Why Model Selection Matters in Pediatric Oncology Vancomycin Dosing
As discussed in Part I of this series, selecting a suitably predictive pharmacokinetic (PK) model for model-informed precision dosing (MIPD) is essential when treating pediatric subpopulations with unique PK characteristics.
Pediatric oncology patients may exhibit augmented renal clearance (ARC), a phenomenon that increases vancomycin clearance and risks subtherapeutic vancomycin concentrations when standard dosing approaches are used. This raises a critical clinical question: should we use specialized PK models tailored to this group, or can well-specified general models perform just as well?
Evaluating Vancomycin PK Models in Pediatric Oncology Patients
At InsightRX we continuously review, validate, improve and compare PK models for use in InsightRX Nova to support safe, effective, and personalized dosing.
In this study, our Data Science team:
- Analyzed 371 pediatric oncology patients receiving vancomycin across 20 clinical sites (median age: 7.2 years; 82% with hematological malignancy)
- Evaluated and compared the predictive performance of several published population PK models, including:
- Five specialized models developed for pediatric oncology patients
- Two general population models that frequently perform well
- Used root mean square error (RMSE), mean percentage error (MPE), and accuracy as primary metrics
Model Highlight: The Colin 2019 Meta-Model
One of the general models we evaluated was the Colin 2019 model—our current default recommendation for all patients receiving vancomycin in InsightRX Nova. The Colin 2019 model is a meta-model developed using pooled data from 14 diverse populations, including adults with hematological malignancies. To account for significantly higher clearance in this subpopulation, this model incorporated a cancer status covariate on clearance (CL).
However, for convenience using the Colin 2019 model with non-oncology patients, we offer two versions of this model in our InsightRX Nova platform:
- Colin 2019: standard version, without the cancer status covariate
- Colin 2019 (Onc): includes the cancer status covariate
We were particularly interested in how these two versions of the model compared in pediatric oncology patients, given that the cancer status covariate was based on additional vancomycin clearance in adults, not children.
Key Findings: General Model Outperformed Specialized Ones
The standard Colin 2019 model without the cancer status covariate outperformed all other published models—both general and specialized—across:
- A priori predictions (population-based)
- A posteriori predictions (Bayesian-informed with TDM)
This suggests that a well-specified general model may be suitable for pediatric oncology patients over specialized models.
Figure 1: Predictive performance of published PK models in pediatric oncology patients dosed with vancomycin. Assessed using root mean square error (RMSE), mean percentage error (MPE), and prediction accuracy. Error bars represent the point and 95% confidence interval estimate for each model. Dotted lines denote clinically acceptable thresholds, and the solid line for MPE represents a target value of zero bias.
Refitting the Colin 2019 (Onc) Model: Any improvement?
Because the cancer status covariate in the Colin 2019 (Onc) model was based on additional vancomycin clearance in adults, we decided to refit this model to:
- Test whether model performance could be improved over the standard Colin 2019 model.
- Re-estimate the effect of cancer status on vancomycin clearance in children using the sample data for pediatric oncology patients and a set of matched non-oncology pediatric patients
After applying our continuous learning framework to refit the Colin 2019 (Onc) model, we found that:
- The refit model generally produced similar PK parameter estimates compared to the published values
- Cancer status was not a statistically significant covariate for vancomycin clearance
Together, these factors resulted in nearly identical predictive performance between the refit Colin 2019 (Onc) model and the standard Colin 2019 model without the cancer status covariate.
Figure 2: Predictive performance of refit Colin 2019 (Onc) Model in comparison to published models. Assessed using root mean square error (RMSE), mean percentage error (MPE), and prediction accuracy. Error bars represent the point and 95% confidence interval estimate for each model. Dotted lines denote clinically acceptable thresholds, and the solid line for MPE represents a target value of zero bias.
Clinical Insight: Evaluating PK Model Performance in Pediatric Subpopulations
This analysis reinforces the importance of external evaluation when selecting a suitable model for guiding vancomycin MIPD decisions:
- In Part I, we previously found that specialized population models delivered the most accurate vancomycin dosing recommendations in pediatric cardiovascular ICU (CVICU) patients
- In contrast, this analysis of pediatric oncology patients showed that a well-specified general model (Colin 2019) outperformed all specialized models, even those developed specifically for oncology, and that refitting the specialized Colin 2019 (Onc) model with real-world oncology data did not improve predictive performance
Key Takeaway: Select Models Based on Real-World Results
Both general pharmacokinetic models and subpopulation-specific models can support precision vancomycin dosing in pediatrics, but their effectiveness must be validated externally. Selection of the most suitable model should be based on empirical performance, not assumptions about patient group characteristics.
This fit-for-purpose evaluation approach helps ensure optimal dosing accuracy, improves clinical outcomes, and supports scalable precision dosing practices across diverse pediatric populations.
Recommendation for Practice
Based on this analysis, InsightRX recommends the Colin 2019 model (without the oncology covariate) for vancomycin dosing in pediatric oncology patients.
Want to Learn More?
Explore our methodology for model evaluation, re-fitting, and dosing recommendations for special populations. Read the full analysis below or contact us to see how InsightRX Nova supports pediatric precision dosing at your institution.
