Predictors of All Cause Mortality in the BARI 2D Trial Identified by Machine-learning
Amit K. Dey2, Zyannah Mallick2, Ruba Shalhoub2, Runqiu Wang2, Tejas Patel1, Xin Tian2, Colin Wu2, Yuan Gu2, Iffat Chowdhury1, Nehal N. Mehta2, Eileen Navarro-almario1, Frank Pucino1, Yves Rosenberg2, Ahmed Hasan2. 1U.S Food and Drug Institution, Rockville, Maryland, United States, 2National Institutes of Health, Bethesda, Maryland, United States
Purpose of Study The Bari Angioplasty Revascularization in Type 2 Diabetes (BARI 2D) Trial tested the effects of different treatment strategies on the rate of death and a composite of death, myocardial infarction, or stroke (MACE) in patients with both coronary artery disease and type 2 diabetes. We sought to identify factors predictive of death, using random survival forests (RSF), a machine-learning methodology.
Methods Used A total of 84 variables were analyzed in a total of 2368 patients as potential predictors of all-cause mortality using RSF, a machine learning approach. The top 10 RSF predictors were then included in both a univariate and multivariate analysis using a Cox proportional hazards model with a stepwise selection approach.
Summary of Results During a 3-year median follow-up, 316/2368 patients died. The top 10 variables from RSF analysis and Cox regression for death are presented in Table 1.The univariate analysis of death showed all 10 predictors as significant while the multivariate analysis showed 8 as significant (Table 1). Important predictors of death include age, serum creatinine, history of congestive heart failure, diuretic use and abnormal ST depression.
Conclusions Age and renal function markers were top predictors of all-cause mortality. As previous studies have not looked into several of these top predictors, this preliminary analysis will be explored further to identify potential risk predictors and risk factors in subgroups of this population and examine relationships with other outcomes of interest.
Back to 2019 Abstracts