Dosing prediction of valproic acid in pediatric patients with epilepsy: population pharmacokinetic model or machine learning model?

Purpose: This study develops and compares population pharmacokinetics (PopPK) models and machine learning methods, including neural networks, to predict steady-state trough concentrations in pediatric patients and provide improved dosing recommendations.

Methods: Valproic acid concentration data were collected from 490 pediatric epilepsy patients treated at Beijing Tiantan Hospital and Beijing Children's Hospital. We developed predictive models employing PopPK, maximum a posteriori Bayesian (MAPB), multiple linear regression (MLR), machine learning (including Random Forest, XGBoost, and LightGBM for feature selection), and neural network techniques. The predictive accuracy of these models was then rigorously tested through external validation using the independent dataset from Beijing Children's Hospital. Upon identifying the optimal model, dosing regimens for various clinical scenarios were derived and presented.

Results: Under the same dataset modeling conditions, the original PopPK models showed limited predictive performance. Transforming these models into multiple linear regression enhanced prediction accuracy. Moreover, when prior data was available, the MAPB method significantly boosted prediction performance. Machine learning and neural networks showed higher accuracy, with neural networks achieving an F₃₀ value above 80%.

Conclusion: This study explored model optimization strategies and compared machine learning and neural network models alongside traditional PopPK. It introduced an advanced method to predict drug concentrations and stable trough dosing regimens in pediatric epilepsy treatment, reducing the need for frequent, invasive blood tests in TDM. These improvements enhanced the efficacy and safety of valproic acid therapy for children, supporting the development of personalized treatment plans.