Use of Machine Learning for Dosage Individualization of Vancomycin in Neonates

Area of Science:

• Pharmacology
• Neonatal Medicine
• Machine Learning Applications

Background:

• Neonatal vancomycin dosing exhibits high variability, necessitating individualized regimens.
• Achieving target trough concentrations (C0) and area-under-curve (AUC0-24) is crucial for effective vancomycin therapy.
• Optimizing vancomycin dosing in neonates requires advanced predictive strategies.

Purpose of the Study:

• To evaluate the efficacy of machine learning (ML) models in predicting vancomycin treatment targets (C0 and AUC0-24) in neonates.
• To assess if ML can facilitate the calculation of optimal individual vancomycin dosing regimens.
• To compare ML-based predictions against traditional pharmacokinetic models for neonatal vancomycin therapy.

Main Methods:

• Utilized a large neonatal vancomycin dataset to develop ML models for C0 and AUC0-24 prediction.
• Employed Bayesian post hoc estimation for individual AUC0-24 calculations.
• Validated predictive performance using an external dataset and various ML algorithms, including Catboost.

Main Results:

• The Catboost-based C0-ML model accurately predicted pre-treatment C0 using dosing regimen and covariates, improving prediction accuracy by 42.5% over population pharmacokinetic models.
• ML-optimized dosing achieved the pharmacodynamic target C0 in 80.3% of virtual neonates, significantly higher than standard doses.
• The Catboost-based AUC-ML model predicted AUC0-24 with 80.3% accuracy after obtaining initial C0 measurements.

Conclusions:

• Developed precise ML models for predicting C0 and AUC0-24 in neonatal vancomycin therapy.
• These models enable accurate individual vancomycin dose recommendations before treatment initiation.
• ML models facilitate dose adjustments after the first therapeutic drug monitoring (TDM) result, optimizing neonatal vancomycin treatment.