AI chest X-ray models underdiagnose rare conditions in vulnerable subgroups

A new preprint from Pham et al. (7 authors) tackles a critical fairness problem in AI chest X-ray (CXR) analysis: even when models achieve strong overall ranking performance, they can systematically miss rare-positive patients in specific subgroups—especially those in the long tail of disease prevalence. The study, submitted to arXiv on July 4, 2026, presents a 'diagnostic ladder' framework that separates class-level long-tail losses, subgroup-aware weighting, group robustness, and threshold selection. On the VinDr-CXR and MIMIC-CXR/CXR-LT datasets, the authors show that aggregate metrics like macro-mAP can hide severe disparities.

Key results: On VinDr-CXR, a combination of group-tail weighting and tail-aware thresholding reduced tail FNR from 0.665 to 0.269, sex worst-group FNR from 0.705 to 0.157, and age worst-group FNR from 0.822 to 0.133—while macro-mAP actually increased from 0.611 to 0.635. On MIMIC-CXR/CXR-LT, tail FNR dropped from 0.866 to 0.741, with improvements across sex, age, race, and insurance subgroups. But residual missed-positive rates remained stubbornly high, and group robustness methods alone (e.g., GroupDRO) didn't eliminate rare subgroup misses. The study supports a narrow audit claim: rare-label fairness in CXR depends jointly on the finding, subgroup, and operating threshold, not just label frequency or ranking metrics.