Subgroup-based model selection to improve the prediction of vancomycin concentrations.

Area of Science:

• Pharmacokinetics and Pharmacodynamics
• Clinical Pharmacy
• Computational Biology

Background:

• Individualized vancomycin dosing is crucial for efficacy and safety.
• Model-informed precision dosing (MIPD) is preferred but faces challenges in model selection and initial dose determination.
• Accurate vancomycin concentration prediction is essential for optimizing patient treatment.

Purpose of the Study:

• To develop and evaluate a model selection tool (MST) for vancomycin dosing.
• To assess the MST's ability to improve concentration prediction precision and reduce bias compared to a universally best-performing model (UBM).
• To determine the optimal number of prior concentrations for accurate vancomycin level forecasting.

Main Methods:

• Retrospective analysis of vancomycin treatment data from adult ICU patients.
• Development of an MST using a genetic algorithm.
• Comparison of MST performance against a UBM using training and validation datasets.
• Evaluation of forecasting accuracy based on previous vancomycin concentrations.

Main Results:

• The MST demonstrated improved precision over the UBM, with lower mean absolute percentage prediction errors (PAPE) in both training (22.8% vs 26.0%) and validation (28.4% vs 30.2%) datasets.
• The UBM showed lower bias, with mean percentage prediction errors (PPE) of 5.8% (training) and -2.8% (validation).
• Predicting the third vancomycin concentration using two prior measured concentrations yielded the highest accuracy (mean PAPE 17.0% training, 18.9% validation).

Conclusions:

• The developed MST can enhance vancomycin dosing accuracy from the initial dose.
• The MST simplifies the model selection process, facilitating broader adoption of MIPD.
• Improved vancomycin dosing strategies can lead to better patient outcomes and reduced toxicity.