Discrepancies Between Bayesian Vancomycin Models Can Affect Clinical Decisions in the Critically Ill

Purpose To assess the agreement in 24-hour area under the curve (AUC24) value estimates between commonly used vancomycin population pharmacokinetic models in the critically ill. Materials and Methods Adults admitted to intensive care who received intravenous vancomycin and had a serum vancomycin concentration available were included. AUC24 values were determined using Tucuxi (revision cd7bd7a8) for dosing intervals with a vancomycin concentration using three models (Goti 2018, Colin 2019, and Thomson 2009) previously evaluated in the critically ill. AUC24 values were categorized as subtherapeutic (<400 mg·h/L), therapeutic (400–600 mg·h/L), or toxic (>600 mg·h/L), assuming a minimum inhibitory concentration of 1 mg/L. AUC24 value categorization was compared across the three models and reported as percent agreement. Results Overall, 466 AUC24 values were estimated in 188 patients. Overall, 52%, 42%, and 47% of the AUC24 values were therapeutic for the Goti, Colin, and Thomson models, respectively. The agreement of AUC24 values between all three models was 48% (223/466), Goti-Colin 59% (193/466), Goti-Thomson 68% (318/466), and Colin-Thomson 67% (314/466). Conclusion In critically ill patients, vancomycin AUC24 values obtained from different pharmacokinetic models are often discordant, potentially contributing to differences in dosing decisions. This highlights the importance of selecting the optimal model.


Introduction
International clinical practice guidelines and position statements suggest that the therapeutic drug monitoring of vancomycin should be guided by area under the curve (AUC) values [1][2][3]. Tis is because AUC-guided dosing is associated with less treatment failure and improved safety compared to trough-guided monitoring in patients with methicillin-resistant Staphylococcus aureus infection [4]. Te ratio of the 24-hour AUC to the minimum inhibitory concentration (AUC 24 /MIC) is the pharmacokinetic (PK) parameter used to guide dosing decisions [1]. An AUC 24 /MIC target of 400-600 mg·h/L is considered to be optimal when MIC is determined by broth microdilution [1,5]. In clinical practice actual MIC is often not available, and an MIC of 1 mg/L is assumed. Tus, an AUC 24 <400 mg·h/L would prompt a dose increase to avoid treatment failure, and an AUC 24 >600 mg·h/L would suggest a dose decrease is required to avoid acute kidney injury or other toxicity. AUC 24 can be estimated by frst-order linear PK equations using two concentrations within a dose interval. Tis can be calculated manually or via readily available online calculators [6]. Alternatively, AUC 24 can be estimated via population PK models using Bayesian software platforms [7]. Te logistical advantage of the latter approach is that the AUC 24 value estimate can be obtained using only one vancomycin concentration [6]. However, previous studies have shown a low precision of an AUC 24 estimation using Bayesian software and a single concentration [8,9]. In a cohort study of hospitalized patients (n � 978), clinical agreement was 76.8% between AUC 24 values predicted by linear and Bayesian oneconcentration [9]. Te study was not specifcally conducted among the critically ill. However, the investigators evaluated one Bayesian model [10] in the critically ill subset of the cohort.
Some of the Bayesian software platforms available ofer multiple population PK models for the same medication, and end-users are required to select the most appropriate model. Tus, studies are needed to compare the clinical agreement between population PK models in the critically ill to inform this decision. It is possible that depending on the population PK model selected, it could lead to diferent dosing decisions. Te objective of this study was to assess the agreement in AUC 24 value estimates between commonly used vancomycin population PK models in the critically ill.

Ethics/Institutional Review Board.
Te study was approved by the Human Research Ethics committee of the Sydney Local Health District-Royal Prince Alfred Hospital zone (Approval number: 2019/ETH12033).

Study Design and Setting.
Tis was a retrospective cohort study using data obtained from the intensive care unit (ICU) of a quaternary care hospital in Sydney, Australia. Tis is a substudy of a previously reported investigation evaluating the predictive performance of three vancomycin population PK models [8]. Te hospital uses trough-based vancomycin monitoring to guide dosing decisions. Usually only one vancomycin concentration is obtained per dosing interval. Bayesian software is not used for clinical care, and AUC 24 based monitoring has not been implemented. Te ICU has a fully electronic medical record that includes all medications and pathology values used for the study (Philips IntelliSpace Critical Care and Anesthesia). Vancomycin concentrations are determined using an immunoassay, as has been previously described [8].

Participants and Data Acquisition.
Te study included adults (>18 years old) who were admitted to the ICU between 1 January 2019 and 31 May 2020, had received intravenous vancomycin via intermittent infusion, and had at least one vancomycin serum concentration available. Patients were excluded if they received continuous renal replacement therapy (CRRT) or a continuous infusion of vancomycin. Patients were also excluded if they were administered vancomycin doses on a non-ICU ward and subsequently transferred to the ICU, as dosing data were not available. All data were obtained from the electronic medical record. Vancomycin concentrations were categorized independently by two investigators as true trough concentrations or true troughs at steady state. A true trough concentration was defned as a concentration taken within 60 minutes before administration of an anticipated dose. A true trough at steady state was defned as a concentration taken within 60 minutes of the anticipated 4 th dose in treatment courses with <60-minute deviations from scheduled times for all prior doses.

Estimation of Vancomycin AUC24
Values. An AUC 24 value was estimated for each dosing interval for which a vancomycin concentration was available using Tucuxi via a command line interface (revision cd7bd7a8 joint development by HEIG-VD, Yverdon-les-Bains, Switzerland and CHUV, Lausanne, Switzerland). All AUC24 value estimates were determined using a single vancomycin concentration. AUC 24 values were estimated using three previously published vancomycin population PK models: (1) Goti [10], (2) Colin [11], and (3) Tomson [12]. Tus, each vancomycin concentration had three corresponding AUC 24 value estimates.

Data Analysis.
AUC 24 values were categorized as subtherapeutic (<400 mg·h/L), therapeutic (400-600 mg·h/L), or toxic (>600 mg·h/L) for each model, assuming a minimum inhibitory concentration of 1 mg/L. Tese categorizations were defned and based on cut-of values from international guidelines [1,3]. AUC 24 value categorization was compared across the three models and reported descriptively as percent agreement. In other words, if an AUC 24 value with one model was nontherapeutic and another model was therapeutic, then the two models would be discrepant and could result in diferent vancomycin dosing decisions. Te concordance between AUC 24 values from models was reported as scatter plots with the concordance correlation coefcient (ρ c ) [13]. Other parameters calculated were Pearson's r, the average diference, the standard deviation of the diference, and Bland and Altman's 95% limits of agreement [14].
Two sensitivity analyses were conducted. First, the discrepancy between models was evaluated in the subset of concentrations that were considered as true troughs. Second, the discrepancy between models were evaluated in the subset of concentrations that were considered as true troughs at steady state. Tis was done as trough levels are commonly used in clinical practice, and it is possible that predictive performance is diferent in this subset [15]. All analyses were conducted in STATA 15 (College Station, Texas), and scatter plots were created using the R software (version 4.0.3, Vienna, Austria).

Study Cohort.
A total of 330 patients received vancomycin during the study time frame and had a vancomycin concentration measured. Of these, 188 patients were included and had 466 vancomycin concentrations. Te selection of the study cohort is shown in Figure 1. Of these, 250 concentrations were true troughs in 136 patients, and 113 concentrations were true troughs taken at steady state in 71 patients. Te baseline characteristics of patients are in Table 1.
Te two sensitivity analyses using the subset of concentrations that were true troughs or true troughs at steady state showed similar or lower agreement than the overall sample (Supplementary Appendix (available here)). Overall, agreement for true trough subset was 48% between all models (n � 121/250), and true trough at steady state subset was 42% between all models (n � 47/113). Parameters for concordance between models and scatterplots for the sensitivity analysis subsets are in the Supplementary Appendix (available here). Tere were 37 concentrations that were therapeutic (15-20 mg/L) and considered true troughs at steady state. Of these, 28 (75%) in the Goti

Discussion
Te key fnding of this investigation was that agreement between vancomycin PK models occurred approximately two-thirds of the time. In other words, clinicians may come to diferent dosing decisions based on standard cut-of values for AUC 24 in 1 of 3 ICU patients, depending on the PK model used. Te Goti model had lower AUC 24 estimates than Colin and Tomson. Tus, using Goti-based AUC 24 estimates, clinicians would be prompted to use higher doses than the other two models.
Te question regarding the most appropriate PK model for the critically ill has not been settled. Some commonly used Bayesian software platforms have recommended the Goti model for ICU patients [9]. Although the three models used in our investigation were not derived using critically ill patients exclusively, they have been subsequently evaluated in ICU cohorts [8,16,17]. In a retrospective study, data from 82 ICU patients was used to evaluate 12 vancomycin PK models [16]. Te investigators considered a model to be clinically acceptable if the relative bias was ±20% and the 95% CI included zero. Te Goti model was the only one considered to be clinically acceptable based on both a priori and a posteriori approaches. However, the defnition for acceptability did not include parameters for precision. In a larger study (n � 188 patients) of the same cohort as our current investigation [8], it appeared that the Goti model may be slightly more suitable based on precision, but the extent of the diferences between PK models were too small to be clinically meaningful. Another smaller investigation with data from 50 patients and using simulation techniques considered the model by Tomson to have the best predictive performance [17]. However, precision appeared to be low based on the relative root mean squared error reported by the authors. In addition, the model by Goti was not evaluated in the aforementioned study. Vancomycin PK models need to be improved for suitable precision of AUC 24 estimated from one concentration. It is possible that sparse sampling has been a contributor to models with poor precision.
Te AUC 24 can also be estimated by taking two concentrations within a dose interval (e.g., peak and trough levels). Te estimation is based on frst-order PK equations. A detailed approach to these calculations has been described  [18]. Although this can be done manually, readily available software programs can be used to conduct these calculations with relatively little training [6]. AUC 24 estimation via frst order PK calculations using two levels is the reference standard to which Bayesian estimates have been compared [9]. Such a comparison has been described in a retrospective cohort study [9] (n � 978 patients) where clinical agreement was assessed between frst-order PK equation-calculated AUC 24 vs Bayesian two-(i.e., peak and trough) and oneconcentration AUC 24 (i.e., trough only). Clinical agreement was defned similarly to the cut-of values in our study. Te PK model used for Bayesian estimation depended on the patient (noncritically ill, critically ill, and obese). Te Goti model was used in the critically ill subset and 69% of the patients were critically ill. Clinical agreement was higher with two-concentration estimates (87.4%) than oneconcentration (76.8%) estimates. In clinical practice, the advantage of the Bayesian AUC 24 is that it can be estimated using just one concentration. Tus, the results pertaining to the one-concentration estimate are more relevant. If two concentrations are taken, then clinicians could use the reference standard frst-order PK equations, and Bayesian estimates are not needed. Te clinical agreement using oneconcentration Bayesian AUC 24 was still higher than the agreement we found between PK models in our study. Tis may be because there was less variability in the noncritically ill patients in the aforementioned study that made up approximately one-third of the patient subset.   Sensitivity analyses were conducted to evaluate subsets of patients with true troughs and true troughs at steady state. We considered that some PK models are developed using routine clinical data rather than richly sampled data with multiple concentrations taken during a dosing interval. Tis is true for the Goti model, where most patients had only one sample taken [10]. Tese are usually trough concentrations (or similar). Tus, prediction using such a PK model may be diferent for troughs. In addition, trough concentrations are more likely to refect what is done in clinical practice. However, we did not show that agreement between PK models changed in the subsets with true troughs.
Given the uncertainty of the evidence, the availability of diferent PK models within Bayesian software platforms, and the possible lack of agreement between PK models, we suggest that if clinicians use a one-concentration Bayesian estimation using currently available PK models, it may be useful to verify the estimation using two diferent PK models available within the software program (e.g., Goti and Tomson). Tis can be done relatively quickly and does not require the re-entry of patient data into the software platforms we are aware of. A mid-interval concentration may also improve estimates [15]. Te discrepancies in the estimate and direction of the discrepancy could help guide dosing decisions, which must be made in the context of the clinical situation. For example, a subtherapeutic AUC 24 with the Goti model but a therapeutic AUC 24 with the Tomson model would still support a dose increase if an infection was severe or the patient was deteriorating. Based on the uncertainty and lack of precision, our institution has not implemented the use of Bayesian software platforms. Instead, there are ongoing eforts to transition to AUC 24 estimation using two concentrations calculated from frstorder PK equations in the ICU. Our study has some limitations. First, we have only compared AUC 24 estimates between PK models. From the data available we were unable to determine the "true" AUC value as a reference. As such, we are unable to comment on the accuracy of AUC value estimates derived from each of the population PK models. Second, we did not assess the implication of drug exposure on clinical outcomes. However, that was not the intent of this study. An evaluation of clinical outcomes would require a comprehensive clinical dataset to adjust for any potential confounders as well as an appropriate assessment of the drug concentration and pathogen susceptibility [19]. Tird, ICU patients are a heterogenous population and it is possible that clinical agreement would be improved in some subsets or phenotypes of patients. However, we do not have the data to meaningfully delineate this.

Conclusion
In critically ill patients, vancomycin AUC 24 values estimated from diferent PK models are often discordant, potentially contributing to diferences in dosing decisions. Tis highlights the importance of selecting the optimal model. Given the lack of agreement between PK models currently used in Bayesian software, it may be useful to use more than one model to guide decisions that are supported by clinical context, especially when estimates are close to decision cutof values for AUC 24 .

Data Availability
Data are available upon reasonable request to the corresponding author.

Additional Points
What is already known about this subject: Monitoring of vancomycin should be guided by 24-hour area under the curve (AUC 24 ). AUC 24 can be estimated via population pharmacokinetic (PK) models using Bayesian software platforms. Diferent vancomycin population PK models are available for AUC 24 estimation. What this study adds: Vancomycin AUC 24 values obtained from diferent PK models were often discordant at decision thresholds. Clinical dosing decisions of vancomycin may be dependent on the PK model used.

Ethical Approval
Te study was approved by the Human Research Ethics committee of the Sydney Local Health District-Royal Prince Alfred Hospital zone (Approval number: 2019/ETH12033).

Conflicts of Interest
Te authors declare that there are no conficts of interest. Table S1: Cross-tabulation of agreement between models at decision cut-ofs in the subset with true troughs. Table S2: Cross-tabulation of agreement between models at decision cut-ofs in the subset with true troughs at steady state. Table  S3: Concordance parameters between models in the sensitivity analysis subsets. Figure S1: Scatterplot of AUC24 between models in subset of concentrations that were true troughs. Figure S2: Scatterplot of AUC24 between models in subset of concentrations that were true troughs at steady state. (Supplementary Materials)