Introduction

There is considerable excitement internationally surrounding the applications of Artificial Intelligence (AI) in healthcare, with the National Health Service (NHS) having ambitions to be world leaders [1]. Developments such as the Academic Health Science Network (AHSN) AI initiative, NHSX and the £250 million investment for the NHS AI lab are important steps in the NHS realising this ambition [2–5]. In addition, over £85 million has been awarded to 80 applicants through the ‘AI in Health and Care Award’ across 70 NHS sites since September 2020 [6].

AI compromises of multiple subsets [7], with machine learning (ML) and deep learning (DL) offering methods to systematically extract insight from the increasing complexity and quantity of healthcare data. The successful integration of AI tools into the delivery of healthcare can help healthcare systems in delivering the quadruple aim [8, 9]: improving experience of care; improving the health of populations; reducing the per capita cost of healthcare; and improving the experience of providing care [10, 11]. Topol outlined how the benefits would be impactful at three levels [8]:

• Clinicians—primarily assisted by rapid diagnostic aid in patient management (particularly image interpretation e.g radiology, pathology, ophthalmology), and automation of repetitive administrative tasks to free up time.
• Health systems- operational applications in backend systems which are non-patient facing that improve workflow, and a reduction in variability and in medical errors.
• Patients- process patients’ health data to empower them in their own health promotion.

The four areas of focus for AI technology in healthcare as highlighted currently by responses in the 2021 NHSX survey are: diagnostics, remote monitoring, triage and population health [12]. This survey included 368 participants across healthcare in the UK, with 197 participants working as developers of AI for health and care.

However, despite increasing investment and research [13], there is surprisingly limited application of AI systems in clinical practice and recent systematic reviews have shown the majority of research studies are retrospective, lacking in transparency and at a high risk of bias [14, 15]. The slow translation and adoption of AI within the NHS and other healthcare systems can be attributed to several factors. Global issues relating to the science and safety of AI systems [16], which also include ethical and medico-legal issues are still being addressed [17, 18]. The NHS also faces further system-level challenges in technology innovation, implementation and adoption, such as fragmentation, underfunding, legacy IT infrastructure and ability in forming partnerships with the private sector [19–22].

In 2018, the AHSN conducted a national survey to identify how AI can be harnessed within the NHS. It included 106 thought leaders and AI pioneers working across the AI ecosystem in England. In addition to ‘capacity and capability to deliver scope’, ‘clarity around ownership of data’ and an ‘ethical framework to build/ preserve trust and transparency’, participants identified the ‘education of healthcare professionals’ and ‘engagement of healthcare professionals’ as two of the top five most important enablers of AI in the NHS; with the latter being the most important [23]. This is congruent with the recommendations from the 2019 Topol Review, which was commissioned to explore the implications of a digital NHS on the workforce and how it would change their way of working. Topol concluded that in order to leverage the full benefits of AI, four conditions have to be met in relation to the workforce [24]:

1. Need for time and willingness to adopt new technology
2. An understanding of the technology
3. Well-designed technology meeting user need
4. Workplace support to maximise the potential of the technology.

The influence doctors and other clinicians may have in the adoption and sustaining of AI tools could also be inferred from literature regarding other technology-supported services in healthcare. Wade et al. conducted a qualitative study of 36 Australian telehealth services with the view of intentionally identifying a single point of intervention, rather than a multifactor theory, to offer practical advice that would have the maximum impact on implementation. They proposed clinicians’ acceptance explains much of the variation in the uptake, expansion and sustainability of telehealth services [25]. Whilst this is not the only factor affecting the outcome, they found the clinicians that accepted telehealth would continue to supply it even if the demand was very low, there was pressures on the clinical workforce or if the technology was problematic. A sociological study conducted by Greenhalgh et al. supports this idea. A key finding from their investigation of why and when healthcare staff do not adopt nationally mandated information and communication technology (ICT) systems, was that the single most important determinant of whether a new technology-supported service succeeds or fails at a local level may be the acceptance by the workforce using it [26]. They reported that local staff, ‘champions’, appear key in persuading their colleagues that the technology-supported service is safe, effective, and professionally appropriate.

Given that doctors should be major stakeholders in this discourse, it is concerning that evidence suggests there are fears among doctors and other health professionals that AI may replace them and undermine trust [27]. Moreover, a web-based survey amongst 720 NHS General Practitioners (GPs) reported a disconnection between the expectations of frontline GPs compared to those of bioinformaticians and experts working with AI for healthcare [28]. Blease et al. proposed this may be due to a level of disengagement with the literature, and suggested further in-depth qualitative work to explore this.

Internationally, the awareness amongst doctors is a mixed picture. Quantitative studies have found that awareness was not widespread amongst doctors and medical students [29–32]. A 2020 survey found that American physicians’ awareness of AI tools had considerably increased since 2016 [33]. A theoretical model developed by Greenhalgh et al. identified 4 elements at which clinician resistance can impede new health care technology: resistance to the policy reflected in technology, resistance to the sociomaterial constraints of the new technology, resistance to compromised professional relationships, and resistance to compromised professional practice [34]. It is unclear if this theory would also apply to AI technology, but the lack of education amongst doctors could foster negative views towards AI and therefore limit its translation.

It is important to note that there is a lack of universal consensus in what engagement of doctors with AI entails. For example, it was not clearly defined in the AHSN survey despite being identified as a key enabler of AI translation. However, since AI is largely in the development phase, rather than widespread clinical use, the authors have chosen to define it as doctors working with AI, as opposed to engaging in educational activities such as conferences or courses to further personal understanding of AI. There is limited literature exploring the barriers for NHS doctors in engaging with AI. However, a number of barriers have been identified in the context of adoption of digital health technology (health tech) innovation, which could also pertain to AI engagement. The Nuffield Trust conducted a report to explore why the adoption of medical technologies is low in the NHS [20]. Similar to Asthana et al., they noted a number of cultural issues amongst clinicians that can act as barriers [19]. These include risk aversion, resistance to change and lack of entrepreneurial spirit. Greenhalgh et al proposed that resistance among clinicians may be due to perceived ‘hidden work’, where technology ultimately increases their workload [35]. These ideas link with the findings of Maguire et al, which suggested that technologies need to fit with the values, priorities and routine of staff for adoption [36]. Doctors may also be wary of the intentions of the private sector, which may discourage some from taking opportunities in industry [19].

Another key issue highlighted was that technology is largely supply-driven and top-down in the NHS, rather than the problems being identified by clinicians, and subsequently, they may see this as a threat to their professional judgement and autonomy [20]. Stakeholders from the Nuffield Trust roundtable suggested co-development with clinicians would resolve this. Collaborative designing involving doctors is also considered to be important for patient safety when developing AI tools, as their ‘soft intelligence’ [37] and medical expertise would provide oversight for the technology professionals who lack the clinical knowledge [23, 38, 39].

Whilst the involvement of doctors is considered to be important in the development of AI, it has been highlighted that it is unclear whose responsibility it is to innovate, and that there are a lack of incentives to encourage innovation from clinicians [20]. There is also a lack of time available for doctors to interact with these technologies and consider how they can improve healthcare [20].

Although being an emerging field, the increase in opportunities for NHS doctors to become involved with AI [40], and more broadly the digital health field, such as the Topol fellowships, NHS Digital Academy roles and the NHS clinical entrepreneurship programme has been encouraging for the development of the field [41–43]. Outside of the NHS, it is unclear what the extent of available opportunities are. There are reports of doctors pursuing academia posts and postgraduate qualifications in AI-related subjects, as well as some working within industry, for larger tech organisations such as Babylon and Deepmind, or as part of small and medium-sized enterprises (SMEs) [44, 45].

The majority of literature on doctors and industry collaboration [46–49] concerns the pharmaceutical industry or are regarding doctors’ views on industry collaboration rather than the motivators, barriers and personal experiences of doctors innovating and developing health tech devices in industry. To date, there are limited qualitative studies that explore the experiences of doctors who are actively working with AI and developing this field. Laï et al conducted a qualitative study on doctors regarding AI, however, they interviewed French stakeholders on their perceptions of AI in healthcare, which included 13 doctors that were purposively selected due to a previous interest in AI rather than direct development [50].

Therefore, this study aims to address this gap by exploring the experiences of doctors working with AI in English healthcare. For the purpose of this study, ‘working with AI’ describes doctors who are directly involved in the development or delivery of AI, within a healthcare or industrial setting, in addition to or instead of their clinical role. The authors have identified these group of doctors as key drivers for the development, implementation and the eventual widespread adoption of AI tools within healthcare.

The research objectives of this study include an exploration of; the facilitators for doctors working within the AI field in the NHS, the challenges they have experienced, and their views on how to increase acceptance and engagement of NHS doctors nationally.

Methods

The following framework (see Table 1) was used to aid recruitment. Potential participants were excluded if they had not worked with AI in English healthcare. The literature discussed above used the term clinicians and doctors interchangeably. However, it is important to note that the term clinicians compromise a range of healthcare professionals, such as nurses and physiotherapists, reflecting different roles and experiences. The authors refer to ‘doctors’ rather than ‘clinicians’ in the rest of this paper, which refers to doctors practicing healthcare on the ‘frontline’ and are or have been patient facing, and best described those who were included in the research.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Recruitment framework.

https://doi.org/10.1371/journal.pone.0282415.t001

Purposive sampling was used in a three-stepped recruitment strategy: 1) Individuals contacted either by social media (Twitter and LinkedIn) or email (n = 8); 2) Snowball sampling (n = 0); 3a) Recruitment posters shared on Twitter and LinkedIn platforms (n = 1) and 3b) Identifying gatekeepers from NHS networks and private sector companies (n = 2). Those registering interest were then provided with a participant information sheet and after collecting written e-consent, participants chose a time and date suitable to them via VoIP (Voice over Internet Protocol) technology mediated platforms. Although the onset of the COVID-19 pandemic heavily disrupted the availability of doctors, a total of 11 participants from varied backgrounds across primary care, medicine and surgery (see Table 2) were recruited.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Participants key characteristics.

https://doi.org/10.1371/journal.pone.0282415.t002

Semi-structured individual interviews were conducted from March to April 2020. The use of a topic guide ensured a degree of consistency but also allowed the researcher to probe further on emerging themes to capture authentic insights from each individual. It included questions on the personal journeys and facilitators for the participants to working with AI, their experience and challenges in this field, their views on AI in relation to the NHS and frontline doctors and the engagement of doctors in this discourse.

Audio-recordings were transcribed and analysed using Braun and Clarke’s six-phase thematic analysis, with Nvivo software being used to facilitate the coding process. The coauthor randomly selected and independently coded one transcript, and codes were compared and any inconsistencies were discussed.

Results

Three themes were generated which will be presented in turn with data extracts to illustrate their significance: the route to working with AI; challenges of working with AI; frontline clinicians in the AI discourse.

Discussion

This study aimed to explore the experiences of doctors working with AI in English healthcare, focussing on: the facilitators for doctors working within the AI field in the NHS, the challenges they have experienced, and their views on how to increase acceptance and engagement of NHS doctors nationally. The facilitators and personal experiences of doctors shaping the discourse in the AI field have not been explored before, and the findings from this study addresses this gap and contributes to the growing field.

Two main themes emerged; 1) an overarching motivator for entering this space was the ability to scale their impact to patients and 2) the current pathway to begin working with AI is unstructured. The other motivators and facilitators differed. Reduced job satisfaction had pushed doctors who were now working in industry full-time to pursue alternative careers. They now enjoyed greater job satisfaction due to a mixture of factors: greater autonomy, improved work-life balance and a career better suiting their interests and skillset. Doctors who were working with AI part-time in addition to their clinical role, believed their clinical role utilised their strengths and entered the AI field due to a sense of responsibility to shape development to protect patients’ interests. With the engagement of doctors identified as a key factor in the translation of AI within healthcare delivery, these findings can help to identify how to attract more doctors to work with AI. As highlighted by the different motivators and facilitators, the NHS and health systems should build a structured pathway and be cognisant when promoting opportunities that a variety of doctors from different backgrounds may be interested in pursuing a role in AI. This will help encourage a diversity of doctors to be included in shaping this discourse.

In addition to the system-level barriers to AI implementation that have been discussed in literature, a significant finding was the challenges that doctors can experience due to operating outside of their typical clinical environment. These include a lack of skillset, lack of collaboration and sharing of knowledge between health tech companies, less job stability and the uncertainty of navigating through an unchartered field. Some doctors conveyed certain skills could only be gained through experience and collaboration with professionals from other disciplines. This further supports the recommendation in the Topol review that the NHS should form partnerships with industry to provide opportunities for the NHS workforce to gain experience and skills, that will benefit the NHS [24]. This may also be important in terms of retaining NHS talent. An increased proportion of doctors are leaving the NHS due to burnout [52]. There is no data on those leaving to join industry, although this is likely to represent a small percentage of those leaving. Despite this, these doctors would represent a minority of those with specialist interdisciplinary knowledge that would be well placed to work on further development and eventual implementation and adoption of AI tools. Therefore, while providing these opportunities would remove staff from clinical roles, the NHS and related bodies must realise that providing flexible jobs for doctors to explore non-traditional clinical routes within the NHS, and work in specialist roles, could help retain the NHS talent that are being lost to full-time industry roles. Moreover, these doctors could be key ‘local champions’ described by Greenhalgh et al to other frontline doctors, and help build trust and acceptance towards use of AI tools [26].

The combination of working in a capitalist commercial setting and working alongside non-healthcare professionals who are used to a ‘move fast and break things’ mantra can bring tension. In particular, clinical values should be at the forefront for patient safety and identifying the biggest value propositions, supporting the view expressed in reports and articles [23, 38, 39, 53, 54]. The findings revealed the building trust with non-healthcare professionals was a solution to manage these culture clashes. This can be fostered in a number of ways: recognising and respecting the different expertise, developing a common language, establishing shared goals and demonstrating an interest to learn. Going forward, healthcare organisations encouraging the development of doctors working with AI should make them aware of this culture clash and share insight from doctors already working in this field to support their transition. Industry organisations should empower doctors working within their setup to raise safety issues and take steps to address a culture clash.

It is important to note the significant hype and misinformation surrounding AI, and in fact, progress is still in its infancy, with limited robust evidence conducted in real clinical practice and limited translation in healthcare. This is concordant with the views of doctors interviewed by Laï et al. [50], and resonates with the sentiment in existing literature [16, 22]. It also supports the reports from NHS-related organisations [55] that have highlighted that AI should not be heralded as a silver bullet.

The findings presented here suggest that frontline NHS doctors need to be significantly engaged with AI for its development, adoption and integration, which supports previous recommendations [24, 55–57]. It is therefore problematic that all participants perceived low awareness and understanding among frontline doctors, as suggested by Blease et al and anecdotal evidence [29–32]. In fact, the findings reveal disengagement can be categorised to three groups: 1) those resistant and sceptical; 2) those unaware or indifferent; and 3) those who may be interested in working in the AI healthcare field but lack time, mentorship and flexible opportunities to enter this field. The first group largely stems from the misconceptions propagated by the hype [27]. These included worries about AI replacing doctors, the risks of AI, resistance to change, and fear over losing their clinical autonomy, mistrust in partnering with industry; all having parallels to previous findings on barriers to adoption of technologies in healthcare [19, 20, 35, 36]. While some in the second group may prefer to focus on clinical work, it is also possible that they are simply unaware. Industry and healthcare organisation should avoid propagating hype and provide more accurate depiction of AI use-cases to promote translation of AI.

A limitation of this study was the small sample size. A total of 24 doctors had either expressed interest or had an interview arranged, but could not participate due to limited time from responding to the COVID-19 pandemic. A number of doctors from particular backgrounds could not be interviewed, for example, the majority of this sample compromised of doctors who were working with AI in a full-time capacity. This sample also did not contain any doctors working with AI in-house of the NHS, which was a group within the recruitment framework (Table 1). Female underrepresentation in tech has been highlighted as an issue [58] and although the researcher intended to have similar ratio of male to female participants, only three females were included in this sample. Due to there not being an established doctor developer role, there is likely a range of roles and experience that could yield important data to contribute to this evidence base. Although there was diversity within the sample, these findings may not be generalisable to the experiences and views of all doctors working with AI.

There are three areas of research that are warranted to better understand how to increase awareness amongst frontline clinicians, how to encourage more doctors to work with AI and how to best support them: 1) As recommended by Blease et al, further qualitative research should be conducted interviewing doctors who are not directly working with AI, to explore their attitudes and perceptions on AI; 2) qualitative research on the views and experiences of other non-clinical professionals developing AI tools with doctors within these interdisciplinary team; 3) qualitative research on the views and experience of the teams managing the implementation of AI tools and the healthcare professionals adopting them within their clinical workflow.

The key recommendations in Table 3 revolve around better education for both current and future doctors, providing more opportunities and fixed time for those interested and, a call to industry to empower doctors within their setup. The findings from this study demonstrate that engagement with AI is still in its infancy and whilst addressing wider translational challenges, the NHS and related bodies must ensure they dispel misconceptions and build a receptive context for adoption by engaging current and future workforce, whilst also providing opportunities for interested doctors to develop new skillsets and shape this discourse.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Recommendations.

https://doi.org/10.1371/journal.pone.0282415.t003

Acknowledgments

The authors thank Professor Mark Exworthy for his support, as well as two anonymous peer reviewers, and journal editors for their valuable comments.