OBJECTIVES : The study aimed to assess the utility of machine learning techniques in identifying predictors of major bleeding among patients with non-valvular atrial fibrillation (NVAF).

METHODS : We identified AF patients who initialized oral anticoagulants between 2015 and 2020 in the Clinical Practice Research Datalink (CPRD)-Hospital Episode Statistics (HES) link database. The main outcome of interest was major bleeding. We applied machine learning technique in a hypothesis-free way to identify predictors of major bleeding events. Major bleeding included major gastrointestinal bleeding, major intracranial bleeding, and other major bleeding events requiring hospitalizations.

RESULTS : A total of 54,762 NVAF patients receiving oral anticoagulants were identified, including 2,513 edoxaban, 25,456 apixaban, 1,908 dabigatran, 17,772 rivaroxaban, and 7,113 warfarin patients. The mean age was 76.0 years (SD = 10.5) and 55.7% of patients were male. Overall, 9.99% of patients experienced major bleeding during the follow-up period. Predictors from the best-performing model included history of bleeding, use of non-vitamin K oral anticoagulants (NOACs), concomitant use of proton pump inhibitors (PPI) or H₂ blockers, and advanced age.

CONCLUSIONS : Unlike the traditional hypothesis-driven regression model approach, this study applied machine learning technique with no pre-selection process for variables in the database. We identified the strongest predictors for major bleeding were use of NOACs, use of PPI or H₂ blockers, and advanced age. The prediction model enumerates the risk of major bleeding in NVAF patients and provides a potential explicit standard for clinicians when making individual patient treatment decisions.