Machine-learning prediction of pre-dose pharmacokinetics optimizes initial vancomycin dosing in critically ill...

Area of Science:

• Pharmacometrics
• Machine Learning in Pediatrics
• Drug Dosing Optimization

Background:

• Accurate vancomycin dosing is crucial for critically ill children.
• Traditional population pharmacokinetic (PPK) models may not fully capture individual variability.
• Predicting pharmacokinetic parameters pre-dose is challenging but essential for initial therapy.

Purpose of the Study:

• To develop and validate a machine-learning/population-pharmacokinetic (ML-PPK) hybrid model.
• To predict individual vancomycin clearance (CL) and volume of distribution (Vd) before the first dose.
• To inform initial vancomycin dosing strategies in critically ill pediatric patients.

Main Methods:

• Retrospective analysis of pediatric data from two tertiary centers.
• Re-estimation of a one-compartment PPK model and use as a Bayesian prior.
• Training and evaluation of ten ML/deep learning algorithms, including XGBoost and CatBoost.
• Validation on a held-out test set and an external cohort.

Main Results:

• Six high-impact predictors identified: body weight, cardiothoracic surgery, eGFR, sex, ICU admission, and post-menstrual-age class.
• CatBoost model achieved R²=0.89 for CL and R²=0.95 for Vd, with 81.8% and 92.1% of predictions within ±30%, respectively.
• External validation confirmed robust performance with R²=0.85 for CL and R²=0.95 for Vd.
• SHAP analysis confirmed body weight, renal function, and cardiothoracic surgery as key determinants of CL.

Conclusions:

• An interpretable CatBoost-based ML-PPK hybrid model accurately estimates CL and Vd pre-dose.
• The model utilizes routinely available clinical data for personalized vancomycin regimens.
• This approach can reduce initial vancomycin under- or overexposure in pediatric critical care.