Multiple Model Optimal Sampling Promotes Accurate Vancomycin Area-Under-the-Curve Estimation Using a Single Sample in Critically Ill Children

Abstract

Background:

Area-under-the-curve (AUC)-directed vancomycin therapy is recommended; however, AUC estimation in critically ill children is difficult owing to the need for multiple samples and lack of informative models.

Methods:

The authors prospectively enrolled critically ill children receiving intravenous (IV) vancomycin for suspected infection and evaluated the accuracy of Bayesian estimation of AUC from a single, optimally timed sample. During the dosing interval, when clinical therapeutic drug monitoring was performed, an optimally timed sample was collected, which was determined for each subject using an established population pharmacokinetic model and the multiple model optimal (MMopt) function of Pmetrics, a nonparametric population pharmacokinetic modeling software. The model was embedded in InsightRx NOVA (InsightRx, Inc.) for individual Bayesian estimation of AUC using the optimal sample versus all available samples (optimally timed sample + clinical samples).

Results:

Eighteen children were included. The optimal sampling time to inform Bayesian estimation of vancomycin AUC was highly variable, with trough samples being optimally informative in 32% of children. Optimal samples were collected by clinical nurses within 15 min of the goal time in 14/18 participants (78%). Compared to all samples, Bayesian AUC estimation with optimal samples had a mean bias of 0.4% (+/− 5.9%) and mean imprecision of 4.6% (+/− 3.6%). Bias of optimal sampling was <10% for 17 of the 18 participants (94%). When estimating AUC using only a peak sample (≤2 h after dose) or only a trough (≤30 min before next dose), bias was <10% for 78% and 86% of participants, respectively.

Conclusions:

Optimal sampling supports accurate Bayesian estimation of vancomycin AUC from a single plasma sample in critically ill children.

Introduction

Vancomycin is commonly used to treat serious methicillin-resistant Staphylococcus aureus (MRSA) infections in critically ill children.¹ Vancomycin’s efficacy and toxicity are both linked to 24-hour area under the curve (AUC₂₄)^2–6 and AUC-directed dosing is recommended for both children and adults with serious MRSA infections.⁷ Although AUC₂₄ monitoring is being more commonly adopted clinically, it remains challenging in children, often because of the need to collect multiple blood samples to estimate AUC₂₄ during a single dosing interval. To maximize efficacy and minimize toxicity, reliable methods for estimating vancomycin AUC₂₄ in critically ill children using limited samples are needed.

One approach to vancomycin therapeutic drug monitoring (TDM) in children is to calculate the AUC₂₄ using two measured concentrations via the Sawchuk–Zaske method.⁸ Limitations of this approach are as follows: a) collection of multiple samples can be uncomfortable for children; b) obtaining multiple samples at specific times can be difficult for clinical teams; c) mistimed samples may be uninformative; d) samples must be collected at steady-state, which is not always possible in critically ill patients with dynamic physiology. Although this approach can facilitate reasonably accurate AUC estimation⁹ its multiple limitations make it less than ideal for critically ill children.

Bayesian dose adaptation is an alternative approach for determining the AUC₂₄ that can provide robust data on drug exposure in individual patients.¹⁰ By using population pharmacokinetics (popPK) models as prior information (Bayesian prior), Bayesian approaches incorporate a patient’s characteristics (e.g., age, sex, renal function), dosing history, and drug concentrations to estimate patient-specific PK parameters (Bayesian posterior) that accurately reflect drug behavior in that patient.¹¹ Importantly, Bayesian methods do not require waiting for steady-state and, depending on the robustness of the Bayesian prior, the AUC₂₄ can be estimated from as little as a single drug concentration. However, to the best of our knowledge, no pediatric studies have demonstrated sufficient accuracy of Bayesian estimation of vancomycin AUC₂₄ in critically ill children using a single sample.

This study aimed to evaluate the ability of a popPK model for intravenous (IV) vancomycin in critically ill children to inform Bayesian estimation of AUC₂₄ from a single, optimally timed sample. This study also sought to assess the feasibility of Bayesian AUC estimation compared to clinical AUC calculations using the Sawchuk-Zaske method.⁸

Materials and Methods

STUDY POPULATION

We performed a prospective observational study in the Pediatric Intensive Care Unit (PICU) of the Children’s Hospital of Philadelphia (CHOP) from November 2022 to November 2023. Patients aged 1–17 years old who received intermittent IV vancomycin dosing for any indication were eligible for inclusion. The patients were identified as soon after initiation of vancomycin as possible. Patients who underwent renal replacement therapy, plasmapheresis, or extracorporeal membrane oxygenation were excluded. At our institution, clinical pharmacists in the PICU deem children as candidates for TDM, as well as when it is performed. Children were required to be candidates for vancomycin AUC₂₄ monitoring per clinical pharmacists to be enrolled in our study.

The study protocol was approved by the CHOP Institutional Review Board (IRB). Documented informed consent and assent were obtained from the participants and/or parents/legal guardians, as appropriate. This study was registered at clinicaltrials.gov (NCT05691309).

VANCOMYCIN DOSING AND SAMPLING

Vancomycin was ordered for clinical care of all participants, and dosages and infusion rates were determined by the clinical team. All decisions regarding the vancomycin therapy duration were also at the discretion of the clinical team. Typical initial dosages were 10–15 mg/kg/dose every 6–8 h, depending on age, weight, and estimated renal function, which was calculated using serum creatinine (SCr) and the bedside Schwartz equation.¹²

Clinical AUC monitoring included the collection of 2 blood samples at 60–120 min after the end of the infusion (i.e. peak) and at ≤30 min before the next dose (trough, C_min); the trough could be collected before the dose to reduce the time between samples. Vancomycin AUC₂₄ was calculated clinically using the Sawchuk–Zaske method and doses adjusted to achieve an AUC₂₄ of 400–600 mg-h/L. AUC monitoring took place at steady-state, defined as ≥4 doses of a specific dosing regimen.

Within the same dosing interval(s) used for clinical AUC monitoring, we collected a single optimally timed vancomycin concentration for Bayesian AUC estimation for this study. The timing of optimal sample collection was determined using the multiple model optimal (MMopt) function of the Pmetrics package (version 1.9.7; Laboratory of Applied Pharmacokinetics and Bioinformatics, Los Angeles, CA)¹³ for R (version 3.6.3; R Foundation for Statistical Computing, Vienna, Austria).¹⁴ We aimed to collect the research samples within 15 min of the optimal sampling time. All samples were collected by bedside nurses.

Both clinical and research vancomycin concentrations were measured using a chemiluminescent microparticle immunoassay (Abbott Diagnostics) at the CHOP Chemistry Laboratory. The lower limit of quantification (LLOQ) of the assay was 3.0 μg/mL. The research concentrations were not made available in the patients’ medical records to avoid misinterpretation by the clinical team.

BAYESIAN PRIOR (POPULATION PK MODEL)

A previously published model of vancomycin in critically ill children¹⁵ was revised by combining model training (n=30) and testing cohorts (n=20), and then refitted to serve as the Bayesian prior for this study. The details of model development are provided in Appendix (A.1). Briefly, nonparametric population PK modeling was performed using the Pmetrics package for R in RStudio (v1.2.5033; RStudio, Inc., Boston, MA, USA).¹⁶ A 2-compartment model was constructed with allometric scaling on all parameters (a power of 0.75 and 1 for clearance and volume terms, respectively). Cystatin C (CysC)-based glomerular filtration rate (GFR) using the Hoek equation¹⁷ and age were included as covariates on vancomycin clearance. The final parameter estimates of the popPK model are shown in A.2; observed versus population and individual predicted concentration plots are shown in A.3.

This popPK model served as the basis for determining the optimal sampling times and for Bayesian estimation in the current study. The model was embedded within the BestDose package of the InsightRX Nova platform (InsightRx, Inc., San Francisco, CA, 2021) for Bayesian estimation.¹⁸

DETERMINATION OF OPTIMAL SAMPLING TIMES

Based on the anticipated time of clinical AUC monitoring for a given subject, we constructed a data file containing all available dosing information and relevant covariate information (CysC-based GFR, age in years). If a participant did not have CysC measured clinically, a blood sample was collected upon enrollment and then daily. CysC was measured by particle-enhanced turbidimetric immunoassay (PETIA) on the VITROS 4600 Chemistry System (QuidelOrtho Corporation, San Diego, CA) in the CHOP Chemistry Laboratory.

For each enrolled subject, we utilized the MMopt algorithm in Pmetrics and our popPK model to determine the optimal sampling time for estimating AUC₂₄. MMopt finds the optimal time based on when all the PK curves generated by the support points in the nonparametric model (i.e., our final model) are most separated. Thus, MMopt determines the most informative time point for minimizing the Bayesian risk of misclassifying an individual with an incorrect set of support points.¹⁹ As a simple example, if possible PK curves in the model have different peak concentrations but similar trough concentrations and, therefore, different AUCs, sampling at the trough time where the curves cluster could result in the wrong PK curve emerging as the most likely, with an incorrectly estimated AUC₂₄. Sampling at an earlier time, where the curves are more distinct, increases the probability of the most representative curve(s) emerging with the highest probabilities. We set the MMopt algorithm to search 30-min intervals from the end of infusion through the anticipated start time of the subsequent dose.

BAYESIAN ESTIMATION OF VANCOMYCIN AUC₂₄

For each subject, de-identified patient, covariate, and dosing data were entered into the InsightRX NOVA Bayesian dosing platform.¹⁸ We included dosing information from the start of the individual’s course through the end of TDM sampling. Bayesian posterior estimates were generated for each individual using a) the optimally timed vancomycin concentration only (optimal sampling) and b) all available vancomycin concentrations (full sampling). Individual fits were initially generated without nonparametric adaptive grid (NPAG) cycles²⁰ allowing for similar weighting of Bayesian prior information and measured concentrations. If the measured concentration(s) did not fit our model well, we implemented medium NPAG cycles, which weighted the measured concentration (s) more heavily. We recorded the estimated AUC₂₄ and C_min based on the dosing interval within which the optimal sample was collected, as well as the estimated PK parameters for that subject (clearance, V1, V2). Figure 1 shows the workflow of Bayesian estimation in this study.

Figure 1. Study workflow.

Cystatin C was measured upon enrollment, if not measured within the preceding 24 h for clinical care, and then each morning until vancomycin therapeutic drug monitoring was performed. Figure created using Biorender.com.

To understand the specificity of optimal sampling using our popPK model we repeated the above Bayesian estimation steps in InsightRx NOVA using only the trough concentration (≤1 h before next dose) and only the peak concentration (≤2 h after the end of an infusion) for each subject. Individuals without available trough concentrations were excluded from the sub-analysis.

STATISTICAL ANALYSES

The primary outcome of interest in our analysis was the predictive performance of our model for AUC₂₄ estimation when fitting a single sample. In our previous work,¹⁵ we found that our population PK model supported accurate Bayesian estimation of vancomycin AUC₂₄ using limited sampling (n=2) compared to estimation from five samples. In the current study, we evaluated the relative bias and imprecision of AUC predictions using the optimal sample only compared with full-sample (≥2) predictions for each subject. Bias and imprecision were calculated as follows:

$$\text{Bias}=\left({\text{AUC}}_{\text{single}}–{\text{AUC}}_{\text{full}}\right)/{\text{AUC}}_{\text{full}}*100$$

$$\text{Imprecision}=\left|{\text{AUC}}_{\text{single}}–{\text{AUC}}_{\text{full}}\right|/{\text{AUC}}_{\text{full}}*100$$

where AUC_single is the Bayesian posterior AUC₂₄ after fitting only a single concentration (e.g., the optimally timed concentration), and AUC_full is the Bayesian posterior AUC₂₄ after fitting all available samples. The data were then summarized as median bias and imprecision across the population. Bias and imprecision were also calculated and summarized using trough- and peak-only samples (versus all samples). Additionally, we determined the fraction of subjects whose AUC_single was within 5%, 10%, 15%, and 20% of AUC_full with each approach.

Finally, we calculated the AUC₂₄ using the Sawchuk–Zaske method for each subject (clinical AUC) and determined the PK parameter estimates (clearance and total volume) from these calculations. We compared the AUCs and PK parameters determined by each method using Pearson’s correlations and one-way analysis of variance (ANOVA) tests. We also determined the proportion of individuals whose AUC₂₄ was within the recommended therapeutic range of 400–600 mg*h/L using each AUC estimation approach.

Results

STUDY POPULATION AND SAMPLING

Overall, 29 patients provided consent to participate; seven were withdrawn or excluded due to discharge from the ICU or discontinuation of vancomycin by the clinical team before the research procedures. Twenty-two subjects underwent determination of the optimal sampling time. The distribution of the optimal sampling times among the 22 participants is shown in Figure 2. Children whose optimal sampling time was a trough had lower CysC-based GFR (median 79 mL/min/1.73 m²) than those with peak (median 137 mL/min/1.73 m²) or other (median 134 mL/min/1.73 m²) optimal times (ANOVA, p=0.03); sampling times were not associated with age or weight. Among 18 participants who underwent optimal sampling time determination and were receiving IV vancomycin doses every 6 h as a 1-h infusion, the median optimal sampling time for AUC estimation was 2.25 h after the end of the infusion.

Figure 2. Distribution of optimal sampling times.

Solid outlined boxes represent individuals receiving IV vancomycin every 6 h; hashed outlined boxes reflect those receiving vancomycin every 8 h. Peak times: 0–2 h after the end of the infusion. Trough times: ≤1 h prior to next dose. Graph created using Biorender.com.

Of the 22 participants, 2 were withdrawn from the study after determining the optimal sampling time because vancomycin was discontinued by the clinical team before sampling. Two others were deemed unevaluable after the collection of vancomycin samples and excluded from the AUC estimation steps: one had clinical TDM samples that were mislabeled and the sampling times were indeterminate, while the other had clinical TDM samples drawn from a venous catheter that was used to infuse vancomycin and the samples were contaminated (i.e., results uninterpretable). Thus, 18 participants underwent optimal sampling and AUC TDM and were deemed fully evaluable for our study. Of these, 8 (44%) had optimal sampling times that were peaks (≤2 h after dose), 6 (33%) were troughs (≥1 h prior to the next anticipated dose), and 4 (22%) were other times.

The characteristics of the study population that underwent AUC estimation are shown in Table 1. Most individuals (n=12, 67%) had three samples collected during the study (peak, trough, optimal sample), while six (33%) had a peak or trough that aligned with the optimal sample and only had 2 samples collected; no vancomycin concentrations were below quantification. The median difference between the optimal sample collection and the goal time was -4 min (IQR: -8, +8). The sample obtained for optimal sampling was >15 min from the goal time in 4 (22%) participants. These 4 samples were collected +41, +28, -16, and -73 minutes from the goal time.

Table 1.

Study population characteristics.

Characteristic	Data
Age in years, median (range)	6.1 (1.1–15.3)
Weight in kg, median (range)	20 (9.5–73.2)
Female sex, n (%)	11 (61)
Indication for vancomycin, n (%)
Rule-out sepsis	10 (56)
Suspected/confirmed meningitis	3 (17)
Othera	5 (28)
Estimated GFRHoek in mL/min/1.73 m2, median (range)b	136 (40–178)
<60 mL/min/1.73 m2, n (%)	3 (17)
Vancomycin dose in mg/kg/dose, median (range)b	15.0 (10.0–20.0)
Duration of vancomycin therapy prior to TDM (in hours), median (range)b	37.3 (7.5–105.2)
≤48 h from vancomycin start, n (%)	10 (60)

^aOther indications for vancomycin included one case each of mastoiditis, fever with neutropenia, pneumonia, skin and soft tissue infection, and post-operative care.

^bAt the time of therapeutic drug monitoring/optimal sampling.

Abbreviations: GFR, glomerular filtration rate; IQR, inter-quartile range; TDM, therapeutic drug monitoring

AREA UNDER THE CURVE AND PK PARAMETER ESTIMATION

When compared using Wilcoxon rank-sum tests, there were no statistically significant differences in the AUCs across any of the estimation approaches. Plots of the AUCs derived using the various approaches are shown in Appendix (A.4, A.5). The correlation between AUC estimation using all samples and the calculations using the Sawchuk–Zaske method⁸ (i.e., clinical TDM calculations) was 0.914.

The performance of AUC estimation using the optimal sample and our model was excellent (Table 2). The bias between AUCs estimated using the optimal sample compared to all samples was ≤10% for all but 1 subject, for whom bias was +11.5%. Although the median bias and imprecision of AUC estimation using troughs or peaks were largely comparable to those of optimal sampling, the bias was within 10% and 15% for fewer subjects. Among the eight individuals whose optimal sample was not a trough, the use of a trough for Bayesian estimation resulted in a median bias of -3.1% (range: -18.9–1.2%), and the median imprecision was 4.0% (range: 1.2–18.9%). Among the 10 children whose optimal sample was not considered a peak, the use of a peak for Bayesian estimation resulted in a median bias of 2.8% (range: -27.3–32.1%) and a median imprecision of 6.4% (1.7–32.1%).

Table 2.

Performance of optimal sampling, trough only, and peak only estimation of area under the curve (AUC) compared to AUC estimation with all available samples

	Optimal sample (n=18)	Trough only (n=14)a	Peak only (n=18)a
AUC24 in mg*h/L, median (range)	494 (347–1008)	494 (326–1028)	547 (261–1008)
Bias, median (range)b	0.3% (−7.0–11.5%)	−2.2% (−18.9–10.3%)	2.0% (−27.3–32.1%)
Imprecision, median (range)b	4.6% (0.2–11.5%)	3.6% (0.7–18.9%)	3.5% (0.7–32.1%)
Subjects with bias:b
<5%, n (%)	10 (56%)	9 (64%)	9 (50%)
<10%, n (%)	17 (94%)	12 (86%)	14 (78%)
<15%, n (%)	18 (100%)	13 (93%)	15 (83%)
<20%, n (%)	18 (100%)	14 (100%)	16 (89%)
Correlation coefficient	0.984	0.984	0.941

Median (range) AUC using Bayesian estimation from all available samples with our model was 495 mg*h/L (336–1084).

^aTrough measurement (≤30 min prior to next dose) was the optimal sampling time for 6 subjects. Peak measurement (≤120 min after a dose) was the optimal sampling time for 8 subjects.

^bCompared to Bayesian estimation using all available samples.

Bayesian AUC estimation using optimal sampling did not approximate as closely with clinical AUC calculations using the Sawchuk-Zaske method. With the clinical AUC calculations as the comparator, the median bias and imprecision of optimal sampling were 8.8% (range: -27.7–48.7%) and 16.5% (range: 0.7–48.7%), respectively. The bias was < 10% in half of the subjects with optimal sampling, with clinical AUC calculations used as the reference.

When categorizing each AUC₂₄ as therapeutic (400–600 mg*h/L), sub-therapeutic (<400 mg*h/L), or supratherapeutic (>600 mg*h/L), there was general agreement (94%) between optimal sampling only and all sampling when using Bayesian estimation with our model (Figure 3). Twelve subjects’ AUCs (67%) were therapeutic, two (11%) were sub-therapeutic, and four (22%) were supra-therapeutic when estimating the AUC using optimal sampling while eleven (61%) were therapeutic, two (11%) were sub-therapeutic, and five (28%) were supra-therapeutic. The one child whose AUC was therapeutic via optimal sampling but supratherapeutic via all samples had estimated AUCs of 545 vs 602 mg*h/L, respectively. Seven AUCs (39%) were in the therapeutic range based on clinical calculations, five (28%) were sub-therapeutic, and six (33%) were supra-therapeutic.

Figure 3. Individual AUC plots using different approaches and models.

Open circles reflect AUC estimates within the recommended therapeutic range of 400–600 mg*h/L. X’s reflect AUCs above the therapeutic range (>600) and boxes reflect AUCs below the range (<400). Lines are used to connect AUC estimates for the same individual using different approaches. Optimal, All samples, Peak and Trough utilized Bayesian estimation with our model. Sawchuk–Zaske calculated AUC using peak and trough vancomycin measurements.

The mean PK estimates among the population using Bayesian estimation with our model were largely similar when only the optimal sample and all samples were used (A.6). The population means for clearance and volume were not significantly different when derived from optimal sampling, all samples, or clinical calculations (ANOVA, p > 0.05).

Discussion

In this study, we found that the bias of Bayesian AUC estimation using a single optimally timed sample was low compared to the AUC estimated from multiple samples. Informed by a robust population PK model, optimal sampling promoted accurate estimation of AUC (bias <10%) in all participants for whom samples were obtained within goal times. Vancomycin remains an important antibiotic for critically ill children with serious gram-positive infections, and AUC-directed therapy is recommended for all children receiving dedicated treatment courses.⁷ Our study demonstrated that vancomycin AUC₂₄ can be reasonably estimated using a single, model-informed, optimally timed sample, even within a highly dynamic, critically ill pediatric population.

Although TDM is common practice within the intensive care unit, AUC estimation can be challenging. At our hospital, vancomycin AUCs are manually calculated from peak and trough concentrations using the Sawchuk–Zaske method. This may not be the preferred method for AUC estimation, however, since it requires collection of multiple samples, which is more costly and time-consuming than collecting a single sample. This also creates increased opportunities for sampling and measurement errors, as observed in our two unevaluable participants. Furthermore, this approach assumes that the patient is at steady-state, which may not be a reasonable assumption for children with critical illnesses. However, the reliance on a single timed sample to inform Bayesian estimation also faces challenges. As shown, the optimal sample was mistimed in 22% of participants. Clinical nurses were responsible for blood sample collection in our study because most samples were drawn from existing arterial or venous catheters. Understandably, clinical care responsibilities can impede collection of precisely timed TDM samples. And, in a clinical setting, the collection of TDM samples at non-conventional times (i.e., not the peak or trough) can lead to confusion and potential errors. Despite this, the AUC was estimated with good overall accuracy in our study population. In fact, even the use of trough samples promoted accurate estimation of AUC, as previously described.²¹

Our model described vancomycin PK in two compartments and relied on cystatin C for GFR estimation, which we previously found to be more informative than serum creatinine for vancomycin clearance in critically ill children.^15,22 Although the performance of Bayesian estimation using our model was excellent, our model may not be ideally suited for use in non-critically ill children. Furthermore, it is important to recognize that the optimal sampling time was specifically informed by MMopt using our model. Different times would likely be optimally informative for Bayesian estimation using other population PK models, which we did not evaluate in this study. Although the bias of AUC estimation using optimal sampling with our model was low, optimal sampling may be of greater or lesser importance when performing Bayesian estimation using other population PK models.

Optimal sampling is not only model-specific but also patient- and occasion-specific. Using MMOpt, all the information leading up to sampling is informative regarding the optimal sampling time. Hence, we observed significant variability in optimal sampling times among the study population (Figure 2). While this could be considered a barrier, timed labs are not unusual in ICU settings. Bedside nurses were able to collect the majority of samples (78%) within 10 min of our goal times. Furthermore, the collection of trough samples is not easier than the collection of timed samples in the ICU. Although the estimation of the AUC₂₄ from a trough was comparable to that from optimal sampling, the trough was usable in only a fraction (78%) of the study population. A larger study is required to assess the adequacy of AUC estimation using only trough sampling.

Importantly, there were notable differences between the individual AUCs estimated using Bayesian methods and clinical calculations. The AUC was more often outside of the therapeutic range (400–600 mg*h/L) according to clinical calculations (61%) than with optimal sampling and our model (33%). Although the “true” AUC could not be determined using either method, fewer patients would have undergone dose adjustments if the AUCs were monitored using our optimal sampling approach. This may affect the safety and efficacy of the drug. The Sawchuk–Zaske method can derive patient-specific PK estimates and may be especially useful when population estimates of PK parameters are unreliable.⁸ However, this approach more closely approximates 1-compartment drug behavior, which may be an oversimplification of true vancomycin PK. Ultimately, additional investigation is needed to determine how targeting the AUC₂₄ using our approach versus traditional calculations affects clinical outcomes.

As mentioned previously, determining the optimal sampling time is model-specific. However, the optimal sampling times in our study were consistent with those in previous studies that evaluated vancomycin. In a prospective trial by Neely et al.,²³ trough samples were found to be optimal for determining the AUC in 29% of adults. This is comparable to the 32% in our study. Interestingly, 50% of the optimal samples fell within the first 2 h of dosing in our study (i.e., peaks). This differs from the Neely trial (12%) but may be reflective of the shorter dosing intervals used in children. However, this may also suggest that early concentration time-points are highly variable among critically ill children, perhaps because of larger variability in volume of distribution. Notably, in our analyses, Bayesian estimation using only peak samples showed a substantially wider range of bias and imprecision values at the individual level than when using optimally timed samples. Thus, although the early time points were more often considered optimal, they were less accurate when used forBayesian AUC estimation.

Our study has some limitations. First, we did not evaluate how well our model empirically predicted measured vancomycin concentrations. Bayesian approaches can be used to guide empirical dose selection; however, this was not the goal of this study. Larger studies are required to evaluate our model for this purpose. Furthermore, we could not collect cystatin C until after enrollment, and AUC estimation occurred at varying times after the start ofvancomycin treatment. Thus, we designed this study to focus on AUC estimation in the context of TDM, rather than dosing. Additionally, despite robust data supporting the use of cystatin C over creatinine for the estimation of vancomycin clearance and renal function in children,^{15,22,24–26} this test may not be available at all centers. Second, the sample size of our study was small. Possibly, there are subpopulations of critically ill children for whom our model does not estimate the AUC well, and there are specific populations (CCRT and ECMO) that were omitted from recruitment for whom our model would not be applicable. Finally, we had intended to evaluate how well the Bayesian estimation on the first sampling day predicted the vancomycin AUC 24–72 h later. Unfortunately, few individuals remained on vancomycin for >24 h beyond the first sampling day, precluding this assessment. Additional studies are required to examine whether the early Bayesian estimation using our model can inform longitudinal dosing.

Conclusion

The present study showed that Bayesian estimation of AUC in critically ill children using a single, optimally timed sample provided accurate AUC estimation compared to Bayesian estimation with multiple samples. The Bayesian AUC from an optimal samplewas within 10% of the AUC estimated using multiple samples in >90% of subjects These findings highlight the fact that our model, derived from a critically ill pediatric population, can facilitate AUC estimation using limited sampling.

Data Availability Statement:

Data can be made available upon request.