Why Model Verification Matters in Clinical Decision Support
We previously wrote about how we select “fit for purpose” models from the literature for implementation in our InsightRX Nova model-informed precision dosing (MIPD) clinical decision support software. Once we’ve picked a model or two, we’re just getting started!
In this blog post, we address an essential question:
“How do we ensure that a published pharmacokinetic (PK) model is implemented accurately in the InsightRX Nova platform?”
Accurate model implementation is foundational to the trust clinicians place in our decision support system. We’ve designed a rigorous, multi-step verification process to mitigate the two most common risks:
- Misinterpretation of model details: Even well-described models may contain ambiguities or lack information on data transformations and scaling.
- Unexpected clinical behavior: Models may produce PK profiles inconsistent with published data or real-world outcomes due to misreported parameters or poor translatability.
Common Challenges When Implementing Published PK Models
First, although authors of pharmacokinetic studies try their best to clearly describe their final published models, sometimes certain characteristics can be interpreted in multiple ways. It’s also possible that even when reading carefully, an important aspect of a model is missed. When authors include their model code as part of their publication, the risk of misinterpretation is greatly reduced. However, data transformations important for model fidelity may only be described in the paper itself, and so misinterpretations are still possible. Second, models might not capture expected pharmacokinetic profiles, whether due to mis-reporting of model parameters or due to challenges translating a model from one software to another. Our model implementation and verification process has been carefully designed to mitigate these risks.
InsightRX’s Five-Step Model Verification Process
Step 1: Dual Review of Published Pharmacokinetic Models
Two data scientists independently review the publication, evaluating the core model structure, covariate relationships, variability terms, and other considerations related to model development and data handling.
Step 2: Dual Coding in NONMEM and PKPDsim for Cross-Validation
Our data scientists code the model into two formats: (1) PKPDsim - the InsightRX developed R package used within the Nova platform and (2) NONMEM - the pharmacometrics industry standard for model estimation. While the NONMEM script is not directly used within the Nova platform, it is important for us to corroborate that our platform aligns with the industry standard modeling software. If a NONMEM-version of the model has already been shared in the publication itself, we will use that and do not dual code the model ourselves.
Step 3: Resolving Discrepancies and Author Clarifications
Once the model has been coded up in PKPDsim and NONMEM, our data scientists compare each other’s implementation of the model. This is the first vital check and balance to ensure a model is accurately encoded. This check helps catch everything from clerical errors to differences in interpretation of the published model. If we cannot resolve an ambiguous aspect of the model, we reach out to the original model authors for clarification.
Step 4: Simulations to Confirm Clinical Plausibility
After we are confident the PKPDsim and NONMEM models have identical implementations, we next create and compare simulation-based plots to ensure similarity with either a figure as presented in the published article and/or with other models in the same drug, route, and population space. This is the second vital check that helps catch issues in parameter or observation scaling and other potential implementation errors. While we expect there to be differences between models, we don’t expect grossly different results so this is a key check to ensure this model will work as intended alongside other models present in a given module.
Step 5: Internal Numerical Equivalence Testing with Realistic Inputs
Finally, we verify the PKPDsim and NONMEM models are numerically equivalent using an internally developed tool that automates this comparison.
While there are no clear best practices for software like ours, other developers have since published very similar approaches for other precision dosing software. (Ravix 2025 ahead of publication, Le Louedec 2021) We input realistic drug administration records, patient covariates, and sampling schemes and output simulated observations and estimated model parameters for both versions of the model. This is the third vital check in the model verification process and helps to ensure the fine-grain details of the model are behaving as described in the model publication.
Due to small differences between ODE solvers and numerical optimizers, there are sometimes slight differences between parameter estimations and level prediction, so we set a tolerance for what is considered numerically equivalent. Additionally, internal verification helps to give a sense of the stability of parameter estimation based on realistic clinical input.
Sometimes we find that parameter estimates are sensitive to the choice of modeling software for limited sampling schemes in very flexible models. In this case, we might constrain variability on certain parameters, which we report to the user.
This diagram illustrates the five-step process InsightRX uses to ensure published pharmacokinetic models are implemented in the Nova platform. The process includes independent coding in PKPDsim and NONMEM, comparison and resolution of discrepancies, simulation-based clinical plausibility checks, and final internal verification for numerical equivalence.
How InsightRX Uses Real-World Data to Refine Model Performance
Once the model is in actual clinical use to help clinicians in their treatment of patients, we are eager to compare its predictive and clinical performance against other available models so our platform can continuously learn and improve. Models demonstrating high predictive utility may be promoted into InsightRX GEMINISM, our intelligent model selection algorithm available for some drug modules.
The End Result: A Transparent, Reproducible Model Verification Workflow
When models are added to InsightRX Nova, our verification process ensures that published models are replicated as described and clinical behavior matches real-world expectations, so that users can confidently apply precision dosing at the bedside. We believe that our rigorous verification isn’t just best practice, it’s essential for safe, scalable model-informed precision dosing.
