This study used a qualitative research design using semistructured, one-to-one expert interviews to explore perceptions of transparency and trust in AI-CDSS for chest X-ray diagnosis. The interview approach was designed to elicit in-depth insights into user interface design considerations and users’ mental models perceived to support transparent communication and trust in AI-CDSS. One-to-one interviews were selected to encourage open discussion, enabling participants to articulate their perspectives based on professional experience.

All participants were asked to reflect on AI-CDSS for chest X-ray interpretation. No common scenario, AI-CDSS interface screenshot, or example AI output was presented; radiologists reflected on systems they had used, whereas AI experts referred to systems they had developed or worked with. The systems discussed were designed to detect or localize suspected abnormalities to assist clinicians in interpreting chest radiographs. Interview questions addressed user interface elements, including heatmaps, confidence information, and feedback mechanisms that allow users to report errors. The interview questions were developed from the research questions and a review of relevant literature on transparency, trust, user interface design, and users’ mental models in human-AI interaction. The guides began with questions about radiologists’ experience with AI-CDSS or AI experts’ goals in developing such systems, followed by questions on transparency and trust, user interface design, and users’ mental models. Questions were tailored to each group’s role: radiologists were asked about clinical use and interpretation, whereas AI experts were asked about development and design considerations. Open-ended questions invited explanations and additional suggestions beyond the specified topics. To ensure clarity, relevance, and content validity, the interview guide was reviewed by domain experts using the Index of Item-Objective Congruence (IOC). Following revision, all interview questions received an IOC score of 1.00. Before data collection, the guide was reviewed by a health care innovation expert and a physician who served as the head of a radiology center. However, it was not formally pilot tested or cognitively pretested with representatives of the target groups. The interview guide was not modified after data collection began. Interviews were conducted in Thai to facilitate the accurate expression of participants’ perspectives. The validated interview questions are provided in Multimedia Appendix 1.

Purposive sampling was used to recruit experts whose professional backgrounds aligned with the objectives of the study. Potential participants were identified based on their professional roles and experience in either the development or clinical use of AI-CDSS for chest X-ray diagnosis. Eligible experts were invited to participate via direct contact (eg, email or professional networks), and participation was entirely voluntary.

Two groups of experts were included to capture complementary perspectives: (1) Medical professionals specializing in radiology for at least 5 years, with at least 3 years of experience using AI-CDSS for chest X-ray in clinical diagnostic practice. (2) AI experts with a minimum of 5 years of experience in developing AI technologies for medical imaging. The selection criteria for experts in this study were established based on the minimum threshold of 5 years of professional experience in their respective fields. This duration ensures that participants have surpassed the initial skill acquisition stages and possess the necessary practical wisdom to provide insightful perspectives [25,26]. For medical professionals, a minimum of 3 years of experience using AI-CDSS for chest X-ray was required, as the widespread adoption of AI-CDSS in clinical practice in Thailand has occurred only in recent years, beginning around 2021. The sample size for the qualitative interviews was guided by recommendations in qualitative research, suggesting that between 5 and 30 participants are generally sufficient for in-depth interviews, depending on the complexity of the research questions. Recruitment continued until theoretical saturation was achieved, that is, when additional interviews no longer yielded new insights relevant to the RQs [27].

Qualitative data were collected through online interviews conducted via Microsoft Teams. Before the interviews, participants received an informed consent form detailing the study objectives, procedures, and ethical considerations, followed by a briefing using Thai-language slides covering the study objectives, interview structure, key concepts, and participants’ rights. Participants then completed a brief questionnaire collecting demographic and professional background information before their interviews began. The briefing slides and their English translation are provided in Multimedia Appendix 2. Separate, but conceptually aligned interview guides were used for AI experts and radiologists. Interviews typically lasted 20‐30 minutes. All interviews were audio-recorded with participants’ informed consent and transcribed verbatim for subsequent analysis.

Interview transcripts were analyzed using thematic analysis following the six-phase approach proposed by Braun and Clarke [28]. This method was selected for its suitability in identifying patterned meanings across expert accounts while remaining grounded in participants’ perspectives and the study’s RQs. Analysis began with familiarization through repeated reading of transcripts to gain an overall understanding of participants’ views. Initial codes were then generated inductively, focusing on statements related to transparency and trust, user interface, and users’ mental models in AI-CDSS use. Coding was conducted across both expert groups (radiologists and AI experts) to identify shared and differing perspectives. Next, codes were examined and grouped into potential themes by identifying conceptual relationships and recurring patterns across participants. These candidate themes were iteratively reviewed and refined to ensure internal coherence and clear distinction between themes. Finally, themes were reported with illustrative quotations to preserve participants’ voices and to demonstrate analytic grounding in the data. The quotations selected for reporting were translated from Thai into English by the principal investigator. The English translations were subsequently checked against the original Thai quotations in collaboration with coresearcher (AJ) to ensure accuracy and preserve the intended meaning.

The study was approved by the Research Ethics Review Committee for Research Involving Human Research Participants, Group I, Chulalongkorn University, Thailand (Certificate of Approval number 079/69). The protocol adhered to the Belmont Report (1979), Declaration of Helsinki (2013), Council for International Organizations of Medical Sciences (2016), Standards of Research Ethics Committee (2017), and National Policy for Human Research (2015).

A total of 21 experts participated in the study, including 11 radiologists and 10 AI experts. Among the radiologists, 10 of 11 (91%) participants were female, and 1 of 11 (9%) participants was male. In terms of workplace, 7 of 11 (64%) participants worked in private hospitals, and 4 of 11 (36%) participants worked in public hospitals. Participants had substantial clinical experience, ranging from 5 to 33 years, with a mean of 16.0 (SD 8.94) years of clinical practice in radiology. All radiologists had direct experience using AI-CDSS for chest X-ray diagnosis, with a mean AI-CDSS usage experience of 4 (SD 0) years. On average, participants interpreted approximately 30.9 (SD 8.31) chest X-ray cases per day, reflecting routine engagement with high-volume diagnostic tasks. Among the AI experts, 3 of 10 (30%) participants were female, and 7 of 10 (70%) participants were male, with professional roles spanning professor, associate professor, assistant professor, lecturer, researcher, and chief technology officer (CTO). In terms of affiliation, 8 of 10 (80%) participants were affiliated with public universities, 1 of 10 (10%) participants was affiliated with a private university, and 1 of 10 (10%) participants was affiliated with an AI-based medical imaging company. The AI experts had a mean of 9.4 (SD 3.5) years of experience in AI and medical imaging–related research or development, ensuring in-depth familiarity with model development, explainability techniques, and user interface design for chest X-ray applications. Participant characteristics for radiologists are presented in Table 1, and for AI experts in Table 2.

The results use a participant coding scheme to attribute quotes and insights while preserving confidentiality. Each participant is identified using a shorthand code (X, Y, and Z), where X denotes the participant’s role (Radiologist [R] or AI expert [A]), Y indicates gender (Male [M] or Female [F]), and Z represents years of professional experience in radiology or AI development in health care. A total of 266 attributed quotations were identified across the interview dataset. Table 3 presents their distribution across the 3 themes and associated subthemes, together with sample quotations. The first theme, Perspective of transparency and trust, describes participants’ perspectives on transparency and trust when interacting with the AI-CDSS. The second theme, User Interface, highlights how specific user interface design elements facilitate clear communication and foster trust in the AI’s recommendations. The third theme, Users’ Mental Models, illustrates how participants’ experience, attitude, and knowledge about AI shape their interpretation of the AI’s outputs and their trust in the system. Figure 1 presents an overview of the 3 themes, their associated subthemes, and the key findings within each subtheme.

Across the themes, our findings suggest that in AI-CDSS for chest X-ray, transparency is interpreted functionally, not as full access to model internals, but as the system’s ability to communicate why it flagged a finding, how certain it is, and where it is likely to fail in a way clinicians can meaningfully use. Its contribution to trust therefore depends not only on the information presented through the interface but also on clinicians’ knowledge, experience, and attitudes when evaluating that information. This interpretation is consistent with a recent systematic review that identified system transparency, interpretability, clinical reliability, and alignment with clinical judgment as factors shaping health care professionals’ trust in AI-CDSS [29]. In this context, when outputs are delivered as unexplained “answers,” users are pushed into accept/reject behavior without defensible justification, which can erode confidence over time and inhibit sustained adoption [30]. Our findings extend prior work by showing that trust develops dynamically through repeated comparisons between AI outputs and clinicians’ own interpretations, rather than from accuracy or transparency alone. In chest X-ray reading, radiologists described forming an independent judgment, consulting AI as a second opinion, and revisiting discordant findings, illustrating how transparency is integrated into clinical reasoning.

A key implication is that transparency should be designed to support trust calibration, not to maximize trust. In safety-critical domains, both extremes are harmful: undertrust produces disuse of potentially valuable support, while overtrust increases susceptibility to automation bias and uncritical acceptance. Recent commentary in clinical AI highlights automation bias as a realistic risk when users defer to AI recommendations despite uncertainty or imperfect reliability, especially under time pressure [31].

Accordingly, transparency mechanisms should be explicitly “calibration-facing”: they should (1) communicate uncertainty, (2) disclose known limitations and boundary conditions, and (3) provide performance expectations that are understandable and clinically relevant. These cues should support clinicians’ judgments about when to rely on, verify, or override AI outputs, rather than serve merely as signals of credibility. To address undertrust specifically, clear information about performance in relevant clinical tasks, together with training using representative cases, may help clinicians calibrate their trust to the system’s demonstrated reliability and recognize situations in which AI is likely to provide valuable support. When an AI output conflicts with an initial interpretation, the user interface could provide a neutral prompt to reexamine the relevant region rather than a persuasive recommendation. This approach may discourage premature rejection without encouraging automatic acceptance. Clinician-facing transparency may also have implications for patient communication. Understanding the basis and limitations of AI outputs may help clinicians explain how AI contributed to a diagnosis and discuss uncertainty during shared decision-making. Such communication may influence patient trust and acceptance of AI-supported diagnoses, although these patient-facing implications were not examined in this study.

Supporting trust calibration requires more than making system information available: clinicians must be able to interpret and use it within their workflow. Considered together, the user interface and users’ mental model findings suggest that transparency depends on aligning what the system communicates with users’ understanding and clinical tasks. A review of AI in radiology similarly emphasizes that integration success depends not only on performance but also on workflow fit and adoption factors; the “right” information must arrive in a form that clinicians will actually use at scale [32,33]. The radiology user interface literature has reported mixed findings regarding preferred interfaces: some users favor detailed displays combining heatmaps, text, and confidence information, whereas others prefer simple and uncomplicated designs [34]. The current study helps explain this apparent inconsistency. Participants did not necessarily prefer either more or less information in absolute terms; rather, they wanted information to be task-dependent, clinically relevant, and accessible without creating screen clutter or disrupting independent image interpretation. Within these shared requirements, radiologists emphasized actionable, context-sensitive cues, whereas AI experts additionally emphasized system-level information about training data, provenance, and limitations. These competing requirements may be partly reconciled through progressive disclosure, whereby essential information is presented initially, and further details are made available on demand [33,35]. In practice, localized findings and concise confidence or uncertainty information could appear on the primary display, while technical details about training data, validation, performance, and limitations could be accessed through secondary views. For critical abnormalities such as pneumothorax, the suspected finding, location, urgency, and uncertainty should be immediately visible to support timely verification and prioritization, whereas routine screening may benefit from a concise initial display with additional information available on demand to minimize workflow burden. Moreover, UI design requires more than simply displaying additional information; it must ensure that users can correctly interpret what each element represents. For example, in usability testing of a chest X-ray AI tool, radiologists were uncertain about how to interpret confidence values and requested guidance on which confidence levels should be considered clinically meaningful [36]. Additionally, AI assistance increased total reporting time by 14 (33%) seconds, although this difference did not reach statistical significance, with the additional decision-making time largely attributed to reviewing heatmaps [36]. These findings highlight a potential trade-off: additional explanations may support interpretability but can also increase cognitive and workflow demands without necessarily improving diagnostic performance. User interface elements should therefore be evaluated according to whether their benefits for understanding and clinical decision-making justify the additional time required. Beyond these usability considerations, the explanation and uncertainty methods themselves have important technical limitations. Current visual explanation methods, including saliency- and attribution-based heatmaps, may produce plausible-looking outputs without reliably explaining the basis or correctness of an individual prediction [37]. In evaluations using chest radiographs, saliency-map methods have also demonstrated limitations in localization utility, sensitivity to model randomization, repeatability, and reproducibility [38]. Similarly, predictive uncertainty estimates may become less reliable when deployment data differ from the training distribution [39]. Heatmaps and uncertainty estimates should therefore be presented as decision-support cues rather than definitive evidence of diagnostic correctness, together with clear information about their intended meaning, validation, and limitations.

Additionally, the user interface should include guardrails that reduce automation bias. Visual cues can improve efficiency, but they can also become persuasive anchors if not paired with uncertainty and boundary messaging. The automation-bias literature in clinical decision support underscores that systems should avoid presenting outputs in ways that encourage default deference; instead, they should support deliberate verification, especially for borderline or high-impact decisions [31]. Consistent with this concern, our findings suggest communicating uncertainty and limitations and presenting the original radiograph before introducing the AI output. For borderline outputs, additional guardrails could include explicit labels such as “uncertain” or “verification required,” avoidance of definitive recommendations, and concise prompts for independent verification. Under time pressure, such guardrails should be visually salient and limited to ambiguous or clinically consequential findings to facilitate verification without producing excessive alerts or workflow burden.

Finally, feedback extends this relationship beyond the interpretation of individual outputs by connecting clinicians’ recognition of errors with opportunities for system improvement. The findings suggest that allowing users to flag errors may also support agency and continued critical engagement with AI. This aligns with implementation recommendations that emphasize feedback loops as part of integrating AI into real-world clinical workflows, rather than treating deployment as a one-time event [32]. However, the presence of a feedback function alone is insufficient; it must be intuitive and integrated into the existing workflow while minimizing cognitive and interaction burden. Our findings revealed different priorities: radiologists emphasized immediate, workflow-integrated feedback, whereas AI experts valued detailed input but cautioned against workflow disruption. These priorities could be reconciled by embedding one-click confirmation or rejection directly into the image-reading interface while making error localization and brief comments optional. This approach would enable rapid routine feedback and more detailed input when needed without requiring a separate reporting process.

The relationship between interface design and trust calibration depends partly on users’ mental models. This research suggests that the same transparency feature may be interpreted differently depending on clinicians’ knowledge, experience, and attitudes. This helps explain why “more explanation” is not automatically better: transparency is only effective when users can correctly interpret it and integrate it into clinical reasoning.

Previous evidence regarding the need for user training has been mixed. A recent scoping review reported that one included study characterized a chest X-ray AI system as easy to understand without requiring specialized training, although the evaluation used no formal scoring system and reported only limited qualitative findings [34]. By contrast, usability testing of another chest X-ray AI tool found that radiologists were uncertain about how to interpret confidence values and considered some form of training necessary before deployment [36]. More broadly, a systematic review identified training and familiarity as factors shaping trust and noted that opaque AI outputs may cause confusion and, in some cases, contribute to overreliance, particularly among less experienced clinicians [29]. Based on our findings, the 2 groups emphasized complementary pathways for developing users’ mental models. AI experts focused on structured information about system scope, validation, performance, and limitations, whereas radiologists described refining their understanding through repeated use and recognition of recurring errors. Together, these findings suggest that initial conceptual training should be complemented by continued learning through clinical use. This interpretation is consistent with previous work emphasizing that explainability and trust depend on users understanding the system’s intended purpose, outputs, and limitations [40]. Therefore, training is not “extras”; it is part of the transparency system. Well-designed training strengthens mental models so that clinicians use transparency cues for calibration (verify/override appropriately) rather than as reassurance signals that accelerate overreliance.

Attitudes, in particular, may shape the relationship between transparency and trust. Previous research has similarly recognized attitudes, perceptions, and positive expectations as factors shaping health care professionals’ trust in AI-CDSS [29]. Our findings add further context by suggesting that openness or resistance shapes whether transparency cues are attended to, dismissed, or treated as confirmation. This suggests a design requirement for plural-user transparency: systems should simultaneously (1) provide enough credibility and evidence to engage skeptics and (2) provide enough uncertainty and boundary messaging to constrain enthusiastic users from overtrust.

To synthesize these findings and highlight their practical implications, Table 4 summarizes key design considerations for transparent and trustworthy AI-CDSS derived from expert perspectives and provides illustrative clinical scenarios showing how these considerations could be applied in practice. Figure 2 presents a conceptual user interface mockup illustrating one possible implementation of the findings from Theme 2, while Table 5 presents a proposed training curriculum based on the interview findings.

First, this study focused on chest X-ray diagnosis; therefore, transparency and trust requirements may differ across other diagnostic contexts with varying information needs, risk levels, or interpretive workflows. Second, this study was conducted primarily in Thailand; thus, participants’ perspectives may reflect local disease patterns, including those associated with tropical settings, as well as clinical workflows, resource availability, and experience with AI-CDSS. Although context-specific preferences may not be directly transferable to all health care systems, the broader principles of understandable AI communication, uncertainty disclosure, workflow-aligned design, and support for users’ mental models may have wider relevance. Multicountry studies across diverse clinical and health care contexts are needed to assess the international applicability of these findings. Third, the purposive sample of experienced radiologists and AI experts may not capture the perspectives of novice users or clinicians without prior AI-CDSS experience. Participants’ existing experience with AI-CDSS may also have been associated with greater technology affinity, potentially underrepresenting barriers faced by less willing adopters. Fourth, the interview guide underwent researcher review, expert review, and assessment using the IOC; however, it was not formally pilot tested or cognitively pretested with representatives of the target groups. Consequently, potential differences in how participants interpreted the questions may not have been identified before data collection. Fifth, interviews shared a chest X-ray AI-CDSS context; however, the absence of a standardized scenario or visual stimulus may have led participants to refer to systems with different functionalities and user interfaces, potentially affecting their mental representations and the comparability of some responses. Future studies could use a shared prototype or standardized scenarios to provide a more consistent basis for eliciting perspectives. Finally, interviews were conducted in Thai, and quotations were reported in English. The Thai and English versions were checked against each other in collaboration with another researcher (AJW) to preserve meaning. Nevertheless, some linguistic nuances and culturally specific meanings may not have been fully conveyed in translation. Future research could adopt a mixed methods design by developing a survey instrument based on the themes identified in this study. Administering the instrument to a larger and more diverse sample of radiologists and AI experts could complement the qualitative findings, quantify the distribution of the identified attitudes and perspectives, and examine their associations with professional role, clinical experience, and prior experience with AI-CDSS.

This qualitative study demonstrates that trust in AI-CDSS is fundamentally shaped by transparency as experienced through human-AI interaction, rather than by algorithmic accuracy alone. Radiologists and AI developers converged on the view that transparent communication through interpretable explanations, visible uncertainty, and clearly defined system boundaries enables clinicians to integrate AI outputs into clinical reasoning while retaining professional autonomy. Our findings further show that trust is dynamic and calibrated, developing through repeated use, feedback, and alignment between system behavior and clinicians’ mental models. Importantly, transparency must be enacted not only through interface design, but also through consideration of users’ experience, knowledge, and attitudes toward AI. Together, these results underscore that designing trustworthy AI-CDSS requires a human-centered approach, in which transparency supports appropriate reliance rather than blind acceptance or rejection. By aligning technical design with clinical workflows and users’ mental models, AI-CDSS can function as a reliable collaborative partner in real-world diagnostic practice.