Wider stakeholder engagement activities, drawing on findings from 2 systematic reviews in this area [18,19], were conducted. These comprised (1) an in-person patient advisory group (PAG) meeting, (2) an online workshop with national leaders, (3) follow-up conversations with national leaders, and (4) hospital staff based at 2 hospitals who would be involved in the subsequent clinical trial (trial registration number: ISRCTN12421486). Summaries containing data gathered from the informal conversations (wider stakeholder work) informed recruitment and topic guide development for formal data collection, while contributing to the joint analysis.

Drawing on insights from this wider stakeholder engagement, we conducted formal data collection (ethnographic observations and semistructured interviews) between October 15, 2025, and May 7, 2026, at a large teaching hospital in Greater London and a smaller teaching hospital in an urban town. The focus of this research was on the roles and responsibilities of hospital staff working to keep hospitals free from airborne disease transmission. Data collected at this stage were triangulated with aggregated summaries from the wider stakeholder engagement activities (informal conversations with hospital staff, national leaders in ventilation, and members of the public) as reflected in Figure 1.

Wider stakeholder engagement, and particularly informal interviews, compensated for relatively less formal semistructured interviews. Ethnographic observations complemented what was being said during semistructured interviews by allowing us to understand staff’s day-to-day work in ventilation. Formal semistructured interviews included questions raised in informal conversations and expanded on current practices, acceptability, and feasibility aspects. Data obtained formally and informally aligned well.

All data were collected before trialing of the AISaT system, so both informal discussions and formal data collection were based on a hypothetical future implementation of AISaT rather than on its actual use or implementation. The COREQ (Consolidated Criteria for Reporting Qualitative Research) checklist was used to support the presentation of methods and results [20] (Checklist 1).

For context, AISaT will require the user to introduce details on room geometry, airflow (including the nature of air conditioning if present), occupancy, and mitigations used (air filter, screens, fans, and so on). AISaT then uses computational fluid dynamics (CFD)–based simulations to generate 2D and 3D representations of the entire room, with a color chart showing ranked recommendations (from highest to lowest) based on the likelihood of chosen mitigations to reduce the number of respiratory particles. AISaT may only need to be used once in rooms where occupancy remains similar over time (eg, consulting rooms), while in areas where occupancy changes, it may need to be rerun accordingly. The ultimate decision to place the mitigation strategies where AISaT recommends lies with the human operating the device, who will be provided with alternatives and will need to avoid hazardous placements.

The PAG was involved throughout the study. The PAG was consulted on the relevance and priority of clean air within hospitals, and their initial reactions to HEPA filters and to AISaT. Members of the group were recruited through advertisement on the National Institute for Health and Care Research (NIHR) People in Research website, charity connections, and established patient networks. The interaction with patients was led by SM, guided by Simon Watt and FS.

National leaders were invited to the online workshop and follow-up conversations were drawn from academic, industry, and health care sectors, through connections with members of the trial team and through online searches, because of their contributions to the field of ventilation. The workshop shared the project aims, sought first reactions about the trial, and helped to identify further national leaders in the field and key hospital job roles to target in ethnographic observations and interviews. It lasted 60 minutes to maximize participation by people working in the health service in high-pressure roles. The online workshop was led by YB-M and SM, with support from FS, ZK, and SEC.

The online workshop was supplemented by follow-up conversations, which allowed stakeholders to share more detail in relation to their views. Follow-up conversations were also conducted with people who could not join the workshop, and those who were identified through snowball sampling (n=8 [YB-M: n= 7; SM: n= 1]). After collecting information during the workshop about relevant hospital roles to engage with, the research team scheduled follow-up conversations with professionals in these roles (n=27 [YB-M: n= 25; SM: n= 2]). These informal conversations clarified the contribution of different job roles in keeping the hospital free from airborne disease transmission. A draft summary of barriers and enablers identified through the PAG, national leaders’ workshop, and preliminary contacts in hospitals is provided in Multimedia Appendix 1, which is the focus of a separate publication reporting on the development of an implementation strategy for rollout. Further description of the data collection process followed for wider stakeholder engagement is provided in Multimedia Appendix 2.

Hospital staff were recruited following a combination of purposive and snowball sampling techniques. Participants in diverse roles (eg, infection prevention and control [IPC], estates, and clinical) were invited to capture the views of all those likely to be involved in the implementation of Air Safety. Potential participants were approached up to 3 times by email. We completed ethnographic observations in the work environment and online semistructured interviews following a predefined internal study protocol. Researchers had no prior relationship with participants. Participants had read a participant information sheet and knew the aims of the study. The researcher introduced themselves and their role in the study and explained the aims of the semistructured interview or ethnographic observation.

Ethnographic observations were used to understand hospital staff’s day-to-day involvement in keeping the hospital air free from infection while complementing and facilitating understanding of the content of interviews. Ethnographic observations were unstructured, but a minimal data collection template was used (Multimedia Appendix 3). These were conducted by female researchers trained in ethnographic research methods and with previous experience in ethnography (FS and YB-M). Hospital staff at both site 1 and site 2 (the two trial sites) were invited to participate in ethnographic observations; however, volunteers were only found for site 1. A total of 8 hours of ethnographic observation were conducted at site 1 with the 4 research participants who volunteered to be shadowed. Field notes were taken in situ and transcribed electronically as soon as possible after the ethnographic observation. Interactions with nonparticipants were not included in field notes.

Each semistructured interview was guided by an interview schedule (Multimedia Appendix 4) which had been pilot-tested in preliminary conversations with stakeholders. Two researchers with expertise in qualitative methods (YB-M and SEC) and a background in psychology and molecular medicine conducted the interviews (YB-M: n=9; SEC: n=9). The interviews were conducted iteratively, building on previous insights.

Semistructured interviews lasted on average 33.9 (SD 6.61) minutes (ranging from 25 to 49 minutes; median 33.09, IQR 28.7-38.1 minutes). All interviews were audio- or video-recorded on Microsoft Teams with automatic transcription. The researchers (YB-M and SEC) cross-checked transcripts for accuracy and pseudonymized the data. Given hospital staff’s limited availability, interviews were not returned to participants to check their content, but during the interview, summarization and clarification were used instead.

Semistructured interviews concluded when thematic sufficiency had been reached, considering input both from semistructured interviews and previous informal conversations with hospital staff from the same hospitals. No interviews were repeated. Participants did not provide feedback on the findings.

Ethnographic observations and semistructured interview transcripts were analyzed using codebook thematic analysis [21], which sits between coding reliability and reflexive approaches of thematic analysis [22], in NVivo (Lumivero; version 12) [23]. YB-M coded 2 interviews and developed a preliminary coding framework inductively that was presented and discussed with FS. The rest of the ethnographic observations and semistructured interviews, and the wider stakeholder engagement work, were coded by YB-M applying this framework. Coding was conducted iteratively, and new codes and refinements were incorporated and discussed with the team while analyzing the remaining data. Subthemes were organized according to the core constructs of NPT [16,17,24]. We selected NPT to understand how the use of AISaT would need to be embedded in the day-to-day routines within the hospital, rather than focusing on a list of elements that could influence its successful implementation. However, we have used the Consolidated Framework for Implementation Research (CFIR) in a separate publication to develop the implementation strategy for rollout and the Non-Adoption, Abandonment, Scale-Up, Spread, and Sustainability (NASSS) framework to understand how AI implementation was affected by different factors in previous research [19].

We completed the following steps: (1) familiarization with the data started during data collection and continued through successive readings of the transcripts and field notes; (2) generating initial codes involved inductive coding of the data; (3) refining the codebook led to the organization of codes into units of meaning, still quite attached to the data and similar to what would have been topic summaries; and (4) the NPT constructs were then brought into the analysis and initial codes were reexamined, searching for more interpretive content relevant for each of the constructs.

Different data sources were analyzed together, adding depth to the analysis. Summaries of data from all informal conversations, PAG meetings, conversations, and workshops with national leaders were aggregated with an indication of the respondents supporting it. Feedback from wider stakeholder engagement was initially analyzed separately, leading to the generation of summaries and informing recruitment and topic guide development for ethnographic observations and semistructured interviews. For instance, informal conversations with stakeholders suggested different hospital staff needed to be involved in deciding on the use of AISaT. Ethnographic observations suggested that those staff would be part of committees that would review and decide whether changes or the introduction of AISaT would go ahead. Subsequent semistructured interviews tried to unpack who would be leading the implementation and cascading those decisions at the local level. The aim of this study was not to quantify responses; instead, we have indicated within the findings section the agreement or disagreement between data sources, or the origin of the data supporting the different themes.

The workshop with national leaders was recorded to facilitate accurate note-taking by researchers (YB-M and SM) and support credibility. Likewise, semistructured interviews were recorded and transcribed to capture hospital staff’s accounts accurately. Field notes were taken during each follow-up conversation with national leaders and informal conversations with hospital staff. Field notes were then summarized via a Rapid Research Evaluation and Appraisal Lab (RREAL) data collection sheet [25] that reflected areas of agreement and disagreement, as well as areas requiring further data collection. Informal conversations were not audio-recorded as they were exploratory and aimed to understand the level of involvement of different professionals and their initial reactions. Once relevant staff had been identified through informal conversations and the study had received ethical approval, semistructured interviews were conducted with further staff in relevant roles to avoid participant burden, enrich data collection, and facilitate confirmatory triangulation. This approach, together with the ethnographic observations, allowed a more diverse and prolonged engagement of the researcher with hospital staff in relevant roles. The resulting sample size of the interviews was smaller but more relevant than those involved in informal conversations. To reinforce dependability, we kept a record of every interaction with stakeholders in a spreadsheet, field notes were made during informal conversations and ethnographic observations, and interview transcripts were checked and stored together to facilitate an audit trail. An entry was added to a reflexive diary after every interaction with the stakeholders, regular discussions were held with the rest of the research team, and quotes from interviews are provided throughout the findings to ease confirmability. Finally, while the aim of this study was not to produce generalizable findings but to inform the development of AISaT and the next stages of the trial, we described in detail the types of data collected and the setting, so other researchers can gauge the contextual similarities.

The implementation research team consists of experts in digital health, implementation science, medical sociology, and qualitative research. Researchers involved in data collection and analysis were not involved in AISaT’s development other than by providing our insights to members of the wider team, including engineers working on the development of AISaT. Hence, our knowledge of AISaT was limited to the information shared with participants and discussed during interviews.

The Air Safety Trial was approved by the Health Research Authority and Health and Care Research Wales Ethics Committee (ref 25/EM/0156). All participants involved in ethnographic observations and semistructured interviews provided informed written consent.

We involved 12 public contributors (through the PAG) living around the Greater London area. All have a long-term condition and experience of care in multiple hospitals. Contributors were aged between 20 and 80 years, included 8 women and 4 men, self-identifying as Black, African, Caribbean, or Black British (n=4); Asian or Asian British (n=4); and White (n=4).

A total of 58 stakeholders were involved in conversations around the project and its potential implementation: 23 were involved in a workshop with national leaders (including senior NHS staff and policymakers, industry experts, and academics), and 8 national leaders took part in follow-up conversations, such as the 27 hospital staff who were identified either using online searches or purposive sampling within the 2 clinical trial hospitals.

Ethnographic observations were conducted with 4 hospital staff and semistructured interviews with 18 staff members, including 11 male and 11 female staff from Black, African, Caribbean or Black British (n=2), Asian or Asian British (n=8), and White (n=12) ethnic backgrounds. Staff interviewed had not been observed. Of 56 hospital staff members initially approached for observation and interviews at site 1, 5 declined to participate (feeling other colleagues were better placed to respond [n=1], having no time [n=4]). A further 22 did not respond to our contact, and with another 7, communication was lost after having expressed initial interest.

The job roles of those observed or interviewed were clinical (n=11), IPC (n=3), estates (n=6), and roles within the management of the hospital (n=2). Among those interviewed who shared job-related details, they had been in post for an average of 10.4 (SD 7.32) years (min=1; max=26), and most ranked the importance of ventilation within their roles as high (min 1 – max 10): 1 or 2 (n=1), 4 (n=1), 8 (n=3), and 10 (n=4).

Hospital staff interviewed or involved in follow-up conversations as part of the wider stakeholder work described several tasks undertaken as part of their role to reduce airborne disease transmission (summarized in Table 1). Their main role was isolating or managing patients with known airborne-transmitted infections, placing them in appropriate spaces (according to their risk level), or advising others on how to do this.

All stakeholders and participants provided their views on the hypothetical implementation of AISaT rather than during or after interacting with it. This work informed further development of AISaT prior to its testing and rollout.

Findings are organized according to the 4 constructs of NPT: coherence, cognitive participation, collective action, and reflexive monitoring (Table 2).

This paper explores key stakeholders’ anticipated barriers and facilitators to the implementation of AISaT to improve its acceptability at the implementation stage [11]. Our findings, organized around NPT, suggest that while AISaT and its recommended mitigation strategies are generally positively received, their wide implementation may be hindered by its associated maintenance and staff costs; the lack of perceived need and immediately observable benefits; and organizational differences within trusts and hospital buildings.

Coherence was variable across hospital staff. While hospital staff in organizational roles anticipated AISaT’s usefulness in their daily work, hospital staff in clinical and IPC roles were more divided on the value of the AISaT tool, with some reluctant to share their views on AISaT’s potential prior to reviewing results from a clinical trial. This may reflect staff not seeing the clinical need required to facilitate AISaT’s implementation [26]. The main risks anticipated to AISaT’s implementation were that existing strategies used within hospitals, such as fixed ventilation systems, are hard to change and may be different from the strategies AISaT may recommend, which are likely to involve portable HEPA filters. Additional anticipated risks to AISaT’s implementation were the potential challenges associated with HEPA devices and the lack of standardized rooms in old NHS buildings. A number of additional potential challenges have also been identified, such as PPE being seen as a more economical alternative; that handwashing and isolation rooms are being prioritized in current guidelines [27]; and that hospital staff felt improving air changes should be prioritized, whereas AISaT’s current prioritization is focused on diminishing aerosol concentrations. The main expected enablers identified related to the implementation of AISaT in outpatient areas; ensuring the AISaT interface is easy to use and responsive to staff’s needs; buy-in from senior staff; and its dissemination through national guidelines. In line with findings from previous research, there is a need to communicate that using AISaT does not remove human involvement from the decision-making process [11,15]. AISaT may be more likely to enhance its coherence in the hospital environment by tailoring its implementation approach to the ways of functioning of different hospitals, while providing staff with proof of the effectiveness of the tool in a way that is relevant for the metrics used within the NHS (eg, justifying cost savings and benefits for staff and patients).

Cognitive participation is anticipated to depend on multidisciplinary decision-making on the implementation of AISaT and its recommendations and having champions to cascade decisions down and encourage implementation on the ground. Anticipated barriers to implementation could be overcome by imposing the use of AISaT rather than relying on staff’s engagement. Imposition, however, would come with the inherent risk of losing the feeling of shared ownership, which would add to a lack of agency among staff, becoming a potential implementation challenge [28]. The potential implementation of AISaT was often not perceived as a priority, in a climate where staff feel overloaded, there are increasing financial constraints, and direct patient care is the most pressing matter. Such a lack of agency, particularly among those in clinical and estates roles, while raising the need to involve multidisciplinary teams may reflect the ways of working within the NHS as well as the nature of the AISaT tool falling in a confluence of responsibilities among different roles.

Collective action is anticipated to be impacted by costs associated with the implementation and maintenance of AISaT and its recommended mitigation strategies; the fragmented implementation of AI tools and devices within trusts; restrictions in accessing data; and variations in approval processes within trusts and buildings. While it was suggested that AISaT could be best integrated within a comprehensive electronic health record, this comes with an added layer of difficulty since electronic health record systems are not universally used across UK hospitals, which would require further adaptations to different systems used. This is a known generalizability issue previously identified in the literature [29,30]. An expected enabler identified was the external implementation of AISaT (eg, by a contractor), which would reduce its impact on staff workload.

Reflexive monitoring is anticipated to derive from expert validation and evidence on effectiveness and impact prior to the implementation of AISaT. However, previous research has also pointed out that such evaluation should continue after implementation [31]. An anticipated risk identified is the need for future refinement support that is unclear after the end of this project.

Our findings are consistent with those of a recent systematic review where AI implementation within hospitals was found to be influenced by understanding and trust in AI, quantification of the benefits of the AI, external validation to facilitate generalization, costs of implementing the AI, training staff and maintaining the AI, and the availability of policies to fund and regulate AI solutions [19]. However, our findings also point toward other critical elements, such as specific environmental characteristics where the AI and mitigation strategies will be implemented (eg, old buildings with busy and crowded wards, with scarce space and overloaded hospital staff), together with the priority given to AISaT and its mitigation strategies within the existing ways of working and measures already available. These findings also align well with a recent criticism of the NHS 10-year plan that AI implementation cannot sort out the whole structural issues within the NHS such as the constrained budgets [32]. Likewise, our findings are aligned with the need to transparently report on how the AI was trained as well as the benefits associated with its use, providing relevant metrics to the end users, and acknowledging the limitations of the tool [15].

Beyond our focus on AISaT’s implementation, but very closely related, difficulties in the use of HEPA devices, such as noise, which had been previously reported in the literature [33], did not preclude their hospital-wide use through ventilation outlets installed in ceilings. This is in line with HEPA filters being the most common mitigation strategy used in hospitals to reduce airborne disease transmission within papers included in a recent review [18], subsequent research [34], and aligned with estates technical guidelines [35]. Although the review reported improvements in patients’ health outcomes [18], some stakeholders included in our study stressed the lack of randomized controlled trials proving HEPA devices’ effectiveness.

This work involved extensive stakeholder engagement, capturing views from the public, subject matter experts, and hospital health care professionals, which helped identify common ground, but also areas of discordance between accounts even among professionals with similar job roles. Hence, it responds to the recommendation to include stakeholders at an early stage to facilitate future acceptability of AI tools [15]. This study is not without limitations; we struggled to access the hospital environment for further ethnographic work, only being able to observe estates work within one building. Difficulties may have been aggravated by the timing of our research during flu season. Furthermore, AISaT was under development at the time of conducting this study, and hence neither researchers nor participants had a tangible idea of what the tool would look like. This may have made it difficult for hospital staff and stakeholders to provide more specific responses and imagine what its implementation could look like in practice. Our formal data collection sample was focused on hospital staff from 2 hospitals, but more so from one of them, due to low response rates and delays in getting the relevant authorizations to do research in the second research site. Hospital staff shared their reservations about the AI tool during interviews; however, there may have been a natural selection bias. Through the wider stakeholder engagement (eg, via a national leaders’ workshop and follow-up conversations with national leaders), experts from around the country were involved, which helps reinforce the applicability of our findings to other hospitals with caution. Areas of concern such as impact on costs, workloads, and diverse ways of working were raised by all parties, supporting the transferability of our findings to other hospitals and health systems working under limited resources, staffing pressures, and in old buildings. This approach supported our ability to provide analytical generalization, to aid the development of the AI tool and future trial plans. Future research should involve relevant stakeholders in designing user-friendly AI tools and seeking their feedback with demonstrations of the product to visualize the end-product and think of ways to implement it in their practice.

Even when AISaT and its recommended mitigation strategies are not considered a priority within the hospital environment now, in a climate of tight budgets, staff shortages, and heavy workloads, staff expressed their openness to accommodate AI innovations if the effectiveness and benefits are proven and shared with them. Therefore, when planning implementation across hospitals, it will be important to frame this according to the tenets of NPT to enable understanding at the level of the structure of services as well as people’s experiences to engender the flexibility that will be necessary to maximize the likelihood of success across different NHS trusts and associated hospitals.