Bias in medical AI: Implications for clinical decision-making
Open accessThe DOI matches its Crossref record; 124 of the 124 references it cites were found in a registry or an index; 30 of the 140 sources its claims cite were checked for support.
Abstract
Biases in medical artificial intelligence (AI) arise and compound throughout the AI lifecycle. These biases can have significant clinical consequences, especially in applications that involve clinical decision-making. Left unaddressed, biased medical AI can lead to substandard clinical decisions and the perpetuation and exacerbation of longstanding healthcare disparities. We discuss potential biases that can arise at different stages in the AI development pipeline and how they can affect AI algorithms and clinical decision-making. Bias can occur in data features and labels, model development and evaluation, deployment, and publication. Insufficient sample sizes for certain patient groups can result in suboptimal performance, algorithm underestimation, and clinically unmeaningful predictions. Missing patient findings can also produce biased model behavior, including capturable but nonrandomly missing data, such as diagnosis codes, and data that is not usually or not easily captured, such as social determinants of health. Expertly annotated labels used to train supervised learning models may reflect implicit cognitive biases or substandard care practices. Overreliance on performance metrics during model development may obscure bias and diminish a model’s clinical utility. When applied to data outside the training cohort, model performance can deteriorate from previous validation and can do so differentially across subgroups. How end users interact with deployed solutions can introduce bias. Finally, where models are developed and published, and by whom, impacts the trajectories and priorities of future medical AI development. Solutions to mitigate bias must be implemented with care, which include the collection of large and diverse data sets, statistical debiasing methods, thorough model evaluation, emphasis on model interpretability, and standardized bias reporting and transparency requirements. Prior to real-world implementation in clinical settings, rigorous validation through clinical trials is critical to demonstrate unbiased application. Addressing biases across model development stages is crucial for ensuring all patients benefit equitably from the future of medical AI.
Overview
CoChat Summary
TL;DR
This narrative review argues that bias in medical AI arises and compounds across the entire lifecycle—training data, model development and evaluation, deployment, and publication—and can worsen clinical decisions and health disparities. It catalogs mitigation strategies including diverse data collection, statistical debiasing, subgroup evaluation, interpretability, bias reporting, and clinical-trial validation.
Key findings
- Bias can enter at multiple stages: imbalanced sample sizes, nonrandomly missing data, uncaptured social determinants of health, biased labels, race/ethnicity in clinical algorithms, overreliance on whole-cohort performance metrics, publication biases, and implementation/end-user biases.
- The review cites specific examples and numbers: over half of published clinical AI models in 2019 used data from the US or China; radiology accounted for over 40% of AI publications in 2019; authors were 75% male; a MIMIC-III mortality model yielded recall rates as low as 25% under imbalanced racial representation; and the Epic Sepsis Model missed two-thirds of sepsis cases after deployment.
- Mitigation strategies discussed include large diverse data sets, oversampling methods such as SMOTE and ADASYN, imputation, SDoH capture, expert consensus labeling, subgroup analysis, bias-centered metrics like equalized odds and predictive parity, interpretability methods such as SHAP and LIME, adversarial debiasing, ongoing monitoring, and clinical trials for AI validation.
- Medical large language models pose unique bias risks because they are autoregressive, can reproduce biased or inaccurate training content, cannot assess their own outputs, and may generate inconsistent recommendations; statistical debiasing methods for clinical LLMs remain limited.
Methods
- Narrative review, not a systematic review, synthesizing literature and real-world or hypothetical examples across diverse medical domains and stages of the AI development pipeline.
- Includes Table 1 summarizing bias types by development stage, illustrative examples, potential clinical consequences, and mitigation strategies.
- No primary data collection, systematic search protocol, or quantitative meta-analysis is reported; the work draws on cited studies and examples.
Limitations
- Stated in the paper: more research is needed; there is a lack of standardized bias reporting guidelines; there are no explicit FDA regulations examining fairness or bias in medical AI outputs; and statistical debiasing methods for real-world clinical LLMs remain limited.
- Evident from the text: as a narrative review, it lacks systematic search and selection criteria and does not quantitatively synthesize evidence; examples are illustrative or hypothetical and not exhaustive; and the proposed mitigations are not empirically validated in this paper.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
17 Valid · 2 Not valid · 65 Not checked · 42 Own results and interpretations
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Fewer studies have concentrated specifically on the implications and consequences of biased AI in healthcare [19–22].”
§ Bias in medical AI: Implications for clinical decision-making > Related work
What the sources say
- Not checked yetcites [19] RB Parikh, 2019, Addressing bias in artificial intelligence in health care, JAMA source
- Supportedcites [20] MD Abràmoff, 2023, Considerations for addressing bias in artificial intelligence for health equity, NPJ Digit Med source
The cited paper is an example of a study focused on biased AI in healthcare and its consequences. Its Introduction describes AI exacerbating health disparities and contributing to worse outcomes for Black patients; the “Bias in the healthcare process” section examines unequal outcomes across groups. The citation supports identifying this paper as one of the studies, rather than establishing a count of how many such studies exist. (LLM, conf=0.86)
- Partly supportedcites [21] R Daneshjou, 2021, Lack of transparency and potential bias in artificial intelligence data sets and algorithms: a scoping review, JAMA Dermatol sourceExperimental — accuracy not yet measured
The Discussion says that few dermatology AI studies report patient race, ethnicity, or skin tone, limiting assessment of potential bias. That supports a narrower gap in studying bias, but the paper does not establish that fewer studies across healthcare have examined the implications and consequences of biased AI. (LLM, conf=0.82)
- Not checked yetcites [22] AS Albahri, 2023, A systematic review of trustworthy and explainable artificial intelligence in healthcare: Assessment of quality, bias risk, and data fusion., Inf Fusion. source
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Standardized reporting checklists such as the Prediction Model Risk of Bias Assessment (PROBAST) tool have been developed to assess the risk of imbalanced data biases, which allows both developers and downstream users (e.g., physicians, hospitals) to better understand the suitability and limitations of AI-based systems in specific clinical settings [54–56].”
§ Bias in medical AI: Implications for clinical decision-making > Biases in training data > Imbalanced sample sizes
What the sources say
- Not checked yetcites [54] RF Wolff, 2019, PROBAST: a tool to assess the risk of bias and applicability of prediction model studies, Ann Intern Med source
- Partly supportedcites [55] T Hernandez-Boussard, 2020, MINIMAR (MINimum Information for Medical AI Reporting): Developing reporting standards for artificial intelligence in health care., J Am Med Inform Assoc sourceExperimental — accuracy not yet measured
The Introduction and General Principles of the MINIMAR Design say that reporting training-data characteristics can reveal potential biases and help assess whether an AI model generalizes to a clinical population. But the paper proposes MINIMAR; the passages found do not identify PROBAST or say that PROBAST was developed to assess biases from imbalanced data. No page numbers are available in our copy. (LLM, conf=0.83)
- Could not verifycites [56] Szczekocka E, Tarnec C, Pieczerak J, editors. Standardization on bias in Artificial Intelligence as industry support. 2022 IEEE International Conference on Big Data (Big Data); 2022: IEEE. sourceExperimental — accuracy not yet measured
The Standardization Specifics section says standards could help identify data-related biases, but it does not discuss PROBAST or a reporting checklist. Searches for PROBAST, imbalanced data, and clinical settings found no passage establishing the claimed purpose or use of that tool. (LLM, conf=0.76)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“The researchers found that imbalanced representation of racial groups had a significant effect on model performance (yielding recall rates as low as 25%), raising critical concerns for the model’s potential application in real clinical settings [44].”
§ Bias in medical AI: Implications for clinical decision-making > Biases in training data > Imbalanced sample sizes
What the sources say
- Partly supportedcites [44] E Röösli, 2022, Peeking into a black box, the fairness and generalizability of a MIMIC-III benchmarking model., Sci Data. sourceExperimental — accuracy not yet measured
The Discussion reports significantly worse prediction performance for Black patients and says their low representation might partly explain the bias; it does not establish that imbalanced racial representation caused the performance gap. The Discussion also says the model identified only every fourth to fifth high-risk patient, but that low recall concerns the high-risk class overall, not a racial group. (LLM, conf=0.90)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The sources with a verdict support this claim; the rest weren’t checked yet, have no full text we could read, or their verdicts aren’t rated yet.
In the paper
“Furthermore, the degree of practitioner misclassification can vary along sociodemographic factors.”
§ Bias in medical AI: Implications for clinical decision-making > Biases in training data > Data labels and misclassification
What the sources say
- Supportedcites [84] GH Rauscher, 2013, Potentially missed detection with screening mammography: does the quality of radiologist’s interpretation vary by patient socioeconomic advantage/disadvantage?, Ann Epidemiol source
The INTRODUCTION states the goal was 'to examine whether variation in the quality of mammography interpretation as performed by the original reading radiologist varied by patient sociodemographic characteristics,' i.e. radiologist (practitioner) error by sociodemographic group. The RESULTS (p. 3, Table 3 discussion) report that potentially missed detection differed by minority ethnicity (+17 pp), income, education and lack of private insurance (+30 pp), and the DISCUSSION concludes the observed racial/ethnic and socioeconomic disparities 'shed light on differences in the quality of interpretation for different racial/ethnic and sociodemographic groups.' The Abstract's CONCLUSIONS affirm the association, so the hedged claim that practitioner misclassification 'can vary' along sociodemographic factors is supported; only the education association was attenuated in the selection-bias sensitivity analysis. (LLM, conf=0.90)
- Not checked yetcites [85] AM Kilbourne, 2005, Care without coverage: too little, too late., J Natl Med Assoc
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“There are several well-documented instances of insufficient evidence for race and ethnicity to be used as predictive factors in models.”
§ Bias in medical AI: Implications for clinical decision-making > Biases in training data > Race and ethnicity in clinical algorithms
What the sources say
- Partly supportedcites [92] MA Marzinke, 2022, Limited evidence for use of a black race modifier in eGFR calculations: a systematic review, Clin Chem sourceExperimental — accuracy not yet measured
The Discussion (page unavailable) concludes that there is little evidence for including a Black race modifier in eGFR calculations. This documents an instance of insufficient evidence for using race in a model, but the paper does not establish the broader claim of several instances involving race and ethnicity across models. (LLM, conf=0.85)
boosted- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
65 more claims · 8 with their cited sources checked
Create a free account to see all 126 claims and to verify the 110 references still open.
- It may state two separate claims as one. 41
- Probably not a substantial claim. 28
- The quote only partly states this claim. 1
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?
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Generating educational resources without clinical context review risks reinforcing data biases [172]”
From Applications, Challenges, and Prospects of Generative Artificial Intelligence Empowering Medical Education: Scoping Review (2025) · cites this paper as [172]
Experimental — accuracy not yet measuredThe check wasn’t confident either way, so it gives no verdict — never disputed.
In the citing paper
“In contrast, ensemble classification, combining multiple models, demonstrated improved accuracy and robustness [21,22].”
From Machine Learning in Maternity: Early Diagnosis of Placenta Accreta Spectrum (2026) · cites this paper as [22]
Experimental — accuracy not yet measured
Citations
Cites
LeCun, Y., Bengio, Y., & Hinton, G. (2015) · Nature · doi:10.1038/nature14539 · Cited by 84,126
SMOTE synthetic minority over-sampling technique
Chawla, N. V., Bowyer, K. W., Hall, L. O., & Kegelmeyer, W. P. (2002) · Journal of Artificial Intelligence Research · doi:10.1613/jair.953 · Cited by 32,865 · cited 1×
“…which generates synthetic examples for the minority class by interpolating between existing minority class instances, and Adaptive Synthetic Sampling (ADASYN), which uses a similar method to SMOTE but focuses more on examples that are harder to classify (near the decision boundary) [49,50].”
MethodA Unified Approach to Interpreting Model Predictions
Lundberg, S., & Lee, S.-I. (2017) · arXiv · doi:10.48550/arxiv.1705.07874 · Cited by 23,707
Gauriat, C.-M., Pencolé, Y., Ribot, P., & Brouillet, G. (2024) · HAL (Le Centre pour la Communication Scientifique Directe) · doi:10.4230/oasics.dx.2024.27 · Cited by 12,586
Crenshaw, K. (2018) · Living with Contradictions · doi:10.4324/9780429499142-5 · Cited by 7,232
Dissecting racial bias in an algorithm used to manage the health of populations
Obermeyer, Z., Powers, B., Vogeli, C., & Mullainathan, S. (2019) · Science · doi:10.1126/science.aax2342 · Cited by 7,137 · cited 1×
“In reality, however, AI models are frequently found to be biased toward certain patient groups, leading to disparities not only in performance but also in potential or actual clinical benefits [7–9].”
BackgroundA Survey on Bias and Fairness in Machine Learning
Mehrabi, N., Morstatter, F., Saxena, N., Lerman, K., & Galstyan, A. (2022) · ACM Computing Surveys · doi:10.1145/3457607 · Cited by 5,812 · cited 1×
“The topic of bias in AI has garnered significant attention in recent years, with multiple review papers addressing different aspects of this critical issue [14–18].”
BackgroundCrenshaw, K. (1998) · Feminism And Politics · doi:10.1093/oso/9780198782063.003.0016 · Cited by 5,672
Social determinants of health inequalities
Marmot, M. (2005) · The Lancet · doi:10.1016/s0140-6736(05)71146-6 · Cited by 5,086
ADASYN: Adaptive synthetic sampling approach for imbalanced learning
He, H., Bai, Y., Garcia, E. A., & Li, S. (2008) · 2008 IEEE International Joint Conference on Neural Networks (IEEE World Congress on Computational Intelligence) · doi:10.1109/ijcnn.2008.4633969 · Cited by 4,521
Research works printed but not linked: 12
- Food, Administration D., 2019, Proposed regulatory framework for modifications to Artificial Intelligence/Machine Learning (AI/ML)-based Software as a Medical Device.
- D Roselli, 2019, editors. Managing bias in AI. Companion proceedings of the 2019, world wide web conference
- KPUL McFarling, 2024, Doctors use problematic race-based algorithms to guide care every day. Why are they so hard to change?, STAT
- Ribeiro MT, Singh S, Guestrin C. Model-agnostic interpretability of machine learning. arXiv preprint arXiv:160605386. 2016.
- Raji ID, Bender EM, Paullada A, Denton E, Hanna A. AI and the everything in the whole wide world benchmark. arXiv preprint arXiv:211115366. 2021.
- Chauhan VK, Clifton L, Salaün A, Lu HY, Branson K, Schwab P, et al. Sample Selection Bias in Machine Learning for Healthcare. arXiv preprint arXiv:240507841. 2024.
- Yang S, Wang Y, Chu X. A survey of deep learning techniques for neural machine translation. arXiv preprint arXiv:200207526. 2020.
- Topal MO, Bas A, van Heerden I. Exploring transformers in natural language generation: Gpt, bert, and xlnet. arXiv preprint arXiv:210208036. 2021.
- Schmidgall S, Harris C, Essien I, Olshvang D, Rahman T, Kim JW, et al. Addressing cognitive bias in medical language models. arXiv preprint arXiv:240208113. 2024.
- Ebrahimi S, Chen K, Asudeh A, Das G, Koudas N. AXOLOTL: Fairness through Assisted Self-Debiasing of Large Language Model Outputs. arXiv preprint arXiv:240300198. 2024.
- Li J, Tang Z, Liu X, Spirtes P, Zhang K, Leqi L, et al. Steering LLMs Towards Unbiased Responses: A Causality-Guided Debiasing Framework. arXiv preprint arXiv:240308743. 2024.
- Zhang Y-F, Yu W, Wen Q, Wang X, Zhang Z, Wang L, et al. Debiasing large visual language models. arXiv preprint arXiv:240305262. 2024.
Cited by
AI Alignment: A Contemporary Survey
Ji, J., Qiu, T., Chen, B., Zhou, J., Zhang, B., Hong, D., Lou, H., Wang, K., Duan, Y., He, Z., et al. (2025) · ACM Computing Surveys · doi:10.1145/3770749 · Cited by 803
Fahim, Y. A., Hasani, I. W., Kabba, S., & Ragab, W. M. (2025) · European Journal of Medical Research · doi:10.1186/s40001-025-03196-w · Cited by 329
“Pro-moting collaborative communication channels such as workshops and forums supports ongoing learning, user feedback, and iterative improvements to AI systems [182].”
AI-driven healthcare: A review on ensuring fairness and mitigating bias
Chinta, S. V., Wang, Z., Palikhe, A., Zhang, X., Kashif, A., Smith, M. A., Liu, J., & Zhang, W. (2025) · PLOS Digital Health · doi:10.1371/journal.pdig.0000864 · Cited by 171
“Evaluation bias arises when performance assessments fail to account for disparities across demographic groups, leading to inequities in clinical outcomes [72].”
Current AI technologies in cancer diagnostics and treatment
Tiwari, A., Mishra, S., & Kuo, T.-R. (2025) · Molecular Cancer · doi:10.1186/s12943-025-02369-9 · Cited by 207
“It can be concluded that inherent biases that are frequently found in training datasets will need to be addressed by AI-based solutions [267].”
Moore, J., Grabb, D., Agnew, W., Klyman, K., Chancellor, S., Ong, D. C., & Haber, N. (2025) · Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency · doi:10.1145/3715275.3732039 · Cited by 204
Metadata
- Authors
- Cross, J. L., Choma, M. A., & Onofrey, J. A.
- Affiliations
- Yale University (US)
- Corresponding author
- Cross, James L.
- Venue
- PLOS Digital Health
- Publisher
- Public Library of Science (PLoS)
- Published
- 2024-11-07
- Volume
- 3
- Issue
- 11
- Pages
- e0000651
- ISSN
- 2767-3170 (ISSN-L)
- PMC
- PMC11542778
- OpenAlex
- W4404134492
- Cited by
- 607
- References
- 120
- Citations by year
- 2026: 382 · 2025: 222 · 2024: 3
- FWCI
- 13.83
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Licence
- Locations
- PLOS Digital Health (publishedVersion, cc-by)
- PubMed Central (publishedVersion, cc-by)
- PubMed (publishedVersion)
- RePEc: Research Papers in Economics (submittedVersion)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
- Topics
- Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
- Physical Sciences › Computer Science › Artificial Intelligence › Ethics and Social Impacts of AI
- Physical Sciences › Computer Science › Artificial Intelligence › Explainable Artificial Intelligence (XAI)
- Keywords
- algorithmic bias, artificial intelligence in healthcare, clinical decision-making, health care disparities, biased training data, missing data, label bias, model evaluation, bias mitigation, model interpretability, social determinants of health, clinical validation
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 (gpt-6-sol and 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 in medical AI: Implications for clinical decision-making. https://app.cochat.ai/papers/10.1371/journal.pdig.0000651
- The paper
- Cross, J. L., Choma, M. A., & Onofrey, J. A. (2024). Bias in medical AI: Implications for clinical decision-making. PLOS Digital Health. https://doi.org/10.1371/journal.pdig.0000651
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.