A Bayesian Dose-Individualization Method for Warfarin

• Published in: Clinical pharmacokinetics 2013-01, Vol.52 (1), p.59-68
• Main Authors: Wright, Daniel F. B., Duffull, Stephen B.
• Format: Article
• Language: English
• Subjects: Adult; Aged; Aged, 80 and over; Anticoagulants - administration & dosage; Anticoagulants - pharmacokinetics; Bayes Theorem; Biological and medical sciences; Dose-Response Relationship, Drug; Female; General pharmacology; Hemorrhage - chemically induced; Hemorrhage - prevention & control; Humans; Internal Medicine; International Normalized Ratio; Male; Medical sciences; Medicine; Medicine & Public Health; Metabolic Clearance Rate; Middle Aged; Models, Biological; Original Research Article; Pharmacokinetics. Pharmacogenetics. Drug-receptor interactions; Pharmacology. Drug treatments; Pharmacology/Toxicology; Pharmacotherapy; Prospective Studies; Software; Warfarin - administration & dosage; Warfarin - pharmacokinetics
• ISSN: 0312-5963; 1179-1926
• DOI: 10.1007/s40262-012-0017-6

Background Warfarin is a difficult drug to dose accurately and safely due to large inter-individual variability in dose requirements. Current dosing strategies appear to be sub-optimal, with reports indicating that patients achieve international normalized ratios (INRs) within the therapeutic range only 40–65 % of the time. The consequences of poor INR control are potentially severe with INRs below 2 carrying an increased risk of clotting while INRs >4 increase the risk of major bleeding events. Bayesian forecasting methods have the potential to improve INR control. Aims The aims of this study were to (1) prospectively assess the predictive performance of a Bayesian dosing method for warfarin implemented in TCIWorks; and (2) determine the expected time in the therapeutic range (TTR) of INRs predicted using TCIWorks. Methods Patients who were initiating warfarin therapy were prospectively recruited from Dunedin Hospital, Dunedin, New Zealand. Warfarin doses were entered into TCIWorks from the first day of therapy until a stable steady-state INR (INR ss ) was achieved. The predicted INR ss values were determined using the first zero to six serially collected INR observations. Observed and predicted INR ss values were compared using measures of bias (mean prediction error [MPE]) and imprecision (root mean square error [RMSE]). The TTR was determined by calculating the percentage of predicted INR ss values between 2 and 3 when zero to six serially collected INR observations were available. Results A total of 55 patients were recruited between March and November 2011. When no observed INR values were available the resulting INR ss predictions were positively biased (MPE 0.52 [95 % CI 0.30, 0.73]); however, this disappeared once observed INR values were entered into TCIWorks. The precision of the predicted INR ss values improved dramatically once three or more observed INR values were available (RMSE