Precision dosing of voriconazole in immunocompromised children under 2 years: integrated machine learning and population pharmacokinetic modeling.

Abstract

Objective: This study aimed to develop an individualized dosing strategy for voriconazole (VRZ) in children under 2 years of age by integrating machine learning (ML) and population pharmacokinetic (PopPK) modeling.

Methods: This retrospective observational study included 76 eligible pediatric patients for model development, analyzing their baseline characteristics and laboratory parameters. A population pharmacokinetic (PopPK) model using NONMEM^® software was performed to assess the clearance (CL) and volume of distribution (V) of VRZ. The individual CL and V were included as input variables. The Boruta algorithm was employed for feature selection, after which six machine learning algorithms were applied. The models were evaluated using Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and coefficient of determination (R²) to identify the optimal algorithm, which then underwent independent external validation. The selected final model was analyzed for interpretability using Shapley Additive Explanations (SHAP).

Results: A total of 76 pediatric patients were enrolled for model development, consisting of 58 males (76.3%) and 18 females (23.7%), with a median age of 11 months and a median weight of 8.05 kg. We analyzed 110 therapeutic drug monitoring (TDM) samples of VRZ from these participants. A one-compartment model with first-order absorption and elimination described the population pharmacokinetics of VRZ. Population estimates for apparent clearance (CL/F) and volume of distribution (V/F) were 17.9 L/h/70kg (RSE, 10.8%) and 788 L/70kg (RSE, 15.4%), respectively. An XGBoost model accurately predicted voriconazole concentrations (R² = 0.81, RMSE = 0.53) with a relative error of ±20% for most observations. In the external validation, the XGBoost model demonstrated an R² of 0.75, RMSE of 0.14. SHAP analysis identified clearance, weight, and laboratory values as significant predictors.

Conclusion: This study emphasized the importance of personalized treatment in utilizing VRZ for children under 24 months. The XGBoost model demonstrated potential in identifying an initial dose recommendation for VRZ.