Fig. 1: Overview of approach.
Our study aims to (1) better understand the effects of technical parameters on AI-based racial identity prediction, and (2) use the resulting knowledge to implement strategies to reduce a previously identified AI performance bias. a We first train AI models to predict self-reported race in chest X-rays. We then assess how the models’ predictions change as a function of factors relating to image acquisition and processing. These factors include the window width, field of view, and view position. b We next train AI models to predict the presence of pathological findings, where an underdiagnosis bias for underrepresented patients has been previously identified1. Based on the results of the technical factor analysis, we devise strategies with a goal of reducing this bias. PA posterior-anterior, AP anterior-posterior.
