Modelo bayesiano de aprendizaje automático para guiar la dosificación iterativa y personalizada de anticoagulantes:...

Anticoagulation in atrial fibrillation often relies on a fixed dose and infrequent dose adjustment and does not incorporate patient preferences for risks of stroke, bleeding, and death.

We developed a Bayesian machine learning model (Adele) to guide individualized long-term dosing accounting for patient preferences.

Adele is a Bayesian competing-risk, multistate hazard model trained on 5,380 edoxaban-treated patients with atrial fibrillation with pharmacokinetic (PK) data from the ENGAGE AF-TIMI 48 (Effective Anticoagulation with Factor Xa Next Generation in Atrial Fibrillation-TIMI 48) trial. Patient PK and baseline data were used to estimate personalized continual risk prediction. The concordance index compared predictive accuracy of Adele with standard Kaplan-Meier estimators. Positive concordance index values (%) indicate better predictive accuracy for Adele. In 2 patient examples, we demonstrate Adele's ability to predict 3-year event frequencies at randomization and following intercurrent events identifying optimal dosing across three edoxaban doses (60mg, 30mg, 15mg) based on hypothetical outcome preferences.

Adele outperformed Kaplan-Meier estimators, improving 3-year prediction for cardiovascular death by +12.1%, disability by +11.8%, major gastrointestinal bleeding by +13.7%, and ischemic stroke (IS) by +3.3%. Adele demonstrated dynamic risk prediction capabilities, with future event probabilities shifting following intracranial hemorrhage or IS. In patient A (80-yr female; 52 kg), preference-weighted event rates were lowest at 15 mg and 30 mg, irrespective of the preference to avoid cardiovascular death, major bleeding, or disability. In patient B (72-yr male; 80 kg), preference-weighted event rates were lowest at 60 mg, irrespective of the preference to avoid death or IS/intracranial hemorrhage, and instead lowest at 15 mg for disability avoidance.

Adele represents a state-of-the-art Bayesian framework, using clinical factors and PK data, to enable longitudinally adaptive, patient-centric, preference-weighted predictions.