A Performance Evaluation of Bayesian Forecasting for Serum Linezolid Concentrations: Impact of Clinical Factors on Prediction Accuracy

BACKGROUND: Pycsim, a dosing software for linezolid (LZD) based on Bayesian estimation, predicts dosing in the therapeutic drug monitoring (TDM) of LZD. However, its predictive performance, examined using an independent data set with a limited sample size, must be assessed for further external validation before clinical implementation. This multi-institutional study assessed the predictive performance of Pycsim for LZD concentration. The impact of clinical factors on the prediction accuracy was also evaluated. METHODS: A multicenter prospective study across three institutions was conducted to assess the predictive performance of Pycsim using 87 LZD trough concentrations from 30 patients. The predictive performance was assessed based on the observed and Bayesian-predicted concentrations using linear regression, observed/predicted ratios, mean bias error, mean absolute error, and root mean square error. Multivariate logistic regression was used to identify factors influencing predictive accuracy. RESULTS: A strong correlation was observed between measured and Bayesian-predicted trough concentrations (R2 = 0.82). The geometric mean observed/predicted ratio was 0.96, with mean bias error, mean absolute error, and root mean square error of -1.2, 3.3, and 6.0, respectively. The sampling interval (1-7 days) did not affect the prediction accuracy. A multivariate analysis identified body weight