Personalized prediction of initial valproic acid dose in children with epilepsy using machine learning techniques

Valproic acid (VPA) is widely prescribed antiepileptic drug in children because of its broad-spectrum efficacy. However, marked inter-individual variability in pharmacokinetics makes dose optimization challenging. Inappropriate initial dosing can lead to subtherapeutic exposure or toxicity, including hepatotoxicity, encephalopathy, and hematological adverse effects. Despite the importance of early dose selection, reliable methods for predicting individualized initial daily doses of VPA in pediatric epilepsy are lacking.

This study aimed to develop and internally validate a predictive model for the initial daily dose of VPA in pediatric epilepsy patients based on a real-world clinical database using artificial intelligence techniques.

A retrospective cohort of pediatric epilepsy patients treated with VPA was recruited from Baoding Children's Hospital between December 2017 and November 2023. Eligible patients were aged 1-16 years, received oral VPA therapy, and underwent therapeutic drug monitoring. Data pre-processing included z-score standardization, one-hot encoding, and Random Forest imputation for missing values. Variable selection was performed by univariate analysis and sequential forward selection. The dataset (306 VPA dose samples from 184 patients) was randomly divided into training (80%) and testing (20%) sets. Ten machine learning and deep learning algorithms-CatBoost, XGBoost, LightGBM, Random Forest, Support Vector Machine, Multilayer Perceptron, Artificial Neural Network, Transformer, AdaBoost, and TabNet-were trained with tenfold cross-validation. The model performance was assessed using the coefficient of determination (R2), root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE).

Age, weight, total protein, creatinine, and lactate dehydrogenase were the five most influential predictors of initial daily VPA dose. Among all the models, the TabNet algorithm achieved the best predictive performance, with R² = 0.730, RMSE = 0.153, MAE = 0.116, and MAPE = 18.19% in the test set. The proportions of predictions within ± 30, ± 40, and ± 50% of the actual dose were 85.48, 91.94, and 95.16%, respectively.

The TabNet-based model demonstrated strong predictive ability for estimating individualized initial VPA doses in pediatric epilepsy in a single-center cohort with internal validation only. By integrating readily available demographic and laboratory parameters, the model and its accompanying online tool have the potential to offer a practical decision-support resource for clinicians and pharmacists to improve dosing precision, enhance safety, and optimize therapeutic outcomes; however, the generalizability of this internally validated model to other hospitals, populations, and ethnic groups is unknown, and external validation is required before clinical implementation.