The CXR-AI program was rolled out in January 2025 in a single National Healthcare Group (NHG) polyclinic in Singapore. This polyclinic performs approximately 300‐450 CXRs per month. The AI tool, Lunit INSIGHT CXR, had received regulatory approval from Singapore’s Health Sciences Authority and underwent local validation and institutional review before implementation.

In the CXR-AI program, all CXRs were analyzed by the AI tool, which displayed a heat map alongside a table of abnormality outputs within the routine CXR viewing platform (Figure 1). By default, CXRs triaged by AI as “normal” could be reviewed by the attending primary care doctor before radiologist reporting, allowing earlier clinical decision-making, and radiologist reports would be completed by the following day. In contrast, CXRs triaged as “abnormal” were prioritized for early radiologist reporting on the same day (Figure 2).

Clinicians could still override the default workflow. AI-triaged “normal” CXRs could be escalated for early radiologist review if the clinician was uncertain, while clinicians who were sufficiently confident could act on AI-triaged “abnormal” CXRs before radiologist reporting.

Regardless of AI triage status, all CXRs continued to undergo formal radiologist review. Radiologist reports completed the following day were reviewed by a vetting doctor, and patients initially discharged after AI-triaged “normal” CXRs could be recalled for further assessment if needed.

Between May 8 and July 18, 2025, a total of 587 CXRs were processed through the program, of which 69.8% (n=410) were triaged by AI as abnormal and 30.2% (n=177) as normal.

We used a qualitative descriptive design [21] to explore primary care doctors’ perspectives on using the CXR-AI program.

Two months after program rollout, email invitations were sent to all doctors working at the polyclinic in which the CXR-AI program was implemented. Those who had used the CXR-AI program at least once in clinical practice were eligible for the study. We used purposive sampling to obtain maximum variation across age groups, gender, and clinical seniority levels, with selected doctors approached directly for participation.

We developed the semistructured interview guide through iterative team discussions informed by relevant literature and three complementary conceptual frameworks (Textbox 1): Proctor’s implementation outcomes framework [13], the Theoretical Domains Framework (TDF) [15], and the Technology Acceptance Model 3 (TAM3) [22]. These frameworks were selected to capture early implementation outcomes, behavioral determinants of clinician use, and technology-specific adoption factors, respectively. Together, these frameworks guided the development of the interview prompts (Multimedia Appendix 1).

We collected data using both in-depth interviews (IDIs) and focus group discussions (FGDs). Participants were assigned to either IDIs or FGDs based on participant preference, scheduling feasibility, and consideration of hierarchical relationships, with junior doctors not grouped with their supervisors. The same semistructured interview guide was used for both.

All IDIs and FGDs were conducted face-to-face in English, audio-recorded, and transcribed verbatim. IDIs lasted 45‐60 minutes and FGDs 60‐90 minutes. Sessions were held in private consultation rooms within the polyclinic, with no nonparticipants present.

The main interviewer (SK), a female family physician, conducted all sessions. A coinvestigator (KPS, QWF, DHMK, or JGDR) attended as an observer to document field notes, nonverbal cues, and group interactions. All researchers had prior training in qualitative interviewing. Although QWF oversaw implementation of the CXR-AI program and several investigators worked in the same polyclinic as the participants, none had supervisory relationships with them. All participants were aware of QWF’s involvement in the implementation of the CXR-AI program.

To support reflexivity, our team maintained reflexive memos and held regular discussions to consider how our clinical roles and prior assumptions might shape data collection and interpretation. The research team reflected on whether QWF’s presence might inhibit participants from expressing critical views of the program. To minimize this risk, QWF ceased attending interviews after two sessions. Participants were also reassured that their responses would remain confidential and were encouraged to share both positive and negative experiences. We conducted postinterview debriefs to reflect on emerging findings, consider whether researcher assumptions may have influenced question framing, and refine the interview guide for subsequent interviews.

Participants also completed a questionnaire capturing demographic and professional characteristics (Table 1).

Data were analyzed using thematic analysis following Braun and Clarke’s 6-step approach [23]. We employed both inductive and deductive coding, with codes primarily derived from participants’ accounts while being guided by the study’s conceptual frameworks. Data were managed in Microsoft Excel (version 16.78.3).

Each transcript was independently coded by two investigators (SK, KPS, QWF, DHMK, or JDR), and discrepancies were resolved through discussion. Codes, subthemes, and overarching themes were iteratively developed and refined through regular team meetings and organized into a coding tree. Emerging interpretations were continually checked against the transcripts and participant quotations during team discussions. We maintained an audit trail documenting coding decisions, theme development, and codebook revisions [24]. To examine whether the data collection method influenced the findings, we compared the representation of subthemes across IDI and FGD transcripts. We also examined whether the representation and emphasis of subthemes varied across participants with different demographic and professional characteristics.

Data collection and analysis proceeded concurrently. We ceased recruitment after 20 purposively sampled participants once the research team determined that the dataset provided sufficient depth and variation to address the research question. The remaining eligible doctors were not recruited.

To enhance trustworthiness, we used data triangulation, reflexive journaling, and member checking by returning transcripts to participants for verification and by clarifying interpretations with them during analysis [24]. Reporting followed the Consolidated Criteria for Reporting Qualitative Research (COREQ) guidelines (Checklist 1).

Ethics approval was obtained from the NHG Domain Specific Review Board (reference: 2024‐3771). All participants provided informed consent. Pseudonyms were used during FGDs. Recordings were transcribed and deidentified before analysis, and participant demographic data were deidentified before compilation. Each participant received a token of appreciation (SGD $30 [US $23.68]).

Adoption was shaped by clinicians’ understanding of the tool’s purpose (coherence), perceived accuracy, clinical utility and limitations, and its impact on workload.

Ease of use was frequently cited as “one of the biggest factors [influencing] adoption” (IDI: A5, 46, male). The tool was praised for its user-friendly interface and easy access within the electronic medical system. “It loads fairly quickly...the features are quite clear. The fonts are big. The heat map is staring at you. Interface is fairly straightforward...just a quick scroll and you can see the next image” (IDI: A4, 36, male).

Clinician beliefs and confidence refer to individual-level determinants of engagement with the AI tool. Three key factors emerged: self-perceived competence in CXR interpretation, attitudes toward AI, and change mindset.

Organizational culture captures how institutional structures and professional norms shaped clinicians’ engagement with the AI tool. The most prominent influences were institutional direction and safeguards, followed by the reassurance of an expert in the loop, and the reinforcing effects of peer influence and professional norms.

Clinicians’ acceptance of the CXR-AI program was contingent on the continued involvement of radiologists. They felt “protected...that [all] x-rays will ultimately be reviewed by a radiologist...it’s a safety net for all of us” (IDI: A6, 27, female). Knowing that “the radiologist will still double-check,” clinicians felt comfortable to “trust [the AI] as kind of like a first triage” (IDI: A4, 36, male).

Beyond patient safety, clinicians valued the shared medicolegal accountability afforded by radiologist involvement. The prospect of removing radiologist reporting was deemed unacceptable because “100 percent of the responsibility will then fall upon the family doctor.... I wouldn’t like that at all. I think that’s quite unfair” (IDI: A3, 30, female).

Across interviews, radiologists remained the gold standard against which the AI’s value was judged, and their endorsement further strengthened its credibility. As one participant explained, “Whatever we learn and interpret, [the radiologists] had taught us in the first place. So, if they realize that the AI is quite trustworthy and credible, it boosts the credibility further” (IDI: D1, 29, female). In effect, the CXR-AI program was accepted not because of the AI alone, but because it was embedded within a workflow in which radiologists remained involved.

Clinicians described how their colleagues’ behavior could shape their willingness to use the AI tool. One explained that “as human beings, there’s this herd effect...if many people are not using it, I might probably be influenced not to use it” (FGD: B2, 41, female). Another felt that “if the rest of the team feels that it is useful, then it lends more credibility and I want to use it more” (IDI: A5, 46, male).

Peer experiences were generally favorable, with clinicians sharing positive examples of the AI’s performance and how these had built their own confidence in the tool. One recalled hearing that “it picked up pneumothorax...then it escalated...then the patient received care” (IDI: D1, 29, female).

Beyond peer influence, clinicians linked professional norms to medicolegal reassurance, noting that the AI tool alone was not yet defensible in the event of errors. As one explained, “if it is something that’s recognized as of high standard and it is widely used, then it can help to protect you...in terms of medicolegal implications. However, if it is a process that’s under trial, definitely it will not be helpful in any defense” (IDI: A2, 41, male).

However, some clinicians cautioned that peer influence and practice culture should not determine adoption decisions, as professional responsibility ultimately rests with the individual doctor: “Just because it’s something that is widely used or accepted doesn’t necessarily mean that it’s correct.... Ultimately, I’m the one that has to call the patient back and spend time and effort to change the plan” (IDI: A3, 30, female).

Together, these accounts suggest that peer influence helped build confidence in the AI tool, but did not override clinicians’ sense of individual responsibility for decision-making.

Different factors appeared more salient when participants described initial uptake versus subsequent use.

At the point of deciding whether to try the AI tool, coherence and change mindset appeared particularly important in participants’ accounts. Coherence refers to the sense-making work clinicians undertake to understand an intervention’s purpose and relevance to their practice [25] and overlaps with Proctor’s construct of “appropriateness” [13] and TAM3’s “job relevance” [22]. Some clinicians described being more willing to engage when they understood the purpose of the intervention and saw it as meaningful to their work. Similarly, a change mindset appeared to contribute to early willingness to try the tool despite uncertainty [26]. This is consistent with prior work showing that individual mindsets shape receptivity to health care innovations [27].

Following initial use, decisions about whether to continue using the tool were more influenced by clinicians’ assessment of its usefulness in clinical practice. This aligns with TDF’s construct of reinforcement, whereby behaviors are more likely to be continued when they produce rewards that are perceived as valuable and immediate [15]. In our study, these rewards took the form of diagnostic accuracy and workflow efficiency, which have also been identified in the literature as important preconditions for AI adoption [16,17]. The operating threshold used by our AI model provided high sensitivity (93.2%) and negative predictive value (91.3%), supporting its role in triaging normal CXRs, but lower specificity (71.7%) and positive predictive value (76.7%), consistent with clinicians’ perceptions of overcalling abnormalities [12]. These findings reinforce the potential role of Lunit in triage while highlighting its limitations as a stand-alone diagnostic tool [6-8,28,29].

Two findings differed from much of the existing literature. First, institutional endorsement emerged as an important enabler of adoption. TAM3 conceptualizes voluntariness as a moderator of social norms’ effects on intention to use, such that in mandatory contexts, individuals may adopt a system due to compliance rather than internalized beliefs about its usefulness [22]. However, in our study, many participants appeared to interpret mandatory implementation not only as something to comply with, but also as a signal of legitimacy. Relatedly, whereas prior studies have described regulatory uncertainty as a source of anxiety [30,31], our participants generally expressed confidence in institutional oversight. This trust in institutional governance may reflect contextual factors specific to the local health care setting. While this could be influenced by broader cultural tendencies, such as a higher power-distance index [32], our study was not designed to examine these explanations directly, and such interpretations should be made cautiously.

Second, although medicolegal concerns are frequently cited as a major barrier to clinicians’ acceptance of AI [30,31,33,34], these concerns appear mitigated in our study. This may be because radiologists remained embedded in the workflow, preserving shared responsibility and reducing anxiety about liability. In the local primary care context, formal radiologist reporting of CXRs is routine [35], and clinicians are accustomed to practicing within this arrangement. Participants regarded radiologists as a crucial second layer of review because of their specialist expertise and greater exposure to CXRs. This is supported by prior studies showing that CXR interpretation by nonradiologists is associated with higher rates of diagnostic error [36]. Together, these factors help explain why the AI tool was viewed as a complement to, rather than a replacement for, radiologist reporting.

A key strength of this study is that it examined clinicians’ perspectives after real-world use of an AI-supported CXR triage program embedded in routine primary care practice. This contrasts with much of the existing literature, which has explored AI acceptance in hypothetical settings. This is important as first-hand experience significantly shapes clinicians’ perceptions [26,30,37]. The use of multiple conceptual frameworks, together with both IDIs and FGDs, also strengthened the analytic depth of the study.

Several limitations should be acknowledged. Given that this study examined a single AI implementation at a single site within one health care system, the transferability of our findings may be limited in settings with different organizational structures, radiology workflows, or AI technologies. The study was also conducted during early implementation and therefore did not capture how clinicians’ perceptions and determinants of adoption may evolve with prolonged use. Participants may also have moderated their responses because interviews were conducted face-to-face by a physician in their workplace, and one investigator was involved in the program’s implementation. As participant allocation to IDIs and FGDs was nonrandom, the data collection method may have influenced how participants expressed their views. However, the subthemes were represented across both IDIs and FGDs, including the socially mediated themes of peer influence and institutional trust. Finally, as a qualitative study, it did not assess objective clinical or safety outcomes of the program.

Our findings underscore the importance of a context-sensitive approach to AI implementation in primary care. Before rollout, implementation leaders could clearly communicate the clinical and operational problem that the AI intervention is intended to address, how it integrates into existing workflows, and its locally validated performance and limitations. This may facilitate a clearer understanding of the rationale for change among end users [25,38]. Beyond this, implementation strategies could consider how local social and organizational dynamics may influence adoption. Where appropriate, approaches such as peer champions could be used to model use, shape norms, and build confidence [26], alongside feedback loops to refine workflows and sustain engagement through a stepwise approach. Finally, the diversity of clinicians’ views in this study reflects heterogeneous barriers, suggesting that one-size-fits-all implementation strategies may not be effective and that a more personalized approach may be needed [27,39].

While these overarching implementation principles may be transferable across settings, the specific barriers encountered may differ according to the AI intervention, its integration into clinical workflows, and the organizational, cultural, and medicolegal context. Implementation strategies should therefore be adapted accordingly. In our setting, this could include accounting for differences in clinicians’ confidence in CXR interpretation and the influence of peer and institutional factors. Specific concerns about AI use could be addressed through targeted strategies such as training on the tool’s intended role and limitations, ongoing performance auditing and monitoring, and clear guidance on clinical accountability and human oversight.

A comprehensive evaluation of this AI intervention could integrate qualitative findings with objective clinical and operational outcomes, including diagnostic accuracy, radiologist reporting turnaround time, clinical decision-making time, workflow efficiency, and safety metrics. Patient-reported experience measures could also explore the intervention’s impact on patients’ experience of waiting and care delivery. Later-stage implementation outcomes, such as penetration, sustainability, and cost-effectiveness, could also be examined [13]. Longitudinal qualitative research could examine how clinicians’ perceptions of AI and its integration into clinical workflows evolve with prolonged routine use. Studies across other health care settings and workflow designs would also help determine whether similar patterns of adoption emerge elsewhere.

This qualitative study suggests that adoption of AI-supported CXR triage in primary care is shaped not only by the technology itself, but also by how it is introduced, understood, and experienced in practice. Clinicians described greater willingness to use the program when they understood its purpose, found it useful within existing workflows, and trusted the organizational safeguards surrounding its use. These findings support the need for implementation strategies that are responsive to end user perspectives, local workflows, and the realities of primary care practice when integrating AI into routine care.