WARFARIN DOSE PREDICTION THROUGH A USER INTERFACE USING CLINICAL, DEMOGRAPHIC AND PHARMACOGENETIC DATA
Jennifer Eliana Montoya Neyra, Júlia Soler and Paulo Santos
We work on a user interface (UI) designed to assist in warfarin therapy by predicting a best therapeutic dose, calculated from the data entered into the UI. It will be able to predict more accurate doses for patients diagnosed with atrial fibrillation, stroke, thrombosis or heart valve prosthesis in whom it is desired to maintain an international normalized ratio (INR) between two and three, using their clinical, demographic and pharmacogenetic data. The prediction models that were considered for the construction of this prediction interface were the International Warfarin Pharmacogenetics Consortium (IWPC), multiple linear regression, regression using regularizers (Lasso regression, Ridge regression), Elastic net regression, regression of selected variables by AIC, Ridge Regression with Variable Selection (foba package in R) and a simple neural network model that consists of 3 hidden layers of 100 neurons each, using data from patients of the Brazilian Heart Institute (InCor – USP). This data include clinical, demographic, and pharmaceutical factors and and risk genotypes of the cytochrome P450 2C9 (CYP2C9), vitamin K epoxy reductase (VKORC1), leukotriene B(4) omega-hydroxylase 1 (CYP4F2) and NAD(P)H dehydrogenase (quinone) 1 (NQO1) genes. The models were trained with the information of 614 individuals, which reached INR values between 2 and 3 when receiving a maintenance dose of warfarin, and tested in a subset of 152 patients. To evaluate the accuracy of the models, the mean absolute error (MAE), root-mean-square error (RMSE) and R-squared were calculated. The best adjusted model was the Ridge Regression with Variable Selection, which obtained the best performance when analyzing both the training group (MAE = 7.54, RMSE = 0.993, R-squared = 0.296) and the evaluation group (MAE = 0.766, RMSE = 1.07, R-squared = 0.282). This tool is still under development, but we have great expectations about its applicability and usefulness for patients who require it.
Read More