Population pharmacokinetic modelling for vancomycin using Bayesian model averaging approach
Daichi Yamaguchi (1, 2), Takayuki Katsube (1), Yoshifumi Nishi (2), Yasuhiro Tsuji (2)
(1) Clinical Pharmacology & Pharmacokinetics, Shionogi & Co., Ltd., (2) School of Pharmacy, Nihon University.
Introduction:
Vancomycin has been used as a first-line standard treatment against methicillin-resistant Staphylococcus aureus for over 60 years [1]. More than 60 studies of population pharmacokinetic (PK) modelling for vancomycin have been reported, but model structures and selected covariates were different from study to study [2,3]. It is known that model selection based on statistical criteria (e.g. objective function value and Bayesian information criterion) is dependent on data richness, such as the number of subjects and blood sampling time points. If the available data are limited, the models selection based on the statistical criteria may lead to models with incorrect inference. In the treatment with vancomycin which has a narrow therapeutic range, therapeutic drug monitoring and individualized dosing adjustment are recommended to maximize effectivities and minimize toxicities [1]. Some Bayesian dose-optimizing software which can estimate vancomycin exposure are available in clinical situation, but the accuracy of the prediction is dependent on one population PK model selected [4]. Bayesian model averaging (BMA) framework [5] could obtain posterior distribution of parameter estimation and posterior selection probability for each of multi-models. The inference with averaging multi-models based on posterior distributions and probabilities is generally more reliable [6]. The objective of this study was to apply BMA approach to the population PK analysis of vancomycin and develop averaged model for population PK of vancomycin.
Methods:
Data were collected from adult patients who measured at least one plasma vancomycin concentration in Kyorin University Hospital from January 2018 to December 2019. The intrinsic or extrinsic factors in the patients (e.g. age, sex, and renal function) were available. BMA was performed by Markov chain Monte Carlo (MCMC) Bayesian estimation algorithm implemented in NONMEM ver.7.4.4 [7]. A switch parameter, a variable (0 or 1) following a Bernoulli distribution, was adapted to each model structure and each covariate. We obtained 10,000 samples from two MCMC chains to make an inference about posterior distributions, and if the switch parameter was estimated as 1 in each sample, the model or covariate were selected in this sample. The posterior selection probability of model structure or covariate was calculated from the posterior distribution of the switch parameter. We tested two candidate model structures, one and two compartment models, and some covariates for each PK parameters as follows: age, sex, body weight (WT), albumin (ALB), aspartate aminotransferase (AST), and creatinine clearance (CrCL) for the clearance (CL), age, sex, serum creatine (SCr), and WT for the volume of distribution in central compartment (Vc), and WT for the inter-compartmental clearance (Q) and the volume of distribution in peripheral compartment (Vp). Prior information for these model structure and covariate selections were set the proportions of each model structure and covariate from 43 models reported for population PK modelling of vancomycin with adult patients.
Results:
Total 1626 observed plasma vancomycin concentrations from 453 patients aged 20 to 103 years old were available. Model structure was selected as two compartment distribution with first order elimination model from all samples (posterior selection probability = 100%). The population mean values of CL, Vc, Q, and Vp were estimated 3.51 L/hr, 110 L, 0.680 L/hr, and 274 L, respectively. CrCL was selected as a covariate for CL in a power function in all samples (100%). The posterior selection probabilities of ALB, WT, age, AST, and sex for CL were 97.7%, 19.0%, 8.9%, 3.5%, and 0.6%, respectively. Those of WT, age, SCr, and sex for Vc were 56.1%, 7.0%, 4.7%, and 2.6%, respectively. Those of WT for Q and Vp were 33.7% and 37.1%, respectively. The convergence of MCMC chains was confirmed by visual diagnosis and statistical methods [8,9].
Conclusions:
We performed population PK analysis using BMA approach and developed averaged model of vancomycin. Based on our averaged model, the influence of each covariate for PK parameters could be evaluated and vancomycin exposures could be estimated depend on various covariate models but not one model. This model averaging method may be applicable to other population PK analyses.
References:
[1] Rybak MJ et al. Pharmacotherapy. (2020) 40, 363-367.
[2] Aljutayli A et al. Clin Pharmacokinet. (2020) 59, 671-698.
[3] Chung E et al. Clin Pharmacokinet. (2021) 60, 985-1001.
[4] Turner RB et al. Pharmacotherapy. (2018) 38, 1174-1183.
[5] Hoeting JA et al. Stat Sci. (1999) 14, 382–417.
[6] Longford NT. J R Statist Soc A. (2005) 168, 469–472.
[7] Beal SL et al. 1989-2011. NONMEM Users Guides. Icon Development Solutions, Ellicott City, Maryland, USA.
[8] Gelman A et al. Statistical Science. (1992) 7, 457–472.
[9] Gelman A et al. Bayesian Data Analysis, Third Edition. Chapman and Hall/CRC. (2013)
