level: research
chest x-ray classification models often perform well on common conditions but can miss rare findings, especially when decisions are made by applying a threshold to model scores. this study examines who gets missed after converting model scores into binary decisions, focusing on subgroups defined by sex, age, race, and insurance. using two large datasets, vindr-cxr and mimic-cxr/cxr-lt, the researchers audit the fairness of long-tailed multi-label models by analyzing false negative rates across different groups.
the work introduces a diagnostic ladder to isolate the effects of class-level long-tail losses, subgroup-aware weighting, group robustness techniques, and threshold selection. on vindr-cxr, combining group-tail weighting with tail-aware thresholding reduced the tail false negative rate from 0.665 to 0.269. the worst-group false negative rate for sex dropped from 0.705 to 0.157, and for age from 0.822 to 0.133, while macro average precision improved from 0.611 to 0.635. on mimic-cxr/cxr-lt, similar methods reduced the tail false negative rate from 0.866 to 0.741 and lowered worst-group false negative rates across all subgroups.
despite these improvements, residual missed-positive rates remain, indicating that threshold-based decisions still leave some patients undiagnosed. the study highlights that ranking performance alone does not guarantee equitable outcomes when thresholds are applied. the findings suggest that careful threshold selection and subgroup-aware training can mitigate but not eliminate underdiagnosis in long-tailed medical imaging tasks.
why it matters: for ai in healthcare, this shows that even accurate models can unfairly miss diagnoses in underrepresented groups, and that threshold tuning is critical for equitable deployment.