Figure 4. Application of AEquity to detect, characterize, and mitigate a previously undetected racial bias in mortality prediction on a commonly used dataset. (A) Schematic representations of AEquity applied to label selection and demographic prioritization of bias in the National Health and Nutrition Examination Survey. (B) AEquity reveals that the least biased label for mortality prediction occurs at 5 years because the groups have similar levels of complexity. (C) AEquity determined that 5-year mortality has complexity bias because the joint AEq exceeds the AEq value of both Black and White patients. (D) AEquity outperformed other modalities of bias reduction, namely BERM, and calibration. (E) AEquity demonstrated that bias reduction is invariant to sample size. AUC: area under the curve; BERM: balanced empirical risk minimization.