PaperlayerView exact revision on arXiv

Source-linked AI summary

Improving the Fairness of Chest X-ray Classifiers

Haoran Zhang, Natalie Dullerud, Karsten Roth, Lauren Oakden-Rayner, Stephen Robert Pfohl, Marzyeh Ghassemi

arXiv:2203.12609v1cs.LGcs.CVcs.CYeess.IV

TL;DR

Chest x-ray classifiers can show performance gaps across protected groups, while it remains unclear when equal group performance is appropriate in clinical settings. The paper examines group and minimax fairness for chest x-ray disease classification. It finds that worst-group methods do not outperform simple balancing, while group-fairness methods tend to worsen performance for all groups.

  • Problem

    Prior work reports performance gaps across protected groups, but it is unclear when equal group performance is appropriate in clinical settings compared with other fairness definitions.

    Source: §1 · p. 2

  • Method

    The paper examines group and minimax fairness for chest x-ray disease classification and benchmarks methods aimed at improving worst-group performance.

    Source: §1 · p. 2

  • Results

    Methods targeting worst-group performance do not outperform simple data balancing, while group-fairness methods tend to reach parity by worsening performance for all groups.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Relevant sensitivity analyses of external biases require mechanism details that cannot be extracted solely from observed data without further assumptions.

    Source: §7.3 · p. 11

    The paper focuses on label bias and leaves other potential sources of bias for future work; external pipeline biases are not considered.

    Source: §6 · p. 7, §7.3 · p. 11

Abstract

from arXiv · show

Deep learning models have reached or surpassed human-level performance in the field of medical imaging, especially in disease diagnosis using chest x-rays. However, prior work has found that such classifiers can exhibit biases in the form of gaps in predictive performance across protected groups. In this paper, we question whether striving to achieve zero disparities in predictive performance (i.e. group fairness) is the appropriate fairness definition in the clinical setting, over minimax fairness, which focuses on maximizing the performance of the worst-case group. We benchmark the performance of nine methods in improving classifier fairness across these two definitions. We find, consistent with prior work on non-clinical data, that methods which strive to achieve better worst-group performance do not outperform simple data balancing. We also find that methods which achieve group fairness do so by worsening performance for all groups. In light of these results, we discuss the utility of fairness definitions in the clinical setting, advocating for an investigation of the bias-inducing mechanisms in the underlying data generating process whenever possible.

1. Introduction

The paper asks which fairness definitions and debiasing methods are appropriate for chest X-ray classifiers, which can show performance disparities across protected groups. It benchmarks group- and minimax-fairness approaches and investigates possible sources of disparities.

1. Introduction

  • Prior work reports significant TPR and FPR disparities across datasets, tasks, and protected attributes, including underdiagnosis of Black females in disease-presence prediction.The cited study notes that underdiagnosis could delay treatment.Source: §1 · pp. 1–2
  • The paper asks when equal group performance is appropriate in clinical settings, how effective debiasing methods are, and how they affect disparities under other fairness definitions.Source: §1 · p. 2
  • The study benchmarks chest X-ray disease classifiers under group fairness and minimax fairness, which focuses on worst-group performance.It also discusses the utility of these definitions in clinical settings.Source: §1 · p. 2
  • The authors report statistically significant performance gaps across many metrics and protected attributes for ERM models predicting varied pathologies.Source: §1 · p. 2
  • Methods seeking better worst-group performance do not outperform simple data balancing, while group-fairness methods tend to achieve parity by worsening performance for all groups.The authors also investigate possible mechanisms behind disparities in No Finding prediction and emphasize probing their origins.Source: §1 · p. 2

2. Background

The background contrasts fairness definitions, reviews training-time debiasing approaches, and situates this work among medical-imaging fairness studies. It identifies the paper as the first, to the authors’ knowledge, to benchmark bias-reduction algorithms in medical imaging.

2.1. Fairness Definitions

  • Group fairness seeks equal performance metrics across groups, whereas minimax fairness focuses on worst-group error; the definitions can conflict with calibration under stated conditions.The paper uses a relative minimax criterion: a classifier is fairer than a baseline when its maximum group error is lower.Source: §2.1 · pp. 2–3

2.2. Debiasing Methods

  • Debiasing approaches include enforcing group fairness with adversaries, distribution-distance penalties, or constrained optimization, while worst-case methods target poorly performing groups or samples.The study benchmarks representative algorithms from these approaches; it focuses on training-time methods and leaves preprocessing and post-processing methods for future work.Source: §2.2 · p. 3

2.3. Fairness in Computational Medical Imaging

  • Prior chest X-ray studies reported disparities across protected attributes and links between training-sample counts and group performance; this work is the first, to the authors’ knowledge, to benchmark medical-imaging bias-reduction algorithms.Source: §2.3 · p. 3

3. Methods

The methods section benchmarks simple baselines, group-fairness methods, and methods targeting worst-case group performance. The compared approaches use different training strategies, including resampling, separate models, adversaries, distribution matching, and sample or group reweighting.

3. Methods

  • The simple baselines are ERM, Balanced ERM, and Stratified ERM, respectively using population risk, minority-group upsampling, and separate models for each protected group.Source: §3 · p. 3
  • Group-fairness methods target equalized odds, use minority-group upsampling in each minibatch, and do not explicitly seek to improve any group’s performance.The methods include an adversary and penalties matching conditional score distributions or their means.Source: §3 · p. 4
  • The group-fairness comparison also includes FairALM, which uses an augmented Lagrangian to enforce fairness constraints.Source: §3 · p. 4
  • Worst-case methods include GroupDRO, which upweights groups with worse loss, and group-unaware ARL and JTT, which reweight samples.ARL uses an adversary to maximize weighted loss; JTT trains an additional classifier to identify samples misclassified by ERM.Source: §3 · p. 4

4. Experiments

The experiments use MIMIC-CXR and CheXpert to evaluate disease predictions across demographic groups and several performance and calibration metrics. They report threshold-based and threshold-free measures, with confidence intervals from bootstrapping.

4. Experiments

  • The study uses all chest X-rays from MIMIC-CXR and CheXpert, with race and ethnicity, sex, and age groups, to predict No Finding, Pneumothorax, and Fracture.Uncertain labels are treated as negative for Fracture and Pneumothorax; labels are derived from radiology notes with the CheXpert labeller.Source: §4 · p. 4
  • Models use an ImageNet-pretrained DenseNet-121, and hyperparameters are selected using validation worst-group AUROC across five folds.Each dataset is split into a 16.7% test set and an 83.3% cross-validation set; five models are trained per hyperparameter setting.Source: §4 · p. 4
  • Evaluation includes AUROC, BCE, ECE, recall and specificity at a fixed threshold, and recall at a specified specificity; confidence intervals are bootstrapped over test samples and five trained models.Recall at a specified specificity still assesses performance at only one point of the risk distribution.Source: §4 · pp. 4–5
  • Recall and specificity are evaluated at a fixed threshold, whose selection depends on application-specific misclassification costs; opposite recall-specificity trade-offs do not identify the disadvantaged group without those costs.Source: §4 · p. 5

5. Results

The main-paper results use No Finding prediction in MIMIC-CXR as a representative example. They show substantial group disparities, no tested worst-group method outperforming simple data balancing, and performance costs when enforcing group fairness.

5.1. Significant Performance Gaps

  • ERM models show significant performance gaps across many metrics and protected attributes, especially worse AUROC in older populations, while gap direction can vary by metric.Black patients have lower specificity but larger recall than White patients, with no significant Black–White differences in AUROC or calibration error.Source: §5.1 · p. 5, Figure 1 · p. 6

5.2. No Method Outperforms Balanced-ERM on Worst-Case Group

  • No tested method outperforms Balanced ERM on any evaluation metric, while some instances show Balanced ERM outperforming ERM and Stratified ERM, especially in AUROC.JTT and ARL produce worse-performing, more poorly calibrated models; GroupDRO does not significantly improve worst-group performance.Source: §5.2 · p. 5, Figure 2 · p. 7

5.3. Group Fairness Worsens All Groups

  • For the three equalized-odds methods, large λ closes gaps but significantly degrades models, with lower AUROC and higher BCE and ECE.Increasing λ does not increase any performance metric for any group; any recall increase is accompanied by decreased specificity.Source: §5.3 · pp. 5, 7

6. Dissecting the Source of Bias

The paper probes whether label bias contributes to disparities in No Finding prediction, focusing on MIMIC-CXR and assuming its test distribution reflects deployment unless evidence suggests otherwise. Radiologist-checked labels and proxy-outcome analyses reveal label-quality and calibration patterns that vary across some groups, while leaving their causes uncertain.

6. Dissecting the Source of Bias

  • The analysis assumes the test-set distribution reflects deployment and focuses on label bias, leaving other potential bias sources for future work.The authors examine this assumption for MIMIC-CXR No Finding prediction.Source: §6 · p. 7

6.1. Radiologist-Labelled Samples

  • The CheXpert labeller is correct only 64.1% of the time when it labels a report No Finding; label quality varies by age, but not significantly by sex, ethnicity, or their intersections.The 80-and-older group has the poorest labels and the 18–40 group the best; age-related comorbidities may complicate both automated and human labeling.Source: §6.1 · p. 9, Figure 4 · p. 9

6.2. Proxy Labels

  • The proxy-label case study links MIMIC-CXR x-rays to hospital records and evaluates No Finding predictions against mortality and two length-of-stay outcomes.The analysis assumes proxy-label distributions do not differ across groups conditional on the true probability of No Finding.Source: §6.2 · p. 9
  • Against proxy labels, No Finding models show global risk overestimation without significant group differences in calibration curves, but mortality and length-of-stay analyses show some group stratification.The study reports higher mortality rates at a given predicted No Finding probability for patients with “Other” ethnicity, and lower probabilities of shorter stays for this group; age-stratified prolonged stays remain causally ambiguous.Source: §6.2 · pp. 9–10, Figure 5 · p. 10

7. Discussion

The discussion finds that current fairness methods do not improve worst-group performance beyond simple balancing, while equalized-odds methods worsen performance for all groups. It argues that clinical fairness requires attention to definitions, data-generating biases, and the consequences of deployment.

7.1. On Performance

  • Methods seeking better worst-case group performance do not outperform simple data balancing, while equalized-odds methods worsen performance for all groups.The authors conclude that current computational methods do not improve chest X-ray classifier fairness usefully over simple baselines.Source: §7.1 · p. 10

7.2. On Definitions of Fairness

  • Minimax fairness avoids the incompatibilities among group-fairness metrics, though selecting its error function still requires care.A minimax-fair classifier can be made group-fair by worsening nonworst-group performance; the reverse conversion is more difficult.Source: §7.2 · p. 10

7.3. On Sources of Bias

  • Label bias in No Finding is significant in MIMIC-CXR and may partially account for observed age-group performance disparities.Improving language-model agreement with radiologists could address label extraction bias, but not biases in report writing or radiological judgment.Source: §7.3 · p. 11

7.4. Best Practices for Fairness in Clinical Settings

  • Evaluate group performance across metrics and thresholds, including calibration, which remains relatively underexplored in clinical fairness research.Different group calibration curves mean a fixed operating threshold implies different true-risk thresholds across groups.Source: §7.4 · p. 11
  • Assess data-generating biases and whether disparities are clinically justified, since blindly equalizing gaps may worsen welfare for all.The authors note that some tasks may be inherently harder for certain groups, such as older populations with comorbidities.Source: §7.4 · p. 11
  • Computational fixes alone cannot ensure equitable care; the authors recommend accurate risk estimates, common thresholds, and exploring Pareto-optimal models.They leave clinical-classifier Pareto optimality and theoretical justifications for their empirical findings to future work.Source: §7.4 · p. 11

Appendix A. Group Fairness Definitions

  • Table A.1 organizes common group-fairness definitions by their conditional-independence statements and the metrics they equalize for binary predictions.Source: Table A.1 · p. 17

Appendix B. Cohort Statistics

  • Tables B.1 and B.2 summarize cohort characteristics and label prevalence by group for MIMIC-CXR and CheXpert.The experiments use all images in both datasets.Source: Table B.1 · p. 17, Table B.2 · p. 18

Appendix C. Additional Training Details

The appendix reports the shared optimizer settings, stopping rule, and image augmentations used during training.

Optimization

  • Models use Adam with batch size 64 and learning rate 10^-4 on NVIDIA Tesla P100 GPUs.Source: Appendix C · p. 18
  • Training stops when worst-case ROC fails to improve across five validation evaluations, which occur every 200 steps.Source: Appendix C · p. 18

Image Augmentations

  • Training images are randomly flipped horizontally, rotated up to 10 degrees, and cropped to random sizes and aspect ratios.Crop size ranges from 75% to 100%, and aspect ratio from 3/4 to 4/3.Source: Appendix C · p. 18

Appendix D. Additional Experimental Results - Main Experimental Grid

The appendix presents model-performance comparisons and differences from Balanced ERM for five task-and-dataset settings. The figure captions report 95% confidence intervals from 250 bootstrap iterations.

D.1. Pneumothorax Prediction in MIMIC-CXR

  • Figures D.1–D.2 compare Pneumothorax models on MIMIC-CXR and show each model's performance difference from Balanced ERM.Error bars indicate 95% confidence intervals from 250 bootstrap iterations.Source: Figure D.1 · p. 19, Figure D.2 · p. 20

D.2. Fracture Prediction in MIMIC-CXR

  • Figures D.3–D.4 compare Fracture models on MIMIC-CXR and show each model's performance difference from Balanced ERM.Error bars indicate 95% confidence intervals from 250 bootstrap iterations.Source: Figure D.3 · p. 21, Figure D.4 · p. 22

D.3. No Finding Prediction in CheXpert

  • Figures D.5–D.6 compare No Finding models on CheXpert and show each model's performance difference from Balanced ERM.Error bars indicate 95% confidence intervals from 250 bootstrap iterations.Source: Figure D.5 · p. 23, Figure D.6 · p. 24

D.4. Pneumothorax Prediction in CheXpert

  • Figures D.7–D.8 compare Pneumothorax models on CheXpert and show each model's performance difference from Balanced ERM.Error bars indicate 95% confidence intervals from 250 bootstrap iterations.Source: Figure D.7 · p. 25, Figure D.8 · p. 26

D.5. Fracture Prediction in CheXpert

  • Figures D.9–D.10 compare Fracture models on CheXpert and show each model's performance difference from Balanced ERM.Error bars indicate 95% confidence intervals from 250 bootstrap iterations.Source: Figure D.9 · p. 27, Figure D.10 · p. 28

E.1. Radiologist-Labelled Samples

The appendix describes a radiologist-labelled No Finding sample and presents its accuracy comparison, alongside proxy-label evaluations in a modified MIMIC-CXR cohort.

E.1. Radiologist-Labelled Samples

  • Figure E.1 compares CheXpert-labeller accuracy with radiologist gold labels across protected groups and intersectional subgroups in 1,200 reports labelled No Finding.Error bounds are 95% confidence intervals from 500 bootstrap iterations.Source: Figure E.1 · p. 29

E.2. Proxy Labels

  • The proxy-label cohort contains 42,877 images, and Figure E.2 compares an ERM model's performance on No Finding and three proxy labels.Table E.1 summarizes group proportions and label prevalence; Figure E.2 reports 95% confidence intervals from 250 bootstrap iterations.Source: Table E.1 · p. 29, Figure E.2 · p. 30
Loading 2203.12609v1…