Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 93 of the 93 references it cites were found in a registry or an index; 53 of the 114 sources its claims cite were checked for support.

Bias recognition and mitigation strategies in artificial intelligence healthcare applications

Open access

Abstract

Artificial intelligence (AI) is delivering value across all aspects of clinical practice. However, bias may exacerbate healthcare disparities. This review examines the origins of bias in healthcare AI, strategies for mitigation, and responsibilities of relevant stakeholders towards achieving fair and equitable use. We highlight the importance of systematically identifying bias and engaging relevant mitigation activities throughout the AI model lifecycle, from model conception through to deployment and longitudinal surveillance.

Overview

CoChat Summary

TL;DR

This critical/narrative review examines the origins of bias in healthcare AI and mitigation strategies across the AI model lifecycle, from conception through deployment and surveillance. It argues that systematically identifying and mitigating bias is necessary to uphold fairness, equity, and equality in clinical AI.

Key findings

  • As of May 13, 2024, the FDA listed 191 new AI-enabled medical devices, reaching 882 total, predominantly in radiology (76%), cardiology (10%), and neurology (4%). The review’s literature search yielded 233 potentially relevant articles, narrowed to 152 after title/abstract screening and 94 after full-text review.
  • Cited evidence shows substantial bias burden: Kumar et al. found 50% of 48 sampled healthcare AI studies had high risk of bias and only 20% low risk; Chen et al. found only 86 of 555 neuroimaging-based psychiatric AI models (15.5%) included external validation, 97.5% included only high-income-region subjects, and 83% were rated high risk of bias.
  • Case examples include Obermeyer et al.’s algorithm using healthcare costs as a proxy: among 43,539 White and 6,079 Black patients, Black patients had 26.3% more chronic illnesses at the same risk score (4.8 vs. 3.8 conditions); recalibration using direct health indicators nearly tripled high-risk Black patient enrollment from 17.7% to 46.5%. In cardiac MRI segmentation (nnU-Net, 5,903 UK Biobank subjects, 80% White), Dice similarity was 93.5% for White subjects but as low as 84.5% for Black and Mixed-race subjects; protected group models reached 92.15% for Black and 93.17% for Mixed-race subjects.
  • Deployment biases are documented: incorrect AI BI-RADS suggestions dropped mammography reporting accuracy from 79.7% to 19.8% for inexperienced radiologists, 81.3% to 24.8% for moderately experienced, and 82.3% to 45.5% for very experienced radiologists; in a COVID-19 decision-support study, only 35% given risk score alone and 52% given risk score plus explanation critically evaluated incorrect recommendations; an ICU alarm study found 1,154,201 unique alarms, 88.8% for arrhythmias, with most false positives.

Methods

  • The authors followed a critical review methodology, searching English articles published from 1993 to 2024 in PubMed, Google Scholar, and publisher platforms such as Elsevier, using Boolean operators and keywords including “Medical AI,” “AI bias,” “AI ethics,” “Fairness in medical AI,” and “Bias mitigation strategies in AI.”
  • The search yielded 233 potentially relevant articles; after title/abstract screening, 152 were considered directly relevant, and 94 were selected for final review. A thematic-based approach was used to identify and categorize recurring themes, patterns, and insights related to bias types and mitigation strategies.
  • The review is structured around bias origins and types—human, data, algorithmic, and model deployment—and mitigation recommendations across six AI lifecycle phases: conception, data collection, pre-processing, in-processing, post-processing, and post-deployment surveillance.

Limitations

  • The paper states that achieving unbiased, broadly representative healthcare datasets remains formidable: data sparsity for underrepresented populations may not be feasible, prospective data quality improvement is time-consuming, sampling and participation biases are difficult to eliminate, standardizing data collection to reduce measurement bias is resource-intensive, vulnerable groups may remain underrepresented due to abundant WEIRD-population resources, and some socio-demographic features such as gender orientation remain poorly captured.
  • The paper states that specific mitigation techniques have limitations: disaggregation can create noisy overly granular datasets; multiple imputation may introduce inaccuracies; data augmentation can generate synthetic data that fails to reflect true diversity; SMOTE can produce unrealistic data points; random under-sampling risks information loss; cost-sensitive learning can overfit; fairness metrics can create a fairness-accuracy trade-off; federated learning limits data pre-processing/quality assurance access; adversarial training can be computationally intensive and reduce accuracy; transfer learning may transfer unrecognized biases; explainability tools can be inconsistent or misleading; external validation data access is often challenging; and shadow deployment can delay implementation.
  • The paper does not state limitations of the review itself; evident from the text, it is a critical/narrative review using thematic synthesis and does not report a formal risk-of-bias assessment or meta-analysis of the 94 included articles.

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

Integrity signals

Registry

Crossref · 10.1038/s41746-025-01503-7

Cites

93 of 93 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 Study selection procedure for identifying relevant articles on AI bias in healthcare. This figure illustrates the process used to filter and review literature from 233 initial articles, narrowing down to 94 articles that directly explored AI bias and mitigation strategies in healthcare.
Fig. 1 Study selection procedure for identifying relevant articles on AI bias in healthcare. This figure illustrates the process used to filter and review literature from 233 initial articles, narrowing down to 94 articles that directly explored AI bias and mitigation strategies in healthcare.CC BY-NC-ND · PubMed Central
Fig. 2 Schematic description of equality and equity in healthcare. This figure illustrates the key differences between Equality and Equity in healthcare support. It presents two scenarios: a “Equality” is depicted where each individual, regardless of their needs, receives the same level of support, symbolized by equal height green platforms for all figures. b “Equity” is shown where supports are varied according to individual needs, represented by green platforms of different heights, ensuring that each person reaches the support they need. This visualization underscores the importance of tailoring healthcare resources to address specific needs to achieve true equity.
Fig. 2 Schematic description of equality and equity in healthcare. This figure illustrates the key differences between Equality and Equity in healthcare support. It presents two scenarios: a “Equality” is depicted where each individual, regardless of their needs, receives the same level of support, symbolized by equal height green platforms for all figures. b “Equity” is shown where supports are varied according to individual needs, represented by green platforms of different heights, ensuring that each person reaches the support they need. This visualization underscores the importance of tailoring healthcare resources to address specific needs to achieve true equity.CC BY-NC-ND · PubMed Central
Fig. 3 The AI model life cycle and common biases across each phase. This figure maps the stages of the AI model life cycle in healthcare, highlighting the common phase at which biases can be introduced. The AI life cycle is divided into six phases: conception, data collection, pre-processing, in-processing (algorithm development and validation), post-processing (clinical deployment), and post-deployment surveillance. Each phase is prone to specific biases that can affect the fairness, equity, and equality of healthcare delivery.
Fig. 3 The AI model life cycle and common biases across each phase. This figure maps the stages of the AI model life cycle in healthcare, highlighting the common phase at which biases can be introduced. The AI life cycle is divided into six phases: conception, data collection, pre-processing, in-processing (algorithm development and validation), post-processing (clinical deployment), and post-deployment surveillance. Each phase is prone to specific biases that can affect the fairness, equity, and equality of healthcare delivery.CC BY-NC-ND · PubMed Central
Fig. 4 Bias mitigation strategies across AI model life cycle. This figure outlines the key strategies for addressing biases at different phases of the AI model life cycle. It highlights critical interventions from the initial conception phase through to post-deployment surveillance, ensuring that each phase in the AI development process incorporates practices aimed at promoting fairness, equity, and effectiveness. The strategies are categorized by the model’s lifecycle phases, providing a roadmap for systematic bias mitigation in AI applications.
Fig. 4 Bias mitigation strategies across AI model life cycle. This figure outlines the key strategies for addressing biases at different phases of the AI model life cycle. It highlights critical interventions from the initial conception phase through to post-deployment surveillance, ensuring that each phase in the AI development process incorporates practices aimed at promoting fairness, equity, and effectiveness. The strategies are categorized by the model’s lifecycle phases, providing a roadmap for systematic bias mitigation in AI applications.CC BY-NC-ND · PubMed Central

Verification

Claims the paper makes

46 Valid · 4 Not valid · 55 Not checked · 23 Own results and interpretations

65 more claims · 35 with their cited sources checked

Create a free account to see all 128 claims and to verify the 61 references still open.

See all claims — free account
  • Probably not a substantial claim. 18
  • It may state two separate claims as one. 46

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?

Citations

Cites

Show all 93 references

Research works printed but not linked: 2

  1. Jackson, M. C. Artificial Intelligence & Algorithmic Bias: The Issues with Technology Reflecting History & Humans Notes & Comments. J. Bus. Tech. L. 16, 299–316 (2021).
  2. Wachter, S., Mittelstadt, B. & Russell, C. Counterfactual explanations without opening the black box: Automated decisions and the GDPR. Harv. JL & Tech. 31, 841 (2017).

Cited by

See all 322

Metadata

Authors
Hasanzadeh, F., Josephson, C. B., Waters, G., Adedinsewo, D., Azizi, Z., & White, J. A.
Affiliations
  • University of Calgary (CA)
  • Morgan State University (US)
  • Mayo Clinic (US)
  • Stanford Medicine (US)
Corresponding author
White, James A.
Venue
npj Digital Medicine
Published
2025-03-11
Volume
8
Issue
1
Pages
154-154
ISSN
2398-6352 (ISSN-L)
OpenAlex
W4408314709
Cited by
320
References
93
Citations by year
2026: 221 · 2025: 97
FWCI
49.28
Type
journal article
Language
en
Open access
Licence
Locations
  • npj Digital Medicine (publishedVersion, cc-by-nc-nd)
  • Europe PMC (PubMed Central) (publishedVersion, cc-by-nc-nd)
  • PubMed Central (publishedVersion, cc-by-nc-nd)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Ethics and Social Impacts of AI
  • Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
  • Physical Sciences › Computer Science › Artificial Intelligence › Explainable Artificial Intelligence (XAI)
Keywords
algorithmic bias, artificial intelligence in healthcare, bias mitigation, health care disparities, algorithmic fairness, model life cycle, bias detection, stakeholder responsibility, long-term monitoring

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 2026-10-05, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-05 (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).

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: Bias recognition and mitigation strategies in artificial intelligence healthcare applications. https://app.cochat.ai/papers/10.1038/s41746-025-01503-7
The paper
Hasanzadeh, F., Josephson, C. B., Waters, G., Adedinsewo, D., Azizi, Z., & White, J. A. (2025). Bias recognition and mitigation strategies in artificial intelligence healthcare applications. npj Digital Medicine. https://doi.org/10.1038/s41746-025-01503-7

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.