CoChat
Sign Up
Cited by 27Colorectal Cancer Screening and DetectionArtificial Intelligence in Healthcare and EducationAI in cancer detection

Appropriate trust in artificial intelligence for the optical diagnosis of colorectal polyps: the role of human/artificial intelligence interaction

Open access
van der Zander, Quirine E.W., Roumans, Rachel, Kusters, Carolus H.J., Dehghani, Nikoo, Masclee, Ad A.M., de With, Peter H.N., van der Sommen, Fons, Snijders, Chris C.P. +1 more·2024·Gastrointestinal Endoscopy
DOI 10.1016/j.gie.2024.06.029
Open this paper in CoChat — free accountOpen PDF

Abstract

BACKGROUND AND AIMS: Computer-aided diagnosis (CADx) for the optical diagnosis of colorectal polyps is thoroughly investigated. However, studies on human-artificial intelligence interaction are lacking. Our aim was to investigate endoscopists' trust in CADx by evaluating whether communicating a calibrated algorithm confidence score improved trust. METHODS: Endoscopists optically diagnosed 60 colorectal polyps. Initially, endoscopists diagnosed the polyps without CADx assistance (initial diagnosis). Immediately afterward, the same polyp was again shown with a CADx prediction: either only a prediction (benign or premalignant) or a prediction accompanied by a calibrated confidence score (0-100). A confidence score of 0 indicated a benign prediction, 100 a (pre)malignant prediction. In half of the polyps, CADx was mandatory, and for the other half, CADx was optional. After reviewing the CADx prediction, endoscopists made a final diagnosis. Histopathology was used as the reference standard. Endoscopists' trust in CADx was measured as CADx prediction utilization: the willingness to follow CADx predictions when the endoscopists initially disagreed with the CADx prediction. RESULTS: Twenty-three endoscopists participated. Presenting CADx predictions increased the endoscopists' diagnostic accuracy (69.3% initial vs 76.6% final diagnosis, P < .001). The CADx prediction was used in 36.5% (n = 183 of 501) of disagreements. Adding a confidence score led to lower CADx prediction utilization, except when the confidence score surpassed 60. Mandatory CADx decreased CADx prediction utilization compared to optional CADx. Appropriate trust-using correct or disregarding incorrect CADx predictions-was 48.7% (n = 244 of 501). CONCLUSIONS: Appropriate trust was common, and CADx prediction utilization was highest for the optional CADx without confidence scores. These results express the importance of a better understanding of human-artificial intelligence interaction.

AI snapshot

Written from the paper’s full text.

TL;DR

In a 2×2 within-subjects study, 23 endoscopists diagnosed 60 colorectal polyps with four CADx communication variants; CADx predictions improved diagnostic accuracy from 69.3% to 76.6%, but appropriate trust and undertrust were both common. CADx prediction utilization was highest for optional CADx without confidence scores, except when the confidence score was high.

Key findings

  • Presenting CADx predictions significantly increased endoscopists’ diagnostic accuracy from 69.3% initial to 76.6% final diagnosis (P < .001).
  • Among 501 disagreements between initial endoscopist diagnosis and CADx, CADx prediction was used in 36.5% (183/501). Appropriate trust was 48.7% (244/501), undertrust 43.5% (218/501), and overtrust 7.8% (39/501).
  • CADx prediction utilization was higher for optional CADx (41.3%) than mandatory CADx (32.7%, P = .049), and higher without confidence scores (40.5%) than with confidence scores (32.5%, P = .065); it was highest for optional CADx without confidence scores (49.1%).
  • Adding a confidence score lowered utilization except when the score was >60; endoscopists were more likely to switch from benign to premalignant (56.0%) than from premalignant to benign (26.7%). Confidence increased with CADx agreement (+0.4) and decreased with disagreement, more so when CADx was not used (−0.7 vs −0.4, P < .001).

Methods

  • Prospective online study using a 2×2 within-subjects repeated-measures design: 23 endoscopists (6 gastroenterology fellows, 17 gastroenterologists) diagnosed 60 colorectal polyps, with a pretest of 30 polyps without CADx.
  • Four algorithm variants were presented: mandatory CADx with or without a calibrated confidence score, and optional CADx with or without a calibrated confidence score. Histopathology was the reference standard; CADx prediction utilization was defined as following CADx when the endoscopist initially disagreed.
  • Analyses included diagnostic accuracy, prediction utilization, confidence levels, and multilevel regression to control for variance at the endoscopist level. The same underlying CADx was presented as four variants; confidence scores ranged 0–100 with threshold 50, and CADx accuracy/NPV were set at 90.0% but decreased to 86.7% because of incorrect labeling of 2 polyps.

Limitations

  • The paper states that only disagreement cases were used for prediction utilization, lowering actual power; presenting four variants as different algorithms could have caused suspicion; the online setting limits generalizability to clinical practice; HDWL images were used rather than image-enhancement modes; and selection bias may have occurred because the endoscopist sample was small and interested in AI/CADx.
  • The text also makes evident that the study used a single CADx with simulated calibration, no feedback on diagnoses, and a relatively small number of endoscopists and polyps, including only 4 sessile serrated lesions in the main test.

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

Integrity signals

Registry

Crossref · 10.1016/j.gie.2024.06.029

Cites

34 of 34 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

34 of 53 cited sources checked · 17 supported · 1 flagged · accuracy not yet measured

  • Premisecitation

    An important determinant for endoscopists whether or not to follow CADx predictions is endoscopists' trust in the validity of CADx predictions.

    “An important determinant for endoscopists whether or not to follow CADx predictions is endoscopists' trust in the validity of CADx predictions.8”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [8] Jorritsma, 2015, Improving the radiologist-CAD interaction: designing for appropriate trust, Clin Radiol source

      The Radiologist–CAD interaction section (p. 3) says that people who trust an automated aid more are more likely to rely on its decisions. The Discussion (p. 7) applies this relationship to radiologists using CAD. The paper does not address endoscopists or CADx predictions, so the claim extends its conclusion to a different clinical setting. (LLM, conf=0.88)

      Experimental — accuracy not yet measured
  • Findingcitation

    Having a negative attitude toward AI significantly decreased CADx prediction utilization, consistent with literature.

    “In contrast, having a negative attitude toward AI significantly decreased CADx prediction utilization, consistent with literature.7,15”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [7] Rezazade Mehrizi, 2023, The impact of AI suggestions on radiologists' decisions: a pilot study of explainability and attitudinal priming interventions in mammography examination, Sci Rep source

      The Introduction (Attitude towards AI; page unavailable) says a strong negative attitude can cause under-reliance on AI suggestions, supporting the general direction of the claim. It does not establish a statistically significant decrease in CADx prediction utilization; the Findings (The moderating effect of attitudinal priming; page unavailable) report no significant difference across the four priming groups in diagnosis errors, a different outcome. (LLM, conf=0.84)

      Experimental — accuracy not yet measured
    • cites [15] Logg, 2019, Algorithm appreciation: people prefer algorithmic to human judgment, Org Behav Hum Decis Process source
  • Premisecitation

    Commercially available CADx systems communicate by textual or colored predictions, prediction probabilities, confidence scores or intervals, and/or position maps of suspicious areas, without justification as to why a specific method is chosen.

    “Commercially available CADx systems communicate by textual or colored predictions, prediction probabilities, confidence scores or intervals, and/or position maps of suspicious areas, without justification as to why a specific method is chosen.18,19”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [18] Byrne, 2019, Real-time differentiation of adenomatous and hyperplastic diminutive colorectal polyps during analysis of unaltered videos of standard colonoscopy using a deep learning model, Gut source

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

      Experimental — accuracy not yet measured
    • cites [19] 2020

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

  • Premisecitation

    How CADx communicates its predictions with endoscopists is important in developing trust.

    “How CADx communicates its predictions with endoscopists is important in developing trust.13”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [13] Siau, 2018, Building trust in artificial intelligence, machine learning, and robotics, Cut Bus Technol J

      In “Practical Implications and Conclusions” (page unavailable), the paper says AI systems that interface well with humans and explain the rationale behind their conclusions facilitate trust. It does not address CADx, endoscopists, or communication of CADx predictions specifically. (LLM, conf=0.84)

      Experimental — accuracy not yet measured
  • Premisecitation

    Seeing an algorithm making an obvious error negatively influences prediction utilization.

    “Furthermore, seeing an algorithm making an obvious error negatively influences prediction utilization, 28”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [28] Dietvorst, 2015, Algorithm aversion: people erroneously avoid algorithms after seeing them err, J Exp Psychol Gen source
  • Classified as probably not a substantial claim. 35
  • May combine two claims. 35
  • The quote only partly states this claim. 3

  • Classified as probably not a substantial claim.

    Study design and conductcitation

    Trust was measured as CADx prediction utilization, defined as an endoscopist's willingness to follow the CADx prediction when the endoscopist initially disagreed with the CADx prediction.

    “Trust was measured as CADx prediction utilization, defined as an endoscopist's willingness to follow the CADx prediction when the endoscopist initially disagreed with the CADx prediction 9,10”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [9] Goodwin, 2013, Antecedents and effects of trust in forecasting advice, Int J Forecast source

      The Literature review (p. 3) explicitly distinguishes stated trust from subsequent behavior rather than defining trust as willingness to follow advice. The paper measured trust with a five-point rating (section 4.3.1, p. 5) and studied adjustments to forecasting advice, not CADx predictions or endoscopists. (LLM, conf=0.87)

      Checked by a reasoning model — may be wrong.

      Experimental — accuracy not yet measured
    • cites [10] Bonaccio, 2006, Advice taking and decision-making: an integrative literature review, and implications for the organizational sciences, Org Behav Hum Decis Process source

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

    Premisecitation

    Human characteristics are known to affect the willingness to follow algorithm advice.

    “Human characteristics are known to affect the willingness to follow algorithm advice.”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites [15] Logg, 2019, Algorithm appreciation: people prefer algorithmic to human judgment, Org Behav Hum Decis Process source
    • cites [28] Dietvorst, 2015, Algorithm aversion: people erroneously avoid algorithms after seeing them err, J Exp Psychol Gen source

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

    Premisecitation

    Providing information that explains how algorithms come to a prediction, thereby increasing transparency, is thought to support the critical interpretation of predictions, providing decision support and facilitating appropriate trust.

    “Providing information that explains how algorithms come to a prediction, thereby increasing transparency, is thought to support the critical interpretation of predictions, providing decision support and facilitating appropriate trust.7,17”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites [7] Rezazade Mehrizi, 2023, The impact of AI suggestions on radiologists' decisions: a pilot study of explainability and attitudinal priming interventions in mammography examination, Sci Rep source
    • cites [17] Benda, 2021, Trust in AI: why we should be designing for APPROPRIATE reliance, J Am Med Inform Assoc source

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

    Premisecitation

    Previously, it was shown that CADx affected nonexperts more than experts.

    “Previously, it was shown that CADx affected nonexperts more than experts.27”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [27] Reverberi, 2022, Experimental evidence of effective human-AI collaboration in medical decision-making, Sci Rep source

  • May combine two claims.

    Premisecitation

    Deep neural networks are generally poorly calibrated and produce overconfident results.

    “Deep neural networks are generally poorly calibrated and produce overconfident results.22”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [22] Guo, 2017, On calibration of modern neural networks source

  • May combine two claims.

    Interpretationcitation

    These results are in concordance with previous results and emphasize that endoscopists take the CADx prediction into consideration, even when they might not change their eventual diagnosis.

    “These results are in concordance with previous results and emphasize that endoscopists take the CADx prediction into consideration (even when they might not change their eventual diagnosis).27”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [27] Reverberi, 2022, Experimental evidence of effective human-AI collaboration in medical decision-making, Sci Rep source

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

    Premisecitation

    No universal definition for expertise or experts versus nonexperts exists, which might be a reason for the differences in results.

    “No universal definition for expertise or experts versus nonexperts exists, which might be a reason for the differences in results.20”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [20] van der Zander, 2023, Assessing the level of expertise of endoscopists in optical diagnosis of colorectal polyps-not every expert is an expert, Gastroenterology source

      In the Dear Editors section (page unavailable), the paper says that definitions of expert and nonexpert endoscopists are not standardized, supporting the first part of the claim. It notes variability in optical diagnosis, but does not say that differing definitions explain differences in results. (LLM, conf=0.82)

      Experimental — accuracy not yet measured

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

    Premisecitation

    Earlier research has mentioned gender, self-confidence, expertise, and general algorithm aversion or appreciation.

    “Earlier research has mentioned gender, self-confidence, expertise, and general algorithm aversion or appreciation.15,28”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites [15] Logg, 2019, Algorithm appreciation: people prefer algorithmic to human judgment, Org Behav Hum Decis Process source
    • cites [28] Dietvorst, 2015, Algorithm aversion: people erroneously avoid algorithms after seeing them err, J Exp Psychol Gen source

      The introduction (p. 1) discusses general algorithm aversion, and a note in Participants (p. 2) mentions prior research on domain expertise. The Confidence section (p. 8) concerns confidence in forecasts, not self-confidence. The passages do not support attributing gender or algorithm appreciation to this paper. (LLM, conf=0.86)

      Experimental — accuracy not yet measured

  • Classified as probably not a substantial claim.

    Premisecitation

    Appropriate trust implies minimizing the negative performance outcomes through either overtrust or undertrust.

    “Hence, appropriate trust implies minimizing the negative performance outcomes through either overtrust or undertrust.8,9”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites [8] Jorritsma, 2015, Improving the radiologist-CAD interaction: designing for appropriate trust, Clin Radiol source
    • cites [9] Goodwin, 2013, Antecedents and effects of trust in forecasting advice, Int J Forecast source

      The performance results (sections 4.3.3, p. 7, and 5.2.3, p. 10) show that insufficient trust in reliable forecasting advice led participants to adjust forecasts and worsen their performance. The literature review (p. 3) also notes a danger of too much trust in systems that may give erroneous advice, but the paper does not establish negative performance outcomes from overtrust. The claim's two-sided performance assertion is therefore broader than the evidence. (LLM, conf=0.82)

      Experimental — accuracy not yet measured

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

    Premisecitation

    Earlier studies have also shown tendencies toward underuse of algorithms.

    “Earlier studies have also shown tendencies toward underuse of algorithms.25,26”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites [25] Wiczorek, 2019, Effects of trust, self-confidence, and feedback on the use of decision automation, Front Psychol source
    • cites [26] Prahl, 2017, Understanding algorithm aversion: when is advice from automation discounted?, J Forecast source

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

    Premisecitation

    Within radiology, it was suggested but not proven that higher levels of appropriate trust can be reached by adding confidence scores to predictions.

    “Within radiology, it was suggested but not proven that higher levels of appropriate trust can be reached by adding confidence scores to predictions.8”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [8] Jorritsma, 2015, Improving the radiologist-CAD interaction: designing for appropriate trust, Clin Radiol source

  • May combine two claims.

    Premisecitation

    Numerous variables influence prediction utilization, including technological characteristics such as transparency (explainable AI), usability, and interpretability, but also the stated and perceived CADx accuracy.

    “Numerous variables influence prediction utilization.These include technological characteristics such as transparency (explainable AI), usability, and interpretability [13][14][15] but also the stated and perceived CADx accuracy.16,17”
    Quote in the paperSubstantial claim2 of 5 sources checked
    • cites [13] Siau, 2018, Building trust in artificial intelligence, machine learning, and robotics, Cut Bus Technol J

      The sections ‘Building Initial Trust in AI’ and ‘Developing Continuous Trust in AI’ identify transparency, explainability, usability, and interpretability as factors affecting trust in AI. The paper links trust to AI acceptance, but does not establish that these features affect utilization of predictions specifically. The CADx-accuracy clause has separate citations and is not attributed to this source. (LLM, conf=0.85)

      Experimental — accuracy not yet measured
    • cites [14] Kizilcec, 2016

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

    • cites [15] Logg, 2019, Algorithm appreciation: people prefer algorithmic to human judgment, Org Behav Hum Decis Process source
    • cites [16] Yin, 2019

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

    • cites [17] Benda, 2021, Trust in AI: why we should be designing for APPROPRIATE reliance, J Am Med Inform Assoc source

  • May combine two claims.

    Premisecitation

    Overreliance may lead to blind trust (overtrust) or overutilization.

    “whereas overreliance may lead to blind trust (overtrust) or overutilization. 11,12”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [11] Parasuraman, 1997, Humans and automation: use, misuse, disuse, abuse, Hum Factors source
    • cites [12] Lee, 2012, Computer-aided detection of malignant lung nodules on chest radiographs: effect on observers’ performance, Korean J Radiol source

      The Results section (page unavailable) reports that one resident rejected many true nodules CAD did not detect and that readers accepted false-positive CAD marks. These findings support a concern about trusting CAD too much, but the paper does not address overutilization as a consequence of overreliance. (LLM, conf=0.78)

      Experimental — accuracy not yet measured

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

    Premisecitation

    The stand-alone performance of CADx systems and comparisons of CADx performances with thresholds for clinical implementation have been evaluated and show promising results.

    “The stand-alone performance of CADx systems and comparisons of CADx performances with thresholds for clinical implementation have been evaluated and show promising results.[1][2][3][4][5][6][7]”
    Quote in the paperNot a substantial claim5 of 7 sources checked
    • cites [1] Rex, 2011, The American Society for Gastrointestinal Endoscopy PIVI (Preservation and Incorporation of Valuable Endoscopic Innovations) on real-time endoscopic assessment of the histology of diminutive colorectal polyps, Gastrointest Endosc source
    • cites [2] Abu Dayyeh, 2015, 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 source

      The Discussion (page unavailable in our copy) reports promising results for real-time optical biopsy using NBI, i-SCAN and FICE against ASGE PIVI thresholds, particularly with expert operators. But the paper evaluates endoscopist assessments with imaging technologies, not the stand-alone performance of CADx systems; that part of the claim is unsupported. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
    • cites [3] Zachariah, 2020, Prediction of polyp pathology using convolutional neural networks achieves “resect and discard” thresholds, Am J Gastroenterol source
    • cites [4] Biffi, 2022, A novel AI device for real-time optical characterization of colorectal polyps, NPJ Digit Med source

      The Results (Performance comparisons; page unavailable) report promising stand-alone CADx performance: white-light accuracy was non-inferior to expert endoscopists and superior to non-experts. The Methods (Statistical analysis; page unavailable) specify thresholds for those study comparisons, but our copy does not establish that CADx performance was compared with thresholds for clinical implementation. (LLM, conf=0.85)

      Experimental — accuracy not yet measured
    • cites [5] Mori, 2018, Real-time use of artificial intelligence in identification of diminutive polyps during colonoscopy: a prospective study, Ann Intern Med source
    • cites [6] Kader, 2023, Polyp characterization using deep learning and a publicly accessible polyp video database, Dig Endosc source
    • cites [7] Rezazade Mehrizi, 2023, The impact of AI suggestions on radiologists' decisions: a pilot study of explainability and attitudinal priming interventions in mammography examination, Sci Rep source

      The Introduction discusses how CAD assistance affects radiologists’ decisions, and the Methods assess radiologists’ decisions in response to AI suggestions. Searches for stand-alone CADx performance and clinical-implementation thresholds found no passage addressing those comparisons or reporting promising results for them; no page numbers were available in our copy. (LLM, conf=0.76)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Findingown data

    Because of incorrect labeling of 2 polyps, diagnostic accuracy and NPV decreased to 86.7%, and the confidence score distribution was not completely uniform.

    “Because of incorrect labeling of 2 polyps, diagnostic accuracy and NPV decreased to 86.7%, and the confidence score distribution was not completely uniform.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Our multilevel analysis showed that the diagnostic accuracy in the pretest and being either a gastroenterologist or a fellow had no influence on CADx prediction utilization.

    “Our multilevel analysis showed that the diagnostic accuracy in the pretest and being either a gastroenterologist or a fellow had no influence on CADx prediction utilization.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    In the study, both appropriate trust and undertrust were common among endoscopists' optical diagnoses of colorectal polyps.

    “In conclusion, both appropriate trust and undertrust were common.”
    • Also stated in: § APPENDIX 3. ENDOSCOPISTS' PERCEPTIONS OF AI · plain
      “7.8% (39 of 501) overtrust, 43.5% (218 of 501) undertrust, and 48.7% (244 of 501) appropriate trust.”
    Quote in the paperSubstantial claim

  • The quote only partly states this claim.

    Findingnone given

    CADx prediction utilization was highest for the optional CADx without confidence scores.

    “CADx prediction utilization was highest for the optional CADx without confidence scores, except when the confidence score was !60.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    For colorectal polyps in which the CADx prediction was in agreement with the endoscopists' diagnoses, the mean difference in endoscopists' confidence level was +0.4 (SD 0.6), meaning that the confidence level increased.

    “For colorectal polyps in which the CADx prediction was in agreement with the endoscopists' diagnoses, the mean difference in endoscopists' confidence level was þ0.4 (SD, 0.6), meaning that the confidence level increased.”
    Quote in the paperSubstantial claim

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

    Findingnone given

    Six gastroenterology fellows and 17 gastroenterologists (N=23) participated.

    “Six gastroenterology fellows and 17 gastroenterologists (N Z 23) participated.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Including confidence scores smaller than 60 had a significant negative effect on CADx prediction utilization, whereas a CADx confidence score of >60 had a significant and substantial positive effect.

    “Including confidence scores smaller than 60 had a significant negative effect on CADx prediction utilization, whereas a CADx confidence score of !60 had a significant and substantial positive effect.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Endoscopists' mean confidence level was 4.1 (SD 0.9) for the initial diagnosis and 4.0 (SD 0.9) for the final diagnosis.

    “Endoscopists' mean confidence level was 4.1 (SD, 0.9) for the initial diagnosis and 4.0 (SD, 0.9) for the final diagnosis (Table 3).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    In this study, although appropriate trust was higher compared to undertrust (48.7% vs 43.5%), undertrusting the CADx prediction was substantial.

    “In this study, although appropriate trust was higher compared to undertrust (48.7% vs 43.5%, respectively), undertrusting the CADx prediction was substantial.”
    Quote in the paperSubstantial claim

  • May combine two claims. The quote only partly states this claim.

    Findingown data

    The CADx prediction was used in 183 colorectal polyps, corresponding to 36.5% of disagreements.

    “The CADx prediction was used in 183 colorectal polyps.”
    • Also stated in: § DISCUSSION · plain
      “The CADx prediction changed the endoscopists' initial opinions in 36.5% of the cases where there was initial disagreement, representing moderate endoscopist trust in CADx predictions.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Both appropriate trust and undertrust were common (48.7% vs 43.5%, respectively).

    “Both appropriate trust (using a correct CADx prediction or disregarding an incorrect prediction by endoscopists) and undertrust were common (48.7% vs 43.5%, respectively).”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Findingown data

    Multilevel analyses taking several covariates into account corroborate the gist of these findings.

    “Multilevel analyses taking several covariates into account corroborate the gist of these findings.”
    Quote in the paperSubstantial claim

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

    Findingown data

    Only 4 of 60 colorectal polyps were sessile serrated lesions.

    “It should be noted that only 4 of 60 colorectal polyps were SSLs and that the CADx predicted all SSLs correctly as being premalignant.”
    Quote in the paperNot a substantial claim

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

    Findingown data

    Ten endoscopists indicated that they had experience with AI before participation in this study.

    “Ten endoscopists indicated that they had experience with AI before participation in this study, of whom 5 had worked with computer-aided detection systems and CADx systems before and 5 solely with computer-aided detection.”
    Quote in the paperNot a substantial claim

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

    Findingown data

    Of the ten endoscopists with AI experience, 5 had worked with computer-aided detection systems and CADx systems before and 5 solely with computer-aided detection.

    “Ten endoscopists indicated that they had experience with AI before participation in this study, of whom 5 had worked with computer-aided detection systems and CADx systems before and 5 solely with computer-aided detection.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    For the optional CADx: 10.3% overtrust, 37.7% undertrust, 52.0% appropriate trust, and 41.3% CADx prediction utilization.

    “B, For the optional CADx: 10.3% (23 of 223) overtrust, 37.7% (84 of 223) undertrust, 52.0% (116 of 223) appropriate trust, and 41.3% CADx prediction utilization.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    For CADx without confidence scores: 9.5% overtrust, 41.3% undertrust, 49.2% appropriate trust, and 40.5% CADx prediction utilization.

    “A, For CADx without confidence scores: 9.5% (24 of 252) overtrust, 41.3% (104 of 252) undertrust, 49.2% (124 of 252) appropriate trust, and 40.5% CADx prediction utilization.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    For the mandatory CADx: 5.8% overtrust, 48.2% undertrust, 46.0% appropriate trust, and 32.7% CADx prediction utilization.

    “A, For the mandatory CADx: 5.8% (16 of 278) overtrust, 48.2% (134 of 278) undertrust, 46.0% (128 of 278) appropriate trust, and 32.7% CADx prediction utilization.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    For CADx with confidence scores: 6.0% overtrust, 45.8% undertrust, 48.2% appropriate trust, and 32.5% CADx prediction utilization.

    “B, For CADx with confidence scores: 6.0% (15 of 249) overtrust, 45.8% (114 of 249) undertrust, 48.2% (120 of 249) appropriate trust, and 32.5% CADx prediction utilization.”
    Quote in the paperSubstantial claim

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

    Findingown data

    Two endoscopists indicated they felt more likely to use CADx predictions from the optional CADx because they specifically asked for the CADx prediction themselves because their own confidence levels were low in these cases.

    “Two endoscopists indicated they felt more likely to use CADx predictions from the optional CADx because they specifically asked for the CADx prediction themselves because their own confidence levels were low in these cases.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Unsolicited advice can be received poorly and considered intrusive, whereas explicitly solicited advice can be considered cooperative and helpful.

    “Unsolicited advice can be received poorly and considered intrusive, whereas explicitly solicited advice can be considered cooperative and helpful.10”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [10] Bonaccio, 2006, Advice taking and decision-making: an integrative literature review, and implications for the organizational sciences, Org Behav Hum Decis Process source

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    A sample size calculation was performed aiming to detect a 10% difference in CADx prediction utilization between different algorithm variants.

    “A sample size calculation was performed aiming to detect a 10% difference in CADx prediction utilization between different algorithm variants.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    Trust was highest for the optional CADx without confidence scores unless the confidence score was >60.

    “Our results show that trust was highest for the optional CADx without confidence scores unless the confidence score was !60.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    Histopathology was used as the reference standard.

    “Histopathology was used as the reference standard.”
    Quote in the paperSubstantial claim

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

    Study design and conductnone given

    All endoscopists signed informed consent.

    “All endoscopists signed informed consent.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductreasoning

    The 4 algorithm variants were presented to endoscopists as being different CADx algorithms to minimize possible positive or negative perceptions of one algorithm variant to another.

    “However, the 4 algorithm variants were presented to endoscopists as being different CADx algorithms to minimize possible positive or negative perceptions of one algorithm variant to another.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    The endoscopists' confidence level was significantly affected by CADx (dis)agreement, suggesting that endoscopists do take the CADx prediction into consideration, even when they do not change their diagnosis in the direction of CADx.

    “Furthermore, the endoscopists' confidence level was significantly affected by CADx (dis)agreement, further suggesting that endoscopists do take the CADx prediction into consideration, even when in the end they do not change their diagnosis in the direction of CADx.”
    Quote in the paperSubstantial claim

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

    Premisecitation

    A previous study showed higher levels of CADx-assisted diagnostic performance compared to performances of endoscopists alone.

    “This result is in line with a previous study that showed higher levels of CADx-assisted diagnostic performance compared to performances of endoscopists alone.24”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [24] Rondonotti, 2022, Artificial intelligence assisted optical diagnosis for resect and discard strategy in clinical practice: Artificial Intelligence BLI Characterization (ABC) study, Endoscopy source

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

    Study design and conductnone given

    Images were checked for quality, and images with specular highlights, those with too-dark colorectal polyp areas, or images that were out of focus were excluded.

    “Images were checked for quality, and images with specular highlights, those with too-dark colorectal polyp areas, or images that were out of focus were excluded.”
    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 conductreasoning

    Selection bias might have occurred because the number of endoscopists was relatively small, and participating endoscopists might hold special interest in AI/CADx, although it is not obvious in which way this would distort the results.

    “Selection bias might have occurred because the number of endoscopists was relatively small, and participating endoscopists might hold special interest in AI/CADx, although it is not obvious in which way this would distort the results.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductreasoning

    Because of the repeated-measures design, multilevel regression analysis was used to control for variance at the endoscopist level.

    “Because of the repeated-measures design, multilevel regression analysis was used to control for variance at the endoscopist level.”
    Quote in the paperSubstantial claim

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

    Study design and conductnone given

    The study was performed in accordance with the Declaration of Helsinki and adhered to the General Data Protection Regulation.

    “The study was performed in accordance with the Declaration of Helsinki and adhered to the General Data Protection Regulation.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    In this repeated-measures design, we corrected for variance at the individual endoscopist level by clustering data of each endoscopist in a multilevel analysis.

    “In this repeated-measures design, we corrected for variance at the individual endoscopist level by clustering data of each endoscopist in a multilevel analysis.”
    Quote in the paperSubstantial claim

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

    Study design and conductnone given

    Presenting the algorithm variants as 4 different CADx algorithms might have led to unwanted results if participants suspected that, in fact, there was only 1 algorithm used.

    “Presenting the algorithm variants as 4 different CADx algorithms might have led to unwanted results if participants suspected that, in fact, there was only 1 algorithm used.”
    Quote in the paperNot a substantial claim

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

    Study design and conductnone given

    The Ethical Review Board of Eindhoven University of Technology approved the study (ARCHIE 1394).

    “The Ethical Review Board of Eindhoven University of Technology approved the study (ARCHIE 1394).”
    Quote in the paperNot a substantial claim

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

    Study design and conductnone given

    The 4 algorithm variants and the colorectal polyps were presented in random order.

    “The 4 algorithm variants and the colorectal polyps were presented in random order.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductreasoning

    To minimize bias, a maximum of 1 colorectal polyp per patient was used.

    “To minimize bias, a maximum of 1 colorectal polyp per patient was used.”
    Quote in the paperSubstantial claim

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

    Study design and conductnone given

    HDWL images were used, whereas endoscopists might be used to using image-enhancement modes such as blue-light imaging.

    “Also, HDWL images were used, whereas endoscopists might be used to using image-enhancement modes such as blue-light imaging.”
    Quote in the paperNot a substantial claim

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

    Study design and conductnone given

    The aim of the study was not communicated with endoscopists to minimize bias; instead, the aim communicated was to investigate the use of CADx in clinical decision making.

    “The aim of the studydthat is, investigating endoscopists' trust in CADxdwas not communicated with endoscopists to minimize bias.Instead, the aim communicated was to investigate the use of CADx in clinical decision making.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    Despite a significant increase in endoscopists' performance measures after reviewing CADx predictions, the joint performance does not meet the thresholds for implementing treatment strategies based on optical diagnosis (NPV of >90%) set by the American Society for Gastrointestinal Endoscopy.

    “Despite a significant increase in endoscopists' performance measures after reviewing CADx predictions, the joint performance does not meet the thresholds for implementing treatment strategies based on optical diagnosis (NPV of >90%) set by the American Society for Gastrointestinal Endoscopy. 1”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [1] Rex, 2011, The American Society for Gastrointestinal Endoscopy PIVI (Preservation and Incorporation of Valuable Endoscopic Innovations) on real-time endoscopic assessment of the histology of diminutive colorectal polyps, Gastrointest Endosc source

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

    Study design and conductreasoning

    For colorectal polyps in which the endoscopist's initial diagnosis was equal to the CADx prediction, trust is undefined; therefore, these cases were excluded from the analysis of prediction utilization.

    “For colorectal polyps in which the endoscopist's initial diagnosis was equal to the CADx prediction, trust is undefined.Therefore, these cases were excluded from the analysis of prediction utilization.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    The joint performance (endoscopist and CADx) realized a relatively small amount of the team potential, given that the CADx was about 90% correct.

    “However, the joint performance (endoscopist and CADx) realized a relatively small amount of the team potential, given that the CADx was about 90% correct.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Study design and conductown data

    Influential data points were taken into account and resulted only in small changes in the model, except for CADx prediction utilization becoming significant for colorectal polyps being adenomas.

    “Influential data points were taken into account and resulted only in small changes in the model, except for CADx prediction utilization becoming significant for colorectal polyps being adenomas (where this was only significant for SSLs in the model with influential data points).”
    Quote in the paperNot a substantial claim

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

    Study design and conductnone given

    The study simulated a calibrated CADx by selecting a subset of cases that led to a low expected calibration error, adherence to the PIVI criteria with the NPV set at 90.0%, and the CADx diagnostic accuracy set at 90.0%.

    “In addition, we simulated a calibrated CADx by selecting a subset of cases that led to (1) a low expected calibration error, (2) adherence to the PIVI criteria with the NPV set at 90.0%, and (3) the CADx diagnostic accuracy set at 90.0%.”
    Quote in the paperSubstantial claim

  • The quote only partly states this claim.

    Study design and conductreasoning

    The CADx confidence scores were calibrated by sampling equal numbers from bins to reflect the approximate correctness likelihood of the CADx.

    “In this way, the CADx confidence scores reflected the approximate correctness likelihood of the CADx.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    The 4 different algorithm variants used were technically the same CADx, only communicating different outputs.

    “The 4 different algorithm variants used were technically the same CADx, only communicating different outputs (a confidence score or not and a mandatory or optional CADx).”
    Quote in the paperSubstantial claim
60 of 60

  • The paper relies on this without support.

    CADx has little value if the predictions are not trusted and the advice is not followed.

    “CADx has little value if the predictions are not trusted and the advice is not followed.”

  • The paper relies on this without support.

    The black-box character of AI makes it difficult to interpret its output.

    “The black-box character of AI makes it difficult to interpret its output.”

  • The paper relies on this without support.

    Overconfident results affect humans' perceptions of algorithms.

    “Overconfident results affect humans' perceptions of algorithms.”

  • The paper relies on this without support.

    It is unknown how endoscopists make use of CADx predictions.

    “Currently, it is unknown how endoscopists make use of CADx predictions.”

  • The paper relies on this without support.

    To date, no studies have investigated the effect of different communication methods on CADx prediction utilization by endoscopists.

    “To date, no studies have investigated the effect of these different communication methods on CADx prediction utilization by endoscopists.”

  • The paper relies on this without support.

    The role of human-AI interaction and its effect on CADx prediction utilization has not been investigated in detail yet.

    “The role of human-AI interaction and its effect on CADx prediction utilization has not been investigated in detail yet.”

  • The paper relies on this without support.

    No systematic studies have investigated the effect of different aspects of CADx algorithm communication on human-AI interaction.

    “Although numerous studies have presented promising results of CADx, no systematic studies have investigated the effect of different aspects of the CADx algorithm communication on human-AI interaction.”
7 of 7

Cites

Flagged first, then by citations

Cited by

Sorted by citations

Related

Authors
van der Zander, Q. E. W., Roumans, R., Kusters, C. H. J., Dehghani, N., Masclee, A. A. M., de With, P. H. N., van der Sommen, F., Snijders, C. C. P., & Schoon, E. J.
Affiliations
  • Eindhoven University of Technology (NL)
  • Maastricht University Medical Centre (NL)
ORCID
Venue
Gastrointestinal Endoscopy
Publisher
Elsevier BV
Published
2024-12
Volume
100
Issue
6
Pages
1070-1078.e10
ISSN
0016-5107 (ISSN-L)
DOI
10.1016/j.gie.2024.06.029
OpenAlex
W4400039194
Cited by
27
References
30
Citations by year
2026: 9 · 2025: 15 · 2024: 3
FWCI
4.59
Type
journal article
Language
en
Open access
Licence
Locations
  • Gastrointestinal Endoscopy (publishedVersion, cc-by-nc-nd)
  • PubMed (publishedVersion)
  • TU/e Research Portal (acceptedVersion, cc-by-nc-nd)
  • Research Publications (Maastricht University) (submittedVersion, cc-by-nc-nd)
Funders
Dutch Cancer Society
Topics
  • Health Sciences › Medicine › Oncology › Colorectal Cancer Screening and Detection
  • Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
  • Physical Sciences › Computer Science › Artificial Intelligence › AI in cancer detection
Keywords
computer-aided diagnosis, colorectal polyps, human-AI interaction, trust in AI, optical diagnostics, polyp classification, diagnostic accuracy
MeSH terms
Algorithms, Artificial Intelligence, Colonic Polyps, Colonoscopy, Diagnosis, Computer-Assisted, Female, Humans, Male, Middle Aged, Precancerous Conditions, Colorectal Neoplasms, Trust

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 2026-11-03.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-11-03.

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