Explainable Machine Learning for Predicting Early Response to Perampanel in Pediatric Epilepsy: A Real-World Study
Xin Rao1,2, Ping Wu3,4, Xi Lin5
1Department of Pharmacy, The First Affiliated Hospital, Fujian Medical University, Fuzhou, Fujian, People's Republic of China.
Objective:
To develop and validate an early prediction model for perampanel (PER) response in children with epilepsy (CWE) using clinical covariates and ensemble learning, to support individualized treatment decisions.
Methods:
A retrospective cohort of 325 CWE (< 15 years) treated with PER from January 2021 to December 2025 was enrolled. Treatment response at 2 months was defined as the study outcome (responder vs non-responder). Key covariates were screened using Light Gradient Boosting Machine (LGBM) feature importance ranking combined with forward feature selection. Five core algorithms were selected from 11 candidate algorithms, and a soft voting ensemble model was constructed, interpreted using SHapley Additive exPlanations (SHAP), and subjected to age-stratified analysis.
Results:
Seven covariates were ultimately included. The ensemble model achieved a test set AUC of 0.861 (95% CI: 0.727-0.963), accuracy of 0.892 (95% CI: 0.815-0.954), and specificity of 0.940 (95% CI: 0.860-1.000); the mean 5-fold cross-validation AUC was 0.915 ± 0.064; external validation yielded an AUC of 0.798 (95% CI: 0.668-0.900). Decision curve analysis (DCA) demonstrated improved net benefit within threshold probabilities of 0.07 to 0.75. Calibration curves showed good calibration (brier score = 0.12). SHAP analysis identified the number of antiseizure medications (ASMs, 43.7%), PER maintenance dose (17.0%), and CYP3A4 enzyme inducers (CYP3A4-I, 13.1%) as the primary predictors, all positively associated with treatment non-response. Exploratory age-stratified SHAP analysis (≤72 months vs >72 months) suggested that the associations of polytherapy, higher PER maintenance dose, and shorter disease duration with predicted treatment non-response may be more pronounced in younger children.
Conclusion:
In this pediatric cohort, the ensemble model demonstrated good discrimination and calibration for predicting early perampanel response at 2 months. The number of ASMs, PER maintenance dose, and CYP3A4-I were the leading predictors of treatment non-response.