The importance of model selection for a priori model informed precision dosing of vancomycin
Sebastiaan D.T. Sassen
Erasmus MC
Objectives:
Vancomycin is a commonly used antibiotic for gram positive infections at the Intensive Care unit. However, in 2020 only 16% of the patients at the ICU in the Erasmus MC reached the intended target of 20-25 mg/L at around 24 hours after the start of therapy, which consisted of a bolus dose followed by a continuous infusion of 3 g/day. Standard dose was adjusted in case of impaired renal function. To improve target attainment, pharmacokinetic (PK) models can be used via dose individualization and optimization. This method has shown to perform well in therapeutic drug monitoring using models and Bayesian forecasting. However, the latter requires plasma concentrations which are not available at initiation of the treatment. Many pharmacokinetic models of vancomycin are available to estimate the optimal starting dose, however, not every model performs equally well in the target population. Using the right PK model is of utmost importance, especially because a priori use of PK models lacks plasma concentrations and is therefore solely based on patients’ characteristics to tailor the models.
The aim of this study was to compare different methods of PK model selection to determine how to achieve the best a priori dose predictions for each individual patient to increase the number of patients within the target range at the start of the treatment.
Methods:
A retrospective analysis was performed in 100 patients at the Intensive Care Unit (n=90) and Orthopedics department (n=10) at Erasmus Medical Center in Rotterdam the Netherlands. All patients received vancomycin as a continuous infusion preceded by a bolus infusion. Patient on continuous renal replacement therapy and extracorporeal membrane oxygenation were excluded from the analysis. A total of 28 vancomycin PK models were replicated in Python (v. 3.7) and validated using NONMEM (v. 7.5). Model predictions were compared to the observed concentrations for each patient and each model. Dose simulations were performed for all patients and models to determine the optimal dose targeted at 22.5 mg/L at time of the first sampling (±24 hours).
The dosages from the dose simulations were used to test different dose selection methods. 1. Equally weighing of all models; 2. Using only one model with the lowest bias; 3. Excluding models based on quartiles of covariates where >75% of the patients were either above or below the target; 4. Weighing of models based on model performance; 5. Model selection based on covariates using machine learning decision trees (using R and rpart).
To compare the methods, the ratio between the administered dose and newly calculated dose was used to recalculate the observed concentration into a new observation. This approximates what the concentration would be with the new dose, assuming a linear relation between dose and concentration as is expected in first order kinetics.
Results
The average (range) of predicted over observed concentration for all models was 93.0% (56.3-135.7%) and the average standard error (range) was 38.3% (33.9-51.5%). The model that performed the worst had a predicted over observed concentration of 56.3% which means a severe underprediction that could lead to unjustly high dose recommendations. The percentage of patients within the target range were 31.8%, 27.3%, 30.7%, 40.8% and 51.1% for respectively methods one through five. The method using decision trees performed best with an improvement from 31.8% to 51.1% of patients on target, followed by the weighing method which improved from 31.8% to 40.8%
Conclusion
This study shows the importance of model selection for a priori model informed precision dosing. The dose prediction using all models equally weighted did not perform better than no models. However, a major improvement in a priori dose prediction was observed after using model selection methods like, performance-based model weighing and decision trees model selection. This endorses the importance of external validation and the importance of model selection prior to clinical implementation.