Objectives Several population pharmacokinetics (popPK) models for polymyxin B have been constructed to optimize therapeutic regimens. However, their predictive performance remains unclear when extrapolated to different clinical centers. Therefore, this study aimed to evaluate the predictive ability of polymyxin B popPK models. Methods A literature search was conducted, and the predictive performance was determined for each selected model using an independent dataset of 20 patients (92 concentrations) from the Third Xiangya Hospital. Prediction- and simulation-based diagnostics were used to evaluate model predictability. The influence of prior information was assessed using Bayesian forecasting. Results Eight published studies were evaluated. In prediction-based diagnostics, the prediction error within ± 30% was over 50% in two models. In simulation-based diagnostics, the prediction- and variability-corrected visual predictive check (pvcVPC) showed satisfactory predictivity in three models, while the normalized prediction distribution error (NPDE) tests indicated model misspecification in all models. Bayesian forecasting demonstrated a substantially improvement in the model predictability even with one prior observation. Conclusion Not all published models were satisfactory in prediction- and simulation-based diagnostics; however, Bayesian forecasting improved the predictability considerably with priors, which can be applied to guide polymyxin B dosing recommendations and adjustments for clinicians.