Warfarin Dose Prediction Research Based On Feature Selection And Improved Stacking Integrated Algorithm

Warfarin is an oral anticoagulant drug that is widely used in postoperative anticoagulation such as heart valve replacement and atrial fibrillation.It is also commonly used in anticoagulation treatment of various hemoembolic diseases.In the course of anticoagulation therapy,warfarin has the problems of narrow treatment window,large side effects,and large differences in individual patient doses.The dose must be accurate.In clinical use,blood must be collected from patients,and the anticoagulant effect should be monitored according to the International Normalized Ratio(INR)in the blood.Patients must face long-term,frequent blood collection,and adjust the warfarin dose according to blood monitoring results until the INR value reaches the standard and the dose is stable.During this process,patients will face the risk of thrombosis due to underdose for a long time,as well as the risk of bleeding due to overdose.Clinically,the existing warfarin dose prediction model is generally used to calculate the warfarin dose within the ideal INR range according to the patient's own relevant indicators,and use this dose as the initial therapeutic dose.If the predicted initial dose is unlimited Close to the stable dose reached in the later period,patients can always use the most effective and safest dose for anticoagulation therapy,effectively avoiding adverse reactions during treatment.The study of warfarin dose prediction is to use patient demographics and clinical factors as input features of the model,and form a dose prediction model through regression algorithms.The internationally recognized warfarin dose prediction model is the IWPC model established by the International Warfarin Pharmacogenetics Consortium(IWPC)based on linear regression algorithm.In the study of the predictive model,IWPC found that cytochrome P4502C9(CYP2C9)and vitamin K epoxide reductase complex 1(VKORC1)gene polymorphisms have a significant effect on the therapeutic dose of warfarin.Since there is no linear relationship between genotype and dose,warfarin dose prediction is actually a multivariate nonlinear regression task.Machine learning has certain advantages in completing nonlinear regression tasks.The model established by machine learning can fuse a large number of input features and discover the nonlinear relationship of variable features.At present,the algorithm used in the warfarin dose prediction model at home and abroad has also been converted from a linear regression algorithm to a machine learning algorithm.In the field of machine learning,integrated learning can greatly improve the accuracy of the algorithm and enhance the stability of the algorithm.Based on the public sample data provided by IWPC,this paper establishes a warfarin dose prediction model using feature selection and Stacking integration algorithm.In view of the fact that the data dimension is large and there are redundant features,the RRelief F algorithm is used to calculate the feature weights and the method of correlation test to complete the feature selection process and form the optimal feature subset.The optimal feature subset formed by feature selection is superior to the original data set in prediction performance,and can better explain the ability of all features in the IWPC data set to affect the dose.In this paper,the traditional stacking integration algorithm is improved.For the k-fold cross validation in the traditional stacking algorithm,the average value is improved to RMSE reciprocal weighted average,and the feature relationship of the optimal feature subset is input to the meta learner,and the prediction results of the combined base learner are used as the feature input of the meta learner,forming the improved stacking algorithm.The experimental results show that the warfarin dose prediction model based on the optimal feature subset combined with the improved stacking algorithm has better performance in all aspects than the model based on the basic learner and the traditional stacking algorithm.At the same time,compared with the current clinical application of iwpc formula and the prediction model based on multiple linear regression(MLR),the performance of our improved method is better than others in variou aspect.