Fig. 8: Macro-ROC-AUCs and sensitivity SDs of DDMs in four application paradigms.
From: Enhancing fairness in AI-enabled medical systems with the attribute neutral framework
a Schematic diagram illustrating the four application paradigms of the AttrNzr. The four paradigms are as follows: no neutralization (no AttrNzr is used), test-stage neutralization (AttrNzr is used only during the test stage), training-stage neutralization (AttrNzr is used only during the training stage), throughout neutralization (AttrNzr is used during both the training and test stages). Macro-ROC-AUCs in ChestX-ray14 (b, c), CheXpert (d, e), and MIMIC-CXR (f–i). d–k Sensitivity SDs in ChestX-ray14 (j, k), CheXpert (l, m), and MIMIC-CXR (n–q). The neutralized attributes for each subfigure are as follows: sex (b, d, f, j, l, n), age (c, e, g, k, m, o), race (h, p), and insurance (i, q). r Detailed layout template for subgraphs b–q.
