Predicting Vancomycin Clearance in Neonates and Infants by Integrating Machine Learning and Metabolomics With...

Area of Science:

• Pharmacology
• Clinical Pharmacy
• Computational Biology

Background:

• Vancomycin pharmacokinetics in neonates and infants show significant variability, complicating therapeutic drug monitoring.
• Achieving target vancomycin exposure is challenging due to inter-individual differences in drug clearance.
• Patient-specific factors are crucial for optimizing vancomycin dosing in vulnerable pediatric populations.

Purpose of the Study:

• To investigate the impact of patient-specific covariates on vancomycin clearance in neonates and infants.
• To evaluate the predictive performance of machine learning (ML) methods for vancomycin clearance using clinical and metabolomics data.
• To identify key clinical and metabolomic predictors of vancomycin clearance.

Main Methods:

• Retrospective population pharmacokinetic (PK) analysis of 42 neonates and infants.
• Intravenous vancomycin administration with 214 concentration measurements analyzed.
• LC-MS/MS-based untargeted metabolomics assay on plasma samples.
• One-compartment PK model with first-order elimination, identifying significant covariates.
• Evaluation of various ML methods, including Gradient Boosting Regressor (GBR).

Main Results:

• A one-compartment model identified serum creatinine (SCr), postmenstrual age (PMA), and weight as significant covariates influencing vancomycin clearance.
• Gradient Boosting Regressor (GBR) demonstrated the highest predictive performance using clinical covariates (MSE: 0.0033; R²: 0.830).
• Metabolomics data did not significantly enhance predictive accuracy for most models, though some metabolites were top predictors. SCr and PMA were key predictors in both PK and ML models.

Conclusions:

• Ensemble ML methods, especially GBR, are valuable tools for predicting vancomycin clearance using readily available clinical covariates.
• Clinical factors like SCr and PMA are primary drivers of vancomycin clearance variability in neonates and infants.
• While metabolomics offered limited added value for clearance prediction, the integrated approach highlights potential for exploring complex drug PK.