Evaluating and Reducing AI Model Group Disparity: An Analysis of COVID Test Outcomes in Children

AI fairness in healthcare has attracted significant attention due to the potential risk of perpetuating health disparity. This study assessed the group parity of a set of machine learning (ML) models trained on the National Health Interview Survey data, with COVID test result as the outcome. We also experimented with the use of synthetic data to reduce group disparity. Our results suggests that group disparity is prevalent in ML models though often not statistically significant, and the use of synthetic data can sometimes enhance group parity.