Machine Learning Prediction of Unbound Ceftriaxone Concentrations in Children: Capturing Developmental Changes in...

The pharmacological activity of antimicrobial agents depends on unbound concentrations, but accurately estimating these free fractions remains challenging in pediatric patients due to developmental protein binding changes. This study aimed to develop a machine learning (ML) model to predict unbound ceftriaxone concentrations using routinely available clinical variables while capturing nonlinear developmental effects. A total of 176 paired total/unbound ceftriaxone concentrations and routine clinical variables from neonates, infants, and children were used to train ten ML algorithms. The optimal model was evaluated in an independent real-world cohort and compared with two established mathematical equations. The ExtraTrees Regressor achieved the best predictive performance, with a root mean square error (RMSE) of 6.95 μg/mL and an R² of 0.87. Using only routine variables, including total ceftriaxone concentration, serum albumin, weight, and age, the model accurately estimated unbound ceftriaxone concentrations across a wide range of pediatric patients (0-12 years). Real-world validation further confirmed its robustness, yielding a mean absolute percentage error (MAPE) of 28.9% and outperforming both reference equations, with reductions in MAPE of 14.2% and 14.7% relative to the empirical In vivo and disease-adapted equations, respectively. Application of the final model to a virtual pediatric population aged 0-12 years revealed a nonlinear age-dependent protein binding pattern, characterized by low and variable binding in neonates, rapid maturation during early infancy, and stabilization after 2 years of age. By incorporating these maturational effects, the ML model provides a practical tool for predicting unbound drug exposure, facilitating individualized dosing and precision medicine in pediatric patients.