Machine Learning Models for Warfarin Dose Prediction: A Systematic Review of Performance and Clinical Utility

Abstract

Warfarin therapy is complicated by its narrow therapeutic index and significant inter-individual variability. This study aimed to systematically review the performance of machine learning (ML) for predicting warfarin doses in clinical settings, compared with conventional methods, and to assess the quality of evidence. We searched PubMed, Embase, and Web of Science until December, 2024, and included studies developing ML models for warfarin dosing in clinical settings. The performance metrics of ML models were evaluated using predictive accuracy, model fit, and clinical utility. Risks of bias were assessed using the Prediction Model Risk of Bias Assessment Tool. Thirty-five studies were included. In terms of ML models, the majority of included studies evaluated artificial neural networks (n = 24), followed by ensemble models (n = 19), and support vector machines (n = 18). Regarding performance, ML models demonstrated superiority across all three metrics: predictive accuracy (16 of 18 studies), model fit (5 of 7 studies), and INR-related clinical utility (all 4 studies). A high risk of bias was identified in 94% of studies, mainly due to analysis-domain limitations, including missing-data handling, insufficient events per predictor, unclear predictor selection, and limited external validation. In conclusion, ML-based approaches may improve warfarin dose prediction compared with conventional methods, but the current evidence should be interpreted cautiously. Future studies should prioritize transparent reporting, robust model development, external validation, and clinically relevant outcomes before these models can be recommended for routine clinical implementation.