Skip to main content

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

  • Matters Arising
  • Published:

Confounding factors need to be accounted for in assessing bias by machine learning algorithms

Nature Medicine volume 28, pages 1159–1160 (2022)Cite this article

Subjects

Matters Arising to this article was published on 16 June 2022

The Original Article was published on 10 December 2021

arising from L. Seyyed-Kalantari et al. Nature Medicine https://doi.org/10.1038/s41591-021-01595-0 (2021).

The recent paper titled “Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations” by Seyyed-Kalantari et al. highlights the important and timely question of bias in medical artificial intelligence (AI)1 models. However, we are concerned that the paper did not adequately account for confounding factors in some of the analyses. We highlight two such instances here. The first stems from focusing on the ‘no-finding’ label for the analysis. Some findings are more difficult to identify than others (either by human radiologists or by AI models); if these difficult findings are disproportionately present in one subgroup versus another, we may observe a disparity in overall performance between the two subgroups, even if the model is perfectly ‘fair’ in the sense that, for an individual finding, the performance of the model is equal across both subgroups. The second instance is assessing the effect of one individual factor on model performance. While considering one factor, say sex, one must control for other confounders such as age and race to avoid spurious correlations. For example, if a model performs much better on older individuals than on the young (maybe because diseases present more severely and the corresponding findings are easier to identify in older individuals), and the female subgroup contains disproportionately fewer older people compared to the male subgroup, we may observe overall worse performance of the model for the female subgroup, even if the model is ‘fair’ and has equal performance when tested on males or females of the same age. We elaborate on each issue below.

This is a preview of subscription content, access via your institution

Access options

Buy this article

  • Purchase on Springer Link
  • Instant access to full article PDF
Buy now

Prices may be subject to local taxes which are calculated during checkout

References

  1. Seyyed-Kalantari, L. et al. Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations. Nat. Med. 27, 2176–2182 (2021).

    Article  CAS  Google Scholar 

  2. Seyyed-Kalantari, L., Liu, G., McDermott, M., Chen, I. Y. & Ghassemi, M. CheXclusion: fairness gaps in deep chest X-ray classifiers. In Pacific Symposium on Biocomputing 2021 (eds Altman, R. B. et al.) 232–243 (World Scientific Publishing, 2021).

Download references

Author information

Authors and Affiliations

  1. Imaging Biomarkers and Computer-Aided Diagnosis Laboratory, Department of Radiology and Imaging Sciences, National Institutes of Health Clinical Center, Bethesda, MD, USA

    Pritam Mukherjee, Thomas C. Shen, Jianfei Liu, Tejas Mathai, Omid Shafaat & Ronald M. Summers

Authors
  1. Pritam Mukherjee
    View author publications

    You can also search for this author in PubMed Google Scholar

  2. Thomas C. Shen
    View author publications

    You can also search for this author in PubMed Google Scholar

  3. Jianfei Liu
    View author publications

    You can also search for this author in PubMed Google Scholar

  4. Tejas Mathai
    View author publications

    You can also search for this author in PubMed Google Scholar

  5. Omid Shafaat
    View author publications

    You can also search for this author in PubMed Google Scholar

  6. Ronald M. Summers
    View author publications

    You can also search for this author in PubMed Google Scholar

Contributions

P.M. contributed to conception, manuscript drafting and manuscript editing. T.C.S. contributed to conception, manuscript drafting and manuscript editing. J.L., T.M. and O.S. performed manuscript drafting and manuscript editing. R.M.S. contributed to conception, manuscript drafting, manuscript editing and supervision.

Corresponding author

Correspondence to Pritam Mukherjee.

Ethics declarations

Competing interests

R.M.S. receives royalties for patents or software licenses from iCAD, Philips, PingAn, ScanMed and Translation Holdings. The laboratory of R.M.S. receives research support through a Cooperative Research and Development Agreement from PingAn. All other authors declare no competing interests.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Mukherjee, P., Shen, T.C., Liu, J. et al. Confounding factors need to be accounted for in assessing bias by machine learning algorithms. Nat Med 28, 1159–1160 (2022). https://doi.org/10.1038/s41591-022-01847-7

Download citation

  • Received:

  • Accepted:

  • Published:

  • Issue Date:

  • DOI: https://doi.org/10.1038/s41591-022-01847-7

This article is cited by

Search

Advanced search

Quick links

Nature Briefing: Translational Research

Sign up for the Nature Briefing: Translational Research newsletter — top stories in biotechnology, drug discovery and pharma.

Get what matters in translational research, free to your inbox weekly. Sign up for Nature Briefing: Translational Research