B1308
Title: Evaluating the algorithmic fairness for cardiovascular risk prediction model
Authors: Juan Zhao - American Heart Assocation (United States) [presenting]
Abstract: Cardiovascular disease (CVD) is the leading global cause of death. To identify high-risk CVD patients for early treatment, traditional clinical algorithms such as Pooled Cohort Equations (PCE) and Machine Learning (ML) models have been employed. However, whether these models provide fair predictions for different racial/ethnic groups remains unexplored. The objectives are to understand the importance of detecting bias and assessing fairness in clinical predictive models, evaluate metrics to quantify fairness and methods to eliminate bias and introduce a national clinical data registry for cardiovascular disease. A large cohort derived from de-identified EHR data between 2007 and 2017 has been used, while pooled Cohort Equations were applied. ML models include logistic regression, random forests, and gradient-boosting trees (GBT). Fairness metrics were equal opportunity difference and disparate impact. The study included 109,490 individuals (9,824 CVDs). Compared to PCE, most ML models are less biased. No disparities were observed among race groups, but models had a significant bias for the women group. GBT has the highest AUROC but is only moderately fair. Nearly all ML models have superior performance in the fairness metrics than the PCE. Models with the highest AUROC may not perform the best in fairness, indicating that fairness should be considered for performance auditions.
