Effect of using warfarin plasma concentrations in Bayesian forecasting of prothrombin-time response

The predictive performance of a Bayesian computer program using prothrombin-time (PT) response data with and without warfarin plasma concentrations to forecast patients' PTs at the time of hospital discharge was evaluated. A log-linear pharmacokinetic-pharmacodynamic model was used to describe and predict warfarin dose response in patients recently started on warfarin sodium. Individual patients' pharmacodynamic variables relating warfarin concentration to clotting-factor synthesis were obtained by Bayesian nonlinear regression analysis. Pharmacokinetic values for warfarin clearance and volume of distribution were either calculated using nonlinear regression from measured plasma warfarin concentrations or estimated based on literature-derived population regression equations. Percent mean absolute prediction error (precision) and prediction error (bias) for PT predictions were compared among and between analysis methods that used only literature data to estimate PT response and methods that used zero to five PTs with or without warfarin plasma concentrations. Eleven women and eight men completed the study. Predictions after four days of warfarin therapy using PT measurements beginning after either the first or third warfarin dose were clinically useful regardless of whether warfarin concentrations were used in the predictions. Predictions using fewer than four PT measurements were imprecise and biased. The Bayesian method in this study provided good predictions of PTs immediately before hospital discharge based on warfarin dosing and PT response after either four or five days of therapy. The use of warfarin plasma concentrations in the pharmacokinetic-pharmacodynamic model used here appears unwarranted.