Evaluation of a Bayesian Regression Program for Predicting Warfarin Response

    loading  Checking for direct PDF access through Ovid

Abstract

Summary: The ability of a Bayesian regression program (Warfcalc) to predict warfarin response was evaluated retrospectively in 48 inpatients and prospectively in 10 inpatients. The prothrombin ratio (PR) on the last day of inpatient therapy was predicted using zero (naive) to five sequential, daily PR feedbacks. Bias and precision were measured using mean error (ME) and mean absolute error (MAE), respectively. Root mean squared error (RMSE) was used as a combined measure of bias and precision. In the retrospective group, the use of five PR feedbacks yielded the lowest ME, MAE, and RMSE (0.22, 0.30, and 0.45, respectively). The use of two and three daily PR feedbacks resulted in larger prediction errors compared with the use of naive parameters. Further evaluation of the retrospective patient data indicated that deletion of PR feedbacks associated with an activated partial thromboplastin time >100 s and exclusion of metabolic inhibitors in the estimation of warfarin clearance resulted in more reliable predictions (ME = 0.07, MAE = 0.20, RMSE = 0.28). Similarly, deletion of such PR feedbacks and metabolic inhibitors from the prospective data and use of PRs for the first 5 days of therapy yielded ME, MAE, and RMSE values of 0.07,0.21, and 0.27, respectively. The variance for prothrombin complex activity (PCA) as a function of the variance in the prothrombin time (PT) was investigated using Monte Carlo simulation assuming four different random error models for the PT measurements. These error models yielded functions that exhibit a maximum coefficient of variation at PCA values of 40–70%. Thus, the use of PR feedbacks during the first few days of warfarin therapy may yield unreliable predictions unless they are appropriately weighted.

Related Topics

    loading  Loading Related Articles