Background
The use of artificial intelligence (AI) in healthcare can potentially enable solutions to some of the challenges faced by healthcare systems around the world [
1‐
3]. AI generally refers to a computerized system (hardware or software) that is equipped with the capacity to perform tasks or reasoning processes that we usually associate with the intelligence level of a human being [
4]. AI is thus not one single type of technology but rather many different types within various application areas, e.g., diagnosis and treatment, patient engagement and adherence, and administrative activities [
5,
6]. However, when implementing AI technology in practice, certain problems and challenges may require an optimization of the method in combination with the specific setting. We may therefore define AI as complex sociotechnical interventions as their success in a clinical healthcare setting depends on more than the technical performance [
7]. Research suggests that AI technology may be able to improve the treatment of many health conditions, provide information to support decision-making, minimize medical errors and optimize care processes, make healthcare more accessible, provide better patient experiences and care outcomes as well as reduce the per capita costs of healthcare [
8‐
10]. Even if the expectations for AI in healthcare are great [
2], the potential of its use in healthcare is far from having been realized [
5,
11,
12].
Most of the research on AI in healthcare focuses heavily on the development, validation, and evaluation of advanced analytical techniques, and the most significant clinical specialties for this are oncology, neurology, and cardiology [
2,
3,
11,
13,
14]. There is, however, a current research gap between the development of robust algorithms and the implementation of AI systems in healthcare practice. The conclusion in newly published reviews addressing regulation, privacy and legal aspects [
15,
16], ethics [
16‐
18], clinical and patient outcomes [
19‐
21] and economic impact [
22], is that further research is needed in a real-world clinical setting although the clinical implementation of AI technology is still at an early stage. There are no studies describing implementation frameworks or models that could inform us concerning the role of barriers and facilitators in the implementation process and relevant implementation strategies of AI technology [
23]. This illustrates a significant knowledge gap on how to implement AI in healthcare practice and how to understand the variation of acceptance of this technology among healthcare leaders, healthcare professionals, and patients [
14]. It is well established in implementation and innovation research that novel technologies, such as AI, are often resisted by healthcare leaders, which contributes to their slow and variable uptake [
13,
24‐
26]. New technologies often fail to be implemented and embedded in practice because healthcare leaders do not consider how they fit with or impact existing healthcare work practices and processes [
27]. Although, understanding how AI technologies should be implemented in healthcare practice is unexplored.
Based on literature from other scientific fields, we know that the leaders’interest and commitment is widely recognized as an important factor for successful implementation of new innovations and interventions [
28,
29]. The implementation of AI in healthcare is thus supposed to require leaders who understand the state of various AI systems. The leaders have to drive and support the introduction of AI systems, the integration into existing or altered work routines and processes, and how AI systems can be deployed to improve efficiency, safety, and access to healthcare services [
30,
31]. There is convincing evidence from outside the healthcare field of the importance of leadership for organizational culture and performance [
32], the implementation of planned organizational change [
33], and the implementation and stimulation of organizational innovation [
34]. The relevance of leadership to implementing new practices in healthcare is reflected in many of the theories, frameworks, and models used in implementation research that analyses barriers to and facilitators of its implementation [
35]. For example, Promoting Action on Research Implementation in Health Services [
36], Consolidated Framework for Implementation Research (CFIR) [
37], Active Implementation Frameworks [
38], and Tailored Implementation for Chronic Diseases [
39] all refer to leadership as a determinant of successful implementation. Although these implementation models are available and frequently used in healthcare research, they are highly abstract and not tailored to the implementation of AI systems in healthcare practices. We thus do not know if these models are applicable to AI as a socio-technical system or if other determinants are important for the implementation process. Likewise, based on a new literature study, we found no AI-specific implementation theories, frameworks, or models that could provide guidance for how leaders could facilitate the implementation and realize the potential of AI in healthcare [
23]. We thus need to understand what the unique challenges are when implementing AI in healthcare practices.
Research on various types of stakeholder perspectives on AI implementation in healthcare has been undertaken, including studies involving professionals [
40‐
43], patients [
44], and industry partners [
42]. However, very few studies have investigated the perspectives of healthcare leaders. This is a major shortcoming, given that healthcare leaders are expected to have a key role in the implementation and use of AI for the development of healthcare. Petitgand et al.’s study [
45] serves as a notable exception. They interviewed healthcare managers, providers, and organizational developers to identify barriers to integrating an AI decision-support system to enhance diagnostic procedures in emergency care. However, the study did not focus on the leaders’ perspectives, and the study was limited to one particular type of AI solution in one specific care department. Our present study extends beyond any specific technology and encompasses the whole socio-technical system around AI technology. The present study thus aimed to explore challenges perceived by leaders in a regional Swedish healthcare setting regarding implementation of AI systems in healthcare.
Discussion
The perspectives of the challenges described by leaders in the present study are an important contribution to improving knowledge regarding the determinants influencing the implementation of AI systems in healthcare. Our results showed that healthcare leaders perceived challenges to AI implementation concerning the handling of conditions external to the healthcare system, the building of internal capacity for strategic change management and the transformation of professional roles and practices. While implementation science has advanced the knowledge concerning determinants for successful implementation of digital technology in healthcare [
53], our study is one of the few that have investigated leaders’ perceptions of the implementation of AI systems in healthcare. Our findings demonstrate that the leaders concerns do not lie so much with the specific technological nuances of AI, but with the more general factors relating to how such AI systems can be channeled into routine service organization, regulation and practice delivery. These findings demonstrate the breadth of concerns that leaders perceive are important for the successful application of AI systems and therefore suggest areas for further advancements in research and practice. However, the findings also demonstrate a potential risk that, even in a county council where there is a high level of investment and strategic support for AI systems, there is a lack of technical expertise and awareness of AI specific challenges that might be encountered. This could cause challenges to the collaboration between the developers of AI systems and healthcare leaders if there is a cognitive dissonance about the nature and scope of the problem they are seeking to address, and the practical and technical details of both AI systems and healthcare operational issues [
7]. This suggests the need for people who are conversant in languages of both stakeholder groups maybe necessary to facilitate communication and collaboration across professional boundaries [
54]. Importantly, these findings demonstrate that addressing the technological challenges of AI alone is unlikely to be sufficient to support their adoption into healthcare services, and AI developers are likely to need to collaborate with those with expertise in healthcare implementation and improvement scientists in order to address the wider systems issues that this study has identified.
Conditions external to the healthcare system
The healthcare leaders perceived challenges resulting from external conditions and circumstances, such as ambiguities in existing laws and sharing data between organizations. The external conditions highlighted in our study resonate with the outer setting in the implementation framework CFIR [
37], which is described in terms of governmental and other bodies that exercise control, with the help of policies and incentives that influence readiness to implement innovations in practice. These challenges described in our study resulted in uncertainties concerning responsibilities in relation to the development and implementation of AI systems and what one was allowed to do, giving rise to legal and ethical considerations. The external conditions and circumstances were recognized by the leaders as having considerable impact on the possibility of implementing AI systems in practice although they recognized that these were beyond their direct influence. This suggests that, when it comes to the implementation of AI systems, the influence of individual leaders is largely restricted and bounded. Healthcare leaders in our study perceived that policy and regulation cannot keep up with the national interest in implementing AI systems in healthcare. Here, concerted and unified national authority initiatives are required according to the leaders. Despite the fact that the introduction of AI systems in healthcare appears to be inevitable, the consideration of existing regulatory and ethical mechanisms appears to be slow [
16,
18]. Additionally, another challenge attributable to the setting was the lack of to increase the competence and expertise among professionals in AI systems, which could be a potential barrier to the implementation of AI in practice. The leaders reflected on the need for future higher education programs to provide healthcare professionals with better knowledge of AI systems and its use in practice. Although digital literacy is described as important for healthcare professionals [
55,
56], higher education faces many challenges in meeting emerging requirements and demands of society and healthcare.
Capacity for strategic change management
The healthcare leaders addressed the fact that the healthcare system’s internal capacity for strategic change management is a hugh challenge, but at the same time of great importance for successful and sustainable implementation of AI systems in the county council. The leaders highlighted the need to create an infrastructure and joint venture, with common structures and processes for the promotion of the capability to work with implementation strategies of AI systems at a regional level. This was needed to obtain a lasting improvement throughout the organization and to meet organizational goals, objectives, and missions. Thus, this highlights that the implementation of change within an organization is a complex process that does not solely depend on individual healthcare professionals’ change responses [
57]. We need to focus on factors such as organisational capacity, climate, culture and leadership, which are common factors within the “inner context” in CFIR [
37]. The capacity to put the innovations into practice consists of activities related to maintaining a functioning organization and delivery system [
58]. Implementation research has most often focused on implementation of various individual, evidence-based practices, typically (digitally) health interventions [
59]. However, AI implementation represents a more substantial and more disruptive form of change than typically involved in implementing new practices in healthcare [
60]. Although there are likely many similarities between AI systems and other new digital technologies implemented in healthcare, there may also be important differences. For example, our results and other AI research has acknowledged that the lack of transparency (i.e. the “black box” problem) might yield resistance to some AI systems [
61]. This problem is probably less apparent when implementing various evidence-based practices based on empirical research conducted according to well-established principles to be trustworthy [
62]. Ethical and trust issues were also highlighted in our study as playing a more prominent role in AI implementation, perhaps more prominently than in “traditional” implementation of evidence-based practices. There might thus be AI-specific characteristics that are not really part of existing frameworks and models currently used in implementation science.
The healthcare leaders perceived that the use of AI in practice could transform professional roles and practices and this could be an implementation challenge. They reflected on how the implementation of AI systems would potentially impact provider-patient relationships and how the shifts in professional roles and responsibilities in the service system could potentially lead to changes in clinical processes of care. The leaders’ concerns related to the compatibility of new ways of working with existing practice, which is an important innovation characteristic highlighted in the Diffusion of Innovation theory [
63]. According to the theory, compatibility with existing values and past experiences facilitates implementation. The leaders in our study also argued that it was important to see the value of AI systems for both professionals and service-users. Unless the benefits of using AI systems are observable healthcare professionals will be reluctant to drive the implementation forward. The importance of observability for adoption of innovations is also addressed in the Diffusion of Innovation theory [
63], being the degree to which the results of an innovation are visible to the users. The leaders in our study conveyed the importance for healthcare professionals of having trust and confidence in the use of AI systems. They discussed uncertainties regarding accountability and liability in situations where AI systems impacts directly or indirectly on human healthcare, and how ambiguity and uncertainty about AI systems could lead to healthcare workers having a lack of trust in the technology. Trust in relation to AI systems is well reflected on as a challenge in research in healthcare [
30,
41,
64‐
66]. The leaders also perceived that the expectations of patient-centeredness and usability (efficacy and usefulness) for service users could be a potential challenge in connection with AI implementation. Their concerns are echoed in a review by Buchanan et al. [
67], in which it was observed that the use of AI systems could serve to weaken the person-centred relationships between healthcare professionals and patients.
In summary, the expectations for AI in healthcare are high in society and the technological impetus is strong. A lack of “translation” of the technology is in some ways part of the initial difficulties of implementing AI, because implementation strategies still need to be developed that might facilitate testing and clinical use of AI to demonstrate its value in regular healthcare practice. Our results relate well to the implementation science literature, identifying implementation challenges attributable to both external and internal conditions and circumstances [
37,
68,
69] and the characteristics of the innovation [
37,
63]. However, the leaders in our study also pointed out the importance of establishing an infrastructure and common strategies for change management on the system level in healthcare. Thus, introducing AI systems and the required changes in healthcare practice should not only be dependent on early adopters at the particular units. This resonates with the Theory of Organizational Readiness for Change [
70], which emphasizes the importance of an organization being both willing
and able to implement an innovation [
71]. The theory posits that, although organizational willingness is one of the factors that may facilitate the introduction of an innovation into practice, both the organization’s general capacities and its innovation-specific capacities for adoption and sustained use of an innovation are key to all phases in the implementation process [
71].
Methodological considerations
In qualitative research, the concepts credibility, dependability, and transferability are used to describe different aspects of trustworthiness [
72].
Credibility was strengthened by the purposeful sample of participants with various experiences and a crucial role in any implementation process. It is considered of great relevance to investigate the challenges that leaders in the county council expressed concerning the implementation of various AI systems in healthcare, albeit the preparation for implementing AI systems is a current issue in many Swedish county councils. Furthermore, the research team members’ familiarity with the methodology, together with their complementary knowledge and backgrounds enabled a more nuanced and profound, in-depth analysis of the empirical material and was another strength of the study.
Dependability was strengthened by using an interview guide to ensure that the same opening questions were put to all participants and that they were encouraged to talk openly. Because this study took place during the COVID-19 pandemic, the interviews were performed either at a distance, using the Microsoft Teams application, or face-to-face, the variation might be a limitation. However, according to Archibald et al. [
73], distance interviewing with videoconferencing services, such as Microsoft Teams, could be beneficial and even preferred. Based on the knowledge gap regarding implementation of AI systems in healthcare, the authors chose to use an inductive qualitative approach to the exploration of healthcare leaders’ perceptions of implementation challenges. It might be that the implementation of AI systems largely aligns with the implementation of other digital technologies or techniques in healthcare. A strength of our study is that it focuses on perceptions on AI systems in general regardless of the type of AI algorithm or the context or area of application. However, one potential limitation of this approach is the possibility that more specific AI systems and or areas of applications may become associated with somewhat different challenges. Further studies specifying such boundaries will provide more specific answers but will probably also require the investigation be conducted in connection with the actual implementation of a specific AI systems and based on participants' experiences of having participated in the implementation process. With this in mind, we encourage future research to take this into account when deciding upon study designs.
Transferability was strengthened by a rich presentation of the results along with appropriate quotations. However, a limitation could be that all healthcare leaders work in the same county council, so transferability to other county councils must be considered with caution. In addition, an important contextual factor that might have an impact on whether, and how, the findings observed in this study will occur in other settings as well, concerns the nature of, and approach to, AI implementation. AI could be considered a rather broad concept, and while we adopted a broad and general approach to AI systems in order to understand healthcare leader’s perceptions, we would, perhaps, expect that more specific AI systems and or areas of applications become associated with different challenges. Taken together, these are aspects that may affect the possibilities for our results to be portable or transferred to other contexts. We thus suggest that the perceptions of healthcare leaders in other empirical contexts and the involvement of both more specific and broader AI systems are utilized in the study designs of future research.
Conclusion
In conclusion, the healthcare leaders highlighted several implementation challenges in relation to AI within the healthcare system and beyond the healthcare organization. The challenges comprised conditions external to the healthcare system, internal capacity for strategic change management, and transformation of healthcare professions and healthcare practice. Based on our findings, there is a need to see the implementation of AI system in healthcare as a changing learning process at all organizational levels, necessitating a healthcare system that applies more nuanced systems thinking. It is crucial to involve and collaborate with stakeholders and users inside the regional healthcare system itself and other actors outside the organization in order to succeed in developing and applying system thinking on implementation of AI. Given that the preparation for implementing AI systems is a current and shared issue in many (Swedish) county councils and other countries, and that our study is limited to one specific county council context, we encourage future studies in other contexts, in order to corroborate the findings.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.