Comparison of the predictive abilities of pharmacogenetics-based warfarin dosing algorithms using seven mathematical models in Chinese patients

Abstract

Aim: This study is aimed to find the best predictive model for warfarin stable dosage. Materials & methods: Seven models, namely multiple linear regression (MLR), artificial neural network, regression tree, boosted regression tree, support vector regression, multivariate adaptive regression spines and random forest regression, as well as the genetic and clinical data of two Chinese samples were employed. Results: The average predicted achievement ratio and mean absolute error of the algorithms were ranging from 52.31 to 58.08% and 4.25 to 4.84 mg/week in validation samples, respectively. The algorithm based on MLR showed the highest predicted achievement ratio and the lowest mean absolute error. Conclusion: At present, MLR may be still the best model for warfarin stable dosage prediction in Chinese population. Original submitted 10 November 2014; Revision submitted 18 February 2015