Optimizing warfarin dosing for patients with atrial fibrillation using machine learning

Area of Science:

• Artificial Intelligence in Medicine
• Pharmacodynamics and Drug Dosing Optimization
• Clinical Decision Support Systems

Background:

• Warfarin, a vitamin K antagonist, is widely used for stroke prevention in atrial fibrillation but poses challenges due to complex pharmacodynamics, often leading to suboptimal anticoagulation.
• Achieving and maintaining the target International Normalized Ratio (INR) range (2.0-3.0) is crucial for warfarin efficacy and safety, yet difficult to achieve consistently in clinical practice.

Purpose of the Study:

• To develop and validate a deep reinforcement learning (DRL) model for optimizing warfarin dosing to maximize time in the therapeutic INR range (TTR).
• To assess the DRL model's effectiveness in improving clinical outcomes, specifically reducing the composite endpoint of stroke, systemic embolism, or major hemorrhage.

Main Methods:

• A novel semi-Markov decision process formulation of the Batch-Constrained deep Q-learning algorithm was employed to create the DRL model.
• The model was trained on data from 22,502 warfarin-treated patients across three major clinical trials (ENGAGE AF-TIMI 48, ARISTOTLE, ROCKET AF).
• External validation was performed using data from 5730 warfarin-treated patients in the RE-LY trial, comparing algorithm-consistent dosing with TTR and clinical outcomes.

Main Results:

• External validation demonstrated a significant positive association between center-level algorithm-consistent dosing and TTR (R² = 0.56).
• A 10% increase in algorithm-consistent dosing correlated with a 6.78% improvement in TTR and an 11% reduction in the composite clinical outcome.
• The DRL algorithm's performance was comparable to a rule-based clinical algorithm, indicating its robustness and potential for clinical application.

Conclusions:

• A deep reinforcement learning algorithm can effectively optimize warfarin dosing to improve time in the therapeutic INR range for patients with atrial fibrillation.
• Implementing a digital clinical decision support system based on this DRL algorithm holds promise for enhancing warfarin management and improving patient outcomes globally.