Prediction of Zonisamide Concentration in Pediatric Patients With Epilepsy: A Machine Learning Approach

Area of Science:

• Pharmacokinetics and Pharmacogenomics
• Machine Learning in Medicine
• Pediatric Neurology

Background:

• Zonisamide (ZNS) is an antiepileptic drug used in pediatric patients.
• Individualized dosing is crucial for optimizing ZNS efficacy and minimizing toxicity.
• Predictive models can aid in tailoring ZNS regimens for children.

Purpose of the Study:

• To construct and validate machine learning models for predicting ZNS concentration in pediatric epilepsy patients.
• To identify the optimal algorithm for ZNS concentration prediction.
• To provide a basis for individualized ZNS dosing strategies in children.

Main Methods:

• Retrospective analysis of clinical data from 532 pediatric patients undergoing ZNS therapeutic drug monitoring.
• Feature selection using correlation analysis, Lasso regression, and Random Forest.
• Construction and evaluation of 12 machine learning regression models, including Random Forest.
• Validation using internal and external cohorts, with performance metrics like R², MSE, RMSE, and MAE.
• Interpretation of model features using the SHAP method.

Main Results:

• The Random Forest (RF) model demonstrated optimal performance, achieving high R² values (0.97% training, 0.78% internal validation, 0.89% external validation).
• Key predictors of ZNS concentration included gender, dosage, age, red blood cell count, concomitant medication, total protein, uric acid, and platelets.
• SHAP analysis identified dosage and gender as primary factors, with dosage positively influencing concentration and gender having a bidirectional effect.
• The model's decision logic aligned with pharmacokinetic principles, supporting its clinical interpretability.

Conclusions:

• A validated Random Forest model accurately predicts zonisamide concentration in pediatric epilepsy patients.
• Gender, dose, uric acid, and total protein are critical variables for ZNS concentration.
• The developed model offers high precision, stability, and generalizability for individualized pediatric ZNS dosing.