Addressing Racial Bias in Cardiovascular Disease Risk Prediction with Fair Data Augmentation

In the past decade, the application of machine learning approaches to clinical tasks such as risk prediction and diagnosis of diseases has offered several potential benefits for both patients and medical professionals. However, in the development of these machine learning models, a factor is often overlooked: algorithms can be unfairly biased against certain populations, presenting disparate performances for different groups, defined by protected attributes such as race/ethnicity, sex or gender, and socioeconomic status, that are often insufficiently represented in datasets. Implementation of these biased models for clinical use could possibly worsen healthcare inequality. In this paper, we investigate the fairness of health-related ML research by using health survey data from the Centers for Disease Control and Prevention (CDC) to predict cardiovascular disease risk. We use two fairness metrics for the CDC data, and analyze the prediction model performance for racial minority groups to evaluate potential racial bias due to imbalanced or unrepresentative data. Then, we utilize Fair Mixup, a data augmentation method, to enhance fairness measures between racial groups and achieve improvements in predictive accuracy.