Ensemble Machine Learning Model for Real-Time Valproic Acid Prediction in Epilepsy Treatment

Area of Science:

• Pharmacogenomics and Computational Chemistry
• Clinical Pharmacology and Therapeutics
• Artificial Intelligence in Medicine

Background:

• Epilepsy management requires maintaining therapeutic valproic acid (VPA) concentrations.
• Predicting VPA levels is crucial for optimizing treatment efficacy and minimizing toxicity.
• Machine learning offers potential for developing precise VPA concentration prediction models.

Purpose of the Study:

• To develop an optimal machine learning model for predicting VPA plasma concentrations.
• To ensure VPA levels remain within the effective therapeutic range for epilepsy patients.
• To enhance the control of epilepsy through accurate VPA concentration prediction.

Main Methods:

• A retrospective study of adult epilepsy patients on VPA therapy.
• Ensemble prediction model built using top-performing algorithms: Light Gradient Boosting, Categorical Boosting, and Gradient Boosted Regression Trees.
• Shapley Additive exPlanations (SHAP) used for model interpretation and feature importance analysis.

Main Results:

• The final ensemble model, using 20 selected variables, achieved an R-squared of 0.621 on external validation.
• External validation showed a mean absolute error of 10.67 and absolute accuracy of 78.98% (±20 mg/L).
• SHAP analysis identified VPA administration and liver function as key predictors of VPA concentration.

Conclusions:

• An advanced multi-algorithm approach effectively forecasts VPA concentrations in adult epilepsy patients.
• SHAP provides transparent insights into the factors influencing VPA prediction.
• This model offers a robust tool for personalized VPA therapeutic drug monitoring.