Abstract

A classifier that reads chest radiographs and reports COVID-19, viral pneumonia, or a nor-mal study does not read every patient equally well. Because no public COVID-19 radiograph collection carries reliable demographic annotation, the gender and ethnicity labels used here are simulated, drawn independently and uniformly at random per image rather than observed or inferred, making this a synthetic-disparity experiment rather than an audit of real groups. We measure how unevenly one such model performs and where in the pipeline that unevenness is best corrected, comparing seven mitigation configurations in Fairlearn and eleven in AIF360 against the untreated model.

The untreated model is 77.27% accurate, with equalized odds gaps of 10.28% (gender), 23.81% (ethnicity), and 36.36% (both attributes combined). Correction usually helped: 25 of 36 Fairlearn configurationmetric combinations beat the untreated model, though only five held the accuracy cost under two points. Whether a fairness-accuracy trade-off appears turns out to depend less on which criterion is chosen than on whether two degenerate configurations, which simply stop issuing one diagnosis, are included in the comparison (r = +0.08 with them, r = -0.75 without). Individual-level inequality, by contrast, moved substantially, with the Theil index ranging from 22.58% to 55.49%.

Chaining worked best in AIF360, where two chained configurations raised accuracy above the untreated baseline, though the upstream stage traded rather than added to the con-strained learner’s gains. Three configurations across both toolkits achieved near-zero dis-parity by ceasing to predict one class rather than through genuine correction. The two toolkits again disagreed sharply about the same nominal intervention, so no fairness figure should be quoted without naming the software that produced it.