Lea Marie Schatz (1,6), Sebastian G. Wicha (2), Georg Hempel (1), Matthijs de Hoog (3), Karel Allegaert (4,5), Catherijne A.J. Knibbe (6,7), Swantje Völler (6)
(1) Department of Pharmaceutical and Medical Chemistry, Clinical Pharmacy, University of Muenster, Muenster, Germany. (2) Department of Clinical Pharmacy, Institute of Pharmacy, University of Hamburg, Germany. (3) Department of Pediatric Intensive Care, Erasmus MC - Sophia Children’s Hospital, Rotterdam, The Netherlands. (4) Department of Hospital Pharmacy, Erasmus University Medical Center, Rotterdam, The Netherlands. (5) Departments of Development and Regeneration and Pharmaceutical and Pharmacological Sciences, KU Leuven, Leuven, Belgium. (6) Systems Pharmacology and Pharmacy, Leiden Academic Centre for Drug Research, Leiden University, Leiden, The Netherlands. (7) Department of Clinical Pharmacy, St. Antonius Hospital, Nieuwegein, the Netherlands
Objectives:
Model-informed precision dosing (MIPD) using Bayesian forecasting is a useful tool for individualizing therapy in children and thus may reduce the risk of serious adverse effects such as nephrotoxicity of vancomycin [1]. Since previously published population pharmacokinetic (PopPK) models were not developed with the intention of being used for MIPD, but rather to describe the PK in a specific population [2], the purpose of this study is to determine whether a multi-model approach is preferable to a single model approach.
Methods:
Six PopPK models developed in distinct populations such as newborn and infants, covering the entire pediatric age range, obese or post-cardiac surgery children, children with renal dysfunction or different types of cancer were encoded in NONMEM®[3–8]. The predictive performance of these single models was compared to two previously developed multi model approaches [9] using a clinical dataset. The dataset consisted of 113 children selected from two previous publications [6, 10] based on age (28 days to 18 years of age) and the availability of at least two vancomycin plasma concentrations. First, population prediction of the models was obtained based on patients’ characteristics and dosing history. Second, to simulate a real-world scenario, using Bayesian forecasting, the first plasma concentration was integrated to predict the second. Lastly, the performance of two multi-model approaches was investigated. The algorithm calculates individual weights for each model using a predefined weighting scheme (e.g., squared prediction error) to quantify the model fit and either automatically selects the best fitting model (model selection-algorithm (MSA)) or calculates an average of the selected models based on the individual weights (model averaging algorithm (MAA)). For the evaluation, the relative Bias (rBias) for accuracy the relative root mean squared error (rRMSE) for precision and goodness-of- fit plots were used. Models were considered as clinically acceptable if the rBias is ± 20 % [9].
Results:
The models by Le et al. 2013 [7] and Moffet et al. 2019 [8] showed the best population prediction, both, graphically and statistically (rBias: 2.98 %, 5.76 %; rRMSE: 88.9 %, 123 %, respectively) of the selected models. All models, except Capparelli et al. 2001 [5] that was developed for neonates only were considered clinically acceptable based on the individual prediction for the second concentration when the first concentration was included. The most accurate model was the model by De Cock et al. 2014 [6] followed by Le et al. 2013 [7] and Moffet et al. 2019 (rBias: 2.99 %,5.15 %, -8.94 %, respectively) [8], all being similarly precise (rRMSE: 37-40 %). MSA and MAA performed similar with respect to accuracy (rBias < 4 %). However, the MAA was slightly more precise (rRMSE: 38,7 % vs. 46,8 %). For the MSA, the models with the best population prediction were selected more often (Le et al. 2013 [7] n=30 and Moffet et al.2019 [8] n= 45) than the model with most accurate forecast (De Cock et al. 2014 [6] n=9). The model with the worst forecast (Capparelli et al. 2001 [5]) was not selected based on its low individual calculated weight and thus has no (MSA) or only little (MAA) impact on the performance of the algorithms.
Conclusions:
The model by Le et al. 2013 [7] showed the best population prediction for this dataset and was therefore the most suitable among the chosen models to calculate the initial dose in children treated with vancomycin. Although the two multi-model approaches were not superior to the single-model approach with respect to the predictive performance, the use of the more precise MAA is still recommended because even if a model was implemented in the software that does not fit the current user population, here the model by Capparelli et al. 2001 [5], the risk of inadequate forecast is lower compared to the single-model approach due to the small influence of this model on the prediction.
Besides, it is noticeable that models with a good population prediction are selected more often in the MSA than those with the most accurate forecast. This could possibly be due to the variability within an individual in the dataset. For more certainty, it should be investigated whether the results remain the same when more samples are included. Since it was shown that the model explicitly developed for neonates is not suitable for children older than one month, it can be assumed that a separate model selection for neonates should be made.
References:
- [1] Han J et al. (2021) Ther Drug Monit 44(2):241-252
- [2] Wicha SG et al. (2021) Clin Pharmacol Ther 109:928–941
- [3] Guilhaumou R et al. (2016) Ther Drug Monit 38:559–566
- [4] Smit C et al. (2021) AAPS J 23:53
- [5] Capparelli EV et al. (2001) J Clin Pharmacol 41:927–934
- [6] De Cock RFW et al. (2014) Pharm Res 31:2643–2654
- [7] Moffett BS et al. (2019) J Pediatr Pharmacol Ther 24:107–116
- [8] Capparelli EV et al. (2001) J Clin Pharmacol 41:927–934
- [9] Uster D et al. (2022) Clin Pharmacol Ther 109:175–183
- [10] Janssen EJH et al. (2016) Antimicrob Agents Chemother 60:1013–1021
Reference: PAGE 30 (2022) Abstr 10028 [www.page-meeting.org/?abstract=10028]
Poster: Clinical Applications