An ensemble learning based framework to estimate warfarin maintenance dose with cross-over variables exploration on incomplete data set

MOTIVATION Warfarin is a widely used oral anticoagulant, but it is challenging to select the optimal maintenance dose due to its narrow therapeutic window and complex individual factor relationships. In recent years, machine learning techniques have been widely applied for warfarin dose prediction. However, the model performance always meets the upper limit due to the ignoration of exploring the variable interactions sufficiently. More importantly, there is no efficient way to resolve missing values when predicting the optimal warfarin maintenance dose. METHODS Using an observational cohort from the Xinhua Hospital affiliated to Shanghai Jiaotong University School of Medicine, we propose a novel method for warfarin maintenance dose prediction, which is capable of assessing variable interactions and dealing with missing values naturally. Specifically, we examine single variables by univariate analysis initially, and only statistically significant variables are included. We then propose a novel feature engineering method on them to generate the cross-over variables automatically. Their impacts are evaluated by stepwise regression, and only the significant ones are selected. Lastly, we implement an ensemble learning based approach, LightGBM, to learn from incomplete data directly on the selected single and cross-over variables for dosing prediction. RESULTS 377 unique patients with eligible and time-independent 1173 warfarin order events are included in this study. Through the comprehensive experimental results in 5-fold cross-validation, our proposed method demonstrates the efficiency of exploring the variable interactions and modeling on incomplete data. The R2 can achieve 75.0% on average. Moreover, the subgroup analysis results reveal that our method performs much better than other baseline methods, especially in the medium-dose and high-dose subgroups. Lastly, the IWPC dosing prediction model is used for further comparison, and our approach outperforms it by a significant margin. CONCLUSION In summary, our proposed method is capable of exploring the variable interactions and learning from incomplete data directly for warfarin maintenance dose prediction, which has a great premise and is worthy of further research.