[Machine learning-based classification of temporal lobe epilepsy subtypes and surgical prognosis evaluation using PET...

Objective: To identify brain metabolic network features for temporal lobe epilepsy (TLE) subtype classification and surgical prognosis prediction using machine learning algorithms, thereby supporting clinical decision-making for TLE subtyping and outcome assessment. Methods: ¹⁸F-FDG PET images from 137 patients with drug-resistant TLE treated at Xiangya Hospital's Comprehensive Epilepsy Center from January 2016 to June 2021 were retrospectively analyzed as the training cohort. Network connectivity data were derived using Kullback-Leibler divergence similarity estimation (KLSE), yielding 6 902 network attributes alongside relevant demographic and clinical features. Eight machine learning models (including decision tree and random forest) were trained. The resulting models classified TLE subtypes and were validated using ¹⁸F-FDG PET metabolic network data from an independent cohort of 92 drug-resistant TLE patients (from July 2021 to August 2023). Decision curve analysis was used to select the most clinically practical model for predicting the surgical prognosis of 138 temporal lobe epilepsy patients, including 105 with mesial TLE (76 in the training group and 29 in the independent test group) and 33 with neocortical TLE (23 in the training group and 10 in the independent test group). Results: There were 84 males and 53 females in the training group, with an age of (22.0±8.0) years; in the independent test group, there were 45 males and 47 females, with an age of (24.2±12.8) years. The area under the receiver operating characteristic curve(AUC) of the 8 machine learning models in the training group ranged from 0.904 to 0.985; the AUC in the independent test group ranged from 0.859 to 0.946. According to the comparison of the performance of the above models, it was found that the prediction result of the random forest model was the most accurate and stable [AUC 0.985 (95%CI: 0.985-0.986), accuracy 0.998(95%CI: 0.995-1.000), sensitivity 0.950 (95%CI: 0.898-1.000), specificity 1.000 (95%CI: 1.000-1.000)]. For patients with mesial temporal lobe epilepsy who underwent surgery, the AUC for predicting surgical prognosis in the training group was 0.838 (95%: 0.753-0.923), and the accuracy was 0.838 (95%CI: 0.836-0.841); the AUC in the independent test group reached 0.783(95%CI: 0.549-1.000), with an accuracy of 0.793 (95%CI: 0.782-0.804). For patients with neocortical temporal lobe epilepsy who underwent surgery, the AUC for predicting surgical prognosis in the training group was 0.962(95%CI: 0.881-1.000), and the accuracy was 0.957 (95%CI: 0.953-0.960); while the AUC in the independent test group also reached 0.800 (95%CI: 0.408-1.000), with an accuracy of 0.900 (95%CI: 0.882-0.918). Conclusion: Machine learning models incorporating metabolic network features extracted from ¹⁸F-FDG PET data effectively support TLE subtype classification and surgical prognosis assessment.