Bayesian‐Estimated Vancomycin AUC24 as a Near‐Term Renal Risk Signal after Therapeutic Drug Monitoring: A Multicenter Cohort Study

Abstract

Vancomycin therapeutic drug monitoring has moved toward area under the concentration–time curve (AUC)‐guided dosing, but the clinical meaning of a Bayesian‐estimated AUC generated at an individual monitoring episode remains uncertain when renal clearance may already be changing. We conducted a retrospective multicenter cohort study using inpatient dosing, pharmacy, therapeutic drug monitoring, and laboratory data from two hospitals in Hai Phong, Viet Nam, between 2019 and 2025. The analysis included 898 adult index vancomycin monitoring episodes with valid Bayesian‐estimated AUC24 and sufficient serum creatinine data to ascertain acute kidney injury (AKI) within 48 h. AUC24 was estimated within the institutional Bayesian monitoring workflow using SmartDoseAI, with BestDose version 2.4.3 used as an independent pharmacokinetic cross‐check in selected clinically uncertain episodes. AKI occurred in 84 episodes (9.4%). Higher continuous AUC24 was associated with AKI in the minimally adjusted model (odds ratio, 1.16 per 100 mg·h/L; 95% confidence interval, 1.02–1.33; P = .023), whereas categorical AUC contrasts were imprecise. The association attenuated after exclusion of 23 episodes with possible pre‐existing AKI before the index monitoring episode (odds ratio, 1.12; 95% confidence interval, 0.96–1.31; P = .159), and restricted cubic spline modeling did not identify a distinct nonlinear threshold. Bayesian‐estimated AUC24 at monitoring may therefore be most useful as a near‐term renal risk signal that integrates vancomycin exposure with evolving renal clearance, rather than as a standalone causal toxicity threshold.

Introduction

Vancomycin remains a central therapy for serious infections caused by methicillin‐resistant Staphylococcus aureus and other susceptible Gram‐positive organisms. Contemporary consensus guidance has shifted vancomycin monitoring away from trough‐concentration targets and toward exposure‐based assessment using the 24‐h area under the concentration–time curve (AUC24), particularly for serious methicillin‐resistant S. aureus infection. ¹ , ² This change reflects a clinical pharmacology principle rather than a simple monitoring preference: Antibacterial response and concentration‐related toxicity are better conceptualized through drug exposure than through a single concentration surrogate.

The efficacy basis for AUC‐guided dosing comes from pharmacokinetic–pharmacodynamic studies linking the AUC24/minimum inhibitory concentration (MIC) ratio with clinical and microbiological response. Early and subsequent analyses supported an AUC/MIC target near 400 when the MIC is assumed to be 1 mg/L, while also showing that source of infection, organism susceptibility, and host factors can influence outcomes. ³ , ⁴ , ⁵ The current AUC/MIC target range of 400 to 600 mg·h/L was therefore designed to preserve efficacy while reducing excessive exposure, not to imply a uniform renal risk across all patients within that interval. ¹

The safety rationale for AUC‐guided dosing arose from limitations of trough‐based practice. Earlier guidance and clinical practice often used trough concentrations of 15 to 20 mg/L as a pragmatic surrogate for exposure. ⁶ Subsequent studies and systematic reviews showed that higher trough concentrations and larger daily doses were associated with greater nephrotoxicity, while troughs did not reliably identify the individual patient's AUC. ⁷ , ⁸ , ⁹ These concerns are especially relevant in acutely ill patients, for whom renal function, volume status, dose timing, and co‐nephrotoxic therapy can change rapidly during hospitalization.

Model‐informed precision dosing extends this exposure‐based approach by combining a population pharmacokinetic prior with patient‐specific dosing and concentration data. Bayesian dose support can estimate exposure from sparse sampling and can translate therapeutic drug monitoring into patient‐specific dose review. ¹⁰ , ¹¹ Implementation studies and meta‐analyses have generally found that AUC‐guided monitoring is associated with lower vancomycin exposure and less nephrotoxicity than trough‐guided monitoring, although the magnitude of benefit varies across patient populations and workflows. ¹² , ¹³ More broadly, model‐informed precision dosing has been promoted as a clinical pharmacology strategy to integrate patient variability, drug exposure, and treatment response into bedside prescribing. ¹⁴ , ¹⁵

However, a separate and clinically important question remains unresolved: How should clinicians interpret a Bayesian‐estimated AUC generated at a specific therapeutic drug monitoring (TDM) episode? Much of the vancomycin exposure–toxicity literature has examined Day‐1 or day‐2 exposure, selected bacteremia cohorts, or institutional comparisons before and after implementation of AUC‐guided monitoring. ¹⁶ , ¹⁷ , ¹⁸ , ¹⁹ , ²⁰ , ²¹ , ²² Those designs are informative for target selection and program evaluation, but they do not fully address the bedside moment at which a pharmacist and physician receive a post‐dose Bayesian AUC estimate and must decide whether it represents excessive administered exposure, early renal clearance decline, or both.

This distinction is clinically relevant because serum creatinine is a delayed and imperfect marker of acute kidney injury (AKI). Creatinine‐defined AKI may appear after renal tubular injury or reduced filtration has already begun, and serum creatinine kinetics are influenced by muscle mass, volume status, and changing creatinine generation. ²³ , ²⁴ , ²⁵ In this setting, an elevated Bayesian‐estimated vancomycin AUC24 may function simultaneously as an exposure measure and as a short‐horizon signal of reduced clearance. The key contribution of the present study is therefore not to propose a new vancomycin therapeutic target, but to evaluate the renal information contained in an AUC estimate generated during routine clinical TDM. We examined the association between Bayesian‐estimated AUC24 at an index TDM episode and creatinine‐defined AKI within 48 h in a multicenter routine‐care cohort. We also compared continuous and categorical AUC representations, explored nonlinearity, and tested whether the association persisted after excluding episodes with possible pre‐existing AKI before the index TDM event.

This episode‐level perspective is also relevant for implementation of model‐informed precision dosing in health systems where Bayesian platforms are introduced gradually rather than through a tightly controlled clinical trial. In routine care, the AUC estimate is embedded in a sequence of decisions: blood sampling, laboratory reporting, pharmacist review, prescriber response, and subsequent renal monitoring. Each step can influence both exposure and outcome ascertainment. A study focused on the index monitoring episode therefore provides information that is complementary to efficacy target studies and before–after implementation studies. It asks whether the value displayed to clinicians at the time of monitoring contains actionable renal risk information, while acknowledging that such information may be partly prognostic rather than purely causal. ¹⁴ , ¹⁵

Methods

The protocol was approved by the Institutional Ethics Committee of Hai Phong International Hospital, Hai Phong, Viet Nam (approval no. 836/IRB‐DKQT), and the requirement for informed consent was waived for the retrospective analysis of routinely collected clinical data. The study sites were Hai Phong International Hospital and Hai Phong International Obstetrics and Pediatrics Hospital, Viet Nam. We conducted a retrospective multicenter cohort study using electronic health records, pharmacy, laboratory, and vancomycin TDM data from January 1, 2019, through December 31, 2025. The study was designed and reported in accordance with the STROBE statement. ²⁶

The objective was to determine whether Bayesian‐estimated vancomycin AUC24 at an index TDM episode was associated with serum creatinine–defined AKI during the subsequent 48 h. The primary unit of analysis was the first eligible TDM episode per encounter, selected to reduce within‐encounter dependence and to align exposure, covariates, and outcome around a single clinically interpretable monitoring event. The cohort flow is shown in Figure 1, and the index–episode analytic framework is shown in Figure 2.

Figure 1.

Cohort flow diagram.

Figure 2.

Episode‐based analytic framework anchored to the index vancomycin TDM event. Exposure, covariate, and outcome windows were defined relative to the index TDM episode.

Candidate records were identified from institutional TDM databases and linked with dosing records, pharmacy exposure files, and laboratory data using patient‐ and encounter‐level identifiers. The full data‐management process is summarized in Figure 1. Briefly, 1928 prospectively collected vancomycin TDM–related records underwent raw data audit, identifier standardization, recalculation of TDM timing, and pharmacokinetic time–quality control. After harmonization, 970 records were excluded because they were non‐index, duplicate, non‐linkable, or otherwise not evaluable as index vancomycin TDM episodes with sufficient dosing and concentration timing for AUC recalculation. This yielded 958 candidate vancomycin TDM records. Records were eligible for the episode‐level cohort when vancomycin AUC24 was available, AKI ascertainment within 48 h was possible, encounter identifiers were valid, renal replacement therapy exclusion criteria were not met, and dosing or observation‐time flags did not indicate unreliable pharmacokinetic timing. Seven candidate records were excluded at this stage, leaving 951 episode‐level data‐quality eligible records. After 53 episodes in patients younger than 18 years were excluded, the final adult analytic cohort included 898 index TDM episodes from 714 unique patients.

Vancomycin AUC24 was estimated within the institutional TDM workflow using Bayesian model‐informed precision dosing. SmartDoseAI, a web‐based Bayesian dosing‐support platform developed by N2TP Technology Solutions JSC, was the primary platform for AUC24 estimation in routine practice. For each index episode, Bayesian estimation used the recorded vancomycin dosing history, infusion time, sampling time, measured vancomycin concentration, and available patient‐level clinical information. The population pharmacokinetic prior embedded in SmartDoseAI at the time of clinical use was used for individual posterior estimation.

A valid Bayesian estimate required a plausible dosing history, interpretable infusion and sampling times, a measured vancomycin concentration suitable for pharmacokinetic interpretation, and internal consistency between dose administration, sampling time, and estimated exposure. Episodes with incomplete, implausible, or unverifiable dosing or sampling information were excluded before cohort construction. BestDose software, version 2.4.3, was used as an independent Bayesian pharmacokinetic cross‐check in selected episodes requiring additional verification, including complex dosing histories, uncertain sampling interpretation, changing renal function, or Bayesian estimates requiring clinical plausibility review. BestDose outputs were not used as an automatic substitute for SmartDoseAI estimates; they served as an additional pharmacokinetic reference.

Final AUC24 values were not accepted solely on the basis of automated software output. Two clinical pharmacists experienced in vancomycin TDM independently reviewed the dosing record, infusion and sampling times, measured concentration, renal function trajectory, and pharmacokinetic plausibility of each estimate. Discrepancies were resolved by consensus, prioritizing the estimate most consistent with the complete clinical, dosing, and laboratory record. Dose individualization was discussed with treating physicians and incorporated renal function, infection severity, microbiological information when available, concomitant nephrotoxins, and clinical judgment.

Local vancomycin TDM practice evolved during the study period from concentration‐based interpretation toward more consistent Bayesian AUC‐guided review. Prospective AUC‐guided dose adjustment was not implemented as a uniform interventional protocol across all calendar years. Earlier episodes may therefore have been influenced by trough‐based or mixed trough/AUC‐based decision‐making, whereas later episodes more consistently incorporated Bayesian AUC‐guided interpretation. The analytic dataset contained concentration values and sampling times, but laboratory result‐release time, pharmacist review time, clinician review time, and exact time from result availability to dose modification were not uniformly captured as structured variables.

The primary exposure was Bayesian‐estimated vancomycin AUC24, analyzed continuously per 100 mg·h/L increase. Secondary categorical analyses classified AUC24 as less than 400, 400 to 600, or greater than 600 mg·h/L. The 400 to 600 mg·h/L group was the primary categorical reference because it corresponds to the conventional target range when MIC is assumed to be 1 mg/L. ¹ A sensitivity categorical model used less than 400 mg·h/L as the reference to evaluate whether conventional‐range and high‐range exposures differed from a lower‐exposure baseline.

The primary outcome was AKI within 48 h after the index TDM episode, defined using serum creatinine–based Kidney Disease: Improving Global Outcomes criteria. ²⁷ The binary outcome included KDIGO Stage 1, Stage 2, or Stage 3 AKI. AKI was present when serum creatinine increased by at least 0.3 mg/dL within 48 h or increased to at least 1.5 times baseline, known or presumed to have occurred within the prior 7 days. Creatinine‐based staging was defined as Stage 1 for an increase of at least 0.3 mg/dL or 1.5–1.9 times baseline, Stage 2 for an increase to 2.0–2.9 times baseline, and Stage 3 for an increase to at least 3.0 times baseline or to at least 4.0 mg/dL. Urine output criteria were unavailable. Baseline serum creatinine was derived from pre‐index measurements within the same hospitalization; the nearest pre‐TDM serum creatinine was retained separately for description. For the 1.5‐fold criterion, pre‐index creatinine values within the preceding 7 days were reviewed when available.

Possible pre‐existing AKI before the index TDM episode was flagged when serum creatinine at TDM had already increased by at least 0.3 mg/dL or by at least 1.5‐fold relative to the baseline creatinine used for AKI assessment. Such an increase could reflect renal dysfunction that predated vancomycin therapy, renal injury developing during vancomycin exposure before TDM, or both. Because this variable could lie on the temporality pathway and represent reverse causation, it was not included as a primary model covariate. It was instead used to define a prespecified conservative sensitivity analysis excluding these episodes; this analysis was not intended to isolate a population necessarily unaffected by prior vancomycin exposure.

Covariates were selected a priori based on clinical relevance and data availability. The minimally adjusted primary model included age, sex, baseline serum creatinine, intensive care unit (ICU) admission, concomitant piperacillin/tazobactam, concomitant aminoglycoside exposure, and time from first vancomycin dose to TDM. Additional descriptive variables included comorbidities, baseline renal dysfunction, septic shock, vasopressor use, other nephrotoxins, infection characteristics, albumin, inflammatory markers, and microbiological variables. Model 3 additionally included a broader set of available baseline and contemporaneous variables, including comorbidities, baseline renal dysfunction, ICU admission, septic shock, vasopressor use, concomitant nephrotoxins, infection characteristics, albumin, C‐reactive protein, and procalcitonin.

Continuous variables were summarized as mean (standard deviation) or median (interquartile range), as appropriate; categorical variables were summarized as counts and percentages. Baseline imbalance was evaluated using standardized mean differences, and P values were treated as descriptive. Logistic regression estimated odds ratios (ORs) and 95% confidence intervals (CIs) for AKI within 48 h. Model 1 was unadjusted. Model 2 was the prespecified minimally adjusted primary model. Model 3 was an exploratory fully adjusted model used only to evaluate directional robustness under broader adjustment; it was not considered confirmatory given 84 AKI events and substantial model complexity. Model 4 analyzed AUC24 categories using 400 to 600 mg·h/L as the reference. A sensitivity categorical model used less than 400 mg·h/L as the reference.

Restricted cubic spline modeling explored departure from linearity in the minimally adjusted framework. Knots were placed at the 5th, 35th, 65th, and 95th percentiles of observed AUC24. The main spline figure was restricted to the 5th–95th percentile of observed AUC24 to reduce instability at sparse tails. Overall association was tested by a likelihood‐ratio test comparing the covariate‐only model with the spline model; nonlinearity was tested by comparing the linear‐AUC and spline‐AUC models. Prespecified exploratory interaction analyses examined effect modification by piperacillin/tazobactam, aminoglycoside exposure, ICU admission, and baseline renal dysfunction. Interaction analyses were interpreted cautiously, especially when exposed strata were sparse; these exploratory tests were not adjusted for multiple comparisons. A 72‐h AKI sensitivity analysis was not performed because a standardized 72‐h serum creatinine ascertainment variable was not available. Analyses were performed using reproducible Python workflows.

Results

Figure 1 summarizes cohort construction, and Figure 2 shows the episode‐centered analytic framework used to anchor exposure, covariates, and outcome ascertainment to the index TDM event. From 1928 prospectively collected TDM‐related records, harmonization and pharmacokinetic quality control yielded 958 candidate vancomycin TDM records. After exclusion of 7 records with insufficient renal outcome ascertainment, non‐evaluable renal status, renal replacement therapy, or unreliable dosing/sampling timing, 951 episode‐level data‐quality eligible records remained. Exclusion of 53 episodes in patients younger than 18 years produced the final adult analytic cohort of 898 index TDM episodes from 714 patients. Overall, 84 episodes (9.4%) met serum creatinine–based KDIGO criteria for Stage 1 or higher AKI within 48 h (Figure 1).

The observed AUC24 distribution was centered within the conventional therapeutic range but was right‐skewed (Figure S1). Median AUC24 was 452 mg·h/L (interquartile range, 373–551 mg·h/L), and 163 episodes (18.2%) exceeded 600 mg·h/L. AUC24 was modestly higher among episodes with AKI than among those without AKI (median, 469 vs 449 mg·h/L; standardized mean difference, 0.25). This difference was directionally consistent with the exposure–risk hypothesis but small relative to the overlap in exposure distributions, reinforcing the need for multivariable and sensitivity analyses rather than interpretation based on unadjusted distributions alone (Table 1 and Figure S1).

Table 1.

Baseline Characteristics by AKI Status within 48 h after Vancomycin TDM

Variable	Level	Overall	No AKI	AKI	SMD	P Value
N		898	814	84
Demographics
Age (years)		56.5 ± 18.2	56.7 ± 18.0	55.2 ± 20.1	0.08	.532
Male sex	Yes	516 (57.5%)	489 (60.1%)	27 (32.1%)	0.56	<.001
Height (cm)		160 [155–167]	160 [155–167]	160 [154–163]	0.25	.021
Weight (kg)		57.0 ± 10.5	57.2 ± 10.6	55.0 ± 9.4	0.22	.039
Body mass index (kg/m2)		22.0 [19.6–24.4]	22.0 [19.7–24.4]	21.4 [19.1–23.7]	0.07	.371
Comorbidities
Diabetes mellitus	Yes	160 (17.8%)	149 (18.3%)	11 (13.1%)	0.14	.235
Hypertension	Yes	274 (30.5%)	249 (30.6%)	25 (29.8%)	0.02	.875
Chronic kidney disease	Yes	59 (6.6%)	56 (6.9%)	3 (3.6%)	0.15	.244
Renal baseline
Baseline serum creatinine (mg/dL)		0.88 [0.73–1.10]	0.89 [0.77–1.11]	0.66 [0.57–0.76]	0.16	<.001
Nearest serum creatinine before TDM (mg/dL)		0.84 [0.70–1.01]	0.85 [0.71–1.01]	0.67 [0.57–0.97]	0.02	<.001
Baseline creatinine used (mg/dL)		0.88 [0.73–1.10]	0.89 [0.77–1.11]	0.66 [0.57–0.76]	0.16	<.001
Baseline renal dysfunction	Yes	143 (15.9%)	133 (16.3%)	10 (11.9%)	0.13	.290
Possible pre‐existing AKI	Yes	23 (2.6%)	6 (0.7%)	17 (20.2%)	0.64	<.001
Severity
ICU admission	Yes	324 (36.1%)	286 (35.1%)	38 (45.2%)	0.21	.066
Septic shock	Yes	33 (3.7%)	30 (3.7%)	3 (3.6%)	0.01	1.00
Vasopressor use	Yes	82 (9.1%)	69 (8.5%)	13 (15.5%)	0.22	.034
Shock or vasopressor	Yes	95 (10.6%)	82 (10.1%)	13 (15.5%)	0.16	.125
Nephrotoxins
Piperacillin/tazobactam co‐administration	Yes	14 (1.6%)	12 (1.5%)	2 (2.4%)	0.07	.382
Aminoglycoside co‐administration	Yes	40 (4.5%)	32 (3.9%)	8 (9.5%)	0.22	.045
Amphotericin co‐administration	Yes	0 (0.0%)	0 (0.0%)	0 (0.0%)	0.00
Polymyxin co‐administration	Yes	20 (2.2%)	12 (1.5%)	8 (9.5%)	0.35	<.001
Furosemide use	Yes	187 (20.8%)	155 (19.0%)	32 (38.1%)	0.42	<.001
NSAID use	Yes	423 (47.1%)	384 (47.2%)	39 (46.4%)	0.01	.896
Any nephrotoxin within 48 h	Yes	521 (58.0%)	465 (57.1%)	56 (66.7%)	0.20	.092
Nephrotoxin count within 48 h		1 [0–1]	1 [0–1]	1 [0–2]	0.41	.003
Vancomycin exposure/treatment
Vancomycin AUC24 (mg·h/L)		452 [373–551]	449 [368–549]	469 [409–590]	0.25	.036
Vancomycin AUC24 per 100 mg·h/L		4.5 [3.7–5.5]	4.5 [3.7–5.5]	4.7 [4.1–5.9]	0.25	.036
AUC24 category (3 groups)	400–600	445 (49.6%)	401 (49.3%)	44 (52.4%)	0.21	.146
	<400	290 (32.3%)	270 (33.2%)	20 (23.8%)
	>600	163 (18.2%)	143 (17.6%)	20 (23.8%)
Vancomycin duration (days)		8 [6–11]	8 [6–11]	7 [6–10]	0.04	.170
Time from first dose to TDM (h)		14.1 [8.7–18.1]	14.0 [8.5–18.0]	14.9 [9.7–18.5]	0.31	.120
Infection variables
MRSA or Staphylococcus aureus infection	Yes	375 (41.8%)	343 (42.1%)	32 (38.1%)	0.08	.474
Deep‐seated infection	Yes	645 (71.8%)	586 (72.0%)	59 (70.2%)	0.04	.734
Infection site group	abscess_ssti	129 (14.4%)	122 (15.0%)	7 (8.3%)	0.21	.122
	bacteremia	4 (0.4%)	4 (0.5%)	0 (0.0%)
	diabetic_foot_infect	12 (1.3%)	10 (1.2%)	2 (2.4%)
	multi_site	622 (69.3%)	565 (69.4%)	57 (67.9%)
	osteomyelitis	1 (0.1%)	1 (0.1%)	0 (0.0%)
	other_or_unknown	120 (13.4%)	105 (12.9%)	15 (17.9%)
	pneumonia_resp	10 (1.1%)	7 (0.9%)	3 (3.6%)
Bacteremia	Yes	100 (11.1%)	88 (10.8%)	12 (14.3%)	0.10	.335
Baseline labs
Albumin, g/L		32.1 ± 5.6	32.3 ± 5.6	30.5 ± 5.6	0.32	.006
White blood cell count		10.7 [8.2–14.6]	10.7 [8.1–14.6]	10.7 [8.6–14.1]	0.06	.653
C‐reactive protein		41.4 [12.1–107.5]	41.0 [12.1–106.3]	43.9 [13.0–110.5]	0.08	.681
Procalcitonin		2.4 [0.9–3.9]	2.5 [0.9–4.0]	1.8 [0.6–3.4]	0.19	.056

SMD, standardized mean difference; TDM, therapeutic drug monitoring.

Values are mean ± SD, median [IQR], or n (%). AKI was defined according to serum creatinine–based KDIGO criteria.

Baseline clinical differences suggested that short‐horizon AKI was not determined by vancomycin exposure alone. The largest imbalance was sex: Males accounted for 32.1% of AKI episodes and 60.1% of non‐AKI episodes. Possible pre‐existing AKI before the index TDM event was uncommon overall but was markedly more frequent among episodes subsequently classified as AKI (20.2% vs 0.7%). Aminoglycosides, polymyxins, furosemide, and greater nephrotoxin count were also more common among AKI episodes, and albumin was lower in the AKI group. These imbalances were clinically important because they indicated potential confounding and raised the possibility that some patients already had evolving renal dysfunction at the time the Bayesian AUC estimate was generated (Table 1).

When AUC24 was categorized according to conventional exposure ranges, 290 episodes (32.3%) were below 400 mg·h/L, 445 (49.6%) were between 400 and 600 mg·h/L, and 163 (18.2%) were above 600 mg·h/L. Episodes in the greater‐than‐600 mg·h/L group had a more clinically vulnerable profile, including older age, more chronic kidney disease, more baseline renal dysfunction, more ICU admission, more vasopressor use, greater furosemide exposure, and more nephrotoxin exposure. AKI within 48 h occurred in 20 of 290 episodes (6.9%) with AUC24 below 400 mg·h/L, 44 of 445 (9.9%) with AUC24 between 400 and 600 mg·h/L, and 20 of 163 (12.3%) with AUC24 greater than 600 mg·h/L. The distribution of serum creatinine–defined KDIGO Stages 0, 1, 2, and 3 is shown in Table 2. AKI Stage 2 or higher was more frequent in the highest AUC24 category, but the number of severe events was small.

Table 2.

Baseline Characteristics and Renal Outcomes by Vancomycin AUC24 Category

Variable	Level	Overall	<400	400–600	>600	Effect Size	P Value
N		898	290	445	163
Demographics
Age, years		56.5 ± 18.2	56.4 ± 17.2	54.4 ± 18.6	62.6 ± 17.3	0.03	<.001
Male sex	Yes	516 (57.5%)	171 (59.0%)	255 (57.3%)	90 (55.2%)	0.03	.737
Height (cm)		160 [155–167]	160 [156–167]	160 [155–167]	160 [155–165]	0.00	.602
Weight (kg)		57.0 ± 10.5	56.9 ± 11.4	57.3 ± 10.1	56.4 ± 9.8	0.00	.609
Body mass index (kg/m2)		22.0 [19.6–24.4]	21.8 [19.6–24.3]	22.2 [19.9–24.4]	21.7 [19.3–24.3]	0.00	.636
Comorbidities
Diabetes mellitus	Yes	160 (17.8%)	55 (19.0%)	63 (14.2%)	42 (25.8%)	0.11	.003
Hypertension	Yes	274 (30.5%)	82 (28.3%)	123 (27.6%)	69 (42.3%)	0.12	.001
Chronic kidney disease	Yes	59 (6.6%)	14 (4.8%)	20 (4.5%)	25 (15.3%)	0.17	<.001
Renal baseline
Baseline serum creatinine (mg/dL)		0.88 [0.73–1.10]	0.88 [0.77–1.06]	0.87 [0.71–1.07]	1.00 [0.73–1.29]	0.03	.003
Nearest serum creatinine before TDM (mg/dL)		0.84 [0.70–1.01]	0.83 [0.72–0.95]	0.83 [0.68–0.97]	0.96 [0.70–1.25]	0.03	<.001
Baseline creatinine used (mg/dL)		0.88 [0.73–1.10]	0.88 [0.77–1.06]	0.87 [0.71–1.07]	1.00 [0.73–1.29]	0.03	.003
Baseline renal dysfunction	Yes	143 (15.9%)	34 (11.7%)	53 (11.9%)	56 (34.4%)	0.24	<.001
Severity/acute illness
ICU admission	Yes	324 (36.1%)	89 (30.7%)	145 (32.6%)	90 (55.2%)	0.19	<.001
Septic shock	Yes	33 (3.7%)	11 (3.8%)	17 (3.8%)	5 (3.1%)	0.01	.901
Vasopressor use	Yes	82 (9.1%)	20 (6.9%)	32 (7.2%)	30 (18.4%)	0.15	<.001
Shock or vasopressor	Yes	95 (10.6%)	26 (9.0%)	36 (8.1%)	33 (20.2%)	0.15	<.001
Concomitant nephrotoxins
Piperacillin/tazobactam co‐administration	Yes	14 (1.6%)	4 (1.4%)	8 (1.8%)	2 (1.2%)	0.02	.842
Aminoglycoside co‐administration	Yes	40 (4.5%)	12 (4.1%)	13 (2.9%)	15 (9.2%)	0.11	.004
Amphotericin co‐administration	Yes	0 (0.0%)	0 (0.0%)	0 (0.0%)	0 (0.0%)
Polymyxin co‐administration	Yes	20 (2.2%)	2 (0.7%)	14 (3.1%)	4 (2.5%)	0.07	.086
Furosemide use	Yes	187 (20.8%)	39 (13.4%)	86 (19.3%)	62 (38.0%)	0.21	<.001
NSAID use	Yes	423 (47.1%)	144 (49.7%)	197 (44.3%)	82 (50.3%)	0.06	.239
Any nephrotoxin within 48 h	Yes	521 (58.0%)	167 (57.6%)	240 (53.9%)	114 (69.9%)	0.12	.002
Nephrotoxins count within 48 h		1 [0–1]	1 [0–1]	1 [0–1]	1 [0–2]	0.02	<.001
Vancomycin exposure/TDM
Vancomycin AUC24 (mg·h/L)		452 [373–551]	339 [311–367]	471 [435–527]	692 [628–819]	0.74	<.001
Vancomycin AUC24 per 100 mg·h/L		4.5 [3.7–5.5]	3.4 [3.1–3.7]	4.7 [4.3–5.3]	6.9 [6.3–8.2]	0.74	<.001
Time from first dose to TDM (h)		14.1 [8.7–18.1]	12.9 [8.1–16.9]	15.1 [8.9–18.6]	14.2 [8.6–19.0]	0.00	.002
Infection variables
MRSA or S. aureus infection	Yes	375 (41.8%)	109 (37.6%)	201 (45.2%)	65 (39.9%)	0.07	.109
Deep‐seated infection	Yes	645 (71.8%)	214 (73.8%)	293 (65.8%)	138 (84.7%)	0.15	<.001
Bacteremia	Yes	100 (11.1%)	34 (11.7%)	45 (10.1%)	21 (12.9%)	0.04	.584
Baseline labs
Albumin (g/dL)		32.1 ± 5.6	32.3 ± 5.4	32.1 ± 5.8	32.0 ± 5.7	0.00	.847
White blood cell count		10.7 [8.2–14.6]	10.2 [8.0–14.2]	10.7 [8.2–14.9]	11.4 [8.3–14.2]	0.00	.334
C‐reactive protein		41.4 [12.1–107.5]	35.4 [12.2–120.1]	41.8 [12.0–96.9]	56.9 [12.8–136.8]	0.01	.177
Procalcitonin		2.4 [0.9–3.9]	2.3 [0.9–3.9]	2.3 [0.9–3.8]	2.8 [0.8–4.5]	0.01	.381
Outcome distribution by AUC group
AKI within 48 h	Yes	84 (9.4%)	20 (6.9%)	44 (9.9%)	20 (12.3%)	0.07	.146
KDIGO AKI stage within 48 h	No AKI (stage 0)	814 (90.6%)	270 (93.1%)	401 (90.1%)	143 (87.7%)	0.095	0.012
	Stage 1	66 (7.3%)	18 (6.2%)	37 (8.3%)	11 (6.7%)
	Stage 2	10 (1.1%)	1 (0.3%)	5 (1.1%)	4 (2.5%)
	Stage 3	8 (0.9%)	1 (0.3%)	2 (0.4%)	5 (3.1%)
AKI stage ≥2	Yes	18 (2.0%)	2 (0.7%)	7 (1.6%)	9 (5.5%)	0.12	.001
SCr ratio within 48 h		0.97 [0.83–1.14]	0.94 [0.82–1.09]	0.99 [0.84–1.16]	0.97 [0.79–1.15]	0.01	.088
ΔSCr within 48 h (mg/dL)		−0.02 [−0.16–0.10]	−0.04 [−0.18–0.07]	−0.01 [−0.15–0.12]	−0.03 [−0.22–0.14]	0.00	.086
Maximum SCr within 48 h (mg/dL)		0.87 [0.75–1.03]	0.85 [0.74–0.98]	0.87 [0.74–1.03]	0.96 [0.79–1.18]	0.05	<.001

AUC24, 24‐h area under the concentration‐time curve; SCr, serum creatinine.

Values are mean ± SD, median [IQR], or n (%). Effect size is eta‐squared for continuous variables and Cramér V for categorical variables.

In the primary regression analyses, higher continuous AUC24 was associated with AKI within 48 h in the unadjusted model (OR, 1.14 per 100 mg·h/L; 95% CI, 1.01–1.28; P = .029) and in the prespecified minimally adjusted model (OR, 1.16; 95% CI, 1.02–1.33; P = .023). The point estimate was similar but no longer statistically significant in the exploratory fully adjusted model (OR, 1.13; 95% CI, 0.98–1.31; P = .094). The forest plot makes this pattern visually clear: The continuous‐exposure estimates were consistently above 1.0, whereas the confidence interval widened and crossed the null in the more complex exploratory model (Table 3 and Figure 3).

Table 3.

Primary, Sensitivity, Spline, and Exploratory Interaction Analyses

Analysis	Exposure/Contrast	N	Events	OR (95% CI)	P Value	Interpretation
Model 1: Unadjusted	AUC24 per 100 mg·h/L	898	84	1.14 (1.01–1.28)	.029	Unadjusted
Model 2: Minimally adjusted	AUC24 per 100 mg·h/L	898	84	1.16 (1.02–1.33)	.023	Primary continuous exposure
Model 3: Fully adjusted (exploratory)	AUC24 per 100 mg·h/L	898	84	1.13 (0.98–1.31)	.094	Exploratory overadjusted model
Model 4: categorical AUC24	<400 vs 400–600	898	84	0.66 (0.37–1.18)	.159	Reference = 400–600 mg·h/L
Model 4: categorical AUC24	>600 versus 400–600	898	84	1.24 (0.68–2.26)	.475	Reference = 400–600 mg·h/L
Sensitivity: categorical AUC24	400–600 vs <400	898	84	1.51 (0.85–2.67)	.159	Reference = <400 mg·h/L
Sensitivity: categorical AUC24	>600 vs <400	898	84	1.88 (0.94–3.73)	.073	Reference = <400 mg·h/L
Sensitivity: excluding possible pre‐existing AKI	AUC24 per 100 mg·h/L	875	67	1.12 (0.96–1.31)	.159	23 episodes excluded; 17 excluded events
Restricted cubic spline	Overall association	898	84	Not applicable	.052	LRT covariate‐only vs spline model
Restricted cubic spline	Nonlinearity	898	84	Not applicable	.188	LRT linear vs spline model
Exploratory interaction	AUC24 × aminoglycoside exposure	898	84	1.70 (1.05–2.75)	.030	Modifier = 1: 40 episodes, 8 events
Exploratory interaction	AUC24 × baseline renal dysfunction	898	84	0.60 (0.40–0.89)	.012	Modifier = 1: 143 episodes, 10 events

AKI, acute kidney injury; AUC24, 24‐hour area under the concentration‐time curve; CI, confidence interval; LRT, likelihood‐ratio test; OR, odds ratio.

All logistic models used maximum likelihood estimation. Model 2 adjusted for age, sex, baseline serum creatinine, ICU admission, piperacillin/tazobactam, aminoglycoside exposure, and time from first vancomycin dose to TDM.

Figure 3.

Forest plot of the association between vancomycin AUC24 and AKI within 48 h. Odds ratios are shown for continuous AUC24 and categorical AUC24 models. Model 3 is exploratory.

Categorical modeling was less informative than continuous modeling. Using 400 to 600 mg·h/L as the reference group, neither AUC24 less than 400 mg·h/L (OR, 0.66; 95% CI, 0.37–1.18; P = .159) nor AUC24 greater than 600 mg·h/L (OR, 1.24; 95% CI, 0.68–2.26; P = .475) was significantly associated with AKI. When less than 400 mg·h/L was used as the reference, the OR was 1.51 (95% CI, 0.85–2.67; P = .159) for 400 to 600 mg·h/L and 1.88 (95% CI, 0.94–3.74; P = .073) for greater than 600 mg·h/L. The categorical estimates therefore suggested increasing point estimates across exposure groups but with wide confidence intervals, particularly in the highest AUC category (Table 3 and Figure 3).

The sensitivity analysis excluding 23 episodes flagged as possible pre‐existing AKI before the index TDM episode retained 875 episodes and 67 AKI events. In this conservative sensitivity analysis, the AUC24 association attenuated and was not statistically significant (OR, 1.12 per 100 mg·h/L; 95% CI, 0.96–1.31; P = .159). The excluded episodes had creatinine evidence of renal deterioration by the time of TDM, but the available data could not determine whether this deterioration predated vancomycin therapy, developed during vancomycin exposure before TDM, or reflected both processes. Accordingly, the attenuation supports possible reverse causality but may also reflect overcorrection if early vancomycin‐related renal injury had already begun before the index TDM episode (Table 3).

Restricted cubic spline modeling did not support a distinct nonlinear exposure‐risk threshold. The overall spline association was not significant at the.05 level (likelihood‐ratio P = .052), and the test for nonlinearity was not significant (P = .188). The predicted probability and odds‐ratio curves showed broad uncertainty, particularly near the lower and upper tails of the observed AUC24 distribution. Accordingly, the spline analysis was interpreted as exploratory and was not used to define a new toxicity threshold (Figure 4 and Table S4).

Figure 4.

Exploratory restricted cubic spline analysis of vancomycin AUC24 and acute kidney injury within 48 h. (A) The adjusted predicted probability of AKI, and (B) the adjusted odds ratio for AKI relative to AUC24 = 500 mg·h/L. Solid curves represent adjusted spline estimates, and shaded bands represent 95% confidence intervals. In (B), the horizontal dashed line denotes the null value of OR = 1. Vertical dashed lines at 400 and 600 mg·h/L indicate the conventional target‐range boundaries. The vertical dotted line at 500 mg·h/L identifies the value used solely as the OR reference and does not represent a toxicity threshold. Rug marks indicate the distribution of observed AUC24 values. Curves are displayed over the 5th to 95th percentile of the observed AUC24 distribution. Overall‐association and nonlinearity P values were obtained using likelihood‐ratio tests.

Exploratory interaction analyses suggested possible effect modification by aminoglycoside exposure and baseline renal dysfunction, but neither finding was considered confirmatory. The aminoglycoside‐exposed stratum included only 40 episodes and 8 AKI events; although the interaction test reached nominal statistical significance, the estimate was highly unstable and was not adjusted for multiple testing. Piperacillin/tazobactam exposure was even sparser (14 episodes, 2 events). These analyses are presented for transparency and require validation in substantially larger datasets; they should not be interpreted as evidence for subgroup‐specific dosing rules (Table S5 and Figures S2–S4).

Discussion

This multicenter routine‐care study examined how a Bayesian‐estimated AUC24 should be interpreted when it is generated after dosing and close to an evolving renal event. The principal finding was that higher post‐dose Bayesian‐estimated AUC24 was modestly associated with creatinine‐defined AKI within 48 h in a parsimoniously adjusted model. However, the association weakened after broader adjustment and, more importantly, after exclusion of episodes with possible pre‐existing AKI before the index TDM episode. Taken together, these findings indicate that an AUC24 estimate available at the time of TDM should be interpreted as a composite pharmacokinetic signal that reflects both administered vancomycin exposure and current renal clearance, rather than as a standalone causal toxicity threshold.

This interpretation is consistent with, but distinct from, the evidence that led to AUC‐guided vancomycin monitoring. The revised consensus guideline recommends AUC/MIC‐guided monitoring, generally targeting an AUC24/MIC of 400 to 600 for serious methicillin‐resistant Staphylococcus aureus infections when the MIC is assumed to be 1 mg/L. ¹ That recommendation was developed to improve the balance between efficacy and safety at the population level. Earlier trough‐based practice, particularly targeting trough concentrations of 15 to 20 mg/L, was associated with excessive exposure in some patients and higher nephrotoxicity risk. ⁶ , ⁷ , ⁸ , ⁹ Subsequent clinical studies and meta‐analyses have generally supported AUC‐guided monitoring as a strategy that can reduce nephrotoxicity compared with trough‐guided monitoring. ¹⁰ , ¹¹ , ¹² , ¹³ The present study does not challenge this therapeutic target framework. Instead, it addresses the narrower bedside interpretation of a Bayesian AUC value returned during routine care, after at least some vancomycin exposure has occurred and when renal function may already be changing. This distinction is important when comparing the present results with prior exposure–toxicity studies.

This pattern has a clear pharmacokinetic explanation. A Bayesian AUC estimate is not determined solely by the prescribed daily dose. It is also influenced by the observed concentration, the timing and reliability of the dosing and sampling record, the pharmacokinetic prior, and the patient‐specific posterior estimate of clearance. ¹⁰ , ¹⁴ , ¹⁵ When renal clearance begins to decline, a measured concentration can increase, and the Bayesian posterior AUC can rise even before serum creatinine satisfies a formal AKI definition. In that setting, a high AUC may be an exposure contributing to kidney injury, a consequence of early clearance decline, or both. This bidirectional relationship is especially plausible because serum creatinine is a delayed and imperfect renal biomarker influenced by creatinine generation, distribution volume, fluid balance, and the timing of sampling relative to kidney injury. ²³ , ²⁴ , ²⁵

The attenuation observed after excluding possible pre‐existing AKI should be viewed as clinically informative, but not as proof that the excluded episodes were unrelated to vancomycin. Their elevated creatinine at TDM may indicate renal dysfunction that existed before vancomycin treatment, early kidney injury that developed during vancomycin exposure before TDM, or a combination of both. Excluding these episodes therefore provides a conservative assessment of temporality and possible reverse causality, but it may also overcorrect the association by removing patients in whom vancomycin had already contributed to evolving injury. This distinction separates a design‐stage predicted AUC, intended to guide dosing prospectively, from a post‐dose Bayesian AUC that is partly conditional on the patient's recent concentration and renal trajectory. Future toxicodynamic studies should analyze initial predicted exposure, post‐dose Bayesian‐estimated exposure, time‐updated clearance, and renal injury markers separately.

Previous cohort studies and meta‐analyses have consistently linked higher vancomycin AUC with nephrotoxicity, although reported risk thresholds have varied below, near, and above 600 mg·h/L according to patient population, infection type, exposure window, AUC estimation method, and AKI ascertainment. ¹⁶ , ¹⁷ , ¹⁸ , ¹⁹ , ²⁰ , ²¹ , ²² These studies support a genuine exposure–toxicity relationship and the clinical importance of avoiding excessive vancomycin exposure. The present findings do not contradict that literature; rather, they address a different temporal construct. Many prior studies evaluated early cumulative, treatment‐day, steady‐state, or maximum AUC in relation to AKI over a subsequent observation period, whereas our exposure was a post‐dose Bayesian estimate generated during routine TDM and anchored closely to AKI ascertainment within 48 h. At that point, the posterior AUC may already incorporate declining renal clearance. Thus, the observed association can contain both a forward toxic effect of exposure and a reverse component in which evolving renal dysfunction increases the estimated AUC. The attenuation after excluding possible pre‐existing AKI should therefore not be interpreted as evidence that vancomycin exposure is non‐nephrotoxic. It indicates that an association based on a post‐TDM AUC cannot be assumed to represent a purely forward causal exposure effect.

The categorical analyses reinforce this point. Classifying AUC24 into less than 400, 400 to 600, and greater than 600 mg·h/L is clinically familiar because these categories align with conventional target discussions, but categorization reduces information and assumes abrupt changes in risk at fixed cut points. The target range of 400 to 600 mg·h/L should remain an efficacy–safety dosing target for appropriate serious infections; it should not be reinterpreted as a diagnostic boundary for imminent AKI. Conversely, an AUC above 600 mg·h/L should prompt careful review, but the decision to reduce, hold, or continue therapy should account for infection severity, organism susceptibility, renal trajectory, hemodynamics, and alternative antimicrobial options.

Clinically, an unexpectedly high AUC should prompt a structured review rather than automatic dose adjustment based on the value alone. When the estimated AUC24 is unexpectedly high, clinicians should verify the dosing history, infusion duration, sample timing, assay result, and Bayesian model plausibility; review the recent serum creatinine trajectory and urine output when available; assess hemodynamic instability and volume status; and reduce avoidable nephrotoxin exposure. This approach is consistent with model‐informed precision dosing, in which software output supports but does not replace clinical pharmacology judgment. ¹⁴ , ¹⁵ It is also particularly relevant in health systems where AUC‐guided monitoring is implemented gradually and where turnaround time, pharmacist review, and prescriber acceptance can vary across calendar periods and clinical units.

Concomitant nephrotoxins remain a key interpretive issue. The aminoglycoside interaction was biologically plausible because combined vancomycin and aminoglycoside therapy has been associated with nephrotoxicity, ²⁸ but it was based on only 40 exposed episodes and 8 AKI events. The nominal interaction P value was therefore unstable, was not corrected for multiple testing, and should be regarded as hypothesis‐generating only. Piperacillin/tazobactam exposure was too sparse to evaluate reliably in this cohort. Observational studies have reported higher AKI risk with vancomycin plus piperacillin/tazobactam, whereas randomized evidence comparing piperacillin/tazobactam with cefepime in hospitalized adults did not show an increase in the highest‐stage AKI or death outcome. ²⁹ , ³⁰ The present data should not be used to define subgroup‐specific dosing rules for these combinations.

The findings also have implications for reporting and evaluating Bayesian dosing programs. Studies should report the population pharmacokinetic prior when available, assay method and turnaround time, sampling strategy, procedures for adjudicating uncertain sampling or dosing records, pharmacist review workflow, prescriber acceptance, and the timing of dose modification after the AUC result. Without these operational details, an AUC‐guided program may be difficult to reproduce, and its renal outcomes may be difficult to compare across institutions. The clinician verification process used in the present study was therefore not a peripheral quality‐control step; it was part of the pharmacometric workflow that made the AUC estimate clinically interpretable.

This study has limitations. The retrospective design leaves residual confounding, confounding by indication, and unmeasured time‐varying severity possible. Model 3 adjusted for available baseline and contemporaneous markers, including ICU admission, septic shock, vasopressor exposure, baseline renal function, and concomitant nephrotoxic medications. However, standardized time‐updated data on fluid‐resuscitation volume, changes in hemodynamic instability, and the exact initiation or discontinuation times of additional nephrotoxic medications between TDM and AKI ascertainment were not consistently available. Residual confounding by these dynamic factors could therefore have exaggerated or attenuated the observed association. Serum creatinine was measured during routine care rather than at protocolized intervals, and urine output criteria were unavailable. The absence of urine output may have caused patients meeting oliguria criteria without a qualifying creatinine increase to be classified as non‐AKI, tending to underestimate AKI incidence and potentially attenuate associations. Conversely, creatinine changes related to fluid balance, hemodilution, hemoconcentration, or altered creatinine generation could cause misclassification in either direction. Prospective AUC‐guided dosing was not implemented as a uniform intervention throughout the entire study period. The embedded population pharmacokinetic prior and software build for every historical estimate could not be fully reconstructed. The first eligible TDM episode per encounter reduced within‐encounter dependence but did not eliminate all patient‐level clustering. The number of AKI events limited the reliability of extensively adjusted, spline, and interaction analyses. Finally, the study was conducted in two private hospitals in one Vietnamese city; external validity to other care settings, microbiology patterns, assays, and Bayesian platforms requires caution.

A prospective study should therefore capture the initial predicted AUC at dose design, each post‐dose Bayesian AUC, exact concentration sampling time, assay turnaround time, prescriber action, time‐updated serum creatinine, urine output, aminoglycoside concentrations when co‐administered, and competing causes of renal dysfunction. Such a design would better separate exposure‐mediated nephrotoxicity from early clearance decline and would allow stronger causal inference. Until such data are available, a post‐TDM Bayesian‐estimated AUC24 should be treated as a renal risk signal that warrants expert clinical review, not as an isolated diagnosis of toxicity or a reason to abandon efficacy‐oriented exposure targets.

Conclusions

Bayesian‐estimated vancomycin AUC24 at a routine TDM episode was a modest near‐term signal of creatinine‐defined AKI within 48 h. Because the association attenuated after excluding episodes with possible pre‐existing AKI, post‐dose AUC24 likely reflects both vancomycin exposure and contemporaneous renal clearance. A high post‐TDM AUC should therefore prompt structured renal and pharmacokinetic review, but should not be interpreted as a standalone causal toxicity threshold.

Author Contributions

Bui Thi Tham and Tran Van Anh contributed to data curation, analysis, investigation, methodology, visualization, and drafting. Tran Thi Ngan, Nguyen Duc Long, and Nguyen Huong Giang contributed to investigation, data curation, validation, resources, and manuscript review. Nguyen Hoang Anh and Vu Dinh Hoa contributed to methodology, pharmacokinetic and statistical interpretation, supervision, and critical revision. Hoang Van Dung contributed to clinical interpretation, resources, and supervision. Nguyen Thi Thu Phuong conceived the study, supervised the project, provided clinical oversight, and critically revised the manuscript. All authors reviewed and approved the final manuscript.

Conflicts of Interest

The authors declare no conflicts of interest. SmartDoseAI and BestDose were used as clinical decision‐support and pharmacokinetic review tools; final clinical interpretation was performed by clinicians and clinical pharmacists.

Funding

No external funding was received for this study.

Principal Investigator Statement

The authors confirm that Nguyen Thi Thu Phuong was the Principal Investigator for this study and had direct clinical responsibility for patients.

Supporting information

Supporting Information: jcph70265‐sup‐0001‐SupMat.docx

Supporting Information:

Supporting Information:

Supporting Information:

Supporting Information:

Data Availability Statement

The de‐identified analytic dataset and analysis code supporting the findings of this study are publicly available in Zenodo: Nguyen TTP, Tran VA, Ngan T. Code and dataset for: Bayesian‐estimated vancomycin AUC24 and early acute kidney injury after therapeutic drug monitoring. Data set. Zenodo. April 4, 2026. https://doi.org/10.5281/zenodo.19415578