Development of a machine learning-based treatment-response prediction model for levetiracetam in pediatric epilepsy: integrating ABCB1 polymorphism and therapeutic drug monitoring

Background Empirical dosing of levetiracetam (LEV) in pediatric epilepsy often yields variable outcomes. This study aimed to develop a machine learning-based treatment-response prediction model integrating therapeutic drug monitoring (TDM) and ABCB1 C3435 T polymorphism. Methods Clinical, TDM, and genetic data from 346 pediatric patients were retrospectively analyzed. Four machine learning algorithms were evaluated. SHapley Additive exPlanations (SHAP) was utilized to interpret feature associations. Results The XGBoost model demonstrated the highest predictive performance (AUC=0.893). SHAP analysis indicated that LEV trough concentration (C min ) and ABCB1 genotype were the strongest associative factors for clinical outcomes. The homozygous TT mutation was associated with treatment resistance despite adequate peripheral C min . SHAP interaction patterns generated an exploratory hypothesis that ABCB1 genotype may modify the associative relationship between systemic exposure and treatment response. Conclusion Integrating ABCB1 genotyping and TDM via the XGBoost model may provides an interpretable approach for predicting LEV treatment response in pediatric epilepsy. Rather than directly generating individualized dosing regimens, these associative findings offer a hypothesis-generating reference for identifying patients at risk of therapeutic failure, warranting further prospective exposure-response validation.