1 Introduction
Despite the use of vancomycin in neonatal intensive care units to treat infections for more than four decades [
1], a lack of consensus remains on optimal dosing. Dosing decisions are particularly challenging in neonates, because patient characteristics such as age [gestation at birth and postnatal age (PNA)], body weight, and maturation of organ function contribute to rapid changes in vancomycin pharmacokinetics. Vancomycin is predominantly renally cleared, and its renal elimination depends mainly on glomerular filtration, which increases as a function of PNA as renal function matures [
2]. Although vancomycin pharmacokinetics have been extensively studied, empiric dosing based on either fixed, body weight, serum creatinine, and/or age-based dosing frequently fail to achieve the exposure targets [
3]. Hence, dosing decisions should be based on vancomycin concentrations, a practice known as therapeutic drug monitoring (TDM). Trough concentrations (
Ctrough) between 10 and 20 mg/L and the ratio of the area under the 24-hour concentration–time curve to the minimum inhibitory concentration (AUC
24/MIC) of 400–600 h have both been considered as exposure targets in adults [
4]. However, it remains unclear whether these exposure targets can be extrapolated to neonates. Previously, we have shown that vancomycin dose adjustments were required in 62% of courses of therapy of greater than 3 days [
5]. Hence, better vancomycin dosing strategies and more effective TDM practices are needed to achieve vancomycin exposure targets in neonates.
Model-informed precision dosing (MIPD) software, a computational tool, has grown to use mathematical models, patients information, and observed concentrations to optimize therapy [
6]. MIPD proved its benefit to predict the required dose to reach target attainment, and broad implementation is limited due to lack of understanding of its benefit and skills, costs, and regulatory issues [
7]. Such software has shown benefits in reducing incidence and/or rate of adverse effects and improving clinical outcomes. Also, it proved its benefits in reducing the costs of treating infections [
8].
To fully prevail these benefits, we have to select the best model to describe the local patient population, as an inaccurate prediction of drug exposure can lead to inappropriate dose recommendations [
9].
Therefore, the aim of our study is to identify vancomycin PopPK models in neonates from literature and evaluate the predictive performance of the selected models using data from our local neonatal intensive care unit (NICU) population. Literature indicated that pharmacokinetics (PKs) in both preterm and full-term neonates are quite similar, and there is no real difference between term and preterm models, as vancomycin PK models consist of various combinations of significant covariates, such as postnatal age, postmenstrual age, and weight, which indicates growth and maturation [
2]. Vancomycin AUC
24/MIC
, rather than trough concentration, is the pharmacokinetic parameter used to inform dosing decisions for vancomycin in adults [
10]. Recent studies have revealed that an AUC
24/MIC ratio 400–600 h is highly predictive of efficacy and toxicity of vancomycin in neonates [
10,
11]. However, AUC
24/MIC calculation is practically limited in individuals due to the need of collecting more than one trough concentration. In adults, a trough concentration range between 15 and 20 mg/L is expected to reach an AUC
24/MIC ratio > 400 h, but in the case of neonates, lower trough concentration ranges could be sufficient to reach an AUC
24/MIC > 400 h [
12]. As a secondary outcome, the correlation between AUC
24/MIC and a posteriori trough concentrations was evaluated using the best possible model.
4 Discussion
The predictive performance of nine PopPK models for vancomycin in neonates were evaluated in a cohort of 69 preterm neonates (Table
2). In our study, we informed on the most appropriate model to be used in dosing software to calculate vancomycin dosing in NICU. All the models used the trough concentration (one reading) to predict concentration, but no peak concentration was collected in our cohort; this was one of our study limitations. For neonates < 29 weeks gestational age, blood sampling was done before the second dose, which could not be extrapolated to steady state concentration. The De Cock model was the best performing model based on the lowest rBias, bias, and rRMSE using all approaches, and it shows minimal data spread around the linear regression between predicted and observed concentration. As shown in Fig.
2, RMSE values were low for the De Cock model in the first two approaches, and a priori and a posteriori approaches 8.9 and 6.02%, respectively; on the other hand, the RMSE value was not low in case of the third approach (38.3%). This can be explained by the fact that the maturation of renal function in neonates, which is highly related to PNA, is expected to affect volume status and clearance and, therefore, change vancomycin pharmacokinetic in the next course.
Table 2
The overall model bias and precision for Bayesian predicted vancomycin concentrations
| 0.848 | 39.6 (31.3, 47.9) | 62.9 | 0.804 | 37.4 (29.2, 45.6) | 58.2 | 0.557 | 59.2 (46.8, 71.6) | 55.9 |
| 0.566 | − 15.7 (−26.6, − 4.8) | 39.2 | 0.284 | − 12.3 (− 18.9, − 5.6) | 38.4 | 0.290 | 19.6 (5.6, 33.6) | 52.1 |
| 0.896 | 47.5 (40.3, 54.7) | 68.9 | 0.638 | 36.3 (30.2, 42.5) | 49.5 | 0.382 | 49 (− 48.2, 146.2) | 59.3 |
| 0.494 | 16.7 (8.9, 24.5) | 40.8 | 0.240 | 7.7 (2.9, 12.6) | 27.3 | 0.199 | 19.1 (9.5, 28.8) | 47.3 |
| 0.426 | 0.8 (− 7.5, 9.1) | 8.9 | 0.035 | − 0.23 (− 1.3, 0.88) | 6.02 | 0.172 | 5.5 (− 8.2, 19.1) | 38.3 |
| 0.796 | 39.6 (31.7, 47.4) | 62.9 | 0.289 | 16.8 (12.3, 21.4) | 30.1 | 0.299 | 39 (29.3, 48.7) | 51.3 |
| 1.285 | 64.5 (57.6, 71.4) | 80.3 | 0.762 | 41.4 (35.3, 47.4) | 52.9 | 0.358 | 42 (31.3, 52.7) | 55.9 |
| 1.135 | 56.1 (48.2, 64.0) | 47.9 | 0.544 | 34.5 (28.9, 40.1) | 45.8 | 0.456 | 54.8 (44.3, 65.3) | 65.6 |
Grimsley and Thomson [ 21] | 0.681 | 16.1 (5.7, 26.6) | 40.2 | 0.414 | 9.5 (1.7, 17.3) | 43.6 | 0.373 | 37.6 (24.8, 50.4) | 57.9 |
In Fig.
4, using a Bayesian method and trough only concentrations, observed concentrations are very close to predicted concentrations. Unfortunately, no peak concentration was collected in our cohort, and this was one of our study limitations. For neonates < 29 weeks gestational age, blood sampling was done before the second dose, which could not be extrapolated to steady state concentration. There were cases with more than one trough concentration in our cohort, ten babies had two trough samples, and three babies had three samples during their treatment course
. For patients with multiple courses of vancomycin, the accuracy and precision of all the model’s predictions did not improve when using vancomycin concentrations from a previous course of therapy. This highlights the importance of obtaining vancomycin concentration during the current course of therapy in preterm neonates and the use of it to guide dosing. This may in part reflect the rapid organ maturation in neonates [
44] and/or dynamic changes in physiology related to their critical/unstable conditions [
45]. Hence, timely collection of a new vancomycin concentration in the ongoing course will provide the most accurate and precise information on drug exposure and is best suited to guide the dose.
Neonates are a diverse group, ranging from extremely premature infants to full-term newborns, and factors such as gestational age, birth weight, ethnicity, and postnatal age can significantly affect the pharmacokinetics and pharmacodynamics of vancomycin. Herein, many factors could affect the process of model building and evaluation, such as study settings, population characteristics, sample size used to build the model, and analytical model. Race and ethnicity can affect drug response and metabolism, because it contributes pharmacokinetic of drugs such as renal secretion (clearance), hepatic metabolism, protein binding, and volume of distribution [
46]. A study in adult patients showed that a race-based based estimated glomerular filtration rate (eGFR) equation performed better to predict vancomycin clearance in Thai population [
47]. This has not yet been confirmed in neonates likely due to underpowered studies [
48]
. Therefore, model built using a wide range of neonatal populations may exhibit a great variability and could be generalized, as it is not restricted to one category of population. For example, our dataset had a diverse ethnicity, as the catchment area of our hospital has the largest ethnic diversity in Australia [
49]. As such, models built based on a specific ethnic group of population may not be adequate to predict PK parameters in a more general population, such as the model by Lo et al., which was built using data from Malaysian preterm neonates. Moreover, differences in study design could affect model building and evaluation process, such as retrospective versus prospective studies. Retrospective studies rely on medical records and may have limited control over data collection and patient management. Prospective studies, on the other hand, allow for standardized data collection, more rigorous monitoring, and controlled dosing regimens, reducing some sources of variability. Furthermore, models built using a smaller population size may not give a complete description of patient’s population pharmacokinetics of vancomycin. This may result in biased or imprecise Bayesian estimation of vancomycin exposure, such as what we noticed in the Mulubwa and Kimura models, which were built using only 19 neonates, which might be the reason that those two models did not perform well in our dataset. Variations in laboratory techniques, calibration standards, and quality control procedures can introduce variability in measured drug concentrations across studies, as such, affecting model building and evaluation.
There are multiple possible reasons why the De Cock model performed well with our data. First, vancomycin exposure after IV infusion is generally described by a two-compartment model [
9]. The original model by De Cock et al. is a two-compartment model, and it was based on a large dataset of 689 neonates, where all neonates were preterm with a birth weight range between 385 and 2550 g and a PNA of 1–28 days. In our cohort, all neonates were born prematurely (< 28 weeks gestation) and started a vancomycin course within a median (range) of 20 (1–91) days of their life. Comparable neonatal populations (i.e., gestational age and birth body weight) between the model building dataset of De Cock and our validation dataset might explain the good performance of the De Cock model compared with other models.
AUC
24/MIC is an important pharmacokinetic parameter used to evaluate the efficacy of vancomycin therapy in adults. For most infection caused by susceptible Gram-positive bacteria, a target AUC
24/MIC ratio of 400–600 h is considered optimal [
50]. In this study, an AUC
24/MIC ratio of 400–600 h correlated well with a trough concentration of 10–15 mg/L, with a high correlation coefficient close to one that is consistent with the high correlation factor found by other studies, which used a large number of neonates, such as Tseng et al. [
51]. It was found that the trough concentration does not have to be higher than 15 mg/L in neonates to achieve AUC
24/MIC ratio > 400 h [
52]. Also a recent study revealed a significant correlation between trough concentration and AUC
24/MIC at different dosing and dosing intervals [
11]. The PMA of our cohort ranged from 25 to 42 weeks, with a weight of 495–2142 g, which can be considered heterogeneous. The vancomycin dosing regimen of our hospital [
3] includes dosing according to PMA and PNA, bodyweight, and renal function, which likely corrected for this heterogeneity, as shown in Fig.
6.
Our study also has some limitations. We applied a predefined criterion developed with the aim to select a model suitable for clinical practice in our patient population. Use of this criterion may have resulted in the exclusion of models that could have performed well. This was seen earlier in a study by Colin et al. [
53], which included a greater age range than considered appropriate for our intended population. In our study, we did not exclude outliers, as bias due to outliers may have impacted model performance. However, in our opinion, the current number of patients in the validation dataset was fit for the purpose in selecting the best model for the local patient population suitable for clinical implementation. Models used for extracorporeal membrane oxygenation (ECMO) patients and positive ventilation pressure were excluded, as it was not representative of our cohort.