freedom
TL;DR
This comprehensive preoperative clinical‐imaging model achieved robust prognostic efficacy that was not significantly improved by the addition of actual postoperative resection volumes, and demonstrates significant synergistic potential for predicting surgical outcomes.
Abstract
Surgical failure in temporal lobe epilepsy with hippocampal sclerosis (TLE‐HS) may reflect insufficient disruption of epileptogenic networks. We investigated whether integrating quantitative spatial patterns of regional atrophy and hypometabolism, along with clinical features, could provide complementary prognostic information for postoperative seizure freedom. T1‐weighted MRI, 18F‐FDG‐PET, and postoperative CT scans of patients with TLE‐HS were retrospectively analyzed against a validated healthy control cohort to compute age‐ and gender‐adjusted W‐score maps. Regional relationships of atrophy and hypometabolism were assessed using bivariate correlations and multiple regression analyses. Machine learning models integrating laterality of seizure onset, preoperative (extra‐)temporal atrophy/hypometabolism extent, seizure frequency, focal to bilateral tonic–clonic seizures during the preceding year, and resection volume were developed to predict seizure freedom 1 year after surgery. Model interpretability was assessed via permutation‐based variable importance analysis. The cohort comprised 101 patients with TLE‐HS (48 left, LHS; 53 right, RHS), including 72 patients who underwent anterior temporal lobectomy. Distinct modality‐specific patterns emerged: hypometabolism was significantly greater than local atrophy in left temporo‐limbic cortex in TLE‐LHS, while atrophy exceeded hypometabolism in bilateral occipital regions in TLE‐RHS. Positive correlations between local atrophy and hypometabolism were more robust in ipsilateral temporo‐limbic cortex in TLE‐LHS. The combined model integrating preoperative atrophy and hypometabolism features (AUC: 0.56–0.64) significantly outperformed single‐modality models (pfdr < 0.05). Notably, incorporating clinical factors further enhanced predictive performance (AUC: 0.63–0.70) and model calibration. This comprehensive preoperative clinical‐imaging model achieved robust prognostic efficacy that was not significantly improved by the addition of actual postoperative resection volumes. Structural MRI and 18F‐FDG‐PET reveal complementary facets of the epileptogenic network in TLE‐HS. While integrating multimodal imaging with clinical metrics demonstrates significant synergistic potential for predicting surgical outcomes, our models remain exploratory. External validation in independent multi‐center cohorts is required before translating these findings into precise clinical tools for surgical planning.