"WarfarinSeer": a predictive tool based on SMOTE random forest to improve warfarin dose prediction in Chinese patients

Warfarin daily dosage prediction for a specific patient is difficult. To solve the data imbalance and improve the predictive accuracy on Warfarin daily dosage, we develop a dosage predictive tool called "WarfarinSeer", which is based on Synthetic Minority Oversampling Technique-Random Forest (SMOTE-RF) model. In STMOE-RF, STMOE is adopted to oversample the data, which have rare genotypes (i.e., *1/*3 and *3/*3 for CYP2C9, AG and GG for VKORCI), to produce new samples by using k-Nearest Neighbor. Random forest produces a group of trees by training them on minority and the masses as well as different combinations of features. It makes uses of correlation of tress to improve the generalization of ensemble model. In the experiment, six machine learning methods and three conventional Warfarin predictive models are as comparators to "WarfarinSeer". A dataset of 589 Han Chinese patients is collected from the data of The First Affiliated Hospital of Soochow University and an open source data of International Warfarin Pharmacogenetics Consortium (IWPC) for training and test. Results showed that "WarfarinSeer" present the highest accuracy on the prediction of warfarin dose in terms of R-squared (R-2) and mean squared error (mse).