Evaluation of a Bayesian Method for Predicting Vancomycin Dosing

Abstract

The purpose of this study is to evaluate the performance of a vancomycin dosing program in predicting dosages necessary to achieve desired serum vancomycin concentrations in a relatively large patient population. With the completion of initial performance evaluation, revised pharmacokinetic parameter estimates derived in the initial evaluation are used to reevaluate program performance. The program uses population estimates of vancomycin's volume of distribution (Vd) and clearance (Cl) to initially predict dosing, then individualizes those estimates by a Bayesian algorithm (iterations) which uses dosing and the resulting serum vancomycin concentration data. Use of the Bayesian forecaster with one iteration significantly increases the calculated Cl value as compared with population estimates; two and three iterations significantly increase both Vd and Cl when compared with population estimates. Absolute values of the predicted minus observed peak serum vancomycin concentrations (accuracy) are 17.7 ± 14.0, 6.1 ± 3.6, and 3.4 ± 2.1 mg/L for dosing using population estimates, Bayesian with one iteration, and Bayesian with two iterations, respectively. Similarly, accuracy of predictions for trough concentrations is 13.8 ± 12.4, 3.5 ± 3.2, and 3.2 ± 2.6 mg/L for each method, respectively. Bias of dosing predictions in achieving desired peak and trough serum vancomycin concentrations is also significantly reduced by using the Bayesian algorithm. Use of the mean Vd and Cl values from three iterations as the starting parameters in a new group of 12 patients significantly improves program performance when compared with use of initial population parameters. Time of sampling for peak serum concentrations has no effect on program performance. In patients with impaired renal function, use of population estimates resulted in less accurate dosing prediction, but this less accurate performance was not observed with use of the Bayesian forecaster. These data demonstrate the accuracy and lack of bias in individualized dosing predictions using the Bayesian dosing method and the ability of revised pharmacokinetic parameter estimates to improve performance.