Bayesian Machine Learning Model Guiding Iterative, Personalized Anticoagulant Dosing Decision-Making: ENGAGE AF-TIMI...

Area of Science:

• Cardiovascular Medicine
• Machine Learning in Healthcare
• Pharmacometrics

Background:

• Current anticoagulation for atrial fibrillation uses fixed dosing, neglecting patient-specific preferences for risks like stroke, bleeding, and death.
• Individualized treatment strategies are needed to optimize anticoagulation therapy in atrial fibrillation.

Purpose of the Study:

• To develop and evaluate Adele, a Bayesian machine learning model for individualized, long-term anticoagulation dosing in atrial fibrillation patients.
• To incorporate patient preferences for risks of stroke, bleeding, and death into dosing decisions.

Main Methods:

• Developed Adele, a Bayesian competing-risk, multistate hazard model, trained on 5,380 edoxaban-treated atrial fibrillation patients from the ENGAGE AF-TIMI 48 trial.
• Utilized patient pharmacokinetic (PK) and baseline data for personalized, continuous risk prediction.
• Compared Adele's predictive accuracy to standard Kaplan-Meier estimators using concordance index.

Main Results:

• Adele demonstrated superior predictive accuracy over Kaplan-Meier estimators, improving 3-year predictions for cardiovascular death (+12.1%), disability (+11.8%), major gastrointestinal bleeding (+13.7%), and ischemic stroke (+3.3%).
• The model showed dynamic risk prediction, adjusting event probabilities after events like intracranial hemorrhage or ischemic stroke.
• Patient examples illustrated Adele's ability to identify optimal edoxaban doses (15mg, 30mg, 60mg) based on varied hypothetical outcome preferences.

Conclusions:

• Adele is an advanced Bayesian framework integrating clinical factors and PK data for adaptive, patient-centered, preference-weighted predictions.
• This model enables personalized anticoagulation management by dynamically adjusting predictions and guiding optimal dosing.