Source-linked AI summary
CheXclusion: Fairness gaps in deep chest X-ray classifiers
Laleh Seyyed-Kalantari, Guanxiong Liu, Matthew McDermott, Irene Y. Chen, Marzyeh Ghassemi
TL;DR
The paper asks whether state-of-the-art chest X-ray classifiers exhibit fairness gaps across protected subgroups, an important question as high-performing clinical models approach deployment. It trains CNNs across public datasets and evaluates subgroup TPR disparities for diagnostic labels and protected attributes. The study finds disparities across all datasets and tasks, with smaller gaps for a multi-source dataset and no consistent relationship with proportional disease burden. It therefore urges careful fairness auditing before clinical deployment.
Problem
High-performing clinical imaging classifiers may perform differently across protected subgroups, making fairness auditing important before deployment.
Method
The authors train CNN classifiers for 14 diagnostic labels on MIMIC-CXR, CheXpert, Chest-Xray8, and an aggregated multi-source dataset, then measure subgroup TPR disparities.
Results
TPR disparities occur across all datasets, diagnostic tasks, and subgroups; the multi-source dataset has smaller disparities, and disparities are not significantly correlated with proportional disease burden.
Takeaways & Limitations
Clinical decision makers should carefully audit algorithmic disparities before deploying chest X-ray classifiers.
Takeaways & Limitations
Diagnostic labels are silver labels automatically extracted from radiology reports, and dataset differences in imaging devices, collection regions, and patient demographics warrant further investigation.
Abstract
from arXiv · showhide
Machine learning systems have received much attention recently for their ability to achieve expert-level performance on clinical tasks, particularly in medical imaging. Here, we examine the extent to which state-of-the-art deep learning classifiers trained to yield diagnostic labels from X-ray images are biased with respect to protected attributes. We train convolution neural networks to predict 14 diagnostic labels in 3 prominent public chest X-ray datasets: MIMIC-CXR, Chest-Xray8, CheXpert, as well as a multi-site aggregation of all those datasets. We evaluate the TPR disparity -- the difference in true positive rates (TPR) -- among different protected attributes such as patient sex, age, race, and insurance type as a proxy for socioeconomic status. We demonstrate that TPR disparities exist in the state-of-the-art classifiers in all datasets, for all clinical tasks, and all subgroups. A multi-source dataset corresponds to the smallest disparities, suggesting one way to reduce bias. We find that TPR disparities are not significantly correlated with a subgroup's proportional disease burden. As clinical models move from papers to products, we encourage clinical decision makers to carefully audit for algorithmic disparities prior to deployment. Our code can be found at, https://github.com/LalehSeyyed/CheXclusion
1. Introduction
The paper examines whether high-performing deep chest X-ray classifiers are fair across protected subgroups, motivated by their potential clinical use amid radiologist shortages. It evaluates TPR disparities across datasets, diagnostic tasks, and attributes, finding extensive disparities and smaller gaps for the aggregated multi-source dataset.
- Radiologist shortages motivate deep-learning chest X-ray classifiers as potential tools for expanding screening and diagnosis.
- The study trains CNN classifiers on MIMIC-CXR, CheXpert, Chest-Xray8, and their aggregated dataset to predict 14 diagnostic labels.
- Protected attributes include sex, age, race, and insurance type as an imperfect proxy for socioeconomic status.
- TPR disparity measures differences in correct positive diagnoses among protected subgroups, even when overall classifier accuracy is high.
- TPR disparities appear extensively across datasets, while the multi-source dataset yields the lowest disparities.
- Disparities are generally not significantly correlated with subgroups’ proportional disease membership, so increasing subgroup patient counts alone may not address vulnerability.
2. Background and Related Work
Related work frames fairness as a task-dependent concern arising throughout the machine-learning pipeline, with ethical risks in healthcare and limited prior auditing of chest X-ray classifiers across several attributes.
- Fairness and Debiasing: Fairness research has documented sex, racial, and intersectional biases in occupation, criminal-risk, and facial-analysis systems.
- Fairness and Debiasing: Bias can enter through input data, labels, cohort composition, feature collection, and subgroup under-sampling.
- Fairness and Debiasing: Fairness definitions can conflict, so disparity metrics should be selected according to the task.
- Fairness and Debiasing: The paper adopts equality of opportunity to evaluate correct diagnoses across protected attribute groups.
- Ethical Algorithms in Health: Healthcare algorithms raise patient-harm concerns, with prior evidence of racial bias and socioeconomic-status-related EHR misclassification.
- Chest X-Ray Classification: Prior chest X-ray work established strong diagnostic performance, but had not examined systematic bias over age, race, and insurance type.
3. Data
The study uses three public chest X-ray datasets and constructs a multi-site aggregation over their shared labels, while recording available demographic and insurance attributes.
- The datasets are MIMIC-CXR, CheXpert, and Chest-Xray8, each associated with 14 diagnostic labels.
- The multi-site dataset aggregates images from all three sources using the eight labels shared across them.
- CXR and CXP combine non-positive labels into an aggregate negative category, and include a No Finding label for images without diagnosed disease.
- Recorded protected attributes include sex, age, race, and insurance type, although attribute availability differs by dataset.
- Race and insurance data are available for CXR, while NIH, CXP, and ALL provide sex and age data.
4. Methods
The methods train CNN-based multi-label classifiers separately on four datasets and evaluate fairness using subgroup TPR disparities, with repeated models and dataset-specific preprocessing and tuning.
- 4.1. Models: Separate CNN models classify 14 labels for CXR, CXP, NIH, and ALL, evaluating sex and age across all datasets and race and insurance for CXR.
- 4.1. Models: A 121-layer DenseNet initialized with ImageNet weights is trained using multi-label binary cross-entropy loss.
- 4.1. Models: The dataset descriptions cover image and patient counts, view types, and subgroup proportions across the chest X-ray datasets and their eight-label aggregation.
- 4.1. Models: Dataset-specific learning rates and rotation settings are tuned, then fixed while four additional random-seed models are trained per dataset.
- 4.2. Classifier Disparity Evaluation: TPR disparity is the primary bias measure: binary attributes compare subgroup TPRs, while non-binary attributes compare subgroup TPRs with the median subgroup TPR.
5. Experiments
The experiments first establish near-state-of-the-art classifier performance, then measure subgroup fairness through TPR disparities and examine whether those disparities track subgroup disease membership.
- 5. Experiments: The classifiers achieve near-state-of-the-art classification performance, reducing concern that fairness findings simply reflect poor overall performance.The experiments therefore assess fairness in models with strong overall diagnostic performance.
- 5. Experiments: TPR disparity is quantified separately for each subgroup and disease, across sex and age in four datasets and race and insurance type in CXR.The analysis covers protected attributes according to data availability.
- 5. Experiments: The study tests whether positive patient proportions within subgroups and labels are associated with TPR disparities.This probes whether data imbalance could explain observed disparities or propagate vulnerability in smaller subgroups.
6. Results
The trained classifiers show near-state-of-the-art diagnostic performance but also widespread TPR disparities across labels, datasets, and protected attributes. Disparities generally do not significantly track subgroup disease membership, while the aggregated dataset yields smaller average gaps for sex and age.
- 6. Results: The classifiers achieve near-state-of-the-art performance across datasets and tasks, with similar performance to published NIH state-of-the-art results.Published CheXpert results are not comparable because they use a private dataset with only 200 images and five labels, unlike this study’s 22,274-image test set.
- 6. Results: TPR disparities occur across labels, datasets, and protected attributes, with both positive and negative disparities indicating favorable and unfavorable subgroups.In a fair setting, subgroup disparities would be near zero within each label.
- 6. Results: 0.226 is the average cross-14-label race gap; No Finding has the smallest gap at 0.119 and Pneumonia the largest at 0.440.Hispanic patients are most unfavorable for 9/13 labels, whereas White patients are favorable for 9/13 labels.
- 6. Results: The most frequent unfavorable subgroups are often socially disadvantaged groups such as women and minorities, but no disease consistently has the highest or lowest disparity.The aggregated ALL model has the smallest average cross-label gap for sex and age.
- 6. Results: Most TPR-disparity correlations with subgroup disease membership are not statistically significant after testing 33 hypotheses with a Bonferroni threshold of p < 0.0015.The study measures Pearson correlation coefficients between TPR disparities and patient proportions per label.
7. Summary and Discussion
Across datasets and tasks, the classifiers show meaningful TPR disparities across protected attributes, while the multi-source dataset produces smaller disparities for age and sex. These disparities are not consistently explained by subgroup disease membership and motivate fairness checks before deployment.
- All datasets and tasks display non-trivial TPR disparities across protected attributes.The analysis covers sex, age, race, and insurance type, with race and insurance examined in CXR.
- Female and 0-20 subgroups show consistently unfavorable outcomes across age and sex settings.Female patients are described as the least favored sex subgroup in every setting.
- In CXR, Hispanic patients and Medicaid-insured patients are the most unfavorable subgroups for race and insurance type, respectively.White patients are the most favorable race subgroup, while Medicaid patients often receive incorrect diagnoses.
- The multi-source dataset led to the smallest disparities for age and sex among the four datasets.The authors hypothesize that combining large datasets reduces data-collection bias.
- Only 6 of 33 subgroups showed statistically significant correlation between subgroup membership and TPR disparities.In NIH, Consolidation, Nodule, and Pneumothorax each had 45% female patients but TPR disparities of -0.155, -0.079, and 0.047, respectively.
- Fairness audits should accompany general-population accuracy evaluation before deployment because disparities could create unequal care.The paper recommends rigorous fairness analyses and notes that dataset and modeling changes may be needed.
8. Limitations and Future Work
The authors identify limitations involving future investigation and the quality and representativeness of the data used to measure fairness. In particular, automatically extracted labels and site-specific confounders may affect both labels and fairness metrics.
- The work is presented as a first step in quantifying model biases, with important future work remaining.
- Diagnostic labels are at best “silver” labels because they are automatically extracted from radiology reports using NLP.Label errors could compound with observed biases or model errors.
- Imaging devices, collection regions, hospital sites, and patient demographics may affect label quality and observed fairness metrics.The datasets represent different clinical settings, including more complicated, tertiary, and emergency-department cases.
9. Conclusion
The paper shows that high-capacity chest X-ray classifiers trained on large datasets do not naturally provide equality of opportunity. It quantifies TPR disparities across four datasets, 14 labels, and several protected attributes, highlighting potential deployment-related disparities in care.
- High-capacity models trained on large datasets do not naturally provide equality of opportunity.The authors report potential disparities in care if such models are deployed without modification.
- The study quantifies TPR disparities across four datasets, 14 diagnostic labels, and sex, age, race, and insurance type.
Appendix A. Distribution of TPR Disparity per Attributes, Subgroups and Labels
The appendix distributes TPR disparities across protected-attribute subgroups and disease labels, showing substantial variation in subgroup favorability and gap size across datasets.
- Measurement: TPR disparity compares a subgroup’s true positive rate with another subgroup’s rate or, for age, race, and insurance, the subgroup median.Negative disparities indicate bias against a subgroup, while positive disparities indicate bias in its favor.
- Sex disparities: 0.045 average cross-label gap separates the least and most favorable sex subgroups in the ALL dataset.Effusion has the smallest gap at 0.001, while Pneumonia has the largest at 0.105.
- Sex disparities: 0.072 average cross-label gap separates the least and most favorable sex subgroups in MIMIC-CXR.Female has negative disparities for 10/13 disease labels, and Enlarged Cardiomediastinum has the largest gap at 0.151.
- Sex disparities: 0.062 average cross-label gap separates the least and most favorable sex subgroups in CheXpert.Edema has a 0.000 smallest gap, whereas Consolidation has the largest gap at 0.139.
- Sex disparities: 0.190 average cross-label gap separates the least and most favorable sex subgroups in ChestXray8.Mass has the smallest gap at 0.001, while Cardiomegaly has the largest at 0.393.
- Age disparities: Age-disparity plots report subgroup-specific negative-gap counts across disease labels for ALL, MIMIC-CXR, CheXpert, and ChestXray8.The supplied captions also encode subgroup membership through circle size and average disparities over five runs with 95% confidence intervals.