Metabolic subtypes and inferred stages in temporal lobe epilepsy identified using FDG-PET–based subtype and stage inference

Abstract Background Temporal lobe epilepsy (TLE) is a common cause of drug-resistant epilepsy, yet surgical outcomes remain heterogeneous. We aimed to characterize laterality-specific metabolic heterogeneity in unilateral TLE by applying Subtype and Stage Inference (SuStaIn) to FDG-PET. Methods We retrospectively enrolled 1763 patients with unilateral TLE (1062 left-sided, 701 right-sided) and 39 healthy controls. FDG-PET images were preprocessed, regional metabolism was extracted from predefined volumes of interest, and separate Subtype and Stage Inference models were fitted for left-sided TLE (LTLE) and right-sided TLE (RTLE). The optimal subtype number was selected using five-fold cross-validation. Postoperative outcome was analyzed in the surgical subcohort using multivariable logistic regression. Results Four model-inferred metabolic subtypes were identified in each hemispheric model. Two showed clear one-to-one cross-hemispheric correspondence, representing bilateral early-hypermetabolic and bilateral early-hypometabolic patterns. The remaining two showed partial, non-unique correspondence, and were characterized by hippocampal–posterior cingulate and caudate–insular mixed patterns. Clinical profiles differed across subtypes, with differences in age, sex, and epilepsy duration in LTLE and in age and postoperative outcome in RTLE. Longer epilepsy duration and later age at seizure onset were associated with lower odds of seizure freedom across all RTLE models. In the final multivariable models, the bilateral early-hypometabolic subtype in LTLE (OR = 0.54, 95% CI 0.30–0.98; P = 0.044) and the hippocampal–posterior cingulate mixed subtype in RTLE (OR = 0.34, 95% CI 0.15–0.79; P = 0.011) were associated with lower odds of seizure freedom than the reference subtype. Exploratory subtype-stratified analyses identified several nominal associations involving epilepsy duration, age at seizure onset, hippocampal sclerosis, and SuStaIn stage. Split-half validation supported the stability of the broad subtype structure, while agreement in detailed event ordering was moderate. Conclusions Application of SuStaIn to FDG-PET in a large, laterality-specific TLE cohort identified model-inferred metabolic progression patterns associated with distinct clinical profiles and postoperative outcomes.