Bayesian Pharmacokinetic/Pharmacodynamic Forecasting of Prothrombin Response to Warfarin Therapy

The ability of a pharmacokinetic/pharmacodynamic Bayesian forecasting computer program to predict prothrombin response to warfarin therapy was investigated. The performance of the program was evaluated retrospectively in an inpatient study population of 45 subjects. Predictions of prothrombin response at discharge, based on zero to five serially measured prothrombin ratios, were compared. Precision of prediction was measured by root mean squared error (rmse), bias was measured by average prediction error, and significance (p less than 0.05) was determined by 95% confidence intervals and correlation coefficients. Eleven (3.8%) predictions exceeded established limitations of the pharmacokinetic/pharmacodynamic model and were excluded from data analysis. Correlations between measured and predicted prothrombin ratios for all methods were significant. The five prothrombin ratio feedbacks provided the most accurate predictions (rmse 0.219). These predictions were significantly better than the population parameter (rmse 0.418), one (rmse 0.401), and two (rmse 0.459) prothrombin ratio feedback predictions. The predictions based on population parameters and one prothrombin ratio feedback were significantly biased. When provided with sufficient feedback, the bias was not apparent and the predictive performance improved with each additional prothrombin ratio. The predictive performance of the four and five prothrombin ratio feedbacks is sufficient to provide clinically useful dosage guidelines early in the course of warfarin therapy. The population parameter estimates require further delineation in order to improve the performance of limited prothrombin ratio feedback predictions.