Student or educator? Sign up with your school email and get 6 months of Pro free.

The DOI matches its Crossref record; 25 of the 25 references it cites were found in a registry or an index; 19 of the 28 sources its claims cite were checked for support.

Non-task expert physicians benefit from correct explainable AI advice when reviewing X-rays

Open access

Abstract

Artificial intelligence (AI)-generated clinical advice is becoming more prevalent in healthcare. However, the impact of AI-generated advice on physicians’ decision-making is underexplored. In this study, physicians received X-rays with correct diagnostic advice and were asked to make a diagnosis, rate the advice’s quality, and judge their own confidence. We manipulated whether the advice came with or without a visual annotation on the X-rays, and whether it was labeled as coming from an AI or a human radiologist. Overall, receiving annotated advice from an AI resulted in the highest diagnostic accuracy. Physicians rated the quality of AI advice higher than human advice. We did not find a strong effect of either manipulation on participants’ confidence. The magnitude of the effects varied between task experts and non-task experts, with the latter benefiting considerably from correct explainable AI advice. These findings raise important considerations for the deployment of diagnostic advice in healthcare.

Overview

CoChat Summary

TL;DR

In an online experiment with physicians reviewing chest X-rays, correct diagnostic advice that was annotated and labeled as AI-generated was associated with higher diagnostic accuracy, especially for non-task experts. Explainability improved accuracy and quality ratings for IM/EM physicians, while AI-labeled advice increased radiologists’ confidence but not their accuracy.

Key findings

  • Overall, annotated advice and AI-labeled advice were associated with higher diagnostic accuracy: annotated OR = 2.30 (95% CI 1.37–3.85, p = 0.002); AI source OR = 2.10 (95% CI 1.22–3.59, p = 0.007); task experts performed better than non-task experts, OR = 2.20 (95% CI 1.40–3.47, p = 0.001).
  • Non-task experts (IM/EM physicians) improved mean diagnostic performance by 5.66% with annotated advice (p = 0.042); radiologists improved by 3.41% but this was non-significant (p = 0.120). AI-labeled advice improved performance by 4.22% for non-experts and 3.15% for experts, both non-significant (p = 0.129 and p = 0.155).
  • Annotated advice raised advice-quality ratings overall (estimate = 0.25, p = 0.001); among non-task experts it was 0.24 points higher (p = 0.011), while for task experts it was 0.11 points higher but non-significant (p = 0.195). Source of advice did not significantly affect quality ratings (p = 0.645 for IM/EM; p = 0.812 for radiology).
  • Radiologists reported 0.32 points higher confidence when advice was labeled as AI vs. human (p < 0.0001); IM/EM physicians’ confidence was only 0.06 points higher (p = 0.497). Highest accuracy occurred with annotated AI advice: non-task experts M = 86.84, SD = 33.89; task experts M = 94.61, SD = 22.64.

Methods

  • Pre-registered online 2 × 2 mixed factorial experiment: explainability manipulated within subjects (annotated vs. non-annotated X-rays) and source manipulated between subjects (AI “CHEST-AI” vs. human radiologist “Dr. S. Johnson”). Participants reviewed eight chest X-ray cases from MIMIC-CXR with patient vignettes and correct diagnostic advice; six cases were analyzed because two could not be unambiguously assigned to the explainability condition.
  • N = 223 physicians finished the experiment and were included: 117 IM/EM physicians classified as non-task experts and 106 radiologists classified as task experts. The mixed-effects models reported NID = 222, NPATIENTID = 6, and 1,332 observations.
  • Three mixed-effects regression models were run: logistic regression for diagnostic accuracy and linear regression for advice-quality ratings and confidence. Predictors included explainability, source, task expertise, their interaction, and control variables, with random effects for participant and patient case.

Limitations

  • The paper states: the experiment was online and diagnoses did not affect real patient treatment; only eight cases were presented and only six analyzed; the sample fell short of the a priori power target (planned 128 IM/EM and 128 radiologists; recruited 117 and 106); task-expert accuracy was high, limiting room for improvement; and prior-study participation could not be fully ruled out, though data were collected two years later.
  • The paper also states that only correct advice was provided, so it could not test how explanations affect reliance when AI advice is incorrect; the authors note that participants strongly relied on any advice and that more research is needed on incorrect advice.
  • Evident from the text: the source manipulation was between-subjects, so source effects may be confounded with individual differences, and the study used a limited set of curated cases, which narrows generalizability.

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

Integrity signals

Registry

Crossref · 10.1038/s41598-023-28633-w

Cites

25 of 25 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 Experimental setup. Every participant reviewed all eight cases. Each case consisted of a brief patient vignette, a chest X-ray, and diagnostic advice (radiologic findings and primary diagnoses). The advice came either with or without annotations on the X-ray. Additionally, the advice was labeled as coming either from an AI system or an experienced radiologist. Physicians were asked to give a final diagnosis, rate the quality of the advice, and judge how confident they were with their diagnosis.
Figure 1 Experimental setup. Every participant reviewed all eight cases. Each case consisted of a brief patient vignette, a chest X-ray, and diagnostic advice (radiologic findings and primary diagnoses). The advice came either with or without annotations on the X-ray. Additionally, the advice was labeled as coming either from an AI system or an experienced radiologist. Physicians were asked to give a final diagnosis, rate the quality of the advice, and judge how confident they were with their diagnosis.CC BY · PubMed Central
Figure 2 Dependent variables across the advice manipulations. The plots show how the advice manipulations affected non-task experts (i.e., IM/EM physicians) and task experts (i.e., radiologists). Plot (a) shows that explainable advice helped non-task experts to be more accurate on average (pIM/EM = 0.042, pRadiology = 0.120). Chart (b) indicates that the source of advice had only statistically non-significant effects on diagnostic accuracy (pIM/EM = 0.129, pRadiology = 0.155). Plot (c) displays that non-task experts rated the quality of annotated advice higher than non-annotated advice (pIM/EM = 0.011, pRadiology = 0.195). In (d), it is shown that there was no evidence that the source of advice had an effect on the quality rating (pIM/EM = 0.645, pRadiology = 0.812). Plot (e) indicates that explainability had little effect on the confidence ratings (pIM/EM = 0.280, pRadiology = 0.202). Finally, (f) shows that task experts reported higher confidence in their decision when receiving AI advice while non-task experts’ confidence was unaffected by the source (pIM/EM = 0.497, pRadiology < 0.0001). The boxplots show 25th to 75th percentiles and the median as the central line; the whiskers extend to a maximum of 1.5 × interquartile range. The error bars represent standard errors. *p ≤ 0.05, **p ≤ 0.01, ***p ≤ 0.001, ****p ≤ 0.0001, ns statistically non-significant.
Figure 2 Dependent variables across the advice manipulations. The plots show how the advice manipulations affected non-task experts (i.e., IM/EM physicians) and task experts (i.e., radiologists). Plot (a) shows that explainable advice helped non-task experts to be more accurate on average (pIM/EM = 0.042, pRadiology = 0.120). Chart (b) indicates that the source of advice had only statistically non-significant effects on diagnostic accuracy (pIM/EM = 0.129, pRadiology = 0.155). Plot (c) displays that non-task experts rated the quality of annotated advice higher than non-annotated advice (pIM/EM = 0.011, pRadiology = 0.195). In (d), it is shown that there was no evidence that the source of advice had an effect on the quality rating (pIM/EM = 0.645, pRadiology = 0.812). Plot (e) indicates that explainability had little effect on the confidence ratings (pIM/EM = 0.280, pRadiology = 0.202). Finally, (f) shows that task experts reported higher confidence in their decision when receiving AI advice while non-task experts’ confidence was unaffected by the source (pIM/EM = 0.497, pRadiology < 0.0001). The boxplots show 25th to 75th percentiles and the median as the central line; the whiskers extend to a maximum of 1.5 × interquartile range. The error bars represent standard errors. *p ≤ 0.05, **p ≤ 0.01, ***p ≤ 0.001, ****p ≤ 0.0001, ns statistically non-significant.CC BY · PubMed Central
Figure 3 Diagnostic accuracy by clinical case. Case-dependent performance amongst non-task experts (i.e., IM/EM physicians) and task experts (i.e., radiologists) across the two advice manipulations (a) explainability of the advice and (b) source of the advice. The x-axis labels are the case ID numbers (see online supplements pages 2–4 for further information about the cases). The error bars represent standard errors.
Figure 3 Diagnostic accuracy by clinical case. Case-dependent performance amongst non-task experts (i.e., IM/EM physicians) and task experts (i.e., radiologists) across the two advice manipulations (a) explainability of the advice and (b) source of the advice. The x-axis labels are the case ID numbers (see online supplements pages 2–4 for further information about the cases). The error bars represent standard errors.CC BY · PubMed Central

Verification

Claims the paper makes

14 Valid · 1 Not valid · 2 Not checked · 76 Own results and interpretations

37 more claims

Create a free account to see all 93 claims and every source verdict.

See all claims — free account
  • Probably not a substantial claim. 40
  • It may state two separate claims as one. 18
  • The quote only partly states this claim. 1

Create a free account to see these claims.

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.

1 of 1

Citations

Cites

Show all 25 references

Cited by

See all 92

Metadata

Authors
Gaube, S., Suresh, H., Raue, M., Lermer, E., Koch, T. K., Hudecek, M. F. C., Ackery, A. D., Grover, S. C., Coughlin, J. F., Frey, D., Kitamura, F. C., Ghassemi, M., & Colak, E.
Affiliations
  • University Hospital Regensburg (DE)
  • MIT Computer Science and Artificial Intelligence Laboratory (US)
  • Massachusetts Institute of Technology (US)
  • Technische Hochschule Augsburg (DE)
  • Ludwig-Maximilians-Universität München (DE)
  • University of Regensburg (DE)
  • St. Michael's Hospital (CA)
  • Universidade Federal de São Paulo (BR)
  • Vector Institute (CA)
Corresponding author
Gaube, Susanne
Venue
Scientific Reports
Published
2023-01-25
Volume
13
Issue
1
Pages
1383-1383
ISSN
2045-2322 (ISSN-L)
OpenAlex
W4317935575
Cited by
92
References
25
Citations by year
2026: 20 · 2025: 29 · 2024: 27 · 2023: 16
FWCI
3.66
Type
journal article
Language
en
Open access
Licence
Locations
  • Scientific Reports (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • UCL Discovery (University College London) (submittedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • University of Regensburg Publication Server (University of Regensburg) (submittedVersion)
Funders
  • Volkswagen Foundation (98525)
  • Ludwig-Maximilians-Universität München
Topics
  • Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
  • Physical Sciences › Computer Science › Artificial Intelligence › Explainable Artificial Intelligence (XAI)
  • Health Sciences › Medicine › Family Practice › Clinical Reasoning and Diagnostic Skills
Keywords
explainable AI, radiology, X-ray interpretation, physician decision-making, diagnostic accuracy, visual annotation, diagnostic confidence, human-AI collaboration
MeSH terms
Radiologists, Artificial Intelligence, Humans, Physicians, Radiography, X-Rays

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (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

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Non-task expert physicians benefit from correct explainable AI advice when reviewing X-rays. https://app.cochat.ai/papers/10.1038/s41598-023-28633-w
The paper
Gaube, S., Suresh, H., Raue, M., Lermer, E., Koch, T. K., Hudecek, M. F. C., Ackery, A. D., Grover, S. C., Coughlin, J. F., Frey, D., Kitamura, F. C., Ghassemi, M., & Colak, E. (2023). Non-task expert physicians benefit from correct explainable AI advice when reviewing X-rays. Scientific Reports. https://doi.org/10.1038/s41598-023-28633-w

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.