Abstract 12519: Identification of Novel Atrial Fibrillation Risk Factors in The Women’s Health Initiative Database Utilizing the Random Generalized Linear Model
Introduction: Atrial fibrillation (AF) is the most common cardiac arrhythmia and confers significant risk of morbidity and mortality. The etiology of AF remains incompletely understood. The Women’s Health Initiative (WHI) provides an opportunity to evaluate an extensive and diverse set of covariates in a large population to identify novel risk factors.
Hypothesis: We hypothesized that we would validate existing predictors of AF and discover novel markers associated with AF using a random generalized linear model (RGLM).
Methods: A training subset of women enrolled into the WHI without pre-existing AF were included. Women were followed for AF using self-report and medical records coding. 104 demographic, medical history and environmental exposure covariates were selected for analysis in a RGLM. RGLM is a bootstrapped aggregated generalized linear model (GLM) that combines the predictive accuracy of random forests, a machine-learning algorithm, with the interpretability of forward selected GLM’s. Variables that are most commonly found in the best-performing regression models with randomly generated covariates are determined the most influential.
Results: Of the 60,000 women included in this training set, 5,047 developed AF over an average of 16.3 years. The women were on average 61.9 years old at enrollment, 1.7% were African American, 0.9% were Hispanic, 35.2% had hypertension and 4.2% had diabetes. The RGLM generated 100 sub-cohorts with 75 covariates through random sampling. The top 50 covariates, based on correlation with incident AF, were evaluated by forward selection to generate GLM’s. The variables identified by the RGLM as being the most influential predictors were history of leukemia, prior pregnancy, systolic blood pressure, waist circumference, history of DVT, depression, and vegetable consumption.
Conclusions: Using RGLM, we validated several established risk factors for AF. We also, for the first time in a large population cohort, identified history of leukemia as an influential risk marker of AF. Additional study will be needed to determine if leukemia-specific chemotherapy mediates this association. Finally, these novel markers will need to be validated in additional cohorts.
Author Disclosures: J. Parizo: None. M. Perez: Research Grant; Significant; Robert Wood Johnson Foundation.
- © 2016 by American Heart Association, Inc.