Causal Machine Learning for Left Atrial Appendage Occlusion in Patients With Atrial Fibrillation.

Abstract

Background: Transcatheter left atrial appendage occlusion (LAAO) is an alternative to lifelong anticoagulation, but optimal patient selection remains challenging.

Objectives: This study sought to apply a novel causal machine learning framework to identify patients who would benefit from LAAO vs a direct oral anticoagulant (DOAC).

Methods: We identified 744,190 adult patients with atrial fibrillation treated with either LAAO or DOAC between March 13, 2015, and December 31, 2019, using data from OptumLabs Data Warehouse. One-to-one propensity score matching was used to create a cohort where patients were similar in 107 baseline characteristics. A causal forest model was used to estimate the heterogeneous treatment effect for a composite outcome of ischemic stroke, systemic embolism, major bleeding, and all-cause mortality.

Results: In the matched cohort of 28,930 patients, the mean age was 76.8 ± 6.3 years; 5,818 patients (40%) were female, and the mean CHA₂DS₂-VASc score was 5.8. LAAO was associated with no difference with the primary composite outcome in comparison to NOAC early on (average treatment effect of -0.68% [-1.4%, 0.06%] at 1 year), but a lower risk at the end of 2 years (average treatment effect of -2.9% [-3.7%, -2.0%]). At the end of 2 years, 30.1% of the overall cohort were classified as potentially benefiting from LAAO, 69.7% were classified as neutral, and 1.4% were potentially harmed by LAAO.

Conclusions: Novel machine learning algorithms were developed to identify patients who are more likely to benefit from LAAO vs DOACs. This information can support clinical decision-making to determine which patients should be referred to subspecialists for further examination and discussion of LAAO.