CoChat
Sign Up
Cited by 4Ethics and Social Impacts of AIArtificial Intelligence in Healthcare and EducationExplainable Artificial Intelligence (XAI)

Frameworks encompassing intersectional perspective of artificial intelligence in healthcare. Scoping review

Open access
Amaya-Santos, Sua, Vargas, Carlos, Bermúdez-Tamayo, Clara·2026·Public Health in Practice
DOI 10.1016/j.puhip.2025.100713PMC PMC12803852
Open this paper in CoChat — free account

Abstract

This study systematically evaluates how existing AI frameworks in healthcare address intersectional bias across the AI lifecycle and explores the mitigation strategies proposed. Scoping review. A scoping review was conducted per PRISMA-ScR guidelines, analyzing studies from 2014-2024. Searches included MEDLINE (Ovid), PubMed, EMBASE (Ovid), SCOPUS, ESCI, IEEE Xplore, and Google Scholar. Data were extracted on bias-related challenges and mitigation strategies across AI lifecycle phases (development, validation, implementation, monitoring). Studies were ranked by inclusivity (high, medium, or low). Of 374 records, 43 studies met inclusion criteria, primarily from high-income countries. Gender/sex (51.2%) and race/ethnicity (44.2%) were the most addressed dimensions, while disability (14%) and citizenship (9.3%) were least addressed. Inclusivity was categorized as high (21 studies, 48.8%), medium (23.2%), or low (27.9%). Overall, 14 biases and 21 mitigation strategies were identified. Significant gaps remain in addressing intersectional biases in AI frameworks, particularly for underrepresented groups such as individuals with disabilities and non-citizens. Despite many frameworks demonstrating efforts toward inclusivity, attention to intersectionality remains uneven and largely inconsistent. Mapping biases to lifecycle phases highlights actionable strategies to improve equity and inclusivity in AI-driven healthcare. These findings provide valuable guidance for researchers, policymakers, and developers to create equitable and responsible AI systems. This paper examines fairness and inequality in healthcare AI, which could improve diagnosis and treatment but may worsen existing disparities. A review of 43 studies reveals that many AI frameworks inadequately address “intersectional bias”—how combined factors like gender, race, and economic status create unique discrimination. Many frameworks overlook factors like disability and citizenship, risking unfair AI systems and unequal care. The paper identifies common AI development biases (e.g., unrepresentative datasets, data labelling errors) and suggests solutions: diverse data collection, fairness checks, and clear guidelines. The authors stress the need for fair, inclusive AI to prevent increased health disparities, and offer recommendations for developers, policymakers, and healthcare providers.

AI snapshot

Written from the paper’s full text.

TL;DR

This scoping review maps how 43 healthcare AI frameworks address intersectional bias across the AI lifecycle, finding uneven attention to dimensions such as gender/sex and race/ethnicity versus disability and citizenship. It identifies 14 biases and 21 mitigation strategies, with gaps especially in implementation and monitoring.

Key findings

  • Among 43 frameworks, gender/sex (51.2%) and race/ethnicity (44.2%) were most addressed; age was 27.9%, economic status 23.3%, disability 14%, and citizenship 9.3%; 16 studies (37.2%) addressed no specific intersectionality dimension.
  • Inclusivity was high in 21 studies (48.8%), medium in 10 (23.2%), and low in 12 (27.9%); the least-addressed characteristics, disability and citizenship, were nearly exclusively considered by high-inclusivity frameworks.
  • The review identified 14 distinct biases and 21 mitigation strategies across the AI lifecycle; design/development was addressed by 69.7%, validation by 34.8%, implementation by 23.2%, and monitoring by 67.4%, though the text later states monitoring was addressed by 4 studies (9.3%).
  • Most publications had primary authors’ institutions in the United States (n = 14), followed by Canada and the United Kingdom (n = 4 each), with evidence skewed toward high-income countries.

Methods

  • Scoping review following Arksey and O’Malley and PRISMA-ScR; searches included MEDLINE (Ovid), PubMed, EMBASE (Ovid), SCOPUS, ESCI, IEEE Xplore, and Google Scholar for January 2014–March 2024, with English, Spanish, and Portuguese sources.
  • Two reviewers screened titles/abstracts and full texts and extracted data; disagreements were resolved by discussion or a third reviewer.
  • Data extraction covered intersectionality characteristics, AI framework methods, healthcare setting, and key findings; frameworks were mapped to four AI lifecycle phases and ranked by inclusivity as high, medium, or low.

Limitations

  • Stated: as a scoping review it focused on breadth over depth; geographic bias toward high-income countries; linguistic bias because only English-written studies were found; the inclusivity index was unvalidated; and newer frameworks after data collection may not be represented.
  • Evident from the text: internal inconsistencies in reported counts, such as 32 articles included versus a final dataset of 43 frameworks, and monitoring addressed by 67.4% versus 4 studies (9.3%), which may affect confidence in phase-specific estimates.

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

Integrity signals

Registry

Crossref · 10.1016/j.puhip.2025.100713

Cites

57 of 57 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

26 of 39 cited sources checked · 5 supported · 0 flagged · accuracy not yet measured

  • Premisecitation

    Lack of representation introduces harmful biases by systematically disadvantaging excluded groups.

    “Lastly, the lack of representation introduces harmful biases by systematically disadvantaging excluded groups [[39], [40], [41]].”
    Quote in the paperNot a substantial claim2 of 3 sources checked
    • cites [39] Hosseini, 2023, The aspects of running artificial intelligence in emergency care; a scoping review, Arch Acad Emerg Med source
    • cites [40] Sakhare, 2023, Ethical considerations of AI applications in medicine: a policy framework for responsible deployment

      Not checked in the free run.

    • cites [41] Van Kolfschooten H. The AI cycle of health inequity and digital ageism: mitigating biases through the EU regulatory framework on medical devices. Journal Of Law And The Biosciences. July 1, 2023;10(2):lsad031. Available: https://doi.org/10.1093/jlb/lsad031. source
  • Premisecitation

    Lack of representation in datasets during the development phase leads to models trained on biased datasets failing to generalize to underrepresented populations.

    “Firstly, models trained on biased datasets fail to generalize to underrepresented populations.”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [35] Kong, 2023, Leveraging responsible, explainable, and local artificial intelligence solutions for clinical public health in the global south, Healthc source

      Section 3.5 (page not provided) says findings shaped largely by Global North research may not generalize to the Global South and calls for diverse, representative, locally relevant databases. This supports the concern about underrepresentation, but the citing sentence turns a possible limitation into a categorical claim that models trained on biased datasets fail to generalize. (LLM, conf=0.85)

      Experimental — accuracy not yet measured
    • cites [36] Mollura, 2020, Artificial intelligence in Low- and middle-income countries: innovating global health radiology, Radiology source
  • Premisecitation

    Social bias arises from societal inequalities and prejudices embedded in the data.

    “Social bias arises from societal inequalities and prejudices embedded in the data [13,29,48].”
    Quote in the paperNot a substantial claim2 of 3 sources checked
    • cites [13] 2021

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [29] Seyyed-Kalantari, 2021, Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations, Nat. Med. source

      The Discussion (page not available) links underdiagnosis disparities to clinical records that may already contain biases in care, and says the disparities probably originate from the data. This supports the claim’s data-embedded bias mechanism in healthcare, but the paper does not establish the broader, unqualified assertion that social bias arises from societal inequalities and prejudices. (LLM, conf=0.82)

      Experimental — accuracy not yet measured
    • cites [48] Hill, 2024, AI in imaging: the regulatory landscape, Br. J. Radiol. source

      The Discussion’s Bias section says imaging datasets can be demographically skewed and describes potential ethnic and socio-economic biases that may hamper generalization. That supports the data-bias aspect of the claim, but the paper does not specifically say those biases arise from societal inequalities or prejudices embedded in the data. No page number is available in our copy. (LLM, conf=0.82)

      Experimental — accuracy not yet measured
  • Premisecitation

    A recent study found that robots trained on large datasets exhibited stereotypical and biased behaviors in terms of gender and race.

    “A recent study found that robots trained on large datasets exhibited stereotypical and biased behaviors in terms of gender and race [14,15].”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites [14] Marotta, 2022, When AI is wrong: addressing liability challenges in women's healthcare, J. Comput. Inf. Syst. source

      The Introduction says biased AI algorithms can perpetuate gender inequities, and the Conclusions explain how bias in training data can alter AI decisions. The paper does not report a study of robots trained on large datasets or findings about racial stereotypes, so those specifics are unsupported. (LLM, conf=0.78)

      Experimental — accuracy not yet measured
    • cites [15] Amaya-Santos, 42024, Health for whom? Intersectionality and biases in the use of artificial intelligence in clinical diagnosis. Anales del Sistema Sanitario de Navarra.

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Premisecitation

    These limitations reinforce existing healthcare disparities, as benefits of AI solutions are predominantly confined to demographic groups already well-represented in training datasets.

    “Secondly, these limitations reinforce existing healthcare disparities, as the benefits of AI solutions are predominantly confined to demographic groups already well-represented in training datasets [37,38].”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites [37] Zhang X, Gu Y, Yin J, Zhang Y, Jin C, Wang W, et al. Development, Reliability, and Structural Validity of the Scale for Knowledge, Attitude, and Practice in Ethics Implementation Among AI Researchers: Cross-Sectional Study. JMIR Formative Research. Septiembre 24, 2023;7:e42202. Available: https://doi.org/10.2196/42202. source

      The Introduction discusses ethical risks of medical AI, while the Discussion’s Limitations section addresses sampling bias in this paper’s survey. Neither passage addresses healthcare disparities or whether AI benefits are concentrated among demographic groups well represented in training data. Searches for those terms and related concepts found no supporting passage in our copy. (LLM, conf=0.72)

      Experimental — accuracy not yet measured
    • cites [38] Kiyasseh, 2024, A framework for evaluating clinical artificial intelligence systems without ground-truth annotations, Nat. Commun. source

      The Results section reports that a dermatology AI system performed better for patients with Fitzpatrick skin tones I–II than for those with V–VI, supporting the point that benefits can differ across demographic groups. The paper does not establish that the better-performing group was better represented in training data, or that AI benefits are predominantly confined to well-represented groups generally. (LLM, conf=0.85)

      Experimental — accuracy not yet measured
  • May combine two claims. 13
  • Classified as probably not a substantial claim. 15

  • May combine two claims.

    Premisecitation

    The landscape of guidance on how to develop safe and effective AI systems for healthcare is fragmented across hundreds of documents, largely with a focus on products that would be regulated as medical devices.

    “Moreover, the landscape of guidance on how to develop safe and effective AI systems for healthcare is fragmented across hundreds of documents, largely with a focus on products that would be regulated as medical devices [21].”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [21] Silva, 2022, An artificial intelligence life cycle: from conception to production, Patterns [Internet] source

      The Results section (‘The CDAC AI life cycle’; page unavailable) describes a diversity of AI ethics frameworks and vague best practices, which supports the general point that guidance is fragmented. It counts 84 ethics frameworks as of 2019, not hundreds of healthcare guidance documents. The Preliminary risk assessment section mentions regulatory requirements for software-based medical devices in healthcare but does not say that most guidance focuses on those products. (LLM, conf=0.78)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Study design and conductcitation

    The scoping review was conducted following the Arksey and O'Malley methodological framework with additional recommendations from Levac et al. and adheres to PRISMA-ScR.

    “A Scoping review was conducted following the Arksey and O'Malley methodological framework [23] with additional recommendations from Levac et al. [24] The review adheres to the Preferred Reporting Items for Systematic reviews and meta-analyses extension for Scoping Reviews (PRISMA-ScR) [25].”
    Quote in the paperSubstantial claim3 of 3 sources checked
    • cites [23] Arksey, 2005, Scoping studies: towards a methodological framework, Int. J. Soc. Res. Methodol. source
    • cites [24] Levac, 2010, Scoping studies: advancing the methodology, Implement. Sci. source
    • cites [25] Tricco, 2016, A scoping review on the conduct and reporting of scoping reviews, BMC Med. Res. Methodol. source

      Citation [25] is attached to the statement that the citing review adheres to PRISMA-ScR. The source paper’s Background says that a scoping-review reporting checklist did not yet exist in the EQUATOR library; it does not discuss PRISMA-ScR or assess the citing review’s adherence to it. The passages on Arksey and O’Malley and Levac et al. concern the separately cited methodological framework, not the claim attached to [25]. (LLM, conf=0.78)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Label bias occurs when labels assigned to data points during training or validation reflect inherent inaccuracies, inconsistencies, or prejudices.

    “Label and Social Bias were recurrent as well (n = 4), for instance, label bias occurs when the labels assigned to data points during training or validation reflect inherent inaccuracies, inconsistencies, or prejudices [9,39].”
    Quote in the paperNot a substantial claim2 of 2 sources checked
    • cites [9] Ronquillo, 2022, The untapped potential of nursing and allied health data for improved representation of social determinants of health and intersectionality in artificial intelligence applications: a rapid review, Yearb Med Inf source

      The Introduction discusses incomplete EHR data, underrepresentation, and AI bias, but does not define label bias. Searches for labels assigned during training or validation and for inaccurate or prejudiced ground-truth labels found no passage addressing that definition; page numbers were unavailable in our copy. (LLM, conf=0.73)

      Experimental — accuracy not yet measured
    • cites [39] Hosseini, 2023, The aspects of running artificial intelligence in emergency care; a scoping review, Arch Acad Emerg Med source

      The Results (§3.15; page not provided) say that decisions about how data are labelled can produce skewed datasets that distort reality. The paper does not define label bias as inaccuracies, inconsistencies, or prejudices in labels assigned during training or validation, so it supports the general concern but not the claim’s specific definition. (LLM, conf=0.78)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductcitation

    An AI-powered tool, Research Rabbit, was used for snowball sampling, and all suggested articles were subjected to the same two-step manual screening by two reviewers to ensure transparency and mitigate selection biases.

    “An AI-powered tool, Research Rabbit, was utilized to facilitate snowball sampling by identifying related literature based on the included studies from the electronic database searches. All suggested articles were subsequently subjected to the same rigorous, two-step manual screening process by two reviewers to ensure transparency and mitigate the risk of introducing new selection biases or missing”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [26] Cole, 2023, ResearchRabbit (product review), J Can Health Libr Assoc J Assoc Bibl Santé Can. source

  • May combine two claims.

    Premisecitation

    Cognitive bias is introduced during data annotation or input processes and often reflects subjective judgments of healthcare providers, such as manual data entry errors or inconsistent diagnoses.

    “Cognitive bias also poses a major challenge (n = 5), introduced during data annotation or input processes. Kudina et al. and Sikstrom et al., [49,51] acknowledge that this bias often reflects the subjective judgments of healthcare providers, such as manual data entry errors or inconsistent diagnoses.”
    Quote in the paperNot a substantial claim2 of 2 sources checked
    • cites [49] Kudina, 2021, Co-designing diagnosis: towards a responsible integration of Machine Learning decision-support systems in medical diagnostics, J. Eval. Clin. Pract. source

      The sections “From Bias to the Question of Expertise” and “Expertise and Medical Diagnosis: Gadamer’s Notion of Fore-Understanding” discuss biased datasets and clinicians’ preconceptions in diagnosis. They do not attribute cognitive bias to data annotation or input processes, or give manual entry errors and inconsistent diagnoses as examples. Page numbers are not available in our copy. (LLM, conf=0.84)

      Experimental — accuracy not yet measured
    • cites [51] Sikstrom, 2022, Conceptualising fairness: three pillars for medical algorithms and health equity, BMJ Health Care Inform source

      The Results > Impartiality section says medical practices can reflect arbitrary assumptions about patients and that human actors help gather and curate datasets. The Discussion also gives misdiagnosis of schizophrenia as an example of a feature that could bias a model. The paper does not specifically describe cognitive bias arising during annotation or input, manual data-entry errors, or inconsistent diagnoses as a general mechanism; page numbers were not available in our copy. (LLM, conf=0.83)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Premisecitation

    Algorithms trained on datasets with a male bias may fail to detect gender-specific conditions or present inaccuracies in diagnosing diseases that exhibit gender differences in symptoms.

    “For instance, algorithms trained on datasets with a male bias may fail to detect gender-specific conditions or present inaccuracies in diagnosing diseases that exhibit gender differences in symptoms [13,45,46].”
    Quote in the paperSubstantial claim2 of 3 sources checked
    • cites [13] 2021

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [45] Reddy, 2021, Evaluation framework to guide implementation of AI systems into healthcare settings, BMJ Health Care Inform source

      The Results (Utility, ‘Generalisability and contextualisation’; page unavailable) say that under-representation in training and validation datasets can create bias and reduce an AI model’s real-world utility for specific groups. The paper does not address male-biased datasets specifically, or say that they cause missed gender-specific conditions or diagnostic inaccuracies involving gender differences in symptoms. (LLM, conf=0.83)

      Experimental — accuracy not yet measured
    • cites [46] Javed, 2024, Ethical frameworks for machine learning in sensitive healthcare applications, IEEE Access source

      Section 2.3, “Fairness and Bias” (page unavailable), says training data can disproportionately represent demographic groups, including by gender, and that this may make predictions less reliable for underrepresented groups or lead to incorrect diagnoses. It does not specifically discuss male-biased datasets, failure to detect gender-specific conditions, or diseases with gender-differing symptoms. (LLM, conf=0.84)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Premisecitation

    Citizenship intersects with race, ethnicity, and socioeconomic status, particularly for migrants and undocumented individuals, who often face additional systemic discrimination and healthcare exclusion.

    “Similarly, citizenship intersects with race, ethnicity, and socioeconomic status, particularly for migrants and undocumented individuals, who often face additional systemic discrimination and healthcare exclusion, further underscoring the necessity of integrating an intersectional lens into AI design [35].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [35] Kong, 2023, Leveraging responsible, explainable, and local artificial intelligence solutions for clinical public health in the global south, Healthc source

      Section 3.2 (page unavailable) identifies migrants, people without citizenship rights, racialized minorities, and socioeconomically deprived people as vulnerable groups whose needs are often overlooked in health policy. It does not specifically say that citizenship intersects with race, ethnicity, and socioeconomic status, or that undocumented individuals face additional discrimination and healthcare exclusion. (LLM, conf=0.82)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Findingown data

    High-inclusivity frameworks explicitly incorporated the definition of intersectionality and addressed multiple dimensions such as gender, race, and other intersecting factors.

    “The 21 studies categorized as having high inclusivity (48.8 % of all included frameworks) explicitly incorporated the definition of intersectionality and addressed multiple dimensions, such as gender, race, and other intersecting factors.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Findingown data

    The first screening step yielded 264 grey documents and 253 were excluded in full-text screening.

    “The first screening step yielded 264 grey documents and 253 were excluded in the full-text screening.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The vast majority of publications had primary authors institutions in the United States (n=14), followed by Canada and the United Kingdom (n=4).

    “The vast majority of publications had primary authors institutions in the United States (n = 14), followed by Canada and the United Kingdom (n = 4) (Fig. 2).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    126 duplicates were removed from 374 records identified across six databases.

    “We identified a total of 374 records across the six databases and removed 126 duplicates.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    From 248 records screened with titles and abstracts, 120 full-text articles were assessed for eligibility, and 32 articles were included.

    “From the 248 records screened with titles and abstracts, 120 full-text articles were assessed for eligibility. A total of 32 articles were included.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Most frameworks address identity factors in isolation, with limited operationalization of multidimensional intersectionality.

    “Most frameworks address identity factors in isolation, with limited operationalization of multidimensional intersectionality.”
    • Also stated in: § Limitations · plain
      “Furthermore, frameworks show limited operationalization of multidimensional intersectionality, often addressing identity factors in a siloed manner”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingcitation

    Concept drift was mentioned in 3 studies.

    “Concept drift emerges when the relationship between input data and outcomes changes. This bias was mentioned in 3 studies, Lekadir et al. acknowledge that this can result from updated clinical guidelines or modifications in diagnostic criteria, causing AI systems to deliver inaccurate results over time [51,61].”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingnone given

    The literature suffers from a lack of standardized indicators for measuring intersectional fairness and fragmented terminology, reducing the comparability of existing guidance.

    “The literature also suffers from a lack of standardized indicators for measuring intersectional fairness and fragmented terminology, reducing the comparability of existing guidance.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data

    Data were extracted by two reviewers using a draft structured data recording form.

    “Data were extracted by two reviewers (SA and CV) using a draft structured data recording form.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductown data

    A two-step screening process was conducted by two reviewers for title/abstract and full-text screening, with disagreements resolved through discussion involving a third reviewer if necessary.

    “A two-step screening process was conducted by two reviewers (SA and CV) for title/abstract and full-text screening. Disagreements were resolved through discussion, involving a third reviewer if necessary.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data

    The search strategy was validated by a librarian specialized in public health.

    “The search strategy adopted for the different databases was validated by a librarian specialized in public health and is described in Supplementary data 1.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Lack of accountability and transparency in monitoring processes undermines trust among clinicians and patients, slowing the adoption of AI technologies.

    “The Guidance on AI and Medical Products of the US FDA [67] states that the lack of accountability and transparency in monitoring processes undermines trust among clinicians and patients, slowing the adoption of AI technologies.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [67] 2024

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    The lack of comprehensive frameworks creates critical challenges at every stage.

    “The lack of comprehensive frameworks creates critical challenges at every stage.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Interpretationnone given

    The findings of this review align with and expand upon the recommendations highlighted by other authors.

    “The findings of this review align with and expand upon the recommendations highlighted by other authors, providing actionable roadmaps across sectors.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductreasoning

    The rapidly evolving nature of AI research means newer frameworks or strategies developed after the data collection period may not be represented.

    “Finally, the rapidly evolving nature of AI research means newer frameworks or strategies developed after the data collection period may not be represented.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    This study, as a scoping review, focused on breadth over depth.

    “In terms of methodological limitations, this study, as a scoping review, focused on breadth over depth.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given

    The implementation phase is undermined when the absence of meaningful stakeholder engagement leads to AI tools failing to meet real-world healthcare needs, resulting in suboptimal outcomes and reduced trust.

    “Similarly, the Implementation phase is undermined when the absence of meaningful stakeholder engagement leads to AI tools failing to meet real-world healthcare needs, resulting in suboptimal outcomes and reduced trust.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    Cognitive bias introduced by existing prejudices related to race, gender, or socioeconomic status can distort validation data and perpetuate systemic inaccuracies.

    “For example, during the Validation phase, Cognitive Bias introduced by existing prejudices related to race, gender, or socioeconomic status can distort validation data and perpetuate systemic inaccuracies.”
    Quote in the paperNot a substantial claim
25 of 25

  • The paper relies on this without support.

    Most current biomedical AI technologies do not account for bias.

    “Despite significant scientific advances, most current biomedical AI technologies do not account for bias.”

  • The paper relies on this without support.

    Biases identified in the implementation phase perpetuate healthcare inequities.

    “Biases identified in this phase perpetuate healthcare inequities.”

  • The paper relies on this without support.

    Unaddressed bias during monitoring degrades the performance of AI systems.

    “Unaddressed bias during monitoring degrades the performance of AI systems.”

  • The paper relies on this without support.

    AI cannot be considered responsible unless it conforms to certain AI ethical principles, underpinned by specific standards and regulations that explicitly address equity.

    “AI cannot be considered responsible unless it conforms to certain AI ethical principles, underpinned by specific standards and regulations that explicitly address equity.”

  • The paper relies on this without support.

    Disability intersects with other axes of marginalization, such as socioeconomic status and gender, to shape unique healthcare needs and barriers.

    “The minimal focus on disability is concerning as this dimension intersects with other axes of marginalization, such as socioeconomic status and gender, to shape unique healthcare needs and barriers.”
5 of 5

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.

  • Monitoring-phase coverage is reported as both 67.4% and 9.3%

    Major

    The lifecycle overview says 67.4% of the 43 studies addressed monitoring, but the monitoring section says only 4 studies (9.3%) did. This changes whether monitoring appears to be one of the most- or least-covered phases.

  • Sujan framework has conflicting publication years

    Minor

    Table 1 dates the Sujan validation framework to 2019, while its matching reference gives 2023.

  • Included FUTURE-AI framework is dated after the search period

    Minor

    Table 1 dates the included FUTURE-AI guideline to 2023, whereas the corresponding reference dates its publication to February 2025. A 2025 publication also falls outside the stated January 2014-March 2024 inclusion period.

3 of 3

Cites

Flagged first, then by citations

Research works printed but not linked: 3

  1. Landers, 2022, Auditing the AI auditors: a framework for evaluating fairness and bias in high stakes AI predictive models, Am. Psychol., 78
  2. Kasula, 2023, Framework development for artificial intelligence integration in healthcare: optimizing patient care and operational efficiency, Transactions on Latest Trends in IoT, 6, 77
  3. Amaya-Santos S. Frameworks encompassing intersectional perspective of artificial intelligence in healthcare. Scoping review. searchRxiv; Available from: https://www.cabidigitallibrary.org/doi/10.1079/searchRxiv.2025.00953. · DOI

Cited by

Sorted by citations

Related

Authors
Amaya-Santos, S., Vargas, C., & Bermúdez-Tamayo, C.
ORCID
Venue
Public Health in Practice
Publisher
Elsevier BV
Published
2026-06
Volume
11
Pages
100713
ISSN
2666-5352 (ISSN-L)
DOI
10.1016/j.puhip.2025.100713
PMC
PMC12803852
OpenAlex
W4417488550
Semantic Scholar
54f7c4c9cd3e897ed0babbab5d67df2eaec24a4a
Cited by
4
References
49
Citations by year
2026: 4
FWCI
0.63
Type
journal article
Language
en
Open access
Licence
Locations
  • Public Health in Practice (publishedVersion, cc-by-nc-nd)
  • PubMed (publishedVersion)
  • PubMed Central (publishedVersion, cc-by-nc-nd)
  • Institutional Repository of the University of Granada (University of Granada) (submittedVersion, cc-by-nc-nd)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
Center for International Business Education and Research, University of Illinois at Urbana-Champaign
Topics
  • Social Sciences › Safety Research › 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
intersectional bias, artificial intelligence in healthcare, AI framework, algorithmic fairness, health disparities, bias mitigation strategies, AI lifecycle, disability, citizenship, data representativeness, health equity

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.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-28.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-28.

Cited sources are checked by 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