Who Gets Missed in the Tail? Thresholded Subgroup Underdiagnosis in Long-Tailed Chest X-ray Classification

Short summary

This arXiv paper addresses fairness gaps in chest X-ray AI classifiers, where rare conditions are missed disproportionately across demographic subgroups. Using weighted training and threshold tuning, researchers reduced false-negative rates by 50%+ while improving overall accuracy. Key finding: diagnostic fairness depends on the joint interaction of finding, subgroup, and decision threshold—not ranking metrics alone.

• •Fairness audit reveals rare conditions missed disproportionately in demographic subgroups of CXR classifiers
• •Tail-aware weighting + threshold tuning reduces false negatives by 50%+ (VinDr: FNR 0.665→0.269)
• •Diagnostic fairness depends on finding × subgroup × threshold interaction, not aggregate model metrics

Generated with AI, which can make mistakes.