Predictive Performance of Bayesian Vancomycin Monitoring in the Critically Ill*

Abstract

OBJECTIVES: It is recommended that therapeutic monitoring of vancomycin should be guided by 24-hour area under the curve concentration. This can be done via Bayesian models in dose-optimization software. However, before these models can be incorporated into clinical practice in the critically ill, their predictive performance needs to be evaluated. This study assesses the predictive performance of Bayesian models for vancomycin in the critically ill. DESIGN: Retrospective cohort study. SETTING: Single-center ICU. PATIENTS: Data were obtained for all patients in the ICU between 1 January, and 31 May 2020, who received IV vancomycin. The predictive performance of three Bayesian models were evaluated based on their availability in commercially available software. Predictive performance was assessed via bias and precision. Bias was measured as the mean difference between observed and predicted vancomycin concentrations. Precision was measured as the sd of bias, root mean square error, and 95% limits of agreement based on Bland-Altman plots. INTERVENTIONS: None. MEASUREMENTS AND MAIN RESULTS: A total of 466 concentrations from 188 patients were used to evaluate the three models. All models showed low bias (–1.7 to 1.8 mg/L), which was lower with a posteriori estimate (–0.7 to 1.8 mg/L). However, all three models showed low precision in terms of sd (4.7–8.8 mg/L) and root mean square error (4.8–8.9 mg/L). The models underpredicted at higher observed vancomycin concentrations (bias 0.7–3.2 mg/L for < 20 mg/L; –5.1 to –2.3 for ≥ 20 mg/L) and the Bland-Altman plots showed a great deviation between observed and predicted concentrations. CONCLUSIONS: Bayesian models of vancomycin show not only low bias, but also low precision in the critically ill. Thus, Bayesian-guided dosing of vancomycin in this population should be used cautiously.

Comparison of Bayesian‐derived and first‐order analytic equations for calculation of vancomycin area under the curve

“…Additionally, Narayan and colleagues assessed the predictive performance of Bayesian modeling for IV vancomycin and found all Bayesian models exhibited low bias but also considerably low precision in the critically ill patient population. 23 Our findings were similar as, on average, the three methods maintained reasonable correlation and low MD but demonstrated fluctuating variability, particularly at the extremes of AUC 24 .…”

“…However, the assumption that one‐concentration Bayesian is equivalent to two‐concentration Bayesian only remains true if one were to assume that the “gold‐standard” for calculation of AUC 24 is the Bayesian two‐concentration method. Additionally, Narayan and colleagues assessed the predictive performance of Bayesian modeling for IV vancomycin and found all Bayesian models exhibited low bias but also considerably low precision in the critically ill patient population 23 . Our findings were similar as, on average, the three methods maintained reasonable correlation and low MD but demonstrated fluctuating variability, particularly at the extremes of AUC 24 .…”

“…Correlations between AUC 24 estimated by all three methods (linear, Bayesian two‐concentration, and Bayesian one‐concentration) were assessed using Pearson's correlation. To compare the calculated AUC 24 values, we employed a similar approach to that of Narayan et al 23 Agreement was estimated using the mean difference (MD) between Bayesian and linear AUC 24 methods while the 95% limits of agreement (LOA) based on Bland–Altman plots were used to determine variability.…”

PK-RNN-V E: A Deep Learning Model Approach to Vancomycin Therapeutic Drug Monitoring Using Electronic Health Record Data

“…After being validated in various patient subpopulations, multiple Bayesian dose-optimizing software programs are available. [5,6]…”

“…[9] Second, these models are often evaluated in patients with stable PK parameters and may not cover critically ill conditions well, in which the volume of distribution and the elimination rates fluctuate acutely. [5,10,11] Finally, these Bayesian methods take only a limited number of patient-specific variables as input, including simple demographics, creatinine levels, vancomycin doses, the infusion time, and vancomycin levels, while there are potentially other relevant patient characteristics, such as other concomitant medications and vital signs, that potentially improve the prediction. [5] Therefore, more powerful and flexible models, such as deep learning models, provide significant advantages, as the models can integrate a wide range of patient-specific features, use flexible time steps, update the model with a local patient population, and cover a wide variety of populations.…”