Van Dong Nguyen 1,2,3, Mikhail-Paul Cardinal 1,2, Alexandre Rivard 2,4, François Bourdeau 2, Daniel Thirion 1,2, Amélie Marsot 1,3,5
1 Faculté de Pharmacie, Université de Montréal (Montreal, Canada), 2 Pharmacy Department, McGill University Health Centre (Montreal, Canada), 3 Laboratoire STP2, Université de Montréal (Montreal, Canada), 4 Faculté de médecine, Université de Montréal (Montreal, Canada), 5 Centre de recherche CHU Sainte-Justine (Montreal, Canada)
Introduction: Vancomycin remains one of the most extensively studied drugs in population pharmacokinetic (popPK) modeling, with a large number of models published in the literature for a wide range of general and specific populations and across diverse clinical contexts[1,2]. Despite this, empirical dosing in clinical practice remains far from optimal, as dosing approaches vary considerably among clinicians due to differences in clinical experience and reliance on specific models embedded in online, easily accessible clinical decision support tools[3,4] that have not been validated locally. Variability in therapeutic approaches may be further compounded by the existence of multiple competing pharmacokinetic–pharmacodynamic (PKPD) targets for vancomycin, with preferences and adoption varying widely across institutions and practitioners[5]. Recent work has explored whether clinical and demographic factors can be considered to guide pharmacokinetic (PK) model selection aligned with local population characteristics to facilitate clinical implementation[6,7]. The present analysis aims to identify demographic and clinical characteristics associated with an increased risk of suboptimal initial dosing when established or robustly validated PK models are applied, across the range of PKPD targets encountered in clinical practice.
Objectives:
• Compare predictive performance and initial dose adequacy across externally validated vancomycin PK models based on trough and AUC PKPD targets (10–15 mg/L and 400–600 mg·h/L)
• Compare demographic and clinical characteristics between correctly predicted and dosed patients (“adequate”) and those who were not (“inadequate”)
Methods: Models were identified through online “vancomycin calculators”[3,4] and by reviewing external evaluation studies[6-10]. An existing local cohort of patients with vancomycin TDM data from 2023-2024 was used to develop a validation dataset. Patients were carefully selected from various clinical units, including the critical care, hemato-oncology and specialized surgical wards, as well as from specific demographic groups, such as obese patients (BMI ≥ 30 kg/m²) and elderly patients (> 65 years) to ensure diversity in demographic and clinical characteristics. Only steady-state trough levels (≤ 1 hour before the next dose) were included, with a single measurement per patient. Thirty-two demographic and clinical covariates were selected based on their reported effects on vancomycin pharmacokinetics[1] and/or their clinical relevance. Predictive performance and model adequacy were evaluated against four pre-defined criteria:
• F2.5 : absolute prediction error (PE) ≤2.5 mg/L for trough predictions
• F100 : absolute PE ≤100 mg·h/L for AUC predictions
• Trough-target-attainment : model-recommended dosing regimen achieving steady-state trough of 10–15 mg/L
• AUC-target-attainment : model-recommended daily doses achieving AUC of 400–600 mg·h/L
PE was defined as the difference between model-predicted and reference values (trough or AUC). Observed concentrations were used as reference values for trough levels, whereas AUC reference values were derived from a fit-for-purpose two-compartment popPK model developed from the local cohort. For each predefined criterion, patients were classified as adequate or inadequate for each model. Standardized differences (SDs) were calculated for each covariate to assess imbalance between groups, with an absolute SD > 0.2 indicating meaningful imbalance[11,12]. SDs were weighted according to the performance of models for each criterion, defined by the proportion of patients meeting the criterion. The four best-performing models were retained for weighting, reflecting the variety of models most likely to be used in local practice based on their performance. Analyses were performed on NONMEM (v.7.5.1) and R software (v.4.5.2).
Results: Eight models were retained to assess predictive performance and adequacy of dose recommendations: two one-compartment PK models developed through linear regression[13,14], two one-compartment popPK models[15,16] and four two-compartment popPK models[17-20]. The validation cohort comprised 205 patients with mean age of 60 years, mean weight of 76.1 kg, mean eGFR of 96.6 mL/min and 35% female. The proportion of patients meeting each criterion varied substantially across models: F2.5 (2.9–32.7%), F100 (19.5–45.9%), Trough-target-attainment (0.5–36.6%) and AUC-target-attainment (7.3–44.4%). Covariate analysis revealed several significant covariate imbalances between groups under the F2.5 criterion, whereas no significant imbalances were observed for AUC-target-attainment. Age and presence of neurocognitive disorders were the covariates most consistently associated with group differences across multiple criteria.
Conclusion: Predictive performance and adequacy of model-informed initial dosing of commonly used or externally validated models remain limited in our local population. Future efforts should prioritize the refinement of vancomycin popPK models and model-informed dosing strategies in elderly and cognitively impaired patients, with greater consideration of age-related covariates to optimize dosing precision in this vulnerable population.
References:
1. Aljutayli A et al. Clin Pharmacokinet. 2020;59(6):671-698.
2. Marsot A et al. Clin Pharmacokinet. 2012;51(1):1-13.
3. https://www.vancocalc.com/
4. https://clincalc.com/Vancomycin/
5. Claeys KC et al. Can J Hosp Pharm. 2020;73(3):232-237.
6. Agema BC et al. Clin Pharmacokinet. 2024;63(10):1449-1461.
7. Gandia P et al. CPT:PSP. 2025 Jan;14(1):142-151. 26.
8. Duong A et al. Pharmacotherapy. 2024;44(6):425-434.
9. Heus A et al. Int J Antimicrob Agents. 2022;59(5):106579.
10. Lee S et al. Pharmaceutics. 2022;14(5):1023.
11. Austin, P. C. Communications in Statistics – Simulation and Computation, 38(6), 1228–1234.
12. Cohen J. Statistical Power Analysis for the Behavioral Sciences. 2nd edn. Hillsdale, NJ; 1988.
13. Matzke GR et al. Antimicrob Agents Chemother. 1984;25(4):433-7.
14. Bauer LA. Applied Clinical Pharmacokinetics, 3rd Ed. New York: McGraw Hill, 2014.
15. Buelga DS et al. Antimicrob Agents Chemother. 2005;49(12):4934-4941.
16. Medellín-Garibay SE et al. Antimicrob Agents Chemother. 2017;61(12):e01249-17.
17. Goti V et al. Ther Drug Monit. 2018;40(2):212-221.
18. Lim HS et al. J Clin Pharm Ther. 2014;39(2):196-203.
19. Okada A et al. J Clin Pharmacol. 2018;58(9):1140-1149.
20. Yamamoto M et al. J Clin Pharm Ther. 2009;34(4):473-483.
Reference: PAGE 34 (2026) Abstr 11904 [www.page-meeting.org/?abstract=11904]
Poster: Clinical Applications