Vancomycin levels for Bayesian dose-optimization in critical care: a prospective cohort study

Abstract

Background: Vancomycin dosing in critically ill patients typically requires monitoring the area under the concentration-time curve/minimum inhibitory concentration (AUC/MIC), often using at least two vancomycin levels (VLs). However, the optimal number of VLs needed for accurate AUC/MIC estimation in this population remains uncertain. This study aimed to determine the minimum number of VLs required to accurately estimate the AUC/MIC in critically ill patients treated with intermittent infusion of vancomycin.

Methods: A prospective cohort study was conducted in critically ill patients, where VLs were obtained at peak, beta, and trough phases. Five AUC estimates were derived using PrecisePK^™, a Bayesian software: AUC-1 [peak, beta (2 h after the end infusion), trough], AUC-2 (beta, trough), AUC-3 (peak, trough), AUC-4 (trough), and AUC-5 (only Bayesian prior, without VL). These estimates were compared for accuracy and bias (mean ± SEM) against the reference AUC calculated via the trapezoidal model (AUC_Ref).

Results: We enrolled 36 adult patients with age of 65 (52-77) years, moderate severity [APACHE II 10 (5-14) and SOFA 5 (4-6)], 6 of them in ECMO and 4 in renal replacement therapy. A total of 108 blood samples for VL were analyzed. The AUC-3 (0.976 ± 0.012) showed greater accuracy compared to AUC-4 (1.072 ± 0.032, p = 0.042) and AUC-5 (1.150 ± 0.071, p = 0.042). AUC-3 also demonstrated lower bias (0.053 ± 0.009) than AUC-4 (0.134 ± 0.026, p = 0.036) and AUC-5 (0.270 ± 0.060, p = 0.003). Bland-Altman analysis indicated better agreement between AUC-3 and AUC-2 with AUC_Ref.

Conclusion: Bayesian software using two vancomycin levels provides a more accurate and less biased AUC/MIC estimation in critically ill patients.

Keywords: antibiotics; area under curve/minimum inhibitory concentration; glycopeptides; intensive care unit; pharmacokinetics; sepsis.

Figure 1

Timing of vancomycin levels in a schematic representation, where peak level was measured 20 min post-infusion completion; beta level was measured 2 h post-infusion completion; and trough level was measured 1 h before the next scheduled vancomycin dose. The distribution phase reflects the rapid decline due to drug distribution into tissues, followed by the beta level, marking the beginning of the elimination phase, where the drug is excreted. The trough level denotes the lowest concentration before the next dose, critical for therapeutic monitoring.

Figure 2

Flowchart illustrating the selection process of patients included in the vancomycin pharmacokinetic analysis. Out of 167 patients who received vancomycin, 131 were excluded because they did not have three plasma level measurements required for AUC/MIC calculation. The remaining 36 patients with complete pharmacokinetic data were included in the analysis and categorized according to their AUC/MIC ratio into subtherapeutic (n = 17), therapeutic (n = 14), and supratherapeutic (n = 5) groups.

Figure 3

Mean values ± standard error of the mean (SEM) for each experimental condition (AUCref, AUC1–AUC5). Statistical comparisons were performed against the reference group (AUCref), with symbols indicating significant differences: ^*p < 0.05, ^**p < 0.01, and ^***p < 0.001.

Figure 4

Correlation between observed and predicted values using different pharmacokinetic models. Graphics A–E show scatter plots comparing observed versus predicted drug concentrations for five different modeling approaches. Each graphic reports the coefficient of determination (R²), Pearson correlation coefficient (r), and significance level (p < 0.001). The dashed line represents the line of identity. Model A shows the strongest correlation (R² = 0.946, r = 0.973), while model E shows the weakest performance (R² = 0.287, r = 0.536), indicating variability in predictive accuracy among the tested models.

Figure 5

Bland–Altman plots comparing AUC estimations across different pharmacokinetic models. Graphics A–E depict Bland–Altman plots assessing the agreement between predicted and reference AUC values for five different pharmacokinetic models. The y-axis represents the percentage difference relative to the reference AUC, calculated as [(AUC_(estimated) − AUC_ref/AUC_ref)]. The solid red line indicates the mean bias, while the dashed red lines represent the 95% limits of agreement. Narrower limits and points closer to zero indicate better agreement. Graphic A shows the best agreement with minimal bias and tight limits, while graphic E displays greater variability and wider limits, indicating poorer concordance.

Figure 6

(A) Accuracy, where ^&p < 0.01 vs. AUC-2, and p < 0.05 vs. AUC-4; ^%p < 0.01 vs. AUC-1, and AUC-3; ^*p < 0.01 vs. AUC-4, and p < 0.05 vs. AUC-5; ^#p < 0.01 vs. AUC-3, and p < 0.05 vs. AUC-1; ^$p < 0.05 vs. AUC-3. (B) Bias, where ^&p < 0.01 vs. AUC-2, AUC-4, AUC-5, and p < 0.05 vs. AUC-3; ^%p < 0.01 vs. AUC-1, and p < 0.05 vs. AUC-3, AUC-4, and AUC-5; ^*p < 0.001 vs. AUC-5, and p < 0.01 vs. AUC-4, and p < 0.05 vs. AUC-1, and AUC 2; ^#p < 0.01 vs. AUC-1, AUC-3, AUC-5, and p < 0.05 vs. AUC-2; ^$p < 0.001 vs. AUC-1, AUC-2, AUC-3, and AUC-4.