Predictive Performance of pyDarwin-Selected Versus Expert-Developed Paediatric Tobramycin Population Pharmacokinetic...

PyDarwin is an open-source Python package integrating machine learning with NONMEM for automated model selection; however, its utility in model-informed precision dosing (MIPD) remains unvalidated. This study aims to evaluate pyDarwin for population pharmacokinetics (PopPK) model development in MIPD compared with expert-developed models, using paediatric tobramycin as a representative case. The model-building dataset included 450 plasma concentrations from 63 children with cystic fibrosis (CF). One group manually developed a model using stepwise covariate selection, while another applied pyDarwin within a predefined 27-dimensional model space, using a fitness function that incorporated a penalty based on 24-h area under the curve (AUC_24h). External validation was performed using 1011 concentrations from 144 children, including evaluation using all therapeutic drug monitoring (TDM) data and clinically relevant MIPD scenarios (a priori [AP], Bayesian 1, Bayesian 2). Both models included weight-based scaling, a maturation function, and estimated glomerular filtration rate as covariates on clearance, yielding comparable clearance estimates. The pyDarwin-selected model was a two-compartment model with fixed allometric scaling exponents, whereas the manual model was a one-compartment model with estimated exponents. Both models showed comparable probability of target attainment for steady-state AUC_24h. In external validation using all TDM data, the pyDarwin-selected model demonstrated better predictive performance for CF patients than manual model (prediction-based MPE -14.6% vs -41.6%; MAPE 34.0% vs 54.4%). In MIPD scenarios, the pyDarwin-selected model showed superior performance in the AP setting, with differences diminishing after Bayesian updating. These findings support pyDarwin-based automated PopPK model selection as a reliable and efficient framework for MIPD.