Abstract
The pharmacokinetics of vancomycin in neonates and infants exhibits significant variability, presenting challenges in achieving target exposures. This study aimed to investigate the influence of patient-specific covariates on vancomycin clearance and evaluate the predictive performance of various machine learning (ML) methods using clinical covariates and plasma metabolomics data. A retrospective population pharmacokinetic (PK) analysis was conducted on 42 neonates and infants treated at the University of Michigan Neonatal Intensive Care Unit from 2019 to 2022. Vancomycin was administered intravenously at doses ranging from 3.5 to 25 mg/kg every 6 to 24 h. A total of 214 vancomycin concentration measurements, including trough, peak, and random levels, were included in the analysis. Plasma samples collected from the patients were analyzed by an LC-MS/MS-based untargeted metabolomics assay. A one-compartment model with first-order elimination best described the pharmacokinetics of vancomycin, with serum creatinine (SCr), postmenstrual age (PMA), and weight identified as significant covariates influencing clearance. Among the ML methods evaluated, Gradient Boosting Regressor (GBR) achieved the highest predictive performance using clinical covariates (MSE: 0.0033; R(2): 0.830). Incorporating metabolomics data did not significantly improve predictive performance for most models based solely on clinical covariates, although certain metabolomics features were among the top predictors. Both PK modeling and ML identified SCr and PMA as the most important covariates. These findings highlight the utility of ensemble ML methods, particularly GBR, in predicting vancomycin clearance using clinical covariates. While metabolomics provided limited added value for vancomycin clearance prediction, this study demonstrated an integrated ML and metabolomics approach capable of exploring PK variability in drugs with complex metabolic pathways.