Predictors of warfarin therapeutic dose across thromboembolic indications: an exploratory early clinical evaluation study using machine learning-integrated explainable artificial intelligence and Bayesian models.

BACKGROUND

Warfarin dosing varies widely due to genetic, demographic, and clinical factors, but it is unknown whether the importance, equilibrium, and prediction uncertainty of established pharmacogenetic predictors (VKORC1 and CYP2C9) differ between arterial (AF/stroke) and venous (DVT/PE) thromboembolic indications.

RESEARCH DESIGN AND

METHODS

This exploratory study (early clinical evaluation) synthesizes findings from the International Warfarin Pharmacogenetics Consortium dataset. We evaluated five machine learning (ML) models, and SHapley Additive exPlanations (SHAP) and Bayesian Additive Regression Trees (BART) analyses were carried out.

RESULTS

Random Forest demonstrated slightly better predictive performance than other ML models. SHAP analysis quantified feature contributions, revealing that VKORC1 G/G genotype as the most influential in AF/Stroke, while VKORC1 A/A genotype followed by age in the DVT/PE group. BART provided probabilistic predictions and identified indication-specific uncertainty drivers. Age possibly has an interaction effect in the requirements of reduced warfarin doses with AF/Stroke.

CONCLUSIONS

Our findings reveal that while core genetic and anthropometric predictors of warfarin dose transcend thromboembolic indication, the equilibrium among these factors and sources of prediction uncertainty possibly differ between arterial and venous disease. Integration of ML with SHAP offers a roadmap for personalized warfarin dosing, though prospective validation is needed before clinical implementation.