Machine Learning Predicts Bleeding Risk in Atrial Fibrillation Patients on Direct Oral Anticoagulant.

Area of Science:

• Cardiology
• Medical Informatics
• Machine Learning in Healthcare

Background:

• Predicting major bleeding in nonvalvular atrial fibrillation (AF) patients on direct oral anticoagulants (DOACs) is critical for tailoring treatment.
• Existing risk scores like HAS-BLED, ATRIA, and ORBIT have limitations in accurately assessing bleeding risk.
• Left atrial appendage closure is an alternative for stroke risk reduction with fewer nonprocedural bleeds.

Purpose of the Study:

• To compare the performance of various machine learning (ML) models against conventional bleeding risk scores (HAS-BLED, ATRIA, ORBIT) for predicting bleeding events in AF patients on DOACs.
• To evaluate the predictive accuracy of ML models for hospitalization due to bleeding events at multiple time points (1, 2, and 5 years).

Main Methods:

• Retrospective cohort study utilizing electronic health records from 2010-2022 at the University of Pittsburgh Medical Center.
• Inclusion of 24,468 nonvalvular AF patients (age ≥18) on DOACs, excluding those with prior significant bleeding or warfarin use.
• Comparison of ML algorithms (logistic regression, classification trees, random forest, XGBoost, k-nearest neighbor, naïve Bayes) against HAS-BLED, ATRIA, and ORBIT for predicting bleeding hospitalization.

Main Results:

• ML models demonstrated superior performance compared to HAS-BLED, ATRIA, and ORBIT in predicting 1-year bleeding events.
• The random forest model achieved an AUC of 0.76, significantly outperforming HAS-BLED's AUC of 0.57 (p < 0.001).
• ML models showed improved accuracy across all follow-up time points and for predicting hemorrhagic stroke, with SHAP analysis revealing novel risk factors like BMI, cholesterol profile, and insurance type.

Conclusions:

• Machine learning models offer enhanced predictive capabilities for bleeding risk in AF patients on DOACs compared to traditional scores.
• ML models can identify novel risk factors, paving the way for more personalized bleeding risk assessment and management strategies.
• The findings support the integration of ML tools into clinical practice for optimizing anticoagulation therapy in AF patients.