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

Area of Science:

• Artificial Intelligence in Medicine
• Pharmacokinetics and Pharmacodynamics
• Pediatric Neurology

Background:

• Valproic acid (VPA) is a common antiepileptic drug for children, but its effectiveness is limited by significant inter-individual pharmacokinetic variability.
• Imprecise initial VPA dosing can result in subtherapeutic levels or toxicity, including hepatotoxicity and encephalopathy.
• Current methods for predicting individualized initial VPA doses in pediatric epilepsy are insufficient.

Purpose of the Study:

• To develop and internally validate an artificial intelligence (AI)-based predictive model for initial daily valproic acid (VPA) doses in pediatric epilepsy patients.
• To leverage a real-world clinical database for model development and assessment.

Main Methods:

• A retrospective cohort of 184 pediatric epilepsy patients (aged 1-16 years) treated with VPA was analyzed.
• Data underwent pre-processing including standardization and imputation, followed by variable selection.
• Ten machine learning and deep learning algorithms were trained and evaluated using tenfold cross-validation, with performance measured by R², RMSE, MAE, and MAPE.

Main Results:

• Age, weight, total protein, creatinine, and lactate dehydrogenase were identified as key predictors for initial VPA dose.
• The TabNet algorithm demonstrated superior performance, achieving an R² of 0.730, RMSE of 0.153, MAE of 0.116, and MAPE of 18.19% on the test set.
• Approximately 85.48% of predictions were within ±30% of the actual VPA dose.

Conclusions:

• The TabNet-based model shows significant potential for predicting individualized initial VPA doses in pediatric epilepsy patients.
• This AI tool, integrating demographic and laboratory data, could aid clinicians and pharmacists in optimizing VPA dosing precision and safety.
• External validation is necessary to confirm the model's generalizability across diverse populations and clinical settings before widespread implementation.