CoChat

The DOI matches its Crossref record; 46 of the 47 references it cites were found in a registry or an index; 6 of the 34 sources its claims cite were checked for support.

Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations

Open access

Published version of a preprint

Abstract

Artificial intelligence (AI) systems have increasingly achieved expert-level performance in medical imaging applications. However, there is growing concern that such AI systems may reflect and amplify human bias, and reduce the quality of their performance in historically under-served populations such as female patients, Black patients, or patients of low socioeconomic status. Such biases are especially troubling in the context of underdiagnosis, whereby the AI algorithm would inaccurately label an individual with a disease as healthy, potentially delaying access to care. Here, we examine algorithmic underdiagnosis in chest X-ray pathology classification across three large chest X-ray datasets, as well as one multi-source dataset. We find that classifiers produced using state-of-the-art computer vision techniques consistently and selectively underdiagnosed under-served patient populations and that the underdiagnosis rate was higher for intersectional under-served subpopulations, for example, Hispanic female patients. Deployment of AI systems using medical imaging for disease diagnosis with such biases risks exacerbation of existing care biases and can potentially lead to unequal access to medical treatment, thereby raising ethical concerns for the use of these models in the clinic.

Overview

CoChat Summary

TL;DR

This study evaluates algorithmic underdiagnosis bias in chest X-ray classifiers across three large public datasets and a combined multi-source dataset. It finds that models consistently and selectively underdiagnose under-served subpopulations, especially intersectional groups, raising concerns for clinical deployment and triage.

Key findings

  • Underdiagnosis rate, defined as the false-positive rate (FPR) for the “no finding” label, differed across all considered subpopulations. Female patients, patients under 20 years old, Black patients, Hispanic patients, and patients with Medicaid insurance had higher algorithmic underdiagnosis rates in the MIMIC-CXR dataset.
  • Intersectional subgroups often had compounded bias. For example, Hispanic female patients had a higher underdiagnosis rate than white female patients, and patients aged 0–20 years who were female, Black, or had Medicaid insurance had the largest underdiagnosis rates.
  • FPR and false-negative rate (FNR) for “no finding” showed an inverse relationship across under-served subgroups, consistent with selective algorithmic underdiagnosis rather than undirected noise. This pattern held across datasets except for age >80 years and 0–20 years in the NIH dataset, possibly due to small sample sizes or dataset selection bias.
  • Dataset sizes and model AUCs: MIMIC-CXR (371,858 images from 65,079 patients; AUC 0.834 ± 0.001), CheXpert (223,648 images from 64,740 patients; AUC 0.805 ± 0.001), ChestX-ray14/NIH (112,120 images from 30,805 patients; AUC 0.835 ± 0.002), and ALL (707,626 images from 129,819 patients; AUC 0.859 ± 0.001).

Methods

  • Retrospective analysis of three public chest X-ray datasets—MIMIC-CXR, CheXpert, and ChestX-ray14 (NIH)—plus a multi-source ALL dataset combining them on shared labels. Subgroups included race/ethnicity, socioeconomic status via insurance type, sex, and age, with intersectional identities also examined.
  • A 121-layer DenseNet initialized with ImageNet weights was trained on 256 × 256 images to predict diagnostic labels, including the “no finding” label. Train-validation-test splits were fixed, no patient was shared across splits, and results were averaged over five models with different random seeds.
  • Fairness metrics were binarized: underdiagnosis rate as FPR for “no finding,” overdiagnosis as FNR for “no finding,” and false discovery rate (FDR) for “no finding.” A single threshold maximizing F1 score was selected for all groups; protected attributes were not used in model training or testing.

Limitations

  • The paper states that prediction labels are extracted from clinical records and are therefore not unbiased ground truth; the labels may already contain the same bias the model then amplifies, a form of bias amplification. It also notes that automatic NLP-based labelers have not been audited across subpopulations and could be a large source of bias.
  • The paper states that race and ethnicity are partially social constructs with unclear boundaries, and self-reported race/ethnicity may vary by age, socioeconomic level, or acculturation. Insurance type is described as an imperfect proxy for socioeconomic status.
  • The paper states that post-hoc threshold adjustments for equal FNR and FPR have flaws: small intersectional groups make accurate thresholds difficult, the number of thresholds grows exponentially with protected attributes, and equalizing metrics may require randomization or worsening performance for some subgroups. The text also makes evident that the NIH dataset may have less diverse samples and small subgroup sizes, and that MIMIC-CXR lacks race/ethnicity and insurance data for around ~100,000 X-rays.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s41591-021-01595-0

Cites

46 of 46 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 The model pipeline. a, We examine chest radiographs across several datasets with diverse populations. b, A deep learning model is then trained from these data (training across all patients simultaneously) to predict the presence of the no finding label, which indicates that the algorithm did not detect disease for the image. c, The underdiagnosis rate (that is, the false-positive rate (FP) of the no finding label) of this model is then compared in different subpopulations (including sex, race/ethnicity, age and insurance type) to examine the algorithm’s underdiagnosis rate. FN, false negative; TN, true negative; TP, true positive. Symbol colors indicate different races of male and female patients.
Fig. 1 The model pipeline. a, We examine chest radiographs across several datasets with diverse populations. b, A deep learning model is then trained from these data (training across all patients simultaneously) to predict the presence of the no finding label, which indicates that the algorithm did not detect disease for the image. c, The underdiagnosis rate (that is, the false-positive rate (FP) of the no finding label) of this model is then compared in different subpopulations (including sex, race/ethnicity, age and insurance type) to examine the algorithm’s underdiagnosis rate. FN, false negative; TN, true negative; TP, true positive. Symbol colors indicate different races of male and female patients.CC BY · PubMed Central
Fig. 2 Analysis of underdiagnosis across subgroups of sex, age, race/ethnicity and insurance type in the MIMIC-CXR (CXR) dataset. a, The underdiagnosis rate, as measured by the no finding FPR, in the indicated patient subpopulations. b, Intersectional underdiagnosis rates for female patients (b(i)), patients aged 0–20 years (b(ii)), Black patients (b(iii)), and patients with Medicaid (b(iv)). c,d, The overdiagnosis rate, as measured by the no finding FNR in the same patient subpopulations as in a and b. The results are averaged over five trained models with different random seeds on the same train–validation–test splits. 95% confidence intervals are shown. Subgroups with too few members to be studied reliably (≤15) are labeled in gray text and the results for these subgroups are omitted. Data for the Medicare subgroup are also omitted, given that data for this subgroup are highly confounded by patient age.
Fig. 2 Analysis of underdiagnosis across subgroups of sex, age, race/ethnicity and insurance type in the MIMIC-CXR (CXR) dataset. a, The underdiagnosis rate, as measured by the no finding FPR, in the indicated patient subpopulations. b, Intersectional underdiagnosis rates for female patients (b(i)), patients aged 0–20 years (b(ii)), Black patients (b(iii)), and patients with Medicaid (b(iv)). c,d, The overdiagnosis rate, as measured by the no finding FNR in the same patient subpopulations as in a and b. The results are averaged over five trained models with different random seeds on the same train–validation–test splits. 95% confidence intervals are shown. Subgroups with too few members to be studied reliably (≤15) are labeled in gray text and the results for these subgroups are omitted. Data for the Medicare subgroup are also omitted, given that data for this subgroup are highly confounded by patient age.CC BY · PubMed Central
Extended Data Fig. 1 Analyzing underdiagnoses over subgroups of sex, age, within ALL dataset (combined CXR, CXP and NIH dataset on shared labels). Fig. S1. Analyzing underdiagnoses over subgroups of sex, age, within ALL dataset (combined CXR, CXP and NIH dataset on shared labels). The results are averaged over 5 trained model with different random seed ± 95% confidence interval (CI). A. The underdiagnosis rate (measured by ‘No Finding’ FPR). B. The overdiagnosis rate (‘No Finding’ False Negative Rate (FNR)) over subgroups of sex, age. C. The intersectional underdiagnosis rates within only female patients. D. Examining the overdiagnosis rate for the intersectional identities. The number of images with actual 0 or 1 ‘No Finding’ label in the age - sex intersections in the test dataset is presented in Supplementary Table 1.
Extended Data Fig. 1 Analyzing underdiagnoses over subgroups of sex, age, within ALL dataset (combined CXR, CXP and NIH dataset on shared labels). Fig. S1. Analyzing underdiagnoses over subgroups of sex, age, within ALL dataset (combined CXR, CXP and NIH dataset on shared labels). The results are averaged over 5 trained model with different random seed ± 95% confidence interval (CI). A. The underdiagnosis rate (measured by ‘No Finding’ FPR). B. The overdiagnosis rate (‘No Finding’ False Negative Rate (FNR)) over subgroups of sex, age. C. The intersectional underdiagnosis rates within only female patients. D. Examining the overdiagnosis rate for the intersectional identities. The number of images with actual 0 or 1 ‘No Finding’ label in the age - sex intersections in the test dataset is presented in Supplementary Table 1.CC BY · PubMed Central
Extended Data Fig. 2 Analyzing underdiagnoses over subgroups of sex, age, within CheXpert (CXP) dataset. Fig. S2. Analyzing underdiagnoses over subgroups of sex, age, within CheXpert (CXP) dataset. The results are averaged over 5 trained model with different random seed ± 95% CI. A. The underdiagnosis rate is FPR in ‘No Finding’. B. Examining the overdiagnosis rate (‘No Finding’ FNR) over sex and age subgroups, C. The intersectional underdiagnosis rates within only female patients, and D. measure the overdiagnosis rate for the intersectional identities. The subgroups labeled in gray text, with results omitted, indicate the subgroup has too few members (<= 15) to be used reliably. The number of images with actual 0 or 1 ‘No Finding’ label in the age - sex intersections in the test dataset is presented in Supplementary Table 1.
Extended Data Fig. 2 Analyzing underdiagnoses over subgroups of sex, age, within CheXpert (CXP) dataset. Fig. S2. Analyzing underdiagnoses over subgroups of sex, age, within CheXpert (CXP) dataset. The results are averaged over 5 trained model with different random seed ± 95% CI. A. The underdiagnosis rate is FPR in ‘No Finding’. B. Examining the overdiagnosis rate (‘No Finding’ FNR) over sex and age subgroups, C. The intersectional underdiagnosis rates within only female patients, and D. measure the overdiagnosis rate for the intersectional identities. The subgroups labeled in gray text, with results omitted, indicate the subgroup has too few members (<= 15) to be used reliably. The number of images with actual 0 or 1 ‘No Finding’ label in the age - sex intersections in the test dataset is presented in Supplementary Table 1.CC BY · PubMed Central
Extended Data Fig. 3 Analyzing underdiagnoses over subgroups of sex, age, within ChestX-ray14 (NIH) dataset. Fig. S3. Analyzing underdiagnoses over subgroups of sex, age, within ChestX-ray14 (NIH) dataset. The results are averaged over 5 trained model with different random seed ± 95% confidence interval (CI). A. The underdiagnosis rate (‘No Finding’ FPR). B. The over diagnosis rate (‘No Finding’ FNR) over subgroups of sex and age. C. The intersectional underdiagnosis rates within only female patients. D. The over diagnosis rate for the intersectional identities. The subgroups labeled in gray text, with results omitted, indicate the subgroup has too few members (<= 15) to be used reliably. The number of images with actual 0 or 1 ‘No Finding’ label in the age - sex intersections in the test dataset is presented in Supplementary Table 1.
Extended Data Fig. 3 Analyzing underdiagnoses over subgroups of sex, age, within ChestX-ray14 (NIH) dataset. Fig. S3. Analyzing underdiagnoses over subgroups of sex, age, within ChestX-ray14 (NIH) dataset. The results are averaged over 5 trained model with different random seed ± 95% confidence interval (CI). A. The underdiagnosis rate (‘No Finding’ FPR). B. The over diagnosis rate (‘No Finding’ FNR) over subgroups of sex and age. C. The intersectional underdiagnosis rates within only female patients. D. The over diagnosis rate for the intersectional identities. The subgroups labeled in gray text, with results omitted, indicate the subgroup has too few members (<= 15) to be used reliably. The number of images with actual 0 or 1 ‘No Finding’ label in the age - sex intersections in the test dataset is presented in Supplementary Table 1.CC BY · PubMed Central

Verification

Claims the paper makes

4 Valid · 1 Not valid · 16 Not checked · 53 Own results and interpretations

24 more claims

Create a free account to see all 74 claims and to verify the 28 references still open.

See all claims — free account
  • Probably not a substantial claim. 29
  • It may state two separate claims as one. 32

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

6 more claims about this paper

Create a free account to see every claim other papers make about this one.

See all claims — free account

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

2 of 2

Citations

Cites

Show all 47 references

Research works printed but not linked: 6

  1. Sun, T. Y., Bear Don’t Walk, O. J. IV, Chen, J. L., Reyes Nieva, H. & Elhadad, N. Exploring gender disparities in time to diagnosis. In Machine Learning for Health (ML4H) at NeurIPS 2020 (eds Alsentzer, E. et al.) abstr. https://arxiv.org/abs/2011.06100 (2020). · arXiv
  2. Cohen, J. P., Hashir, M., Brooks, R. & Bertrand, H. On the limits of cross-domain generalization in automated X-ray prediction. Proc. Mach. Learn. Res. 121, 136–155 (2020).
  3. Akbarian, S., Seyyed-Kalantari, L., Khalvati, F. & Dolatabadi, E. Evaluating knowledge transfer in neural networks for medical images. Preprint at https://arxiv.org/abs/2008.13574 (2020). · arXiv
  4. Hardt, M., Price, E. & Srebro, N. Equality of opportunity in supervised learning. In Advances in Neural Information Processing Systems 29 (NIPS 2016) (eds Lee, D. et al.) 3315–3323.
  5. del Barrio, E., Gordaliza, P. & Loubes, J.-M. Review of mathematical frameworks for fairness in machine learning. Preprint at https://arxiv.org/abs/2005.13755 (2020). · arXiv
  6. Johnson, A. et al. MIMIC-IV (version 0.4). PhysioNet https://doi.org/10.13026/a3wn-hq05 (2020). · DOI

Cited by

See all 988

Metadata

Authors
Seyyed-Kalantari, L., Zhang, H., McDermott, M. B. A., Chen, I. Y., & Ghassemi, M.
Affiliations
  • University of Toronto (CA)
  • Massachusetts Institute of Technology (US)
  • Vector Institute (CA)
Corresponding author
Seyyed-Kalantari, Laleh
Venue
Nature Medicine
Published
2021-12
Volume
27
Issue
12
Pages
2176-2182
ISSN
  • 1078-8956 (ISSN-L)
  • 1546-170X
OpenAlex
W4200191097
Cited by
928
References
47
Citations by year
2026: 298 · 2025: 239 · 2024: 199 · 2023: 146 · 2022: 42 · 2021: 4
FWCI
27.63
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Medicine (publishedVersion, cc-by)
  • Research Square (acceptedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
Funders
  • Microsoft Research
  • Canadian Institute for Advanced Research
  • Natural Sciences and Engineering Research Council of Canada (PDF-516984)
Topics
  • Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
  • Physical Sciences › Computer Science › Artificial Intelligence › Ethics and Social Impacts of AI
  • Health Sciences › Medicine › Radiology, Nuclear Medicine and Imaging › COVID-19 diagnosis using AI
Keywords
algorithmic bias, underdiagnosis, chest X-ray classification, artificial intelligence in medical imaging, underserved populations, health disparities, algorithmic fairness, health care access, Black patients, Hispanic patients, female patients, low socioeconomic status
MeSH terms
Adolescent, Algorithms, Artificial Intelligence, Child, Child, Preschool, Female, Humans, Infant, Infant, Newborn, Male, Radiography, Thoracic, Vulnerable Populations, Young Adult, Datasets as Topic

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-04, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-04 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (qwen3.8-omni-flash and gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations. https://app.cochat.ai/papers/10.1038/s41591-021-01595-0
The paper
Seyyed-Kalantari, L., Zhang, H., McDermott, M. B. A., Chen, I. Y., & Ghassemi, M. (2021). Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations. Nature Medicine. https://doi.org/10.1038/s41591-021-01595-0

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.