Mehdi El Hassani1, Daniel J.G. Thirion1,2, Amélie Marsot1
1Faculty of pharmacy, University of Montreal, 2Department of Pharmacy, McGill University Health Centre
Introduction: Optimizing drug dosing in vulnerable patient populations can be achieved through two main approaches: developing a population pharmacokinetic (PopPK) model directly from available patient data or externally evaluating existing published models. While model development ensures that the model reflects the target population, it is time consuming and may be constrained by data limitations, especially in geriatric populations where large datasets are difficult to collect. Alternatively, external evaluation of published models is more practical but may not provide optimal predictions if the model does not generalize well to the intended population (1). It remains unclear which approach is preferable in the context of small datasets, and despite the clinical relevance, very few PopPK models have been developed for piperacillin in geriatric patients. This study aims to investigate whether a small dataset can support the development of a fit-for-purpose PK model for initial dosing of piperacillin and to compare its recommendations with those of externally evaluated published models. Methods: Patient data were collected from a prospective clinical study in patients aged 75 years and older at the McGill University Health Center (MUHC), Montreal, Canada. Hospitalized patients receiving intermittent infusions of piperacillin/tazobactam at standard dosing were eligible for inclusion, while those in the intensive care unit were excluded. Blood samples were drawn 3 times once steady state was reached: 1) 30 min after the end of the infusion (peak), 2) in the second half of the dosing interval (middle), 3) before the start of the next infusion (trough). Blood samples were collected at steady state at three time points: (1) peak—30 minutes after the end of infusion, (2) middle—in the second half of the dosing interval, and (3) trough—immediately before the next infusion. Total plasma piperacillin concentrations were quantified using a previously developed and validated ultra-high-performance liquid chromatography method with diode-array detection (2). Population PK model development was conducted using nonlinear mixed effects modeling in NONMEM Version 7.6 (ICON Development Solutions, Ellicott City, MD). The first-order conditional estimation with eta-epsilon interaction was used for parameter estimation. A nonparametric bootstrap (n=1000) was used to evaluate the stability of the final model. Data visualization and statistical analyses were performed with R v3.6.1 (R Project for Statistical Computing) using the RStudio v1.3.1073 graphical user interface. Additionally, published population PK models were selected from a previously published review on piperacillin population PK models in the older population and were externally evaluated with NONMEM using the MUHC data (3). Predictive performance was assessed using goodness-of-fit plots, median prediction error (MDPE), median absolute prediction error (MDAPE), and prediction-corrected visual predictive checks. Monte Carlo simulations using virtual patients representative of the MUHC population were conducted to compare initial dosing recommendations obtained with the developed model versus published models. Tested piperacillin dosing regimens ranged from 2 to 4 g every 4 to 12 hours, administered via a 30-minute infusion. The a priori probability of target attainment (PTA), stratified by renal function, was assessed by comparing the proportion of time the free piperacillin concentration (assuming 30% protein binding (4)) remained above the minimum inhibitory concentration (MIC) for 100% of the dosing interval across all models. Results: The study included 13 patients (5 males, 8 females) with a median age of 89 years (76–93). Median weight was 63.3 kg (49.0–90.5 kg), height was 172 cm (148–178 cm), and BMI was 22.8 kg/m² (20.4–29.6 kg/m²). Most patients were White (9), while 2 were Asian, 1 Black, and 1 Indian. Median serum creatinine was 72.0 µmol/L (53.0–148), with creatinine clearance (CLCr) values of 44.0 mL/min (29.8–120.4) using Cockcroft-Gault and 75.0 mL/min/1.73m² (40.7–97.6) using CKD-EPI. In total, these patients contributed 28 piperacillin concentrations for the analysis. All patients were dosed at 4 g of piperacillin every 6 or 8 hours (30-min infusion). A one-compartment model with first-order elimination best described the observed data. Creatinine clearance, calculated using the CKD-EPI equation, was a significant covariate for total clearance (CL). In the final model, CL was estimated as 4.92 × (CLCr/75)^0.624 L/h, with a relative standard error (RSE) of 6% for base CL and 45% for the exponent. The volume of distribution was estimated at 13.3 L, with an RSE of 14%. Interindividual variability was estimated only for CL at 22% (with 31% shrinkage). Residual variability was modeled using a proportional error and was 32.3%. A total of two population PK models were identified from the literature (4, 5). The Hemmersbach-Miller et al. model (7) is a one-compartment model developed with total piperacillin concentrations with clearance of 4.09 L/h (14.4% interindividual variability, IIV) and volume of distribution of 31.8 L (25% IIV). Its residual variability is 35.8% (proportional) and 0.0488 mg/L (additive). The Ishihara et al. (6) model is a two-compartment model, developed using total piperacillin concentrations, with clearance of 4.58 L/h (26.6% IIV), central volume of distribution of 5.39 L (69% IIV), and peripheral volume of distribution of 6.96 L. Its residual variability is 3.04% (proportional) and 5 mg/L (additive). The Ishihara et al. model (5) performed well, with MDPE of 2.4% for population predictions and 0.5% for individual predictions, and MDAPE of 23.8% and 3.2% for population and individual predictions, respectively. The Hemmersbach-Miller et al. (7) model showed high bias and imprecision, with population MDPE of -37.8% and MDAPE of 43.2%, and individual MDPE of -21.4% and MDAPE of 31.3%. Monte Carlo simulations showed that the developed model provided optimal daily dosing recommendations consistent with Ishihara’s model, which demonstrated the best predictive performance. At a MIC of 2 mg/L, the developed model matched Ishihara’s recommended dose (6 g/day), whereas Hemmersbach-Miller suggested a higher dose (8 g/day). For a MIC of 4 mg/L, both the developed model and Ishihara recommended a daily dose of 12 g/day, which was double the 6 g/day recommended by Hemmersbach-Miller. At a MIC of 8 mg/L, this pattern persisted, with the developed and Ishihara models again recommending a higher daily dose (16 g/day) compared to Hemmersbach-Miller’s lower 8 g/day. Conclusions: This study shows that a fit-for-purpose popPK model can be developed using a small dataset of geriatric patients while achieving consistent initial dosing recommendations to an externally evaluated published model with acceptable predictive performance. Thus, when data are limited, both model development and evaluation approaches can be considered appropriate for guiding initial piperacillin dosing. However, reliance on a popPK model with poor predictive performance could lead to inappropriate dosing recommendations, as illustrated by the Hemmersbach-Miller model in this analysis.
1. El Hassani M, Marsot A. External Evaluation of Population Pharmacokinetic Models for Precision Dosing: Current State and Knowledge Gaps. Clin Pharmacokinet. 2023;62(4):533-40. 2. El-Haffaf I, El Hassani M, Marsot A. Determination of 31 Antimicrobials in Human Serum Using Ultra-High Performance Liquid Chromatography With Diode Array Detection for Application in Therapeutic Drug Monitoring. Ther Drug Monit. 2024. 3. Liu HX, Tang BH, van den Anker J, Hao GX, Zhao W, Zheng Y. Population pharmacokinetics of antibacterial agents in the older population: a literature review. Expert Rev Clin Pharmacol. 2024;17(1):19-31. 4. Hemmersbach-Miller M, Balevic SJ, Winokur PL, Landersdorfer CB, Gu K, Chan AW, et al. Population Pharmacokinetics of Piperacillin/Tazobactam Across the Adult Lifespan. Clinical Pharmacokinetics. 2023;62(1):127-39. 5. Ishihara N, Nishimura N, Ikawa K, Karino F, Miura K, Tamaki H, et al. Population Pharmacokinetic Modeling and Pharmacodynamic Target Attainment Simulation of Piperacillin/Tazobactam for Dosing Optimization in Late Elderly Patients with Pneumonia. Antibiotics (Basel). 2020;9(3).
Reference: PAGE 33 (2025) Abstr 11421 [www.page-meeting.org/?abstract=11421]
Poster: Methodology - Model Evaluation