Development of a system to support warfarin dose decisions using deep neural networks

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Pharmacogenomics

Background:

• Accurate prothrombin time international normalized ratio (PT INR) monitoring is crucial for effective warfarin therapy.
• Existing warfarin dosing methods can be imprecise, leading to suboptimal patient outcomes.
• Developing advanced prediction models can improve the precision of PT INR management.

Purpose of the Study:

• To develop a predictive model for prothrombin time international normalized ratio (PT INR).
• To create a warfarin maintenance dose decision support system for precise warfarin dosing.
• To evaluate the performance of an AI-driven PT INR prediction algorithm against expert physicians.

Main Methods:

• Analysis of PT INR data from 19,719 inpatients across three institutions.
• Development of a PT INR prediction algorithm using dense and recurrent neural networks.
• Training the algorithm on data from one hospital and testing on datasets from two others.
• Comparison of algorithm's 5th-day PT INR predictions with those of 10 expert physicians.
• Generation of individualized warfarin dose-PT INR tables using the developed model.

Main Results:

• The AI algorithm achieved higher accuracy (84.0%) in predicting 5th-day PT INR within ±0.3 of the actual value compared to expert physicians (81.9%).
• The algorithm demonstrated superior performance (P=0.014) in PT INR prediction accuracy.
• Individualized warfarin dose-PT INR tables generated by the algorithm showed acceptable performance for 8th-day PT INR predictions (53.4% within ±0.3).

Conclusions:

• An artificial intelligence-based warfarin dosing algorithm utilizing recurrent neural networks significantly outperforms expert physicians in predicting future PT INRs.
• The developed AI model provides a precise warfarin dosing platform.
• The individualized warfarin dose-PT INR table generator based on the AI algorithm is a viable tool for clinical decision support.