Predictors of warfarin therapeutic dose in cancer patients: a preliminary comparative study using machine learning-integrated explainable artificial intelligence and Bayesian models

BACKGROUND: Warfarin dosing in cancer patients is challenging due to altered pathophysiology and variable responses. While machine learning offers predictive potential, clinical adoption requires interpretability. This study is an initial attempt at addressing whether different dosing strategies are required in this vulnerable population. RESEARCH DESIGN AND METHODS: This observational study analyzed 1,746 patients (124 with cancer) from the International Warfarin Pharmacogenetics Consortium database. Five machine learning algorithms were developed and compared with the best-performing model interpreted using SHAP analysis. Bayesian Additive Regression Trees provided uncertainty quantification, and the Virtual Twins method estimated cancer's individualized treatment effect on dose requirements. RESULTS: Genetic variants, particularly VKORC1 and CYP2C9, dominated dose prediction in both cohorts. However, cancer introduced complexity: clinical predictors showed greater variability and prediction errors, while Bayesian analysis revealed wider uncertainty intervals for inaccurate estimates. The Virtual Twins method demonstrated substantial heterogeneity in cancer's effect, with mean individualized treatment effect of 3.35 mg/week, yet 62.5% of patients exhibited positive effects indicating higher dose requirements. These findings suggest uniform dose adjustments for cancer patients may be inappropriate. CONCLUSIONS: Cancer patients may potentially require personalized warfarin dosing, as genetic and clinical factors interact unpredictably. Explainable artificial intelligence and Bayesian models potentially enable individualized anticoagulation in oncology.