CoChat
Sign Up
Cited by 65Artificial Intelligence in Healthcare and EducationEthics and Social Impacts of AIArtificial Intelligence in Healthcare

Artificial intelligence bias in medical system designs: a systematic review

Open access
Kumar, Ashish, Aelgani, Vivekanand, Vohra, Rubeena, Gupta, Suneet K., Bhagawati, Mrinalini, Paul, Sudip, Saba, Luca, Suri, Neha +8 more·2023·Multimedia Tools and Applications
DOI 10.1007/s11042-023-16029-x
Open this paper in CoChat — free accountOpen PDF

Abstract

Inherent bias in the artificial intelligence (AI)-model brings inaccuracies and variabilities during clinical deployment of the model. It is challenging to recognize the source of bias in AI-model due to variations in datasets and black box nature of system design. Additionally, there is no distinct process to identify the potential source of bias in the AI-model. To the best of our knowledge, this is the first review of its kind that addresses the bias in AI-model by categorizing 48 studies into three classes, namely, point-based, image-based, and hybrid-based AI-models. Selection strategy using PRISMA is adopted to select the 72 crucial AI studies for identifying bias in AI models. Using the three classes, bias is identified in these studies based on 44 critical AI attributes. Bias in the AI-models is computed by analytical, butterfly, and ranking-based bias models. These bias models were evaluated using two experts and compared using variability analysis. AI-studies that lacked sufficient AI-attributes are more prone to risk-of-bias (RoB) in all three classes. Studies with high RoB loses fins in the butterfly model. It has been analyzed that the majority of the studies in healthcare suffer from data bias and algorithmic bias due to incomplete specifications mentioned in the design protocol and weak AI design exploited for prediction.

Integrity signals

Registry

Crossref · 10.1007/s11042-023-16029-x

Cites

Statistics

Nothing to recompute.

Claims the paper makes

0 of 70 cited sources checked · 0 supported · 0 flagged

  • Findingreasoning

    Most healthcare studies suffer from data bias and algorithmic bias because of incomplete design protocol specifications and weak AI design used for prediction.

    “It has been analyzed that the majority of the studies in healthcare suffer from data bias and algorithmic bias due to incomplete specifications mentioned in the design protocol and weak AI design exploited for prediction.”
    Quote in the paperSubstantial claim
  • Findingown data

    AI studies that lacked sufficient AI attributes are more prone to risk of bias in all three classes.

    “AI-studies that lacked sufficient AI-attributes are more prone to risk-of-bias (RoB) in all three classes.”
    Quote in the paperSubstantial claim
  • Findingown data

    There is minor disagreement between experts: studies [62] and [111] were categorized as low-bias by Expert 2 but moderate-bias by Expert 1.

    “Studies such as ([62] & [111]) are categorized in low-bias region by expert 2 but expert 1 evaluated them in moderate-bias region.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Higher mean attribute score corresponds to lower risk of bias.
  • Findingnone given

    Bias mitigation steps such as feature selection and feature augmentation were rarely performed in AI studies.

    “In addition, the bias mitigation steps such as feature selection and feature augmentation are rarely performed in AI-studies.”
    Quote in the paperSubstantial claim
  • Findingown data

    There is strong overlap between both experts' outcomes, with studies [8,34,43,88,99,142], [28,40,94,97,98], and [1,10,11,14,20,26,49,78,108,110,123,143,148,153] categorized by both experts in low, moderate, and high bias zones, respectively.

    “In this, there is a strong overlap between the outcomes of both the expert where studies such as ([8, 34, 43, 88, 99, 142]), ( [28,40,94,97,98]) and ([1, 10, 11, 14, 20, 26, 49, 78, 108, 110, 123, 143, 148, 153]) are categorized by both the expert in three zones namely, low-bias, moderate-bias and high-bias, respectively.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Higher mean attribute score corresponds to lower risk of bias.
  • Classified as probably not a substantial claim. 45
  • May combine two claims. 23

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

    Findingnone given

    Studies with high risk of bias lose fins in the butterfly model.

    “Studies with high RoB loses fins in the butterfly model.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Expert 2 categorized 26 studies as high-bias (1.9 to 1.0), 13 as moderate (2.4 to 1.9), and 9 as low-bias (2.9 to 2.4).

    “Expert 2 categorized 26 studies in high-bias (ranging from 1.9 to 1.0), 13 studies in moderate RoB (ranging from 2.4 to 1.9) and 9 studies under low-bias region (ranging from 2.9 to 2.4), having the highest mean value for the attributes scores.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Higher mean attribute score corresponds to lower risk of bias.

  • May combine two claims.

    Findingown data

    For Expert 1, remaining studies were in the low-bias region (3.5 to 2.5) with highest mean attribute scores.

    “The remaining studies are under low-bias region (ranging from 3.5 to 2.5) and have the highest mean value for the attributes scores.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Higher mean attribute score corresponds to lower risk of bias.

  • May combine two claims.

    Findingown data

    For Expert 1, 12 studies were in moderate-bias and 27 in high-bias zones.

    “Similarly, studies under moderate-and high-bias zones are, 12 [15, 34, 40, 62, 92, 97, 109, 111, 119, 131, 141, 143] and 27 [1, 6, 10, 11, 14, 20, 23, 26, 43, 49, 73, 74, 81, 88, 96, 98, 108, 110, 115, 123, 125, 127, 139, 148, 151, 153, 154].”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Higher mean attribute score corresponds to lower risk of bias.

  • Classified as probably not a substantial claim.

    Findingown data

    Raw cut-offs of 2.1 and 1.9 were determined from the intersection of frequency plot curve and mean score cumulative plot to select studies for RoB by Expert 1 and Expert 2, respectively.

    “On the basis of intersection of frequency plot curve and mean score cumulative plot of the studies, the raw cut-off of 2.1 and 1.9 is determined to select the AI-studies for RoB by both the experts, respectively.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Higher mean attribute score corresponds to lower risk of bias.

  • Classified as probably not a substantial claim.

    Findingown data

    Studies were organized into low-bias, moderate-bias, and high-bias bins for both experts based on mean score.

    “Based on the mean score value, the studies are organized in three bins namely, low-bias, moderate-bias and high-bias for both the experts as depicted in Figs.7 and 8, respectively.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Higher mean attribute score corresponds to lower risk of bias.

  • May combine two claims.

    Findingown data

    The estimated p-values were 0.0033 for t-test and 0.0010 for Spearman's rank coefficient for the studies.

    “The estimated p-values are 0.0033 and 0.0010 for the studies using t-test and Spearman's rank coefficient, respectively.”
    Quote in the paperSubstantial claim

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

    Findingown data

    For Expert 2, the RBA method placed 11 studies (22.9%) in the low-bias zone.

    “However, Expert 2 categorize studies under low-bias for RBM, RBA, RBS, and ANAL as 9 (18.7%)[34,43,88,92,98,99,111, 131, 142], 11 (22.9%)[8, 43, 88, 92, 98, 99, 111, 126, 131, 142, 162], 10 (20.8%) [8, 34, 43, 62, 88, 92, 99, 111, 142, 162], and 11 (22.9%)[14, 15, 40, 43, 49, 96, 98, 123, 131, 142, 148], respectively.”
    Quote in the paperNot a substantial claim

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

    Findingown data

    For Expert 2, the ANAL method placed 11 studies (22.9%) in the low-bias zone.

    “However, Expert 2 categorize studies under low-bias for RBM, RBA, RBS, and ANAL as 9 (18.7%)[34,43,88,92,98,99,111, 131, 142], 11 (22.9%)[8, 43, 88, 92, 98, 99, 111, 126, 131, 142, 162], 10 (20.8%) [8, 34, 43, 62, 88, 92, 99, 111, 142, 162], and 11 (22.9%)[14, 15, 40, 43, 49, 96, 98, 123, 131, 142, 148], respectively.”
    Quote in the paperNot a substantial claim

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

    Findingown data

    For Expert 2, the RBM method placed 9 studies (18.7%) in the low-bias zone.

    “However, Expert 2 categorize studies under low-bias for RBM, RBA, RBS, and ANAL as 9 (18.7%)[34,43,88,92,98,99,111, 131, 142], 11 (22.9%)[8, 43, 88, 92, 98, 99, 111, 126, 131, 142, 162], 10 (20.8%) [8, 34, 43, 62, 88, 92, 99, 111, 142, 162], and 11 (22.9%)[14, 15, 40, 43, 49, 96, 98, 123, 131, 142, 148], respectively.”
    Quote in the paperNot a substantial claim

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

    Findingnone given

    There is no unique distribution of evaluation metrics in AI models to analyze and compare performance.

    “There is no unique distribution of evaluation metrics in the AI-model to analyze and compare the performance.”
    Quote in the paperNot a substantial claim

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

    Findingown data

    For Expert 2, the RBM method placed 12 studies (25%) in the moderate-bias zone.

    “For moderate-bias, expert 2 categorizes studies for RBM, RBA, RBS and ANAL as 12 (25%) [8, 10, 14, 23, 40, 62, 81, 96, 126, 127, 139, 163], 11 (22.9%)[1, 6, 10, 14, 34, 40, 62, 81, 96, 127, 139], 12 (25%)”
    Quote in the paperNot a substantial claim

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

    Findingown data

    For Expert 2, the RBS method placed 12 studies (25%) in the moderate-bias zone.

    “For moderate-bias, expert 2 categorizes studies for RBM, RBA, RBS and ANAL as 12 (25%) [8, 10, 14, 23, 40, 62, 81, 96, 126, 127, 139, 163], 11 (22.9%)[1, 6, 10, 14, 34, 40, 62, 81, 96, 127, 139], 12 (25%)”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingcitation

    The majority of studies neglected kinship (more than 95%) and insurance (90%) attributes.

    “The majority of the studies have neglected kinship (more than 95%) and insurance (90%) attributes which are essential to identify socioeconomic status of the patient to prevent the underdiagnosis [127].”
    Quote in the paperSubstantial claim

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

    Findingown data

    For Expert 2, the RBA method placed 11 studies (22.9%) in the moderate-bias zone.

    “For moderate-bias, expert 2 categorizes studies for RBM, RBA, RBS and ANAL as 12 (25%) [8, 10, 14, 23, 40, 62, 81, 96, 126, 127, 139, 163], 11 (22.9%)[1, 6, 10, 14, 34, 40, 62, 81, 96, 127, 139], 12 (25%)”
    Quote in the paperNot a substantial claim

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

    Findingown data

    For Expert 2, the RBS method placed 10 studies (20.8%) in the low-bias zone.

    “However, Expert 2 categorize studies under low-bias for RBM, RBA, RBS, and ANAL as 9 (18.7%)[34,43,88,92,98,99,111, 131, 142], 11 (22.9%)[8, 43, 88, 92, 98, 99, 111, 126, 131, 142, 162], 10 (20.8%) [8, 34, 43, 62, 88, 92, 99, 111, 142, 162], and 11 (22.9%)[14, 15, 40, 43, 49, 96, 98, 123, 131, 142, 148], respectively.”
    Quote in the paperNot a substantial claim

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

    Findingown data

    For Expert 1, the RBA method placed 10 studies (20.8%) in the low-bias zone.

    “The number of studies categorized in the low-bias zone for RBM, RBA, RBS and ANAL by Expert 1 are 10 (20.8%) [8, 43, 88, 96, 98, 99, 115, 126, 142, 154], 10 (20.8%) [8, 43, 88, 96, 98, 99, 115, 126, 142, 154], 10 (20.8%) [8, 34, 43, 88, 99, 112, 115, 127, 141, 142] and 9 (18.7%)[8, 28, 41, 78, 99, 112, 144, 159, 162], respectively.”
    Quote in the paperNot a substantial claim

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

    Findingown data

    For Expert 1, the ANAL method placed 9 studies (18.7%) in the low-bias zone.

    “The number of studies categorized in the low-bias zone for RBM, RBA, RBS and ANAL by Expert 1 are 10 (20.8%) [8, 43, 88, 96, 98, 99, 115, 126, 142, 154], 10 (20.8%) [8, 43, 88, 96, 98, 99, 115, 126, 142, 154], 10 (20.8%) [8, 34, 43, 88, 99, 112, 115, 127, 141, 142] and 9 (18.7%)[8, 28, 41, 78, 99, 112, 144, 159, 162], respectively.”
    Quote in the paperNot a substantial claim

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

    Findingcitation

    Specific medical domain applications include mechanical ventilation, COVID CT-scans, ECG, and COVID X-rays; generic applications assess bias at different stages of AI model in medical domain.

    “Applications to specific medical domains include mechanical ventilation [45], COVID CT-Scans [104], ECG [97] and COVID X-rays [144], whereas generic refers to assessing bias at different stages of AImodel in medical domain [19,52,107].”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [45] Gallifant J, Zhang J, M.d.P.A. Lopez, T. Zhu, L. Camporota, L.A. Celi, and F. Formenti, (2022) Artificial intelligence for mechanical ventilation: Systematic review of design, reporting standards, and bias. British Journal of Anaesthesia 128(2):343–351 source

      Not checked in the free run.

    • cites [104] Palatnik de Sousa I, Vellasco MM, Costa da Silva E (2021) Explainable artificial intelligence for bias detection in covid ct-scan classifiers. Sensors 21(16):5657 source

      Not checked in the free run.

    • cites [97] Noseworthy PA, Attia ZI, Brewer LC, Hayes SN, Yao X, Kapa S, Lopez-Jimenez F (2020) Assessing and mitigating bias in medical artificial intelligence: The effects of race and ethnicity on a deep learning model for ECG analysis. Circ Arrhythm Electrophysiol 13(3):e007988 source

      Not checked in the free run.

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

    Findingown data

    For Expert 1, the ANAL method placed 27 studies (56.2%) in the high-bias zone.

    “27 (56.2%)[1,6,10,11,14,20,23,26,43,49,73,74,81,88,96,98,108, 110, 115, 123, 125, 127, 139, 148, 151, 153, 154]”
    Quote in the paperNot a substantial claim

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

    Findingown data

    For Expert 1, the RBS method placed 16 studies (33.3%) in the moderate-bias zone.

    “Similarly, number of studies under moderate-bias zone for RBM, RBA, RBS and ANAL are 12 (25%) [6, 14, 15, 34, 40, 62, 73, 92, 109, 112, 131, 159], 15 (31.2%) [6, 14, 15, 34, 40, 62, 73, 81, 92, 109, 112, 119, 154, 159, 162], 16 (33.3%)[6, 15, 28, 40, 62, 73, 92, 96-98, 109, 111, 119, 154, 159, 162], and 12 (25%) [15, 34, 40, 62, 92, 97, 109, 111, 119, 131, 141, 143], respectively.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Findingnone given

    None of the present studies specify the structural details of the dataset explicitly.

    “However, none of the present studies specify the structural details of the dataset explicitly.”
    Quote in the paperSubstantial claim

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

    Premisenone given

    Unbalanced datasets and dataset quality used for training affect the generalizability and interoperability of the model.

    “In addition, the unbalanced datasets and quality of datasets utilized for training the model impacts the generalizability and interoperability of the model.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Models trained on different public datasets generate biased outcomes because the data collection and processing are hidden.

    “Models trained on different public datasets generate biased outcomes as the process of data collection and processing is hidden [129].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [129] Shimron E, Tamir JI, Wang K, Lustig M (2022) Implicit data crimes: Machine learning bias arising from misuse of public data. Proc Natl Acad Sci 119(13):e2117203119 source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Bias in AI models can be observed through the type of input data used and data gaps during data collection.

    “Primarily, bias in the AI-model can be observed through the type of input data exploited and the data gaps during data collection [41].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [41] Ezzeddine H, Awad M, Abi Ghanem AS, Mourani B (2021) On data Bias and the usability of deep learning algorithms in classifying COVID-19 based on chest X-ray. In: 2021 IEEE 3rd International Multidisciplinary Conference on Engineering Technology (IMCET). IEEE, pp 136–143 source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Bias in healthcare AI prediction models can be introduced by incorrect data assessment, model limitations, small training size, less community participation, and other unseen factors.

    “In the healthcare industry, bias in AI-based prediction models can be introduced due to incorrect data assessment, model limitations, small training size, less community participation, and other unseen factors [39].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [39] Esteva A, Chou K, Yeung S, Naik N, Madani A, Mottaghi A, Socher R (2021) Deep learning-enabled medical computer vision. NPJ digital medicine 4(1):1–9 source

      Not checked in the free run.

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

    Study design and conductnone given

    The bias models were evaluated by two experts and compared through variability analysis.

    “These bias models were evaluated using two experts and compared using variability analysis.”
    Quote in the paperNot a substantial claim

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

    Study design and conductnone given

    The review used a PRISMA selection strategy to choose 72 AI studies for identifying bias in AI models.

    “Selection strategy using PRISMA is adopted to select the 72 crucial AI studies for identifying bias in AI models.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Bias in AI algorithms may be due to lack of diversity in data collection methods or limited knowledge of the AI-model developer.

    “This may be either due to lack of diversity in data collection methods or limited knowledge of the AI-model developer [35].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [35] Dastin J (2018) Amazon scraps secret AI recruiting tool that showed bias against women. In: Ethics of data and analytics. Auerbach Publications, pp 296–299

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    The high-bias result may be because studies did not consider imperative evaluation parameters for bias analysis.

    “It may be due to that the studies have not considered the imperative evaluation parameters for bias analysis.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    Studies in low and moderate bias regions have strong AI design and are more suitable for clinical practice than high-bias studies with low normalized mean scores.

    “Studies under low and moderate region have strong AI-design and can be considered suitable for clinical practices in comparison to studies under high-bias having low normalized mean score of the attributes.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Higher mean attribute score corresponds to lower risk of bias.

  • May combine two claims.

    Premisecitation

    Post-processing mitigation focuses on bias due to AI-model and discrimination-aware classification.

    “Bias due to AI-model is focused and mitigation efforts are determined at the post-processing step for discrimination-aware classification [31].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [31] d’Alessandro B, O’Neil C, LaGatta T (2017) Conscientious classification: A data scientist’s guide to discrimination-aware classification. Big data 5(2):120–134 source

      Not checked in the free run.

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

    Interpretationreasoning

    The higher Expert 2 mean values are due to insufficient transparency and explanations by authors, which Expert 2 captured.

    “It is due to the insufficient transparency and explanations by the authors in the paper which is captured by the expert 2.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Bias in clinical AI models can be introduced by biased features that specify the disease or by the hardware module used for image data collection.

    “Bias in clinical AI-model can be introduced either by the biased features that specify the disease or the hardware module used for image data collection [43,150].”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [43] Fernandez-Quilez A (2023) Deep learning in radiology: Ethics of data and on the value of algorithm transparency, interpretability and explainability. AI and Ethics 3(1):257–265 source

      Not checked in the free run.

    • cites [150] Tommasi T, Patricia N, Caputo B, Tuytelaars T (2017) A deeper look at dataset bias. Domain adaptation in computer vision applications. Springer, pp 37–55 source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    PD-ai models predict false positive outcomes on specific datasets and cannot be fundamentally implemented in clinical models.

    “Hence, these models predict the false positive outcomes on specific datasets and the implication to clinical models is not fundamentally possible.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    Deep learning feature extraction from medical images affects image quality and feature importance.

    “Deep learning methods exploited to extract features from medical images impacted the image quality and so the feature importance [79].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [79] Li M, Hsu W, Xie X, Cong J, Gao W (2020) SACNN: Self-attention convolutional neural network for low-dose CT denoising with self-supervised perceptual loss network. IEEE Trans Med Imaging 39(7):2289–2301 source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Loss functions such as mean square error introduce bias in training weights, impacting image feature quality and prediction.

    “In addition, loss function such as mean square error introduces bias in the training weights impacting the image feature quality and affecting the prediction [94].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [94] Nagare M, Melnyk R, Rahman O, Sauer KD, Bouman CA (2021) A Bias-reducing loss function for CT image Denoising. In: ICASSP 2021–2021 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, pp 1175–1179 source

      Not checked in the free run.

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

    Interpretationreasoning

    Expert 2 may have overrated AI attributes in particular studies, leading to high scores.

    “The expert 2 may have overrated the AI attributes in particular studies due to which high scores are deliberated.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Oversampling training data, underrepresentation of population, and testing on non-specific data automate bias in predictive model outcomes.

    “In addition, oversampling of training data, underrepresentation of population, and testing of a model on a non-specific data will automate bias in the outcome of predictive models [107].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [107] Parikh RB, Teeple S, Navathe AS (2019) Addressing bias in artificial intelligence in health care. JAMA 322(24):2377–2378 source

      Not checked in the free run.

  • May combine two claims.

    Interpretationown data

    The studies and AI attributes are strongly correlated because the computed p-values are statistically significant.

    “The results infer that the studies and AI attributes are strongly correlated as the computed p-value is statistically significant.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    Algorithmic bias in AI systems arises from black-box design, sensitive feature choice, classifier/loss/optimizer, lack of context, and concept specifications.

    “Reasons for algorithmic bias in the system are the black box nature of the AI design, choice of sensitive input features, bias due to classifier, loss function and optimizer, lack of context, and concept specifications in the development of the algorithmic design [105].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [105] Panch T, Mattie H, Atun R (2019) Artificial intelligence and algorithmic bias: Implications for health systems. J Glob Health 9(2) source

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Study design and conductown data

    Ranking bias scores were extracted by two AI scientists.

    “The scores are extracted by two AI-scientists.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    Attribute weights were determined by experienced AI professionals capable of identifying each attribute's contribution to bias.

    “Attribute weights are determined by experienced AI professionals capable of identifying the contribution of each attribute toward bias.”
    Quote in the paperSubstantial claim

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

    Interpretationreasoning

    Expert 2's domain knowledge could not extract hidden AI attributes due to less transparency and elaboration by authors.

    “Also, the expert 2 domain knowledge in AI cannot able to extract the hidden AIattributes due to less transparency and elaboration by the authors in their studies.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    Values for AI attributes were assigned based on AI scientists' domain expertise.

    “The values are assigned based on the expertise of the AI-scientist's experience in the domain.”
    Quote in the paperSubstantial claim

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

    Premisecitation

    In-processing mitigation deals with factors to mitigate algorithmic bias.

    “In-processing step deals with factors to mitigate algorithmic bias [16].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [16] Bellamy RK, Dey K, Hind M, Hoffman SC, Houde S, Kannan K, Zhang Y (2019) AI fairness 360: An extensible toolkit for detecting and mitigating algorithmic bias. IBM J Res Dev 63(4/5):4–1 source

      Not checked in the free run.

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

    Interpretationreasoning

    The steep difference may be because Expert 2 could not analyze crucial AI attributes from those studies.

    “It may be due to the reason that expert 2 not able to analyze the crucial AI-attributes from the mentioned studies.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Legal laws for data protection are country-specific and deal with personal data usage, re-usage, and modification.

    “Legal laws for data protection are country-specific which deal with personal data usage, re-usage, and modification [47].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [47] Geis JR, Brady AP, Wu CC, Spencer J, Ranschaert E, Jaremko JL, Shields WF (2019) Ethics of artificial intelligence in radiology: Summary of the joint European and North American multisociety statement. Canadian Association of Radiologists Journal 70(4):329–334 source

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    The strength of the AI attributes was evaluated using four bias computation models: RBA, RBM, RBS, and ANAL.

    “To analyze the risk of bias in a study, the strength of these AI attributes is evaluated using four efficient bias computation models namely, RBA, RBM, RBS and ANAL.”
    Quote in the paperSubstantial claim

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

    Interpretationnone given

    Legal issues concerning data sharing, privacy, and security also worsen bias in AI models.

    “In addition, legal issues related to data sharing, privacy and security are also responsible to exacerbating bias in AI-models.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    The prevalence of data and algorithmic bias may result from the black box nature of algorithmic design and incomplete specification of training data.

    “It may be due to the black box nature of the algorithmic design and incomplete specification of the training data.”
    Quote in the paperSubstantial claim

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

    Premisecitation

    The method aimed to generate a balanced dataset including protected and unprotected attributes information.

    “The method aimed to generate a balanced dataset including protected and unprotected attributes information.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Calmon et al. [21] Calmon F, Wei D, Vinzamuri B, Natesan Ramamurthy K, Varshney KR (2017) Optimized preprocessing for discrimination prevention. Adv Neural Inf Proces Syst 30

      Not checked in the free run.

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

    Premisecitation

    Nagare et al. proposed a loss function to minimize bias in training weights.

    “Nagare et al. [94] proposed a loss function to minimize bias in training weights.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Nagare et al. [94] Nagare M, Melnyk R, Rahman O, Sauer KD, Bouman CA (2021) A Bias-reducing loss function for CT image Denoising. In: ICASSP 2021–2021 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, pp 1175–1179 source

      Not checked in the free run.

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

    Premisecitation

    A ranking framework was proposed to mitigate the impact of bias due to classification error.

    “In another line of work, ranking framework was proposed to mitigate the impact of bias due to classification error [59].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [59] Kallus N, Zhou A (2019) The fairness of risk scores beyond classification: Bipartite ranking and the xauc metric. Adv Neural Inf Proces Syst 32

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Premisecitation

    Kinship and insurance attributes are essential to identify socioeconomic status of the patient to prevent underdiagnosis.

    “The majority of the studies have neglected kinship (more than 95%) and insurance (90%) attributes which are essential to identify socioeconomic status of the patient to prevent the underdiagnosis [127].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [127] Seyyed-Kalantari L, Zhang H, McDermott M, Chen IY, Ghassemi M (2021) Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations. Nat Med 27(12):2176–2182 source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Most public dataset repositories are not diverse, interoperable, and generalizable in their applications.

    “Most of the public dataset repositories are not diverse, interoperable, and generalizable in their applications [50].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [50] Greshake Tzovaras B, Tzovara A (2019) The personal data is political. The ethics of medical data donation:133–140 source

      Not checked in the free run.

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

    Premisecitation

    Adaptive reweighting of training weights was proposed for fairness-aware classification.

    “Authors proposed adaptive reweighting of the training weight for fairness-aware classification [63].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [63] Krasanakis E, Spyromitros-Xioufis E, Papadopoulos S, Kompatsiaris Y (2018) Adaptive sensitive reweighting to mitigate bias in fairness-aware classification. In: Proceedings of the 2018 world wide web conference, pp 853–862

      Not checked in the free run.

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

    Premisecitation

    Adversarial learning with variable of group learning was presented to mitigate undesired bias from a well-trained model.

    “Adversarial learning along with variable of group learning was presented to mitigate undesired bias from the well trained model [161].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [161] Zhang BH, Lemoine B, Mitchell M (2018) Mitigating unwanted biases with adversarial learning. In: Proceedings of the 2018 AAAI/ACM conference on AI, ethics, and society, pp 335–340 source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Black box nature of AI model architecture restricts generalizability and interoperability, which may cause incorrect diagnosis.

    “In addition, the black box nature of AI-model architecture restricts the generalizability and interoperability of the model which may cause incorrect diagnosis in the prediction [55,104].”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [55] Jain R, Kumar A, Nayyar A, Dewan K, Garg R, Raman S, Ganguly S (2023) Explaining sentiment analysis results on social media texts through visualization. Multimed Tools Appl:1–17 source

      Not checked in the free run.

    • cites [104] Palatnik de Sousa I, Vellasco MM, Costa da Silva E (2021) Explainable artificial intelligence for bias detection in covid ct-scan classifiers. Sensors 21(16):5657 source

      Not checked in the free run.

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

    Premisecitation

    Iterative training was recommended to address changes in in-training weights for input classes.

    “Also, iterative training was recommended to address the change in in-training weights for the input classes of the model [63].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [63] Krasanakis E, Spyromitros-Xioufis E, Papadopoulos S, Kompatsiaris Y (2018) Adaptive sensitive reweighting to mitigate bias in fairness-aware classification. In: Proceedings of the 2018 world wide web conference, pp 853–862

      Not checked in the free run.

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

    Premisecitation

    Social and ethical barriers related to big data can be evaluated for differential diagnosis.

    “Social and ethical barriers in relation to big data can be evaluated for differential diagnosis [86].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [86] McGregor C, Dewey C, Luan R (2021) Big data and artificial intelligence in healthcare: Ethical and social implications of neonatology. In: 2021 IEEE international symposium on technology and society (ISTAS). IEEE, pp 1–1

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    Seventy-two key studies were selected using the PRISMA model.

    “To study the bias, 72 key studies were selected using PRISMA model.”
    Quote in the paperSubstantial claim

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

    Study design and conductnone given

    ANAL methods validate results obtained by other methods.

    “ANAL methods validate the results obtained by the other methods.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    Integrating a regularizer in the model reduces correlation between sensitive features and input class labels.

    “In [60], authors proposed integrating a regularizer in the model to reduce the correlation between the sensitive features and input class labels.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [60] Kamishima T, Akaho S, Asoh H, Sakuma J (2012) Fairness-aware classifier with prejudice remover regularizer. In: Machine learning and knowledge discovery in databases: European conference, ECML PKDD 2012, Bristol, UK, September 24–28, 2012. Proceedings, part II 23. Springer, Berlin Heidelberg, pp 35–50 source

      Not checked in the free run.

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

    Premisecitation

    Calmon et al. proposed a mapping approach to map original data points with newly created datasets to incorporate demographic information.

    “Calmon et al. [21] proposed a mapping approach to map the original data points with the newly created datasets for incorporating demographic information.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Calmon et al. [21] Calmon F, Wei D, Vinzamuri B, Natesan Ramamurthy K, Varshney KR (2017) Optimized preprocessing for discrimination prevention. Adv Neural Inf Proces Syst 30

      Not checked in the free run.

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

    Premisecitation

    Anti-discrimination rules to define legal policies may prevent dissemination of discrimination against certain groups in the algorithm.

    “Anti-discrimination rules to define legal policies that may prevent dissemination of discrimination against certain group in the algorithm [163].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [163] Zhang L, Wu X (2017) Anti-discrimination learning: A causal modeling-based framework. International Journal of Data Science and Analytics 4(1):1–16 source

      Not checked in the free run.

66 of 66

  • The paper relies on this without support.

    Unawareness or ignorance of sociodemographic and clinical information during training and test data specification causes misclassification and biased outcomes.

    “Unawareness and ignorance of sociodemographic information namely, gender, age, ethnicity, race, socioeconomic information and clinical information such as medical history, during training and test data specification are accountable for misclassification and biased outcome.”

  • The paper relies on this without support.

    Bias in problem formulation for predictive models is introduced by limited domain knowledge of the developer.

    “The bias in problem formulation for the predictive model is introduced due to the limited domain knowledge of the developer.”

  • The paper relies on this without support.

    Automated and black box design of AI-based predictive models is responsible for algorithmic bias.

    “Automated and black box design of the AI-based predictive model is responsible for algorithmic bias.”

  • The paper relies on this without support.

    Data exploration including synthetic data, surrogate data, and data manipulation for model training is responsible for selection bias.

    “Data exploration that includes the generation of synthetic data, surrogate data and data manipulation for model training is responsible for selection bias.”

  • The paper relies on this without support.

    AI-based algorithms learn and train from historical data that encodes bias in the model.

    “AI-based algorithms learn and train from historical data that encodes the bias in model.”

  • The paper relies on this without support.

    Corrupted and stale data from unreliable sources introduce bias in the outcome because the model learns unfairly.

    “Corrupted and stale data from unreliable sources introduce bias in the outcome as the model learns unfairly.”

  • The paper relies on this without support.

    Corrupted data may be incomplete and inaccurate, leading to hard-to-quantify biased output.

    “The corrupted data may be incomplete and inaccurate that will lead to hard to quantify biased output.”

  • The paper relies on this without support.

    Excessive training of a model causes overfitting and prediction errors.

    “However, excessive trained model causes overfitting of the algorithm and prediction errors.”

  • The paper relies on this without support.

    Classifiers can generate biased and unfair outcomes based on input data.

    “Also, classifiers could generate biased and unfair outcomes based on the input data.”

  • The paper relies on this without support.

    Higher mean attribute score corresponds to lower risk of bias.

    “The higher the mean value, the lower is the RoB.”

  • The paper relies on this without support.

    Synthetic data increases training data and minimizes risk associated with patient data sharing and privacy.

    “Synthetic data not only increases the amount of training data but also minimizes the risk associated with patient data sharing and privacy.”

  • The paper relies on this without support.

    Synthetic data addresses bias in training data due to unbalanced datasets.

    “In addition, synthetic data also addresses the bias in training data due to unbalance datasets.”

  • The paper relies on this without support.

    If original data suffers from bias, synthetic data will intensify it and make the prediction model unstable.

    “However, if the original data suffers from bias then, synthetic data will intensify it in the training dataset and make the prediction model unstable.”

  • The paper relies on this without support.

    Public datasets are inconsistent in data collection, aggregation, and truthfulness.

    “In addition, these datasets are inconsistent in their data collection, aggregation, and truthfulness process.”

  • The paper relies on this without support.

    Discrimination in healthcare designs leads to underdiagnosis or misdiagnosis for underrepresented subgroups.

    “Discrimination in healthcare designs lead to underdiagnosis or misdiagnosis to underrepresented subgroups of the population.”

  • The paper relies on this without support.

    Real-time validation of AI models in healthcare is not possible to prevent harmful outcomes.

    “The real-time validation of the AI-model in healthcare is not possible to prevent harmful outcomes.”

  • The paper relies on this without support.

    Developers generally do not publicize exact training, validation, and testing data split information.

    “Generally, developers do not publicize the exact training, validation, and testing data split information.”

  • The paper relies on this without support.

    Algorithmic bias can be triggered through inbuilt bias in feature extraction and selection.

    “Algorithmic bias can also be triggered through the inbuilt bias in the feature extraction and selection process.”

  • The paper relies on this without support.

    Feature distribution in medical imaging is non-uniform, more informative for one gender.

    “Sometimes feature distribution in medical imaging is non-uniform i.e., it is more informative for one gender in comparison to others.”

  • The paper relies on this without support.

    Algorithmic processing design including hidden layers, training weights, loss function, and classifier design exhibits significant bias in prediction.

    “The algorithmic processing design that includes the number of hidden layers, training weights, loss function, and classifier design exhibits significant bias in the prediction.”

  • The paper relies on this without support.

    Interpreting certain predictions is impossible due to AI model's hidden design architecture.

    “Interpreting certain predictions is impossible due to the AI-model's hidden design architecture.”

  • The paper relies on this without support.

    Lack of contextual standards and causal reasoning for AI model could lead to algorithmic bias.

    “In another line of work, lack of contextual standards and causal reasoning for the AI-model could lead to algorithmic bias.”

  • The paper relies on this without support.

    Usage of algorithm for non-specific and inappropriate context predicted unwanted results.

    “Usage of the algorithm for a non-specific and inappropriate context predicted unwanted results.”

  • The paper relies on this without support.

    The 72 selected studies are representative of the broader population of AI-based medical systems.

    “Our study have also revealed that the majority of the studies in the healthcare suffered from data bias and algorithmic bias.”

  • The paper relies on this without support.

    The 44 AI attributes and four bias computation models adequately capture the sources and risk of bias in AI-based medical systems.

    “Data bias and algorithmic bias in design of AI-based medical systems in these three categorizes is identified using 44 AI-attributes.”

  • The paper relies on this without support.

    The expert scoring used to categorize studies into bias zones is reliable and valid.

    “The number of studies categorized in the low-bias zone for RBM, RBA, RBS and ANAL by Expert 1 are 10 (20.8%) [8, 43, 88, 96, 98, 99, 115, 126, 142, 154], 10 (20.8%) [8, 43, 88, 96, 98, 99, 115, 126, 142, 154], 10 (20.8%) [8, 34, 43, 88, 99, 112, 115, 127, 141, 142] and 9 (18.7%)[8, 28, 41, 78, 99, 112, 144, 159, 162], respectively.”
26 of 26

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?

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.

  • The 2.5 low-bias cutoff points in opposite directions

    Minor

    One passage places Expert 1's low-bias studies between 2.5 and 3.5, while the figure text says scores below 2.5 are in the low-bias zone.

  • Expert 1 assigns study 154 to two RBA zones

    Minor

    In the RBA results, study 154 appears in both Expert 1's low-bias and moderate-bias lists, although the zones are presented as separate classifications.

  • Study 142 is not low-bias under all four Expert 1 methods

    Minor

    The variability section says both experts put study 142 in the low-bias zone under all four methods. Expert 1's stated analytical low-bias list does not contain study 142.

3 of 3

The citation lists haven’t been fetched yet

References 143 · Cited by 65

Research works printed but not linked: 20

  1. Ashraf A, Khan S, Bhagwat N, Chakravarty M, Taati B (2018) Learning to unlearn: Building immunity to dataset bias in medical imaging studies. arXiv preprint arXiv:181201716
  2. Benjamin M, Gagnon P, Rostamzadeh N, Pal C, Bengio Y, Shee A (2019) Towards standardization of data licenses: the Montreal data license. arXiv preprint arXiv:190312262
  3. Calmon F, Wei D, Vinzamuri B, Natesan Ramamurthy K, Varshney KR (2017) Optimized preprocessing for discrimination prevention. Adv Neural Inf Proces Syst 30
  4. Danks D, London AJ (2017) Algorithmic Bias in autonomous systems. IJCAI 17(2017):4691–4697
  5. El-Baz A, Suri JS (eds) (2019) Big data in multimodal medical imaging. CRC Press
  6. Kallus N, Zhou A (2019) The fairness of risk scores beyond classification: Bipartite ranking and the xauc metric. Adv Neural Inf Proces Syst 32
  7. Krasanakis E, Spyromitros-Xioufis E, Papadopoulos S, Kompatsiaris Y (2018) Adaptive sensitive reweighting to mitigate bias in fairness-aware classification. In: Proceedings of the 2018 world wide web conference, pp 853–862
  8. Lee NT, Resnick P, Barton G (2019) Algorithmic bias detection and mitigation: Best practices and policies to reduce consumer harms. Brookings Institute, Washington, DC, USA
  9. Loftus JR, Russell C, Kusner MJ, Silva R (2018) Causal reasoning for algorithmic fairness. arXiv preprint arXiv:180505859
  10. McDuff D, Cheng R, Kapoor A (2018) Identifying bias in AI using simulation. arXiv preprint arXiv:181000471
  11. Mia MR, Hoque ASML, Khan SI, Ahamed SI (2022) A privacy-preserving national clinical data warehouse: Architecture and analysis. Smart Health 23:100238
  12. Pfau J, Young AT, Wei ML, Keiser MJ (2019) Global saliency: Aggregating saliency maps to assess dataset artefact bias. arXiv preprint arXiv:191007604
  13. Saleiro P, Kuester B, Hinkson L, London J, Stevens A, Anisfeld A, Ghani R (2018) Aequitas: A bias and fairness audit toolkit. arXiv preprint arXiv:181105577
  14. Santa Cruz BG, Bossa MN, Sölter J, Husch AD (2021) Public covid-19 x-ray datasets and their impact on model bias–a systematic review of a significant problem. Med Image Anal 74:102225
  15. Sun TY, Walk OJ IV, Chen JL, Nieva HR, Elhadad N (2020) Exploring gender disparities in time to diagnosis. arXiv preprint arXiv:201106100
  16. Suresh H, Guttag JV (2019) A framework for understanding unintended consequences of machine learning. arXiv preprint arXiv:190110002 2(8)
  17. Suri JS, Puvvula A, Biswas M, Majhail M, Saba L, Faa G, Naidu S (2020) COVID-19 pathways for brain and heart injury in comorbidity patients: A role of medical imaging and artificial intelligence-based COVID severity classification: A review. Comput Biol Med 124:103960
  18. Suri JS, Rangayyan RM, Imaging B, Mammography, and Computer-Aided Diagnosis of Breast Cancer. (2006) SPIE: Bellingham. WA, USA
  19. Wu Y, Zhang L, Wu X (2018) Fairness-aware classification: Criterion, convexity, and bounds. arXiv preprint arXiv:180904737
  20. Zhou N, Zhang Z, Nair VN, Singhal H, Chen J, Sudjianto A (2021) Bias, fairness, and accountability with AI and ML algorithms. arXiv preprint arXiv:210506558

Authors
Kumar, A., Aelgani, V., Vohra, R., Gupta, S. K., Bhagawati, M., Paul, S., Saba, L., Suri, N., Khanna, N. N., Laird, J. R., Johri, A. M., Kalra, M., Fouda, M. M., Fatemi, M., Naidu, S., & Suri, J. S.
Affiliations
  • Bennett University (IN)
  • Bharati Vidyapeeth's College of Engineering, Delhi (IN)
  • North Eastern Hill University (IN)
  • University of Cagliari (IT)
  • Indraprastha Apollo Hospitals (IN)
  • St. Helena Hospital (US)
  • Queen's University (CA)
  • Massachusetts General Hospital (US)
  • Idaho State University (US)
  • Mayo Clinic (US)
  • University of Minnesota, Duluth (US)
Corresponding author
Suri, Jasjit S.
ORCID
Venue
Multimedia Tools and Applications
Publisher
Springer Science and Business Media LLC
Published
2023-07-22
Volume
83
Issue
6
Pages
18005-18057
ISSN
  • 1380-7501 (ISSN-L)
  • 1573-7721
DOI
10.1007/s11042-023-16029-x
OpenAlex
W4385074164
Semantic Scholar
512ddb4f691c604bab0c4debfdc2bf066f2366a7
Cited by
65
References
143
Citations by year
2026: 16 · 2025: 40 · 2024: 7 · 2023: 2
FWCI
0.76
Type
journal article
Language
en
Open access
Licence
Locations
Multimedia Tools and Applications (publishedVersion, public-domain)
Topics
  • Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
  • Social Sciences › Safety Research › Ethics and Social Impacts of AI
  • Health Sciences › Health Professions › Health Information Management › Artificial Intelligence in Healthcare
Keywords
algorithmic bias, data bias, risk of bias, hybrid AI models, clinical implementation of AI, medical prediction

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 2026-10-05.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-05.

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