AI-WAR software significantly improved median time in therapeutic range to 81.3% compared to 48.7% with conventional management in patients receiving long-term warfarin therapy.
Cohort (n=400)
Open-label
Yes
Does an AI-based warfarin management system (AI-WAR) improve time in therapeutic range in patients after heart valve replacement?
An AI-based warfarin management system using a Bi-LSTM model significantly improves time in therapeutic range and reduces adverse events compared to conventional management in patients after heart valve replacement.
Absolute Event Rate: 81.3% vs 48.7%
p-value: p=<0.001
Background Warfarin remains the preferred anticoagulant in patients after heart valve replacement. However, its narrow therapeutic window, substantial interindividual variability, and dependence on frequent INR monitoring make effective management challenging. The quality of anticoagulation in real-world practice is often suboptimal, particularly in primary care settings with limited resources. Methods We developed an artificial intelligence-based warfarin management system (AI-WAR) integrating remote follow-up, systematic data management, and individualized dosing prediction models. A randomized controlled trial (n = 624) and a prospective registry study (n = 176) were used for model training, while an independent real-world cohort (n = 200) served for external validation. We compared anticoagulation quality and clinical outcomes between conventional management and AI-WAR, and evaluated prediction performance of long short-term memory (LSTM) and bidirectional LSTM (Bi-LSTM) models. Results Compared with conventional management, AI-WAR significantly improved median TTR (48.7% vs. 81.3%, P 0.001), increased time in target INR (36.3% vs. 61.9%, P 0.001), and reduced both elevated INR percentage (12.0% vs. 16.8%) and overall adverse event rate (9.6% vs. 19.7%, P 0.001). The Bi-LSTM model exhibited superior dose prediction accuracy (80.3% vs. 66.6%) and achieved 93.2% accuracy for stable dose prediction, while significantly reducing overdose predictions (11.9%, P 0.001). Subgroup analyses demonstrated robust performance across different genotypes and in scenarios with missing genetic information. Conclusion The integration of AI-WAR software and Bi-LSTM prediction models provides an effective and safe decision-support tool for warfarin individualized dosing. This system not only improves anticoagulation quality and reduces adverse events but also shows practical advantages in remote management and primary healthcare settings, supporting broader implementation of precision anticoagulation therapy.
Li et al. (Wed,) conducted a cohort in Long-term warfarin therapy (n=400). AI-WAR software vs. Conventional management was evaluated on Percentage of time in therapeutic range (%TTR) (p=<0.001). AI-WAR software significantly improved median time in therapeutic range to 81.3% compared to 48.7% with conventional management in patients receiving long-term warfarin therapy.