Warfarin is an anticoagulant that prevents venous and arterial clots but doubles the risk of major hemorrhage. Over the past decade, warfarin use has caused more medication-related emergency department visits among older Americans than any other drug. Our long-term goal is to improve the safety and effectiveness of antithrombotic therapy. To advance this goal, we have developed clinical and pharmacogenetic (PGx) dosing algorithms to guide days 1–5 of warfarin dosing and placed them on a non-profit web application (WarfarinDosing.org) that has been accessed more than 1.8 million times. This success provides the rationale for the proposed study: that algorithm-based dosing of warfarin reduces the risks of overdose and iatrogenic hemorrhage compared to trial-and-error dosing. Because the rate of warfarin overdose is highest between approximately 6-28 days of therapy, we will use penalized regression and machine learning (ML) to develop warfarin-dosing algorithms for this interval. To promote use of the new algorithms, we will configure the electronic health record (EHR) used at 23 medical centers in MO and IL to export data into WarfarinDosing.org, seamlessly providing clinicians with clinical decision support to guide warfarin initiation. Aim 1: To use linear regression to develop clinical and PGx dosing models for days 6–28 of warfarin therapy. To balance accuracy and parsimony, we will use penalized regression to elucidate relationships between the therapeutic dose and anthropomorphic, clinical, demographic, laboratory, and genetic variables. We have data collected from 3107 participants in 3 randomized clinical trials. Aim 2: To use ML to develop clinical and PGx dosing models for days 6–28 of warfarin therapy. Aim 3: To validate the models developed in Aims 1 and 2. We will quantify the accuracy of the models developed in Aims 1 and 2 in a set-aside 20% testing sample using mean absolute error (MAE) and secondary metrics (e.g. R2). We hypothesize that the best clinical and PGx models developed in Aims 1 or 2 will predict the therapeutic warfarin dose with MAEs < 1.0 mg/d in the testing sample. Aim 4: To update and expand WarfarinDosing.org to provide clinical decision support based on the best clinical and PGx models validated in Aim 3. WarfarinDosing.org will be revised to interface with Epic (Epic Systems Corp, WI), the EHR used across our 23 medical centers in MO and IL. We hypothesize that integrating its use with Epic will decrease the rate of the composite outcome of a warfarin overdose or a hemorrhage as compared to historic rates among patients starting warfarin. The proposed research is innovative and significant because it uses penalized regression and ML to derive novel PGx and clinical algorithms that will reduce the risk of overdose and iatrogenic hemorrhage from warfarin initiation. The integration of Epic and WarfarinDosing.org will be a sustainable and scalable intervention to improve the safety of anticoagulant therapy.

Warfarin (Coumadin™ and others) is a blood thinner that prevents venous and arterial blood clots but doubles the risk of bleeding. To reduce the risks of warfarin overdose and bleeding, the investigators will use personalized medicine and machine learning to develop algorithms to guide the initiation of warfarin therapy. They will put these algorithms on a free, non-profit website (WarfarinDosing.org) and integrate them into electronic medical records.