Laura B. Zwep 1, Tadao Tamura 2, Lounis Kaid 1,2, Sunny Shao 2, Arsenii Zats 2, Mirjam van de Velde 3, Aniek Uittenboogaard 3, Jugal Suthar 2, Coen van Hasselt 1, Alaric Taylor 2
1 Division of Systems Biomedicine and Pharmacology, Leiden Academic Centre for Drug Research, Leiden University (Leiden, The Netherlands), 2 Vesynta Ltd, Innovation Gateway, The London Cancer Hub (Cotswold Road, Sutton, London, United Kingdom), 3 Emma Children's Hospital, Pediatric Oncology, Amsterdam UMC (Amsterdam, The Netherlands)
Objectives:
Model-Informed Precision Dosing (MIPD) leverages population pharmacokinetic (PK) models to predict optimal dose adjustments that incorporate patient covariates to account for inter-individual variability in PK. When applying MIPD to a new patient, it is assumed that this patient is adequately represented by the population used to train the underlying PK model. While it is straightforward to evaluate if a patient’s covariates (e.g., age, weight, creatinine) fall within the marginal ranges of the training data, this approach fails to identify multivariate outliers: patients with covariate combinations unrepresented in the training data, even if individual values are within marginal ranges. This challenge is compounded in longitudinal care, where evolving covariates can render an initially appropriate model unsuitable over time. A method is therefore needed to assess whether the training covariate distribution of a PK model is representative of a new patient.
In the current study, we developed an innovative vine copula-based workflow aimed to: 1) establish a novel representation score (RS) framework to quantify patient-model consistency; 2) determine the stability and data requirements for the RS framework; and 3) evaluate the relationship between the RS and model predictive performance using a clinical vincristine PK model, demonstrating a critical step towards patient-centric, reliable MIPD across diverse clinical populations.
Methods:
We established a workflow to define the joint probability distribution of covariates from a specific training population. First, a vine copula is fitted to the patient covariates, capturing their dependence structure [1,2]. A reference probability density is then derived via Monte Carlo sampling of the fitted copula. For any new patient or treatment occasion, covariates are mapped to this joint density to calculate the RS, defined as the probability that the patient’s covariate set falls within the model’s training space. We evaluated this workflow in three domains:
1. Intuition: A low-dimensional simulation (two continuous, one categorical covariate) was conducted to visualize the RS’s ability to identify in- and outliers in mixed-data space.
2. Stability: We performed a resampling analysis on the NHANES database [3], training copulas on subsets of varying sizes (n=20 to n=2000) to determine the minimum sample size required for stable RS estimation.
3. Clinical validation: The framework was applied to real-world pediatric vincristine PK data [4]. A PK model was fit to a clinically distinct subset of the population and RS performance was assessed by comparing prediction residuals between patients classified as well- versus poorly-represented individuals (high versus low RS).
Results:
We demonstrate the newly developed RS can be directly interpreted as a probability, ranging from 0 to 1, with a value of 1 indicating that the patient has the most typical covariate profile, and a value of 0 indicating no overlap between the training population and the new patient. The intuition example confirmed that the RS clearly distinguished patients outside the training population from those within it, behaving reliably for larger datasets.
The stability analysis showed that stable RS estimation required approximately 40 patients for three covariates, demonstrating feasibility on smaller datasets. Required sample size increased substantially with covariate dimensionality.
In the clinical validation example, we compared a population with high RS values (RS_mean=0.51) to a population with a RS<0.01. The RS corresponded to higher median residual errors in the PK predictions from the estimated PK model (rRMSE_highRS=40, rRMSE_lowRS=80). Poorly-represented individuals yielded significantly lower PK predictive performance, confirming the link between multivariate representation and model suitability. Conclusions: We present a novel RS framework that quantifies how well a patient is represented in a PK model's training data, supporting model selection in clinical MIPD. We show that predictive performance is directly linked to a patient's multivariate covariate representation. As MIPD expands to diverse and underrepresented patient populations, including pediatric, elderly, and critically ill patients, ensuring model-patient compatibility becomes essential. Integrating this copula-based approach into MIPD software would give clinicians a transparent, patient-specific metric to guide model selection, flagging when a patient falls outside a model's training space, at any point during longitudinal care, enabling safer, more reliable dose individualization across heterogeneous populations. References: 1. Czado, C. & Nagler, T. Vine Copula Based Modeling. Annu. Rev. Stat. Its Appl. 9, 453–477 (2022). https://doi.org/10.1146/annurev-statistics-040220-101153 2. Zwep, L. B. et al. Virtual Patient Simulation Using Copula Modeling. Clin. Pharmacol. Ther. 115, 795–804 (2024). https://doi.org/10.1002/cpt.3099 3. Centers for Disease Control and Prevention (CDC), National Center for Health Statistics (NCHS). National Health and Nutrition Examination Survey Data. https://wwwn.cdc.gov/nchs/nhanes/Default.aspx 4. Uittenboogaard, A. et al. Vincristine exposure in Kenyan children with cancer: CHAPATI feasibility study. Pediatr. Blood Cancer 71, e31160 (2024). https://doi.org/10.1002/pbc.31160
Reference: PAGE 34 (2026) Abstr 12140 [www.page-meeting.org/?abstract=12140]
Poster: Oral: Methodology - New Tools