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

To develop and validate a machine-learning/population-pharmacokinetic (ML-PPK) hybrid model that predicts individual vancomycin clearance (CL) and volume of distribution (V_d) before the first dose, thereby informing initial dosing in critically ill children.

We retrospectively analysed children from two tertiary centers in China (2013-2023). A previously published one-compartment PPK model was re-estimated with the pooled dataset and used as a Bayesian prior to derive individual CL and V_d as training targets. Ten machine-learning and deep-learning algorithms were trained, and an XGBoost-based sequential forward-selection procedure was applied to identify a minimal predictor set. Model performance was evaluated on a held-out test set and an external cohort.

Data from 821 children and 1,767 vancomycin concentrations were included. 29 candidate variables were screened, and six high-impact predictors - body weight, cardiothoracic surgery, estimated glomerular filtration rate, sex, ICU admission, and post-menstrual-age class - maximized performance. CatBoost outperformed the other evaluated algorithms and, under this study design, more closely approximated PPK-Bayesian posterior PK estimates than the original parametric PPK covariate model, yielding for CL: R² = 0.89, and 81.8% of predictions within ±30%; and for V_d: R² = 0.95, with 92.1% within ±30%. Performance remained robust in both the test set and the external validation cohort, with external validation R² values of 0.85 for CL and 0.95 for V_d. SHAP analysis highlighted body weight, renal function, and cardiothoracic surgery status as the main determinants of CL, consistent with covariate effects in the PPK model.

An interpretable CatBoost-based ML-PPK hybrid can estimate CL and V_d pre-dose using routinely available data, enabling patient-specific initial vancomycin regimens and reducing early under- or overexposure in pediatric critical care.