Individual [${}^{18}$F]FDG PET and functional MRI based on simultaneous PET/MRI may predict seizure recurrence after temporal lobe epilepsy surgery

To investigate the individual measures of brain glucose metabolism, neural activity obtained from simultaneous ${}^{18}$[F]FDG PET/MRI, and their association with surgical outcomes in medial temporal lobe epilepsy due to hippocampal sclerosis (mTLE-HS). Thirty-nine unilateral mTLE-HS patients who underwent anterior temporal lobectomy were classified as having completely seizure-free (Engel class IA; n = 22) or non-seizure-free (Engel class IB-IV; n = 17) outcomes at 1 year after surgery. Preoperative [${}^{18}$F]FDG PET and functional MRI (fMRI) were obtained from a simultaneous PET/MRI scanner, and individual glucose metabolism and fractional amplitude of low-frequency fluctuation (fALFF) were evaluated by standardizing these with respect to healthy controls. These abnormality measures and clinical data from each patient were incorporated into a machine learning framework (gradient boosting decision tree and logistic regression analysis) to estimate seizure recurrence. The predictive values of features were evaluated by the receiver operating characteristic (ROC) curve in the training and test cohorts. The machine learning classification model showed [${}^{18}$F]FDG PET and fMRI variations in contralateral hippocampal network and age of onset identify unfavorable surgical outcomes effectively. In the validation dataset, the logistic regression model with [${}^{18}$F]FDG PET and fALFF obtained from simultaneous [${}^{18}$F]FDG PET/MRI gained the maximum area under the ROC curve of 0.905 for seizure recurrence, higher than 0.762 with ${}^{18}$[F]-FDG PET, and 0.810 with fALFF alone. Machine learning model suggests individual [${}^{18}$F]FDG PET and fMRI variations in contralateral hippocampal network based on ${}^{18}$[F]-FDG PET/MRI could serve as a potential biomarker of unfavorable surgical outcomes. Individual [${}^{18}$F]FDG PET and fMRI obtained from preoperative [${}^{18}$F]FDG PET/MR were investigated. Individual differences were further assessed based on a seizure propagation network. Machine learning can classify surgical outcomes with 90.5% accuracy.