Pith. sign in

REVIEW 3 major objections 5 minor 77 references

Visual-Conversational Interface for Evidence-Based Explanation of Diabetes Risk Prediction

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read A conversational chatbot plus cited evidence helped clinicians understand and trust AI diabetes risk predictions.

desk verdict A solid, open-sourced HCI systems paper with a real user study, but its central trust-calibration claim is weakened by a percentile window that was tuned to match the medical ranges it later confirms. read the letter →

arxiv 2507.02920 v1 pith:TTX7SDOW submitted 2025-06-25 cs.HC cs.AIcs.LG

classification cs.HCcs.AIcs.LG
keywords clinicaldecisionsupportsystemsexplainableAIconversationaluserinterfacesdiabetesriskpredictiontrustcalibrationscientificevidencegroundingfeaturerangeanalysismixed-methodsstudy
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that healthcare professionals can understand and validate AI-based diabetes risk predictions when a decision-support dashboard pairs interactive visualizations with a conversational agent and grounds explanations in verified scientific evidence. In a mixed-methods study with 30 healthcare professionals, nearly all participants reported that chatbot follow-up questions clarified the AI's assessments, and that seeing scientific evidence alongside AI explanations calibrated their trust rather than pushing them toward blind acceptance. The paper's three mechanisms are a hybrid prompt router that sends analytical queries to a specialized model and broader medical questions to a general language model, a manually verified evidence base that supplies pre-checked references instead of freshly generated ones, and a feature range analysis that places the model's observed value ranges next to clinically established ranges. The authors conclude that this integrated visual-conversational design is what lets clinicians evaluate both the risk and the recommended next steps.

What carries the argument

Three mechanisms carry the argument. The first is a hybrid query-processing architecture: a semantic matcher routes queries that match supported analytical operations to a small fine-tuned model that triggers specific backend functions, while everything else goes to a general-purpose large language model supplied with the patient's data, the active visualization, and the relevant scientific evidence, so responses stay context-aware. The second is an evidence-grounding pipeline: references retrieved from medical literature for each risk factor and range are manually verified and stored in a dedicated knowledge base, and the system serves those pre-verified references rather than generating new ones at query time. The third is feature range analysis: for each prediction class, the system filters the samples the model puts in that class, takes the 25th to 75th percentiles of each top factor, calls that window the AI-observed range, and displays it alongside scientifically established ranges from clinical guidelines; the degree of overlap is what participants used to judge whether the model agrees with medical knowledge.

What would settle it

Compute direct feature-attribution estimates, for example by perturbing a feature within its AI-observed range and measuring the change in predicted risk or by comparing against local attribution methods, across the same dataset and model; if factors whose percentile windows overlap scientific ranges produce no meaningful change in output when varied inside those windows, then the trust-building overlap is a property of the percentile definition rather than of the model's actual behaviour.

Watch

Extended reading notes

Core claim

The central claim is that explanation quality in a clinical AI system depends on two things at once: conversation and evidence. The authors found that conversational interactions helped healthcare professionals build a clear understanding of model assessments, with chatbot interaction receiving the highest understandability ratings among all tasks, and that integrating scientific evidence calibrated trust, with ratings rising across tasks and 17 of 30 participants citing the overlap between AI-observed ranges and scientific ranges as trust-building. They interpret this as evidence that a diabetes risk decision-support system should let clinicians ask follow-up questions about each visual and should present the model's findings next to established medical knowledge so the two can be compared directly. They also report that most participants used both modalities for risk evaluation and recommendation development, while identifying personalization of recommendations and more complete patient data as unmet needs.

Load-bearing premise

The load-bearing premise is that the AI-observed range, defined as the 25th to 75th percentile of a factor's values among patients the model puts in the same risk class, actually captures how much that factor influences the model's predictions; if those percentile windows do not track genuine feature influence, the range-overlap results and the trust calibration built on them rest on a distributional artifact.

Editorial extensions

If this is right

  • Clinical decision support systems should pair each visualization with a conversational channel, because follow-up questions, not the chart alone, are what resolved confusion about feature ranges in the study.
  • Showing AI findings next to verified scientific ranges is a usable trust-calibration mechanism: participants explicitly cited the overlap as the reason their confidence increased.
  • A hybrid prompt router is a viable division of labour, with specialized models handling analytical queries reliably and a general-purpose language model with contextual grounding covering the long tail of medical questions.
  • Recommendations should be presented as stepwise, feasible counterfactual changes with feasibility indicators and a timeline, because that component drew the highest actionability ratings even while participants asked for more personalization.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The percentile-window definition of the AI-observed range is a distributional summary, not a proven measure of feature influence; whether the trust-building overlap reflects genuine model behaviour is a testable question the paper leaves open.
  • If the reported pattern holds, adaptive explanation systems could reduce conversational scaffolding as users gain visual literacy, since several participants said the chatbot would become unnecessary once charts were familiar.
  • A natural next experiment is to measure trust behaviourally, for example by observing whether clinicians accept or override AI recommendations in simulated cases, because the study measured trust through self-report and task ratings.
  • The evidence base required manual verification, so scaling this approach to other conditions depends on curation effort; an automated retrieval plus clinician-verification pipeline would test whether the benefits survive at scale.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper presents a decision support system for diabetes risk prediction that combines interactive visualizations with a conversational agent, a hybrid prompt-handling architecture (fine-tuned T5 for analytical queries, Claude for general queries), a pre-verified scientific evidence base, and a "feature range analysis" that displays class-conditional percentile bands beside clinically established ranges. The authors report a mixed-methods study with 30 healthcare professionals, concluding that conversational interactions improved understanding of AI assessments and that scientific evidence integration calibrated trust. The paper claims contributions including an open-source artifact, the hybrid prompt approach, the evidence-grounding methodology, the feature range analysis technique, and empirical findings on clinician interaction.

Significance. If the empirical claims hold, the integrated visual-conversational pattern with evidence-grounded explanations would be a useful design contribution for clinical decision support and explainable AI. The manuscript has genuine strengths: the code is open-sourced, the medical evidence base was manually verified, the qualitative analysis used dual coding with a reported Cohen's kappa of 0.847, and co-design sessions informed the redesign. The participant quotes and interaction-log analysis (e.g., Feature Range Analysis receiving the most visualization interactions) are illustrative and plausible. However, the central trust-calibration claim rests on the "AI-observed range" visualization, which is defined as a tuned class-conditional percentile band rather than a measure of feature influence; the quantitative support is descriptive with no inferential statistics or control condition. The significance of the empirical contribution is therefore currently limited and in need of methodological revision.

major comments (3)
  1. [3.1.3] The "AI-observed range" in Section 3.1.3 is defined as the 25th–75th percentile of raw feature values among samples in the predicted class, and the text states this range is "the source of the factor's influence on predictions." This equivalence is not justified: a percentile band describes where the class lies in feature space, not how the model weights or uses the feature, and it ignores interactions, confounding, and the shape of the decision boundary. Because Section 5.2.1 reports that 17/30 participants cited range overlap as trust-building, the trust-calibration evidence is built on a distributional summary rather than a verified model-behavior measure. The paper should either replace this with a genuine influence-based measure (e.g., SHAP dependence, partial dependence, or individual conditional expectation) or explicitly reframe the visualization as a distributional summary and soften the corresponding claims.
  2. [3.1.3 and 5.2.1] The choice of the 25th–75th percentile window was not neutral: the paper states that this range was selected as "the largest range that aligns well with established medical ranges" after experimenting with 10th–90th percentiles. The overlap between AI-observed and scientific ranges that participants found reassuring (17/30 in Section 5.2.1) is therefore partly a consequence of a design decision, not an independent confirmation that the model aligns with clinical evidence. Please quantify the actual overlap, report the scientific ranges used, and treat the overlap finding as a perception result rather than evidence of model-clinical alignment.
  3. [5 and Figure 4] The quantitative results are reported as means and standard deviations without inferential statistics (significance tests, effect sizes, or confidence intervals), and the study has no control condition that isolates the scientific-evidence feature from the chatbot and visualizations. The claim that evidence "calibrated trust" (abstract, RQ2, Section 6.2) is therefore not established quantitatively; the observed increases across tasks could be due to task order or general system exposure. Please add appropriate statistical analyses or explicitly frame these conclusions as descriptive and qualitative.
minor comments (5)
  1. [6.2 vs 5.2.1] The number of participants positively influenced by evidence is reported as 28/30 in Section 5.2.1 but 29/30 in Section 6.2; please reconcile this inconsistency.
  2. [4.1 vs 5] Section 4.1 reports a 55-minute average completion time, while Section 5 reports an average 41-minute session duration; please clarify whether these refer to different measures (e.g., total study vs. active interaction).
  3. [1 and 3.1.1] Typographical and language errors include "an decision support system" and "decison-making" in Section 1, and "that using semantic matching" in Section 3.1.1.
  4. [3.3 vs 1] The contributions in Section 1 mention "multiple co-design sessions," but Section 3.3 describes two sessions; please make the wording consistent.
  5. [5.2.1] The term "AI-observed range" appears in quotes in the manuscript but is never formally defined for readers after the procedure in Section 3.1.3; consider a one-sentence recap at first use in the findings.

Circularity Check

2 steps flagged · score 6.0 of 10

Feature-range 'AI-observed' bands are self-definitional and were tuned to match the medical ranges they are later said to align with, partly constructing the trust-calibration result.

  1. self definitional [Section 3.1.3, Feature Range Analysis (steps 1-3)]
    "(2) Identify the 25th and 75th percentiles of values for each important factor (3) Define the range between these percentiles as the source of the factor's influence on predictions"

    The paper claims the technique identifies ranges of values 'most influential in predictions' and 'the source of the factor's influence,' but the range is computed purely as the 25th-75th percentile of raw feature values among samples predicted in the same class. No model-output perturbation, gradient, or attribution enters the range calculation (apart from the prior feature-importance ranking, which is computed separately). Consequently, any statement that the displayed band is where the factor's influence lies is true by construction; the visualization's apparent insight into model behavior is an artifact of the definition rather than a measured property of the model.

  2. fitted input called prediction [Section 3.1.3 (range tuning) and Section 5.2.1 (trust finding)]
    "While we experimented with broader ranges (10th to 90th percentiles), we observed that the 25-75 percentile range provided optimal results, representing the largest range that aligns well with established medical ranges while minimizing outlier influence. ... The overlap between AI-observed ranges and scientific ranges had a particularly positive effect on building trust for many participants (17/30): ... 'with the overlap shown it enhances my confidence as it shows that the AI is aligned with previous scientific research.' (P22)"

    The 25-75 window was selected 'as the largest range that aligns well with established medical ranges,' i.e., tuned against the same scientific ranges with which it is later compared. The feature-range visualization then displays the overlap between this constructed 'AI-observed range' and the scientific ranges, and the study reports that 17/30 participants found this overlap trust-building, with P22 concluding 'the AI is aligned with previous scientific research.' The alignment/trust signal is therefore partly manufactured by the choice of percentile window; reporting it as evidence of model-evidence agreement presents a fitted input as an empirical confirmation.

full rationale

The paper's main contribution is an integrated visual-conversational DSS plus a 30-participant user study. Most system components—the chatbot, the LIME/SHAP-based feature importance analysis with faithfulness scoring, the counterfactual recommendations, and the manually verified evidence base—are independent of the target claims and are not circular. Citations to prior work by the same group (e.g., Bhattacharya et al. [6]) are used for design principles and evaluation-task scaffolding, but they are not load-bearing for the central trust-calibration conclusion. However, two linked steps compromise the evidence-based trust-calibration finding. First, Section 3.1.3 defines the 'AI-observed range' as a 25th-75th percentile band and then calls this band 'the source of the factor's influence on predictions'—an influence claim that is true only by stipulative definition. Second, the same section admits the band was chosen because it 'aligns well with established medical ranges,' and Section 5.2.1 then reports that participants' trust was boosted by the overlap between the AI-observed and scientific ranges, including a participant quote concluding that 'the AI is aligned with previous scientific research.' The displayed overlap is thus partly a design choice, so the strongest trust-calibration evidence is partially circular. The conversational-understanding findings (RQ1) and usefulness findings (RQ3) do not depend on this constructed overlap and retain independent empirical content. The paper honestly discloses the tuning, which helps transparency, but the reduction remains: a fitted range is presented as an independent confirmation of model-evidence alignment. Overall, this is partial circularity in the central trust-calibration path, giving a score of 6.

Assumptions & free parameters 3 free parameters · 4 assumptions · 1 invented entities

The central empirical claims rest on two key constructs: the AI-observed range, which is a percentile-based summary with tuned bounds, and the curated scientific evidence base, which is treated as ground truth. The similarity threshold for query routing is a hidden free parameter. No new physical entities are introduced.

free parameters (3)
  • Percentile range for AI-observed ranges = 25th to 75th percentiles
    The authors experimented with 10th to 90th percentiles and selected 25-75 because it 'provided optimal results, representing the largest range that aligns well with established medical ranges while minimizing outlier influence' (Section 3.1.3).
  • Semantic similarity threshold = not reported
    The semantic matcher uses a 'carefully calibrated similarity threshold' to route queries to the T5 model or the general LLM; the value is not disclosed (Section 3.1.1).
  • Perturbation scale sigma = 0.05
    Set in the faithfulness computation inherited from TalkToModel, chosen to maintain local perturbations (Section 2.5.2).
assumptions (4)
  • ad hoc to paper The 25th-75th percentile range of feature values among samples in a predicted class represents the range of values that contribute most to the model's predictions.
    Section 3.1.3 defines the 'AI-observed range' in this way; no evidence is provided that percentile windows correspond to feature influence. This assumption drives the feature range visualization and the trust-calibration finding.
  • domain assumption Scientific ranges from WHO/ADA guidelines and the manually verified evidence base are a valid ground truth for evaluating AI alignment.
    Section 3.1.1 describes the curated evidence repository; the study treats overlap between AI-observed and scientific ranges as trust-building, assuming the guidelines are correct and applicable to the Pima population.
  • standard math Faithfulness metrics (Fudge score, Faith) from TalkToModel correctly measure explanation quality.
    Section 2.5.2; inherited from Slack et al. without re-derivation or validation in this context.
  • domain assumption Participants recruited via Prolific who pass the knowledge check are representative of healthcare professionals for evaluating the system.
    Section 4.1; the paper acknowledges the lack of formal verification of professional background and limited demographic diversity.
invented entities (1)
  • AI-observed range
    purpose: Defines the range of values for each important factor that supposedly contributes most to model predictions, to be compared against scientific ranges.
    Defined as the 25th-75th percentile of feature values among samples predicted in the same class (Section 3.1.3). No independent validation that percentile windows capture feature influence; the bounds were tuned to align with medical ranges.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Visual-Conversational Interface for Evidence-Based Explanation of Diabetes Risk Prediction." pith.science (2026). https://pith.science/paper/TTX7SDOW

@misc{pith2026250702920,
  author       = {Pith},
  title        = {Pith review of: Visual-Conversational Interface for Evidence-Based Explanation of Diabetes Risk Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TTX7SDOW}},
  note         = {Machine review of arXiv:2507.02920}
}
read the original abstract

Healthcare professionals need effective ways to use, understand, and validate AI-driven clinical decision support systems. Existing systems face two key limitations: complex visualizations and a lack of grounding in scientific evidence. We present an integrated decision support system that combines interactive visualizations with a conversational agent to explain diabetes risk assessments. We propose a hybrid prompt handling approach combining fine-tuned language models for analytical queries with general Large Language Models (LLMs) for broader medical questions, a methodology for grounding AI explanations in scientific evidence, and a feature range analysis technique to support deeper understanding of feature contributions. We conducted a mixed-methods study with 30 healthcare professionals and found that the conversational interactions helped healthcare professionals build a clear understanding of model assessments, while the integration of scientific evidence calibrated trust in the system's decisions. Most participants reported that the system supported both patient risk evaluation and recommendation.

Figures

Figures reproduced from arXiv: 2507.02920 by the authors.

Figure 1
Figure 1. Hybrid query processing architecture for conver [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. The DSS interface components. (a) Shows the main dashboard with patient records, analysis visualizations, and chat [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. The initial prototype featuring a conversational UI with a chat interface and visualization canvas. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Mean and standard deviations of participants’ ratings across all tasks [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Analysis of user interactions: (a) shows the frequency of different visualization types used, with Feature Range [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 4
Figure 4. Figure 4: , for Task 3 (Recommendation System), which showed improvements across all dimensions compared to previous tasks, with particularly high ratings for Actionability (𝑀 = 4.04, 𝑆𝐷 = 0.87) and Understandability (𝑀 = 3.98, 𝑆𝐷 = 0.87), as shown in [PITH_FULL_IMAGE:figures/f…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

77 extracted references · 64 canonical work pages

  1. [1]

    S. A. Alowais, S. S. Alghamdi, N. Alsuhebany, T. Alqahtani, and A. I. Alshaya

  2. [2]

    Shuroug A Alowais, Sahar S Alghamdi, Nada Alsuhebany, Tariq Alqahtani, Abdul- rahman I Alshaya, Sumaya N Almohareb, Atheer Aldairem, Mohammed Alrashed, Khalid Bin Saleh, Hisham A Badreldin, et al. 2023. Revolutionizing healthcare: the role of artificial intelligence in clinical practice. BMC medical education 23, 1 (2023), 689

  3. [3]

    Julia Amann, Alessandro Blasimme, Effy Vayena, Dietmar Frey, Vince I Madai, and Precise4Q Consortium. 2020. Explainability for artificial intelligence in healthcare: a multidisciplinary perspective. BMC medical informatics and decision making 20 (2020), 1–9

  4. [4]

    O. Asan, E. Bayrak, and A. Choudhury. 2020. Artificial Intelligence and Human Trust in Healthcare: Focus on Clinicians. Journal of Medical Internet Research 22, 6 (2020), e15154. doi:10.2196/15154

  5. [5]

    Rubaiyat Hossain Mondal, and Prajoy Podder

    Subrato Bharati, M. Rubaiyat Hossain Mondal, and Prajoy Podder. 2024. A Review on Explainable Artificial Intelligence for Healthcare: Why, How, and When? IEEE Transactions on Artificial Intelligence 5, 4 (April 2024), 1429–1442. doi:10.1109/tai.2023.3266418

  6. [6]

    Aditya Bhattacharya, Jeroen Ooge, Gregor Stiglic, and Katrien Verbert. 2023. Directive Explanations for Monitoring the Risk of Diabetes Onset: Introducing Directive Data-Centric Explanations and Combinations to Support What-If Ex- plorations. In Proceedings of the 28th International Conference on Intelligent User Interfaces. 204–219

  7. [7]

    Aditya Bhattacharya, Simone Stumpf, Robin De Croon, and Katrien Verbert. 2024. Explanatory Debiasing: Involving Domain Experts in the Data Generation Process to Mitigate Representation Bias in AI Systems. arXiv preprint arXiv:2501.01441 (2024)

  8. [8]

    Aditya Bhattacharya, Simone Stumpf, Lucija Gosak, Gregor Stiglic, and Katrien Verbert. 2024. EXMOS: Explanatory Model Steering Through Multifaceted Ex- planations and Data Configurations. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems . 1–27

Show all 77 references
  1. [9]

    Tianyu Cao, Natraj Raman, Danial Dervovic, and Chenhao Tan. 2024. Character- izing multimodal long-form summarization: A case study on financial reports. arXiv preprint arXiv:2404.06162 (2024)

  2. [10]

    Victor Chang, Jozeene Bailey, Qianwen Ariel Xu, and Zhili Sun. 2023. Pima Indi- ans diabetes mellitus classification based on machine learning (ML) algorithms. Neural Computing and Applications 35, 22 (2023), 16157–16173

  3. [11]

    Hao-Fei Cheng, Ruotong Wang, Zheng Zhang, Fiona O’connell, Terrance Gray, F Maxwell Harper, and Haiyi Zhu. 2019. Explaining decision-making algorithms through UI: Strategies to help non-expert stakeholders. In Proceedings of the 2019 chi conference on human factors in computin...

  4. [12]

    Victoria Clarke and Virginia Braun. 2017. Thematic analysis. The journal of positive psychology 12, 3 (2017), 297–298

  5. [13]

    Xinyue Dai, Mark T Keane, Laurence Shalloo, Elodie Ruelle, and Ruth MJ Byrne

  6. [14]

    Dazeley, K

    A. Dazeley, K. Karpowicz, and T. Menzies. 2021. Levels of explainable artificial intelligence for human-aligned conversational explanations. Artificial Intelligence 298 (2021), 103525. doi:10.1016/j.artint.2021.103525

  7. [15]

    Nils Feldhus, Qianli Wang, Tatiana Anikina, Sahil Chopra, Cennet Oguz, and Sebastian Möller. 2023. InterroLang: Exploring NLP models and datasets through dialogue-based explanations. arXiv preprint arXiv:2310.05592 (2023)

  8. [16]

    Jennifer Fereday and Eimear Muir-Cochrane. 2006. Demonstrating rigor using thematic analysis: A hybrid approach of inductive and deductive coding and theme development. International journal of qualitative methods 5, 1 (2006), 80–92

  9. [17]

    Randi E Foraker, Bobbie Kite, Marjorie M Kelley, Albert M Lai, Caryn Roth, Marcelo A Lopetegui, Abigail B Shoben, Michael Langan, Nicole L Rutledge, and Philip RO Payne. 2015. EHR-based visualization tool: adoption rates, satisfaction, and patient outcomes. eGEMs 3, 2 (2015), 1159

  10. [18]

    Elizabeth Gibney. 2024. Has your paper been used to train an AI model? Almost certainly. Nature 632, 8026 (2024), 715–716

  11. [19]

    Oscar Gomez, Steffen Holter, Jun Yuan, and Enrico Bertini. 2020. Vice: Visual counterfactual explanations for machine learning models. In Proceedings of the 25th international conference on intelligent user interfaces . 531–535

  12. [20]

    Mateusz Grechuta, Arun Patel, Meera Singh, et al . 2025. Exandra: A clinical decision support system for pharmacological management in type 2 diabetes. Journal of Medical Systems 49, 2 (2025), 45–56

  13. [21]

    Masooma Hassan, Andre Kushniruk, and Elizabeth Borycki. 2024. Barriers to and facilitators of artificial intelligence adoption in health care: scoping review. JMIR Human Factors 11 (2024), e48633

  14. [22]

    Robert R Hoffman, Shane T Mueller, Gary Klein, and Jordan Litman. 2018. Metrics for explainable AI: Challenges and prospects. arXiv preprint arXiv:1812.04608 (2018). Visual-Conversational Interface for Evidence-Based Explanation of Diabetes Risk Prediction CUI ’25, July 8–10, ...

  15. [23]

    Andreas Holzinger, Bernd Malle, Peter Kieseberg, Peter M Roth, Heimo Müller, Robert Reihs, and Kurt Zatloukal. 2017. Towards the augmented pathologist: Challenges of explainable-ai in digital pathology. arXiv preprint arXiv:1712.06657 (2017)

  16. [24]

    Robert Kaufman and David Kirsh. 2023. Explainable AI And Visual Reasoning: Insights From Radiology. arXiv preprint arXiv:2304.03318 (2023)

  17. [25]

    Kent, Nigam Shah, and colleagues

    Davida M. Kent, Nigam Shah, and colleagues. 2022. EHR-based prediction of Type 2 Diabetes in prediabetes patients using machine learning. Journal of Biomedical Informatics 132 (2022), 104121

  18. [26]

    Kerstan, N

    S. Kerstan, N. Bienefeld, and G. Grote. 2023. Choosing human over AI doctors? How comparative trust associations and knowledge relate to risk and benefit perceptions of AI in healthcare. Risk Analysis 44, 4 (2023), 939–957

  19. [27]

    Papageorgiou, and Dimitrios I

    Konstantina Kourou, Eleni I. Papageorgiou, and Dimitrios I. Fotiadis. 2021. Inte- gration of decision support systems into electronic health records: A review of recent efforts. Health Informatics Journal 27, 1 (2021), 840–857

  20. [28]

    A. W. Kushniruk et al. 2011. Issues and challenges in designing user interfaces for healthcare applications. In Studies in Health Technology and Informatics

  21. [29]

    Bum Chul Kwon, Min-Je Choi, Joanne Taery Kim, Edward Choi, Young Bin Kim, Soonwook Kwon, Jimeng Sun, and Jaegul Choo. 2018. Retainvis: Visual analytics with interpretable and interactive recurrent neural networks on electronic medi- cal records. IEEE transactions on visualizat...

  22. [30]

    J Richard Landis and Gary G Koch. 1977. The measurement of observer agreement for categorical data. biometrics (1977), 159–174

  23. [31]

    Liliana Laranjo, Adam G Dunn, Huong Ly Tong, Ahmet Baki Kocaballi, Jessica Chen, Rabia Bashir, Didi Surian, Blanca Gallego, Farah Magrabi, Annie YS Lau, et al. 2018. Conversational agents in healthcare: a systematic review. Journal of the American Medical Informatics Associati...

  24. [32]

    Q Vera Liao, Daniel Gruen, and Sarah Miller. 2020. Questioning the AI: informing design practices for explainable AI user experiences. In Proceedings of the 2020 CHI conference on human factors in computing systems . 1–15

  25. [33]

    Hui Wen Loh, Chui Ping Ooi, Silvia Seoni, Prabal Datta Barua, Filippo Molinari, and U Rajendra Acharya. 2022. Application of explainable artificial intelligence for healthcare: A systematic review of the last decade (2011–2022). Computer Methods and Programs in Biomedicine 226...

  26. [34]

    Scott M Lundberg and Su-In Lee. 2017. A unified approach to interpreting model predictions. In Advances in Neural Information Processing Systems . 4765–4774

  27. [35]

    Dimitry Mindlin, Fabian Beer, Leonie Nora Sieger, Stefan Heindorf, Elena Esposito, Axel-Cyrille Ngonga Ngomo, and Philipp Cimiano. 2025. Beyond one-shot explanations: a systematic literature review of dialogue-based xAI approaches. Artificial Intelligence Review 58, 3 (2025), 81

  28. [36]

    Van Bach Nguyen, Jörg Schlötterer, and Christin Seifert. 2022. Explaining Machine Learning Models in Natural Conversations: Towards a Conversational XAI Agent. arXiv preprint arXiv:2209.02552 (2022)

  29. [37]

    Navid Nobani, Fabio Mercorio, Mario Mezzanzanica, et al . 2021. Towards an Explainer-agnostic Conversational XAI.. In IJCAI. 4909–4910

  30. [38]

    National Institute of Diabetes, Digestive, and Kidney Diseases. 1990. Pima Indi- ans Diabetes Database. https://www.kaggle.com/datasets/uciml/pima-indians- diabetes-database

  31. [39]

    Jeroen Ooge, Gregor Stiglic, and Katrien Verbert. 2022. Explaining artificial intelligence with visual analytics in healthcare. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery 12, 1 (2022), e1427

  32. [40]

    Dympna O’Sullivan, Paolo Fraccaro, Ewart Carson, and Peter Weller. 2014. De- cision time for clinical decision support systems. Clinical medicine 14, 4 (2014), 338–341

  33. [41]

    Cliodhna O’Connor and Helene Joffe. 2020. Intercoder reliability in qualitative research: debates and practical guidelines. International journal of qualitative methods 19 (2020), 1609406919899220

  34. [42]

    Ravi Patel, Angus Brayne, Rogier Hintzen, Daniel Jaroslawicz, Georgiana Neculae, and Dane Corneil. 2024. Retrieve to Explain: Evidence-driven Predictions with Language Models. arXiv preprint arXiv:2402.04068 (2024)

  35. [43]

    Rob Procter, Peter Tolmie, and Mark Rouncefield. 2023. Holding AI to account: challenges for the delivery of trustworthy AI in healthcare. ACM Transactions on Computer-Human Interaction 30, 2 (2023), 1–34

  36. [44]

    Prolific. 2014. Prolific. https://www.prolific.com Accessed: February 20, 2025

  37. [45]

    Niroop Channa Rajashekar, Yeo Eun Shin, Yuan Pu, Sunny Chung, Kisung You, Mauro Giuffre, Colleen E Chan, Theo Saarinen, Allen Hsiao, Jasjeet Sekhon, et al

  38. [46]

    Amy Rechkemmer and Ming Yin. 2022. When confidence meets accuracy: Explor- ing the effects of multiple performance indicators on trust in machine learning models. In Proceedings of the 2022 chi conference on human factors in computing systems. 1–14

  39. [47]

    Reddy, W

    S. Reddy, W. Rogers, V.-P. Makinen, E. Coiera, P. Brown, M. Wenzel, E. Weicken, S. Ansari, P. Mathur, A. Casey, and B. Kelly. 2021. Evaluation framework to guide implementation of AI systems into healthcare settings. BMJ Health Care Inform 28, e100444 (2021). doi:10.1136/bmjhc...

  40. [48]

    Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. 2016. Why should I trust you? Explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining . ACM, 1135–1144

  41. [49]

    Tabea E Röber, Rob Goedhart, and Sİ Birbil. 2024. Clinicians’ Voice: Fundamental Considerations for XAI in Healthcare. arXiv preprint arXiv:2411.04855 (2024)

  42. [50]

    i wonder if my years of training and expertise will be devalued by machines

    M. Rony. 2024. “i wonder if my years of training and expertise will be devalued by machines”: concerns about the replacement of medical professionals by artificial intelligence. Sage Open Nursing 10 (2024). doi:10.1177/23779608241245220

  43. [51]

    Neda Rostamzadeh, Sheikh S Abdullah, and Kamran Sedig. 2021. Visual analytics for electronic health records: a review. In Informatics, Vol. 8. MDPI, 12

  44. [52]

    David L Sackett, William MC Rosenberg, JA Muir Gray, R Brian Haynes, and W Scott Richardson. 1996. Evidence based medicine: what it is and what it isn’t. 71–72 pages

  45. [53]

    Zahra Sadeghi, Roohallah Alizadehsani, Mehmet Akif Cifci, Samina Kausar, Rizwan Rehman, Priyakshi Mahanta, Pranjal Kumar Bora, Ammar Almasri, Rami S Alkhawaldeh, Sadiq Hussain, et al. 2023. A brief review of explainable artificial intelligence in healthcare. arXiv preprint arX...

  46. [54]

    Jessica M Schwartz, Maureen George, Sarah Collins Rossetti, Patricia C Dykes, Simon R Minshall, Eugene Lucas, and Kenrick D Cato. 2022. Factors influencing clinician trust in predictive clinical decision support systems for in-hospital deterioration: qualitative descriptive st...

  47. [55]

    Secinaro, D

    S. Secinaro, D. Calandra, A. Secinaro, V. Muthurangu, and P. Biancone. 2021. The role of artificial intelligence in healthcare: a structured literature review. BMC Medical Informatics and Decision Making 21, 125 (2021). doi:10.1186/s12911-021- 01488-9

  48. [56]

    Shinners, C

    L. Shinners, C. Aggar, S. Grace, and S. Smith. 2019. Exploring healthcare profes- sionals’ understanding and experiences of artificial intelligence technology use in the delivery of healthcare: An integrative review. Health Informatics Journal 26, 2 (2019), 1225–1236. doi:10.1...

  49. [57]

    Ronal Singh, Tim Miller, Liz Sonenberg, Eduardo Velloso, Frank Vetere, Piers Howe, and Paul Dourish. 2024. An Actionability Assessment Tool for Explainable AI. arXiv preprint arXiv:2407.09516 (2024)

  50. [58]

    Dylan Slack, Satyapriya Krishna, Himabindu Lakkaraju, and Sameer Singh

  51. [59]

    Roberto Spoladore, Laura Rossi, Marco Gatti, and et al. 2024. OnT2D-DSS: An Ontology-Based Clinical Decision Support System for Personalized Manage- ment of Type 2 Diabetes. In Proceedings of the IEEE International Conference on Bioinformatics and Biomedicine (BIBM) . IEEE

  52. [60]

    Mukund Sundararajan, Ankur Taly, and Qiqi Yan. 2017. Axiomatic attribution for deep networks. In Proceedings of the 34th International Conference on Machine Learning. PMLR, 3319–3328

  53. [61]

    Reed T Sutton, David Pincock, Daniel C Baumgart, Daniel C Sadowski, Richard N Fedorak, and Karen I Kroeker. 2020. An overview of clinical decision support systems: benefits, risks, and strategies for success.NPJ digital medicine 3, 1 (2020), 17

  54. [62]

    Nature Machine Intelligence (27 Jul 2023)

    Explaining machine learning models with interactive natural language conversations using TalkToModel. Nature Machine Intelligence (27 Jul 2023). doi:10.1038/s42256-023-00692-8

  55. [63]

    Danding Wang, Qian Yang, Ashraf Abdul, and Brian Y Lim. 2019. Designing theory-driven user-centric explainable AI. In Proceedings of the 2019 CHI confer- ence on human factors in computing systems . 1–15

  56. [64]

    Qianli Wang, Tatiana Anikina, Nils Feldhus, Josef van Genabith, Leonhard Hennig, and Sebastian Möller. 2024. LLMCheckup: Conversational examination of large language models via interpretability tools. arXiv preprint arXiv:2401.12576 (2024)

  57. [65]

    Zifeng Wang, Lang Cao, Benjamin Danek, Qiao Jin, Zhiyong Lu, and Jimeng Sun

  58. [66]

    Sana Tonekaboni, Shalmali Joshi, Melissa D McCradden, and Anna Goldenberg

  59. [67]

    Anjana Wijekoon, Nirmalie Wiratunga, David Corsar, Kyle Martin, Ikechukwu Nkisi-Orji, Belen Díaz-Agudo, and Derek Bridge. 2024. XEQ Scale for Evaluating XAI Experience Quality. arXiv preprint arXiv:2407.10662 (2024)

  60. [68]

    Yao Xie, Melody Chen, David Kao, Ge Gao, and Xiang’Anthony’ Chen. 2020. CheX- plain: enabling physicians to explore and understand data-driven, AI-enabled medical imaging analysis. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems . 1–13

  61. [69]

    Zhaopeng Xing, Fei Yu, Jian Du, Jennifer S Walker, Claire B Paulson, Nandita S Mani, and Lixin Song. 2019. Conversational interfaces for health: bibliometric analysis of grants, publications, and patents. Journal of medical Internet research 21, 11 (2019), e14672

  62. [70]

    Qian Yang, Yuexing Hao, Kexin Quan, Stephen Yang, Yiran Zhao, Volodymyr Kuleshov, and Fei Wang. 2023. Harnessing biomedical literature to calibrate CUI ’25, July 8–10, 2025, Waterloo, ON, Canada Samimi et al. clinicians’ trust in AI decision support systems. In Proceedings of ...

  63. [71]

    arXiv preprint arXiv:2406.17755 (2024)

    Accelerating clinical evidence synthesis with large language models. arXiv preprint arXiv:2406.17755 (2024)

  64. [72]

    Bo Wen, Raquel Norel, Julia Liu, Thaddeus Stappenbeck, Farhana Zulkernine, and Huamin Chen. 2024. Leveraging Large Language Models for Patient Engagement: The Power of Conversational AI in Digital Health.arXiv preprint arXiv:2406.13659 (2024)

  65. [77]

    Qian Yang, Aaron Steinfeld, and John Zimmerman. 2019. Unremarkable AI: Fitting intelligent decision support into critical, clinical decision-making processes. In Proceedings of the 2019 CHI conference on human factors in computing systems . 1–11

  66. [2019]

    In Machine learning for healthcare conference

    What clinicians want: contextualizing explainable machine learning for clinical end use. In Machine learning for healthcare conference . PMLR, 359–380

  67. [2022]

    In Pro- ceedings of the 2022 AAAI/ACM Conference on AI, Ethics, and Society

    Counterfactual explanations for prediction and diagnosis in XAI. In Pro- ceedings of the 2022 AAAI/ACM Conference on AI, Ethics, and Society . 215–226

  68. [2023]

    BMC Medical Education 23, 1 (2023)

    Revolutionizing healthcare: the role of artificial intelligence in clinical practice. BMC Medical Education 23, 1 (2023). doi:10.1186/s12909-023-04698-z

  69. [2024]

    In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems

    Human-algorithmic interaction using a large language model-augmented artificial intelligence clinical decision support system. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems . 1–20

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.