Title:Thinking Outside the Box: Orthogonal Approach to Equalizing Protected Attributes

Abstract:There is growing concern that the potential of black box AI may exacerbate health-related disparities and biases such as gender and ethnicity in clinical decision-making. Biased decisions can arise from data availability and collection processes, as well as from the underlying confounding effects of the protected attributes themselves. This work proposes a machine learning-based orthogonal approach aiming to analyze and suppress the effect of the confounder through discriminant dimensionality reduction and orthogonalization of the protected attributes against the primary attribute information. By doing so, the impact of the protected attributes on disease diagnosis can be realized, undesirable feature correlations can be mitigated, and the model prediction performance can be enhanced.