Performance of a Theory-Based Mechanistic Model for Predicting the Target Dose of Warfarin

Background:

Warfarin is widely used as a treatment of venous thromboembolism and its capability to reduce the hazard of thromboembolic events has been unequivocally demonstrated. Variability between individuals, as well as a narrow therapeutic range are barriers to safe and effective warfarin therapy. Inadequate dose individualization contributes to under-utilization, 18-55% of patients who would benefit from warfarin do not receive it, and the amount of time spent within the therapeutic range is sub-optimal for many that do receive warfarin.

Bayesian dose individualization using a target concentration intervention dosing strategy may be superior to (conventional) empirical warfarin dose individualization and may lead to improved clinical outcomes (1, 2). A theory-based mechanistic model was recently used to describe the pharmacokinetics and pharmacodynamics of S- and R- warfarin using the International Normalized Ratio (INR) (3). Whether this model can accurately and precisely predict warfarin dose has not yet been examined.

Objective:

Use simulation and estimation techniques to evaluate the predictive performance and potential clinical utility of the theory-based mechanistic model.

Methods: 

A simulation-estimation procedure implemented in NONMEM 7.4.1 was used to individualize warfarin doses in 1000 simulated patients. All patients were initiated on an initial warfarin dose of 6 mg on day 1 and 3 mg on days 2 and 3. The model was used to individualize doses after each INR measurement on days 3, 7, 10, 14, 21, 28, 35, 42, 49, and 56. The method is available at https://www.nextdose.org.

The predictive performance of the model was quantified using measures of bias (mean prediction error, ME) and imprecision (root mean square error, RMSE). The clinical utility of the model was determined using the percentage of time within, above and below the therapeutic range (INR 2.0-3.0). This was determined by numerical integration of the model predicted time-course of INR.

Results:

Model predictions of the target dose were initially biased (ME: -0.56 mg/day; 95% CI: -0.59, -0.52 mg/day) and imprecise (RMSE: 0.4 mg/day). This diminished following INR measurements and dose adjustments. After six INR measurements and dose updates over 28 days, predictions were both accurate (ME: -0.06 mg/day; 95% CI: -0.18, 0.07 mg/day) and precise (RMSE: 0.66 mg/day). During days 4 to 14 the percentage of time within therapeutic range was 42% while during days 15 to 28 the percentage was 75%.

Conclusion:

The findings of this investigation suggest that warfarin dose individualization using the evaluated model and the target concentration intervention may be a useful method for warfarin management. Investigations using an external data set and randomized control trial are needed for the model to be recommended for routine use in the clinic.

References:

1. Saffian SM, Duffull SB, Roberts RL, Tait RC, Black L, Lund KA, et al. Influence of Genotype on Warfarin Maintenance Dose Predictions Produced Using a Bayesian Dose Individualization Tool. Therapeutic drug monitoring. 2016;38(6):677-83.

2. Motykie GD, Mokhtee D, Zebala LP, Caprini JA, Kudrna JC, Mungall DR. The use of a Bayesian forecasting model in the management of warfarin therapy after total hip arthroplasty. The Journal of arthroplasty. 1999;14(8):988-93.

3. Xue L, Holford N, Ding XL, Shen ZY, Huang CR, Zhang H, et al. Theory-based pharmacokinetics and pharmacodynamics of S- and R-warfarin and effects on international normalized ratio: influence of body size, composition and genotype in cardiac surgery patients. British journal of clinical pharmacology. 2017;83(4):823-35.

Guangda Ma

  • The University of Auckland