Monitoring of dual-drug combination therapy in pediatric epilepsy patients: a machine learning model for simultaneous VPA-LEV concentration-dose prediction.

BACKGROUND

To better meet the therapeutic demand in pediatric epilepsy patients (PEPs) receiving combined valproic acid (VPA) and levetiracetam (LEV) (therapeutic ranges: VPA 50–100 μg/mL; LEV 6–20 μg/mL.), We developed a dual-predictive machine learning (ML) model that integrates concentration monitoring with dose recommendation capabilities, serving as an adjunct tool for therapeutic drug monitoring (TDM).

METHODS

A retrospectively collected dataset comprising 497 paired concentration samples from 402 PEPs was used to train (i) classification models that predict adequacy of VPA and LEV concentrations and (ii) regression models that recommend individualised daily doses. Model explainability was interrogated with SHapley Additive exPlanations (SHAP) analysis to delineate concentration-dose covariates.

RESULTS

Among nine nonlinear ML algorithms, the Extra-Trees classifier achieved optimal performance for concentration adequacy prediction, delivering test-set accuracies and AUCs of 0.74 and 0.75 for VPA, and 0.75 and 0.75 for LEV, respectively. SHAP analysis elucidated body weight-normalized daily dose, blood urea nitrogen/creatinine (BUN/CREA) ratio, and platelet count (PLT) were identified as shared critical covariates. Serum uric acid (UA) exhibited LEV-specific positive regulation (SHAP rank 2, contribution 16.2%). For dose recommendation, an XGBoost multi-output regressor yielded test-set R2 values of 0.60 for both VPA and LEV daily dose. Body weight dominated dose predictions (VPA 70.2%; LEV 62.8%), followed by respective trough concentrations (VPA 7.1%; LEV 14.7%).

CONCLUSION

This dual-prediction model enables simultaneous concentration monitoring and dose recommendation of VPA and LEV, offering a data-driven decision-support tool for personalised therapy in PEPs.