CoChat
Sign Up
Cited by 241Artificial Intelligence in Healthcare and EducationHuman-Automation Interaction and SafetyClinical Reasoning and Diagnostic Skills

Experimental evidence of effective human–AI collaboration in medical decision-making

Open access
Reverberi, Carlo, Rigon, Tommaso, Solari, Aldo, Hassan, Cesare, Cherubini, Paolo, GI Genius CADx Study Group, Antonelli, Giulio, Awadie, Halim +25 more·2022·Scientific Reports
DOI 10.1038/s41598-022-18751-2PMC PMC9440124

Published version of a preprint

Open this paper in CoChat — free accountOpen PDF

Abstract

Artificial Intelligence ( ai ) systems are precious support for decision-making, with many applications also in the medical domain. The interaction between md s and ai enjoys a renewed interest following the increased possibilities of deep learning devices. However, we still have limited evidence-based knowledge of the context, design, and psychological mechanisms that craft an optimal human– ai collaboration. In this multicentric study, 21 endoscopists reviewed 504 videos of lesions prospectively acquired from real colonoscopies. They were asked to provide an optical diagnosis with and without the assistance of an ai support system. Endoscopists were influenced by ai ( $$\textsc {or}=3.05$$ O R = 3.05 ), but not erratically: they followed the ai advice more when it was correct ( $$\textsc {or}=3.48$$ O R = 3.48 ) than incorrect ( $$\textsc {or}=1.85$$ O R = 1.85 ). Endoscopists achieved this outcome through a weighted integration of their and the ai opinions, considering the case-by-case estimations of the two reliabilities. This Bayesian-like rational behavior allowed the human– ai hybrid team to outperform both agents taken alone. We discuss the features of the human– ai interaction that determined this favorable outcome.

AI snapshot

Written from the paper’s full text.

TL;DR

In a multicentric within-subject experiment, 21 endoscopists diagnosed 504 colonoscopy lesion videos with and without AI support. They were influenced by AI but selectively followed correct more than incorrect advice, producing better hybrid human–AI accuracy than either alone.

Key findings

  • AI influenced endoscopists’ decisions (OR = 3.05 [2.76, 3.39]); on average, for every three lesions where endoscopists disagreed with AI in Session 1, only one disagreement remained in Session 2.
  • AI improved diagnostic accuracy (OR = 1.39 [1.28, 1.51]); every five lesions correctly evaluated without AI became seven correctly evaluated with AI. Effectiveness was OR = 3.48 [3.07, 3.98], while safety was OR = 0.54 [0.48, 0.62], indicating some negative impact when AI was wrong but a stronger tendency to accept correct AI advice.
  • Endoscopists discriminated good from bad AI advice: they followed AI more when it was correct than when it was incorrect, and their own confidence and perceived AI confidence predicted accuracy and modulated opinion switching.
  • Expertise modulated effects: non-experts benefited more in accuracy (OR = 1.61 [1.44, 1.79]) than experts (OR = 1.15 [1.01, 1.30]); both groups showed influence, effectiveness, and safety, though experts had lower safety (0.45 [0.37, 0.54]) than non-experts (0.63 [0.54, 0.75]).

Methods

  • Mixed design: within-subjects factor was treatment (no AI vs. AI); between-subjects factor was expertise (experts vs. non-experts). Twenty-one endoscopists participated: 10 experts and 11 non-experts from Austria, Israel, Japan, Portugal, and Spain.
  • Participants reviewed 504 videos of real colonoscopies, each presenting one lesion, in two sessions separated by at least two weeks per batch. Session 1 had no AI advice; Session 2 used GI Genius v3.0 in cade+cadx modality to dynamically offer an optical diagnosis. Histopathological diagnosis was ground truth.
  • Main analysis used logistic regression models with lesion- and endoscopist-specific random effects via lme4, estimating four odds-ratio endpoints: influence of AI, diagnostic accuracy, effectiveness, and safety. The study was preregistered.

Limitations

  • The paper does not include a dedicated limitations section. Evident from the text: the sample was 21 endoscopists using one AI device in one clinical domain (colonoscopy optical diagnosis), limiting generalizability.
  • AI accuracy depended on interpretation of its dynamic output: human-interpreted standard accuracy was 84.9% but 72.9% when “uncertain” or “not sure/not notice” outputs were counted as errors; automated interpretation gave 84.5% excluding “uncertain” and 79.3% including it.
  • The AI device was non-transparent about its reasons, and the authors note that absence of algorithmic transparency did not disrupt performance; this leaves the mechanisms by which endoscopists inferred AI reliability only indirectly examined.

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

Integrity signals

Registry

Crossref · 10.1038/s41598-022-18751-2

Cites

47 of 47 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

16 of 36 cited sources checked · 8 supported · 0 flagged · accuracy not yet measured

  • Premisecitation

    There is still limited evidence-based knowledge of the context, design, and psychological mechanisms that would craft an optimal human-AI team.

    “Notwithstanding a clear agenda on the issues that we should explore further, we still have limited evidence-based knowledge of the context, design, and psychological mechanisms that would craft an optimal human–ai team15–17.”
    Quote in the paperSubstantial claim2 of 3 sources checked
    • cites [15] Cabitza, F., Campagner, A. & Simone, C. The need to move away from agential-AI: Empirical investigations, useful concepts and open issues. Int. J. Hum. Comput. Stud. 155, 1–11 (2021). source
    • cites [16] Okamura, K. & Yamada, S. Adaptive trust calibration for human–AI collaboration. PLoS One 15 (2020). source

      The Discussion’s “Limitations and future research” section calls for further experiments on different tasks, effective cue designs, and dispositional and situational trust. That supports a narrower claim that aspects of human–AI trust calibration need more evidence; it does not establish a general lack of evidence-based knowledge about all the context, design, and psychological mechanisms needed for an optimal human–AI team. No page number is available in our copy. (LLM, conf=0.82)

      Experimental — accuracy not yet measured
    • cites [17] Gu, H., Huang, J., Hung, L. & Chen, X. A. Lessons learned from designing an AI-enabled diagnosis tool for pathologists. In Proceedings of the ACM on Human–Computer Interaction, Vol. 5, 1–25 (2021). source

      The Introduction says little work has addressed how physicians can use imperfect AI in diagnosis, and §2.1 identifies a lack of research on how medical human–AI collaboration can provide new knowledge to AI. Those passages support a narrower research gap, but not the claim of limited evidence-based knowledge about context, design, and psychological mechanisms for optimal human–AI teams generally. Page numbers are unavailable in our copy. (LLM, conf=0.84)

      Experimental — accuracy not yet measured
  • Interpretationreasoning

    The successful extraction of two critical task parameters (endoscopists' own accuracy and the ai's accuracy) was likely at the core of the endoscopists' ability to select and follow the best ai advice.

    “The successful extraction of two critical task parameters was likely at the core of this ability: endoscopists could intuitively but reliably predict for each lesion both their accuracy (not obvious37,38) and the accuracy of the ai (not obvious39).”
    Quote in the paperSubstantial claim1 of 3 sources checked
    • cites [37] Naguib, M. et al. Anesthesiologists’ overconfidence in their perceived knowledge of neuromuscular monitoring and its relevance to all aspects of medical practice: An international survey. Anesth. Analg. 128, 1118–1126. https://doi.org/10.1213/ANE.0000000000003714 (2019). source
    • cites [38] Meyer, A. N. D., Payne, V. L., Meeks, D. W., Rao, R. & Singh, H. Physicians’ diagnostic accuracy, confidence, and resource requests: A vignette study. JAMA Intern. Med. 173, 1952–1958. https://doi.org/10.1001/jamainternmed.2013.10081 (2013). source
    • cites [39] Benda, N. C., Novak, L. L., Reale, C. & Ancker, J. S. Trust in AI: Why we should be designing for APPROPRIATE reliance. J. Am. Med. Inform. Assoc. 29, 207–212. https://doi.org/10.1093/jamia/ocab238 (2022). source
  • Premisecitation

    Knowing when to trust or distrust the ai allows the human expert to apply its knowledge appropriately, improving decision outcomes in cases where the ai is likely to perform poorly.

    “Knowing when to trust or distrust the ai allows the human expert to apply its knowledge appropriately, improving decision outcomes in cases where the ai is likely to perform poorly4,40.”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites [4] Zhang, Y., Liao, Q. V. & Bellamy, R. K. Effect of confidence and explanation on accuracy and trust calibration in AI-assisted decision making. In Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency, 295–305 (2020). source
    • cites [40] Bansal, G. et al. Updates in human–AI teams: Understanding and addressing the performance/compatibility tradeoff. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 33, 2429–2437 (2019). source
  • Findingown data

    Endoscopists' confidence in an optical diagnosis is strongly predictive of its accuracy in both sessions: the higher the confidence, the higher the accuracy.

    “We found that endoscopists’ confidence in an optical diagnosis is strongly predictive of its accuracy in both sessions: the higher the confidence, the higher the accuracy (Table 3).”
    Quote in the paperNot a substantial claim
  • Findingown data

    Endoscopists were influenced by the AI opinion, with an odds ratio of 3.05 (95% CI 2.76-3.39).

    “Endoscopists were influenced by the ai opinion ().”
    • Also stated in: § Statistical analyses · plain
      “| 1. Influence of the ai | | 3.05 [2.76, 3.39] |”
    Quote in the paperSubstantial claim
  • Classified as probably not a substantial claim. 17
  • May combine two claims. 16
  • The quote only partly states this claim. 2

  • Classified as probably not a substantial claim. May combine two claims.

    Premisecitation

    Complementarity between AI and MDs advises a collaboration between human and artificial minds, fostering a hybrid intelligence that could deliver outcomes superior to those reached by each mind alone.

    “Such complementarity advises a collaboration between human and artificial minds, fostering a “hybrid intelligence” that could deliver outcomes superior to those reached by each mind alone2,3.”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [2] Dellermann, D. et al. The future of human–AI collaboration: A taxonomy of design knowledge for hybrid intelligence systems. arXiv:2105.03354 (2021).

      Not checked in the free run.

    • cites [3] Akata, Z. et al. A research agenda for hybrid intelligence: Augmenting human intellect with collaborative, adaptive, responsible, and explainable artificial intelligence. Computer 53, 18–28 (2020). source

  • May combine two claims.

    Premisecitation

    Previous studies on CADx systems for colonoscopy focused on assessing the performances of AI against the accuracy of physicians or against criteria established by gastroenterology societies for implementing the technology in a clinical setting.

    “Previous studies on cadx systems for colonoscopy focused on assessing the performances of ai against the accuracy of physicians19,20,27–30, or against the criteria established by gastroenterology societies for implementing the technology in a clinical setting31–34.”
    Quote in the paperSubstantial claim5 of 10 sources checked
    • cites [19] Kudo, S.-E. et al. Artificial intelligence and colonoscopy: Current status and future perspectives. Dig. Endosc. 31, 363–371 (2019). source
    • cites [20] Larsen, S. L. V. & Mori, Y. Artificial intelligence in colonoscopy: A review on the current status. DEN Open 2 (2022). source

      The ‘Computer-aided polyp diagnosis’ section (page unavailable) reports a comparison of CADx accuracy with observers’ diagnoses and describes a study assessing whether CADx met the NPV ≥ 90% threshold for a diagnose-and-leave strategy. It does not identify that threshold as a criterion established by a gastroenterology society or describe it as a general criterion for clinical implementation. (LLM, conf=0.84)

      Experimental — accuracy not yet measured
    • cites [27] Mori, Y. et al. Real-time use of artificial intelligence in identification of diminutive polyps during colonoscopy: A prospective study. Ann. Intern. Med. 169, 357–366 (2018). source
    • cites [28] Byrne, M. F. et al. Real-time differentiation of adenomatous and hyperplastic diminutive colorectal polyps during analysis of unaltered videos of standard colonoscopy using a deep learning model. Gut 68, 94–100 (2019). source

      Split into 2 atoms — 1 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

      Experimental — accuracy not yet measured
    • cites [29] Xu, Y. et al. Comparison of diagnostic performance between convolutional neural networks and human endoscopists for diagnosis of colorectal polyp: A systematic review and meta-analysis. PLoS One 16, e0246892. https://doi.org/10.1371/journal.pone.0246892 (2021). source
    • cites [30] Kudo, S.-E. et al. Artificial intelligence and computer-aided diagnosis for colonoscopy: where do we stand now? Transl. Gastroenterol. Hepatol. 6 (2021). source

      Split into 2 atoms — 0 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 1 unverifiable. Expand the claim to see per-atom verdicts.

      Experimental — accuracy not yet measured
    • cites [31] ASGE Technology Committee et al. ASGE Technology Committee systematic review and meta-analysis assessing the ASGE PIVI thresholds for adopting real-time endoscopic assessment of the histology of diminutive colorectal polyps. Gastrointest. Endosc. 81 (2015). source

      The Discussion states that ASGE PIVI established performance thresholds for adopting real-time optical assessment of diminutive colorectal polyps, and this review assesses imaging technologies against those thresholds. It does not establish that previous AI-based CADx studies evaluated their performance against those criteria; the technologies analyzed here are endoscopic imaging methods such as NBI, i-SCAN, and FICE. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
    • cites [32] Berzin, T. M. et al. Position statement on priorities for artificial intelligence in GI endoscopy: A report by the ASGE Task Force. Gastrointest. Endosc. 92, 951–959. https://doi.org/10.1016/j.gie.2020.06.035 (2020). source
    • cites [33] Barua, I. et al. Real-time artificial intelligence-based optical diagnosis of neoplastic polyps during colonoscopy. NEJM Evid. 1 (2022). source
    • cites [34] Hassan, C., Balsamo, G., Lorenzetti, R., Zullo, A. & Antonelli, G. Artificial intelligence allows leaving-in-situ colorectal polyps. Clin. Gastroenterol. Hepatol.https://doi.org/10.1016/j.cgh.2022.04.045 (2022). source

  • Probably not a substantial claim: may not pass S2 (checkable).

    Premisecitation

    What is envisaged is teaming humans together with artificial intelligence to exploit the advantages of hybrid intelligence.

    “what is envisaged is teaming humans together with artificial intelligence to exploit the advantages of hybrid intelligence2,3.”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites [2] Dellermann, D. et al. The future of human–AI collaboration: A taxonomy of design knowledge for hybrid intelligence systems. arXiv:2105.03354 (2021).

      Not checked in the free run.

    • cites [3] Akata, Z. et al. A research agenda for hybrid intelligence: Augmenting human intellect with collaborative, adaptive, responsible, and explainable artificial intelligence. Computer 53, 18–28 (2020). source

  • Classified as probably not a substantial claim.

    Premisecitation

    Over-reliance or under-reliance on ai is wrong when decoupled from considerations on the relative informativeness of the ai.

    “over-reliance or under-reliance on ai, regard a general attitude towards support systems, which is wrong when decoupled from considerations on the relative informativeness of the ai39,41,42.”
    Quote in the paperSubstantial claim3 of 3 sources checked
    • cites [39] Benda, N. C., Novak, L. L., Reale, C. & Ancker, J. S. Trust in AI: Why we should be designing for APPROPRIATE reliance. J. Am. Med. Inform. Assoc. 29, 207–212. https://doi.org/10.1093/jamia/ocab238 (2022). source
    • cites [41] Medow, M. A., Arkes, H. R. & Shaffer, V. A. Are residents’ decisions influenced more by a decision aid or a specialist’s opinion? A randomized controlled trial. J. Gen. Intern. Med. 25, 316–320. https://doi.org/10.1007/s11606-010-1251-y (2010). source

      The Discussion says disregarding advice solely on the basis of confidence may be ill-advised, and cautions that the management decision studied had no indisputably correct answer. This supports caution about judging reliance in isolation, but the paper does not establish the broader claim that both over- and under-reliance on AI are wrong unless assessed against the AI’s relative informativeness. (LLM, conf=0.78)

      Experimental — accuracy not yet measured
    • cites [42] Rubin, D. L. Artificial intelligence in imaging: The radiologist’s role. J. Am. Coll. Radiol. 16, 1309–1317. https://doi.org/10.1016/j.jacr.2019.05.036 (2019). source

      In “Maintaining radiologist expertise” (page not provided), Rubin warns that blindly accepting AI output can lead radiologists astray, particularly when an algorithm performs poorly on unusual cases. “Evaluation of AI tools” also urges testing performance on local data. The paper does not address under-reliance or frame the issue as a general attitude that must be judged against AI’s informativeness relative to the user’s. (LLM, conf=0.85)

      Experimental — accuracy not yet measured

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

    Premisecitation

    Algorithmic transparency has sometimes been advocated as pivotal for promoting an effective interaction with ai.

    “algorithmic transparency has been sometimes advocated as pivotal for promoting an effective interaction with ai6,9,36”
    Quote in the paperNot a substantial claim1 of 3 sources checked
    • cites [6] Gretton, C. Trust and transparency in machine learning-based clinical decision support. In Human and Machine Learning 279–292 (Springer, 2018). source
    • cites [9] Ghassemi, M., Oakden-Rayner, L. & Beam, A. L. The false hope of current approaches to explainable artificial intelligence in health care. Lancet Digit. Health 3, e745–e750 (2021). source
    • cites [36] WHO. Ethics and Governance of Artificial Intelligence for Health: WHO Guidance (World Health Organization, 2021).

      Not checked in the free run.

  • The quote only partly states this claim.

    Findingown data

    When AI was correct, AI assistance increased the probability of a correct human diagnosis, with an odds ratio of 3.48 (95% CI 3.07-3.98).

    “| 3. Effectiveness | | 3.48 [3.07, 3.98] |”
    Quote in the paperSubstantial claim

  • The quote only partly states this claim.

    Findingown data

    When AI was incorrect, AI assistance deteriorated endoscopists' diagnostic performance, with an odds ratio of 0.54 (95% CI 0.48-0.62).

    “| 4. Safety | | 0.54 [0.48, 0.62] |”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingreasoning

    The benefits on accuracy were stronger for non-experts, making their performance with ai assistance similar to that of experts without assistance.

    “The benefits on accuracy were stronger for non-experts (also given their lower baseline), making their performance with ai assistance similar to that of experts without assistance.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Experts were less able than non-experts to discriminate between good and bad ai advice: both efficacy and safety are higher in non-experts than in experts.

    “experts were less able than non-experts to discriminate between good and bad ai advice: both efficacy and safety are higher in non-experts than in experts (Table 5).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Endoscopists consider and are substantially influenced by ai opinion.

    “We showed that endoscopists consider and are substantially influenced by ai opinion.”
    Quote in the paperSubstantial claim

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

    Findingown data

    Similar results on the influence of confidence were obtained in a related analysis on confidence scores.

    “Similar results were obtained in a related analysis on confidence scores (see Appendix B.3).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    No differences were present in the relative trust towards ai between experts and non-experts; both groups had a slightly higher average confidence towards ai as compared to themselves.

    “no differences were present in the relative trust towards ai in the two groups: both groups had a slightly higher average confidence towards ai as compared to themselves (see Fig. B.11 in Appendix).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Endoscopists can separate the good from the bad ai advice, accepting selectively more the former than the latter, as shown by a higher efficacy than safety index.

    “endoscopists can separate the good from the bad ai advice, accepting selectively more the former than the latter, as shown by a higher efficacy than safety index.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    In S2 non-experts increased both their specificity and sensitivity, whereas experts increased only their sensitivity.

    “in S2 non-experts increased both their specificity and sensitivity, whereas experts increased only their sensitivity (see Appendix B.5).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    When an automated algorithm interprets the output, the ai accuracy is 84.5% when the label 'uncertain' is excluded, and 79.3% when 'uncertain' is considered an error.

    “when an automated algorithm interprets the output, the ai accuracy is 84.5% when the label “uncertain” is excluded, while it is 79.3% when “uncertain” is considered an error.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingreasoning

    In the present case study, the use of ai proved effective and safe.

    “In the present case study, the use of ai proved effective and safe.”
    Quote in the paperSubstantial claim

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

    Premisecitation

    Artificial Intelligence systems are increasingly recognized as a precious tool for improving medical decision-making.

    “Artificial Intelligence systems are increasingly recognized as a precious tool for improving medical-decision making1.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [1] Topol, E. J. High-performance medicine: The convergence of human and artificial intelligence. Nat. Med. 25, 44–56 (2019). source

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

    Study design and conductnone given

    The study's predictions were pre-registered before data gathering, together with the study plan, statistical models, and analyses.

    “Our predictions were pre-registered before the data gathering, together with the study plan, the statistical models, and the analyses (preregistration is available at https://osf.io/y9at5).”
    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 conductnone given

    The study was approved by the local Institutional Review Board and conducted following the Declaration of Helsinki.

    “The study was approved by the local Institutional Review Board (Comitato Etico Lazio 1, prot. 611/CE Lazio1) and conducted following the Declaration of Helsinki.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Premisenone given

    In qualitative Bayesian terms, the AI output is a further piece of information that should be integrated to revise the endoscopist's diagnosis, not differently from considering an informed opinion volunteered by a human colleague with a slightly different expertise profile.

    “In qualitative Bayesian terms, the ai output is a further piece of information that should be integrated to revise the endoscopist’s diagnosis, not differently as it would happen by considering an “informed opinion” volunteered by a human colleague with a slightly different expertise profile.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    Histopathological diagnosis of each lesion was considered ground truth for evaluating human and AI performance.

    “For evaluating human and AI performance, we considered the histopathological diagnosis of each lesion as ground truth.”
    Quote in the paperSubstantial claim

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

    Study design and conductnone given

    All participants provided written informed consent.

    “All participants provided written informed consent.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Premisecitation

    Incoming visual information is compared to previously stored knowledge about a finite set of possible diagnoses in a Bayesian-like procedure that revises the endoscopist's confidence toward those diagnoses.

    “Incoming visual information is compared to previously stored knowledge about a finite set of possible diagnoses in a Bayesian-like procedure that revises the endoscopist’s confidence toward those diagnoses.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [25] Griffiths, T. L., Kemp, C. & Tenenbaum, J. B. Bayesian models of cognition. In The Cambridge Handbook of Computational Psychology (Cambridge University Press, 2008).

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

    • cites [26] Anderson, J. R. The adaptive nature of human categorization. Psychol. Rev. 98, 409 (1991). source

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

    Study design and conductnone given

    All methods were carried out and results were reported following STROBE guidelines and regulations.

    “All methods were carried out, and results were reported following STROBE guidelines and regulations.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    At least two weeks passed between the evaluation of the same batch in S1 and S2 to avoid memory effects.

    “At least two weeks passed from the conclusion of the evaluation of one batch in S1 and the evaluation of the same batch in S2 to avoid memory effects.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductown data

    The four sets of odds ratios were obtained from logistic regression models that account for lesion- and endoscopist-specific random effects.

    “The above four sets of odds ratios are obtained as the result of logistic regression models that account for lesion- and endoscopist- specific random effects.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    Endoscopists could not fully escape the negative consequences of wrong AI advice, but this was more than compensated by their stronger tendency to accept correct AI advice.

    “In other words, endoscopists could not fully escape from the negative consequences of a wrong ai advice (i.e., is less than 1), but this was more than compensated by their stronger tendency to accept a correct ai advice.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    Belief revision by the endoscopists turned out to be a sound practice in Bayesian terms, resulting in overall increased accuracy of the human-ai hybrid team.

    “In this way, belief revision turned out to be a sound practice in Bayesian terms, resulting in overall increased accuracy of the human–ai hybrid team.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: A Bayesian decision-maker that weighs each opinion over its estimated reliability is an appropriate normative model for how endoscopists should integrate their own opinion with the ai's.

  • May combine two claims.

    Interpretationreasoning

    When the enabling conditions (ai provides an intuitive yet valid clue to its reliability) are met, hybrid decision-making is an effective and appropriate diagnostic approach.

    “When these enabling conditions are met, hybrid decision-making is an effective and appropriate diagnostic approach.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationnone given

    The results could generalize to different medical settings.

    “Our results could generalize to different medical settings.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

    Premisecitation

    In many scenarios ai cannot (or should not) substitute the human professional.

    “in many scenarios ai cannot (or should not) substitute the human professional1,36.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [1] Topol, E. J. High-performance medicine: The convergence of human and artificial intelligence. Nat. Med. 25, 44–56 (2019). source
    • cites [36] WHO. Ethics and Governance of Artificial Intelligence for Health: WHO Guidance (World Health Organization, 2021).

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

31 of 31

  • The paper relies on this without support.

    No previous experimental study addressed the inner dynamics of AI-supported MDs' beliefs revision.

    “To our knowledge, no previous experimental study addressed the inner dynamics of ai-supported mds’ beliefs revision.”

  • The paper relies on this without support.

    The critical parameters collected (MDs' opinion, confidence, interpretation of AI output, and perceived reliability of each AI advice) were ignored in previous studies on AI medical devices.

    “These parameters were ignored in previous studies on ai medical devices.”

  • The paper relies on this without support.

    A Bayesian decision-maker that weighs each opinion over its estimated reliability is an appropriate normative model for how endoscopists should integrate their own opinion with the ai's.

    “An intuitive Bayesian decision-maker would weigh each opinion over its estimated reliability.”

  • The paper relies on this without support.

    Endoscopists' optical diagnosis is a psychological categorization process in which confidence is revised in a Bayesian-like manner.

    “Endoscopists’ optical diagnosis is a psychological categorization process.”
4 of 4

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?

  • Previous generations of non-LLM AI systems can improve human management decisions in some situations, especially when a human user treats an AI suggestion as a second opinion.

    “Previous generations of non-LLM AI systems can improve human management decisions in some situations, especially when a human user treats an AI suggestion as a second opinion.14”

    From Large Language Model Influence on Management Reasoning: A Randomized Controlled Trial (2024) · cites this paper as [14]

    The Discussion (page unavailable) reports that endoscopists using AI advice achieved better diagnostic accuracy than endoscopists alone. The Introduction describes treating AI output as an additional informed opinion when revising an initial diagnosis. The study concerns medical diagnoses during colonoscopy, however, not management decisions more generally. (LLM, conf=0.85)

    Experimental — accuracy not yet measured

  • The effect of AI advice on experts was smaller but still statistically significant due to the study's high power.

    “The effect on experts was smaller, but it still held significance thanks to the study's high power.”

    From Enhancing human-AI collaboration: The case of colonoscopy (2024) · cites this paper as [66]

    The Results (Effects of expertise, Table 5; page unavailable) show a smaller AI effect on experts than non-experts, while the experts’ diagnostic-accuracy odds ratio is 1.15 with a 95% confidence interval of 1.01–1.30. The Methods (Study design and participants; page unavailable) say the sample size provided adequate power for the main endpoints, but do not establish that high power explains the significant result in experts. (LLM, conf=0.85)

    Experimental — accuracy not yet measured

  • In that study, all endoscopists were more likely to accept AI advice when it was correct than when it was wrong.

    “Crucially, all endoscopists were more likely to accept AI advice when it was correct than when it was wrong.”

    From Enhancing human-AI collaboration: The case of colonoscopy (2024) · cites this paper as [66]

    The Results state that endoscopists followed correct AI advice more often than incorrect advice. That finding is reported for the group; the passage does not establish that every individual endoscopist showed this pattern, as “all endoscopists” claims. (LLM, conf=0.86)

    Experimental — accuracy not yet measured

  • Performance of available CADx systems is comparable, if not inferior, to that of experienced endoscopists.

    “Furthermore, the performance of available CADx systems is comparable, if not inferior, to that of experienced endoscopists [66 , 94-96] .”

    From Enhancing human-AI collaboration: The case of colonoscopy (2024) · cites this paper as [66]

    The Results section (‘Effects of expertise’; page unavailable) says expert endoscopists’ accuracy was close to the AI’s accuracy, supporting comparability for the system studied. The paper does not establish that available CADx systems generally perform comparably to, or worse than, experienced endoscopists. (LLM, conf=0.86)

    Experimental — accuracy not yet measured

  • Such intuitive assessment predicted AI accuracy on the specific lesion, allowing endoscopists to accept only good advice.

    “Such intuitive assessment predicted AI accuracy on the specific lesion, thus allowing endoscopists to accept only good advice [66] .”

    From Enhancing human-AI collaboration: The case of colonoscopy (2024) · cites this paper as [66]

    The Discussion says endoscopists could predict AI accuracy for each lesion, and the Results say their reliability estimates predicted AI accuracy. But the Results also say they followed correct advice more often than incorrect advice, not that they accepted only good advice; they could not fully avoid the effects of wrong advice. (LLM, conf=0.88)

    Experimental — accuracy not yet measured

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.

  • Perceived AI diagnosis has three response options in the procedure but four in Table 1

    Minor

    The procedure specifies three forced-choice responses for perceived AI diagnosis, whereas Table 1 lists an additional "I am not sure/I did not notice" response for that variable.

  • Two differently defined accuracy variables share the same name

    Minor

    Table 1 assigns "Accuracy of the perceived ai diagnosis" to both a measure that excludes uncertain and unnoticed outputs and a measure that counts those outputs as errors.

2 of 2

Cites

Flagged first, then by citations

Research works printed but not linked: 1

  1. Griffiths, T. L., Kemp, C. & Tenenbaum, J. B. Bayesian models of cognition. In The Cambridge Handbook of Computational Psychology (Cambridge University Press, 2008).

Cited by

Sorted by citations

Related

Authors
Reverberi, C., Rigon, T., Solari, A., Hassan, C., Cherubini, P., GI Genius CADx Study Group, Antonelli, G., Awadie, H., Bernhofer, S., Carballal, S., Dinis-Ribeiro, M., Fernández-Clotett, A., Esparrach, G. F., Gralnek, I., Higasa, Y., Hirabayashi, T., Hirai, T., Iwatate, M., Kawano, M., Mader, M., Maieron, A., Mattes, S., Nakai, T., Ordas, I., Ortigão, R., Zúñiga, O. O., Pellisé, M., Pinto, C., Riedl, F., Sánchez, A., Steiner, E., Tanaka, Y., & Cherubini, A.
Venue
Scientific Reports
Publisher
Springer Science and Business Media LLC
Published
2022-09-02
Volume
12
Issue
1
Pages
14952-14952
ISSN
2045-2322 (ISSN-L)
DOI
10.1038/s41598-022-18751-2
PMC
PMC9440124
OpenAlex
W4294243134
Semantic Scholar
1bfc69cd9a06be740ea6a0f421e0852a90856220
Cited by
235
References
46
Citations by year
2026: 70 · 2025: 86 · 2024: 47 · 2023: 30 · 2022: 1
FWCI
8.97
Type
journal article
Language
en
Open access
Licence
Locations
  • Scientific Reports (publishedVersion, cc-by)
  • Research Square (acceptedVersion, cc-by)
  • PubMed (publishedVersion)
  • PubMed Central (submittedVersion, cc-by)
  • BOA (University of Milano-Bicocca) (submittedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • ARCA (Università Ca' Foscari Venezia) (submittedVersion, cc-by-nc-nd)
Topics
  • Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
  • Social Sciences › Psychology › Social Psychology › Human-Automation Interaction and Safety
  • Health Sciences › Medicine › Family Practice › Clinical Reasoning and Diagnostic Skills
Keywords
human-AI collaboration, clinical decision-making, colonoscopy, endoscopists, AI-assisted diagnosis, lesion classification, trust in AI, Bayesian integration, reliability estimation, deep learning
MeSH terms
Clinical Decision-Making, Artificial Intelligence, Bayes Theorem, Humans

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 2026-10-09.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-09.

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