Jakob Kolar 1, Mojca Kerec Kos 1, Mateja Črček 1, Štefan Grosek 2,3, Iztok Grabnar 1
1 University of Ljubljana, Faculty of Pharmacy (Ljubljana, Slovenia), 2 University Medical Centre Ljubljana, Department of Perinatology (Ljubljana, Slovenia), 3 University of Ljubljana, Faculty of Medicine (Ljubljana, Slovenia)
Objectives: Gentamicin is widely used to treat bacterial infections, including those in neonatal patients. Elevated trough concentrations frequently lead to nephrotoxicity and ototoxicity which may be prevented by model-informed precision dosing [1]. Selecting the most appropriate pharmacokinetic model for individualised dosing remains challenging. Model selection or averaging provides a solution by identifying the most suitable model or combination of models when at least one concentration measurement is available [2]. The objective of this study was to evaluate the model selection and averaging approach for gentamicin population pharmacokinetic (PopPK) models in neonates using retrospective routine clinical care data.
Methods: A PubMed search was conducted to identify gentamicin PopPK models for neonates, as well as for neonates and children. Demographic, clinical, and treatment data for the neonates included in the study were collected retrospectively from the University Clinical Centre Ljubljana database. Patients who received intravenous gentamicin infusion and had at least one gentamicin serum concentration measurement were selected. The selected models were evaluated in NONMEM® using published final or bootstrap parameter estimates. The first concentration measurement for each neonate was used to estimate individual parameter values (learning set), while the remaining measurements were used to evaluate the accuracy and precision of the predictions (testing set).
The objective function value (OFV) and squared prediction error (SSE) weighting schemes were applied to each neonate based on the first concentration measurement, as previously described [2]. The resulting weights were used for all predictions in neonates with multiple concentration measurements. The model selection algorithm (MSA) and the model averaging algorithm (MAA) were evaluated as previously described [2]. Weights below 0.1% were set to zero.
Accuracy and precision for each published model, MSA, and MAA were assessed using the median relative prediction error (MdPE) and median relative absolute prediction error (MdAPE) in both the testing and learning sets, using R software.
Results: Fourteen published gentamicin PopPK models were evaluated using 240 gentamicin serum concentration measurements from 167 neonates. A single concentration measurement was collected from 124 neonates, while 43 neonates contributed between two and six measurements.
The model by Germovsek et al. [3] demonstrated the best predictive performance among individual models, with MdPE values of -0.4% and -8.4%, and MdAPE values of 0.4% and 40.3% in the learning and testing datasets, respectively. Using the OFV weighting scheme, the MSA selected 13 models at least once, with individual selection frequencies ranging from 1 to 31 neonates. Under the SSE scheme, 9 models were selected at least once. Two models dominated the SSE weighting scheme selections: the Germovsek model (for 125 neonates) and the model by Gastine et al. [4] (for 34 neonates). All other models were selected for at most two neonates.
Individual model MdPE values ranged from -31.2% to 0.2% in the learning dataset and from -64.4% to 9.2% in the testing dataset. For the OFV weighting scheme, the MdPE values were -4.9% and -15.2% for the MSA, and -10.0% and -20.2% for the MAA in the learning and testing datasets, respectively. Using the SSE weighting scheme, MdPE values were -0.3% and -12.0% for the MSA, and -11.8% and 13.9% for the MAA in the learning and testing datasets, respectively.
Individual model MdAPE values ranged from 0.4% to 35.6% in the learning dataset and from 31.9% to 66.5% in the testing dataset. For the OFV weighting scheme, the MdAPE values were 5.9% and 34.2% for the MSA, and 10.6% and 40.7% for the MAA in the learning and testing datasets, respectively. For the SSE weighting scheme, the MdAPE values were 0.3% and 43.3% for the MSA, and 13.2% and 43.0% for the MAA in the learning and testing datasets, respectively.
Conclusions: MSA generally outperformed MAA. The precision of the SSE weighting scheme was better, but its accuracy was worse than that of the OFV weighting scheme. The SSE scheme favoured two of the fourteen evaluated models, whereas the OFV weighting scheme selected models more evenly.
References:
References:
[1] Tong DMH, Hughes JH, Keizer RJ. Evaluating and Improving Neonatal Gentamicin Pharmacokinetic Models Using Aggregated Routine Clinical Care Data. Pharmaceutics. 2022 Sep 30;14(10):2089
[2] Uster DW, Stocker SL, Carland JE, Brett J, Marriott DJE, Day RO, Wicha SG. A Model Averaging/Selection Approach Improves the Predictive Performance of Model-Informed Precision Dosing: Vancomycin as a Case Study. Clin Pharmacol Ther. 2021 Jan;109(1):175-183
[3] Germovsek E, Kent A, Metsvaht T, Lutsar I, Klein N, Turner MA, Sharland M, Nielsen EI, Heath PT, Standing JF. Development and Evaluation of a Gentamicin Pharmacokinetic Model That Facilitates Opportunistic Gentamicin Therapeutic Drug Monitoring in Neonates and Infants. Antimicrob Agents Chemother. 2016 Jul 22;60(8):4869-77
[4] Gastine S, Obiero C, Kane Z, Williams P, Readman J, Murunga S, Thitiri J, Ellis S, Correia E, Nyaoke B, Kipper K, van den Anker J, Sharland M, Berkley JA, Standing JF. Simultaneous pharmacokinetic/pharmacodynamic (PKPD) assessment of ampicillin and gentamicin in the treatment of neonatal sepsis. J Antimicrob Chemother. 2022 Feb 2;77(2):448-456
Reference: PAGE 34 (2026) Abstr 12103 [www.page-meeting.org/?abstract=12103]
Poster: Drug/Disease Modelling - Paediatrics