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

Individualized dosing of vancomycin is recommended, model-informed precision dosing (MIPD) being the preferred method to improve efficacy and limit toxicity. However, its implementation poses challenges, including model selection and initiation dose determination. We developed a model selection tool (MST) and evaluated its potential to improve concentration prediction precision and reduce bias. Retrospective data from adult intensive care unit patients receiving intravenous vancomycin were collected and divided into training and validation data sets. Population predictions from published one-compartment models were computed, and the universally best-performing model (UBM) was selected. A genetic algorithm was used to create an MST. The ability to forecast the third concentration based on previous concentrations was evaluated. A total of 148 vancomycin treatment episodes were included in training and 67 in the validation data set. The MST showed 12% and 6% improved precision compared to the UBM in training and validation data sets, respectively (mean absolute percentage prediction error [mean PAPE] 22.8% vs 26.0% and 28.4% vs 30.2%). The UBM exhibited lower bias in both training and validation data sets (mean percentage prediction error [mean PPE] 5.8% vs 4.7% and -2.8% vs -1.5%, respectively). The MST showed improved performance in predicting the third concentration based on previous concentrations. In both data sets, accuracy was the best/highest when two prior measured concentrations were used (mean PAPE and PPE 17.0% and -3.0% in training and 18.9% and -1.0% in validation data set). Overall, the MST has the potential to enhance vancomycin dosing accuracy from the first dose and simplify model selection, facilitating the utilization of MIPD in clinical practice.