Abstract Objective: Software incorporating warfarin pharmacokinetics and pharmacodynamics, as well as a Bayesian regression method, was evaluated for accuracy in predicting steady-state dosages of warfarin during initialization of anticoagulation therapy. Design: A cohort study was used to compare computer predictions with physician-determined doses during a retrospective chart review of 42 patients. Patients: Forty-two patient charts were selected for monitoring anticoagulation initialization therapy. Of the 42 patients reviewed, 22 were excluded on the following bases: uncontrolled congestive heart failure; ongoing treatment with medications that interfered with warfarin metabolism or displaced warfarin from protein binding sites; recent treatment with fresh frozen plasma or vitamin K; or any bleeding disorders, sepsis, malabsorption or significant changes in liver and/or renal function. Main Outcome Measures: Using physician-determined International Normalized Ratios (INRs) as target levels for the software, the number of INR feedbacks needed to stabilize blood drug levels during initialization were compared between physicians and computer forecasting. Population parameters and sequentially measured INRs were used for evaluation. Results: Current oral anticoagulation protocols require significantly more measured INRs than computer forecasting to achieve steady-state dosages (9.5 v 4.4, respectively, p < 0.01). Physicians showed a statistically significant pattern of underdosing patients (4.3 v 1.7 subtherapeutic INRs, respectively, p < 0.01). Computer predictions tended to overdose patients but did not significantly increase the number of supratherapeutic INRs. This study showed that five INR inputs consistently gave accurate steady-state dosage predictions, and in many ways computer modeling was more effective than clinician-determined steady-state dosing in warfarin initialization. Conclusions: Computer modeling provides a reliable means of predicting initialization dosages for warfarin anticoagulation therapy with five INR inputs and therefore has merit in the clinical setting if used with sound clinical judgment.