Data mining to assess variations in oral anticoagulant treatment

Variations in International Normalized Ratio's (INR) are closely related to bleeding and thrombosis incidents in patients on oral anticoagulation treatment. This study investigates predictive factors that affect INR values. Data sampled with relatively high frequency allows for detection of local INR variations, and hence also allows detection and evaluation of predictive factors where time is taken into consideration. Univariate linear regression was applied and different models were reduced into a final predictive model. F-tests were utilized to test whether or not a model reduction would benefit INR predictions, in terms of decreasing observed variance. In addition to an INR submodel, the final model includes individual interaction from the last three days change in mean warfarin intake and three days change in mean vitamin K intake. Prediction residual error was mainly reduced by the INR submodel, while the warfarin model and the vitamin K submodel did not benefit predictions to same extend compared to the INR submodel. However, more studies on the temporal aspects of the effect of warfarin seem to be relevant.