Survey topics.

Potential topics for a topic guide for FGIs were derived from a review conducted by Sony et al. (2023) [17], which originally focused on Health 4.0. The identified factors were then adapted to better align with the context of our research, which is centred around implementation of AI technologies in (smart) hospital settings. Prior to inclusion of the factors in the topic guide for the FGIs, these factors were validated on their relevance beforehand. To achieve this, a survey in English language was distributed among all seven work package leaders within the AIDPATH-consortium. The survey contained 10 societal factors related to the implementation of AI technologies in smart hospitals, namely:

1. Digital integration and interconnectedness of the healthcare ecosystem;
2. Utilization of big data and analytics;
3. Management of digital healthcare supply chains;
4. Strategies for promoting the use of AI technologies in medicine;
5. Promotion of a culture for the use of AI technologies in medicine;
6. Leadership in healthcare innovation;
7. Development of skills among healthcare employees;
8. Adoption of new business models;
9. Regulatory aspects of AI technologies in medicine;
10. Ethical aspects of AI technologies in medicine.

Please see S1 File for the survey questions.

Data collection and analysis.

Participants were asked to assess the relevance of each factor using a rating scale from 1 to 5, where 1 signified irrelevance and 5 signified relevance. Throughout the survey, ‘relevant’ meant bearing significance for the societal aspects for implementing AI technologies in smart hospitals, while ‘irrelevant’ meant bearing no significance. The ‘N/A’ option was available as well for each factor. For each factor, it was possible to provide additional comments and/or suggestions related to their assessment. Furthermore, participants had the opportunity to suggest any additional factors that are relevant for the successful implementation of AI technologies in smart hospitals.

The survey was distributed to the work package leaders via email, which contained a weblink to the survey on the Survey Monkey® platform in September 2023. All responses were recorded anonymously, and participants had a right to withdraw from the study at any time. The survey data were collected in October 2023. Survey data analysis was performed in Microsoft Excel. Percentage of experts indicating a particular level of relevance on the 1–5 scale were calculated for each factor. Factors were included in the topic guide if at least 75% of the participants found the factors relevant. Factors between the scores 1–3 were considered to be of low relevance. Factors between the scores 4–5 were considered to be of high relevance. The open-question responses were taken into account when interpreting the numerical data.

Sampling.

Participants were selected and sampled using purposive sampling methods, via the professional networks of the AIDPATH consortium members. Participants were selected based on their active involvement in the development and/or implementation of AI-related initiatives in healthcare. Although no strict minimum years of experience were required, most had several years of relevant expertise, as evidenced by their roles and responsibilities. Familiarity with the topic was assessed through their demonstrated involvement in AI development or implementation within their organizations. Maximum variation within the stakeholder groups was aimed to be achieved by including individuals from various professional backgrounds (e.g., HCPs, policymakers, technology developers, researchers)and level of familiarity with ethical or social aspects of AI [26]. Invitations to participate in a FGI were extended to participants via email. Participants were recruited in the period October 2023 till December 2023.

Data collection and analysis.

The FGIs followed a topic guide, compiled based on the factors identified from the survey. The FGIs started with an introductory round, a brief overview and explanation of the study objectives, followed by the main questions, and a closing. The main questions covered the societal factors related to the implementation of AI-technologies in (smart) hospitals. Per factor, questions were set up focusing on the possible related facilitators, concerns/barriers, possibilities, and preconditions (please see S2 File for the topic guide). Participants were provided with the opportunity to share their own insights or experiences, ask questions, and react to other perspectives. ‘Participatory learning and action’ techniques were employed to facilitate equal stakeholder input, promoting active engagement among participants [27]. All FGIs were led by a researcher (YdG or AA) and were conducted in the presence of at least one other researcher (YdG or AA). The FGIs took place online via Microsoft Teams in January and February 2024 and lasted 60 minutes on average. The FGIs were recorded after obtaining consent from the participants. The recordings were transcribed verbatim using the Amberscript software. The responses of participants were thematically coded and analysed by two researchers independently. The researchers (YdG and AA) assigned a code to text segments that relate to specific themes and then compared these codes. Data were analysed both descriptively (summarizing the main characteristics of the data) and exploratively (search for more complex patterns, relationships, or insights within the data).

Regulatory aspects of AI technologies in medicine.

Participants mentioned that a challenging aspect is the absence of clear, AI-specific guidelines in existing regulations, such as those from the European Medicines Agency (EMA). They noted that this regulatory gap, coupled with the lack of concrete procedural guidance, creates uncertainty in the implementation of AI technologies. The European AI Act, while a step in the right direction, also seems to suffer from a lack of specificity. Moreover, the use of disconnected information technology (IT) systems across hospitals is considered a major obstacle for data-driven system development. Harmonizing regulations within hospitals and across European Union (EU) Member States was deemed essential for successful AI technology introduction and obtaining requisite amounts data of AI learning.

“I think there is not much clarity yet to when you can share data for AI development. And I think some physicians might not be that comfortable with just sharing data on a large scale. But it’s not always up to them, of course, but it is very hospital specific.” (P1)

Participants viewed the current period of innovation and advancement of AI as an opportunity for developers and researchers. Due to the high level of uncertainty and quick pace of innovation, the direction and clarity of regulation surrounding (decentralized) AI technologies can be influenced by the field. They recommended proactively addressing ethical, logistical, and regulatory challenges associated with AI. Central to this effort is the recognition that AI is becoming increasingly prevalent and emphasis on the need to manage these issues effectively.

Digital integration and interconnectedness of the healthcare ecosystem.

Digital interconnectedness within the healthcare system, facilitating data sharing among stakeholders, plays a pivotal role in the success of AI implementation in smart hospitals.

The following sections elaborate on the various aspects of this interconnectedness from different perspectives micro, meso, and macro-level.

Macro-level (system level).

Since most EU healthcare systems are still transitioning from paper-based to digital systems and lack electronic patient files, participants felt that adopting AI based on integrated systems as a distant prospect. Differences in both electronic systems used and protocols for handling of sensitive patient data hinder interconnectedness and data sharing. Integration is however needed for personalized medicine, which requires extensive, standardized data gathering.

“Basically, for personalized medicine you need to generate more data than what we’re currently doing. Everybody would have to generate the same data with platforms that are so similar that you can actually lump all the data together.” (P2)

Meso-level (organizational level).

Participants believed that comfort with traditional, paper-based methods hinders progress towards digital integration. Inconsistencies in agreements between systems, such as interpretation of when data is (pseudo)anonymized, pose challenges. Uncertainty about when data can be shared for AI development can also be a barrier, as some HCPs feel they have insufficient knowledge on the General Data Protection Regulation (GDPR) to engage in large-scale data sharing initiatives.

Micro-level (individual level).

A recommendation made was to carefully consider the desirability of sharing specific data in an interconnected manner before initiating a project. Rather than sharing all available data, it was advised to define the type of data needed and connections that need to be made within a hospital.

Ethical aspects of AI technologies in medicine.

Concerns include the ownership and privacy of genetic and cellular data, particularly when used for treatment. Participants also worried about access to personal data. They stressed the importance of informed consent for data use in personalized treatments and registries and emphasized that posthumous data use consent should not be assumed.

A major concern was the lack of transparency in AI algorithms, which leads to misunderstanding and mistrust. This also raises security concerns due to the risk of not being aware of potential hacking or manipulation. The dynamic, self-learning nature of some AI complicates understanding its decision-making process, highlighting the need for developing AI systems that provide reasoning behind their decisions or prediction, allowing for intellectual oversight by humans. Participants also worried about data breaches or cyber-attacks that could affect treatment safety and continuity, underscoring the necessity of contingency plans. Further, AI has the potential to exacerbate existing health inequalities, particularly when it is trained on data from majority populations but applied across diverse populations, including underrepresented minorities.

“Regarding algorithmic fairness and bias there is a healthcare-system wide deficiency. Clinical trials, dose escalations for drugs, are all adjusted to the middle-aged average white male.” (P3)

Participants outlined several conditions for ethical implementation of AI technology. They emphasized that ethical considerations should be interwoven in the design phase of projects to avoid biases and ensure fairness, especially for underserved populations. Access to diverse data is crucial for algorithmic fairness. Furthermore, the recognition and understanding of potential biases in AI systems by different stakeholder groups can help mitigate them, making education on ethical aspects vital.

It was agreed that currently, an ethically acceptable role of AI is to serve as a support tool for decision-making. Critical medical decisions should remain the responsibility of HCPs. Ensuring human oversight helps maintain trust from HCPs and patients.

Ethical issues arising from AI development and utilization often go unaddressed. Participants promoted the view of ethics not as a standalone concept or “checkbox”, but rather as an integral part of AI development and implementation.

Development and promotion of strategies for adopting an AI culture.

Creating a shared culture and vision is needed for stakeholder cooperation and ultimately AI adoption. Strategies must explain and promote AI innovation amongst all end-users. Engaging stakeholders and ensuring HCPs embrace automation are key. Many organizations lack a cohesive AI culture due to limited experience, leading to extreme optimism or scepticism. Demonstrating an AI’s tangible benefits can build a unified culture and foster acceptance for successful integration.

To effectively promote AI adoption, organizations should adopt a gradual approach, incorporating evaluation points throughout implementation and development. This allows for careful assessment at each step to ensure alignment with initial goals.

Participants highlighted the need for better accessibility of information. While scientific papers contribute to understanding, they often do not reach the wider public. Thus, translating complex information into accessible formats is an important mean to increase awareness for -and a pragmatic view on- AI technologies.

Participants noted that meetings can be instrumental in bringing together stakeholders to address resistance to change. By initiating discussions and promoting collaboration, meetings can address concerns and facilitate in finding common ground.

Leadership in healthcare innovation.

It was believed that traditionally, a healthcare leader aligns efforts amongst various HCPs. However, participants emphasized that implementing AI technology in healthcare requires collaboration among additional stakeholders, including hospital management and IT staff. Due to the complexity of the healthcare ecosystem, no single leader can oversee the entire network effectively. Thus, successful implementation depends on a collaborative effort.

HCPs among the participants noted that while a flat hierarchy is desirable, it is not yet realistic. They expressed concerns that HCPs might be held solely responsible for AI-assisted healthcare outcomes due to the absence of a specific role for this responsibility. This responsibility is not something that most HCPs are willing to take up, according to the sample in our FGI’s, certainly not without the technology knowledge needed to be responsible.

Successful adoption of AI technology depends on effective communication and collaboration among stakeholders. Encouraging HCPs to talk to each other and understand the problems they face can facilitate the successful implementation of AI systems.

“When introducing AI based systems into a clinical environment, you have to do a lot of teaching in the beginning and answer a lot of questions and accompany every first step in the usage. Because otherwise it will fail, even if it’s a good system.” (P4)

Stakeholders’ skills.

To maximize effectiveness of AI technology, continuous de-, re-, and up-skilling of those applying the system is important.

Training HCPs in medical technologies is essential for implementation, but many HCPs already feel overwhelmed. Learning new skills and adapting to new technologies require time that may not be readily available. A transition to digitalization may require additional personnel, like clinical technologists, who have expertise in both medicine and technology.

Participants outlined several conditions for successfully integrating AI in HCPs’ skills. They emphasized AI optimizing or simplifying processes rather than replacing human roles. Additionally, (clinical) studies should be conducted to assess changes in HCPs’ skills and the impact of new technology on patient care. Concerns were raised about HCPs deskilling due to AI introduction. Participants emphasized the need for training and support to mitigate skill loss and ensure new skill development. Additionally, it is important to consider that if many different AIs are introduced, users may struggle to develop the necessary skills to use all of them effectively.

FGI results showed that not only HCPs but also patients need education for successful AI implementation. Patients often lack awareness about AI, its uses, benefits, and safeguards. To bridge this gap, patient education initiatives should provide clear and accessible information about AI and its implications for healthcare.

Many tools currently lack clear evidence of clinical benefits, making it hard to convince users to adopt them. Demonstrating clear benefits, such as improved task efficiency, speed, overall performance, or user confidence, is key to gaining user support. Participants noted that multiple intermediaries between decision-makers and end-users can lead to information loss and reduced feelings of ownership. This was identified as a cause of resistance to new technologies. Overcoming this requires direct communication and stakeholder involvement in decision-making. It also requires gradually introducing IT solutions and involving individuals in the change process. Flat hierarchies and inclusive approaches can help mitigate resistance and facilitate smoother transitions.

Adoption of new business models.

Adequate financial conditions must be in place for the successful implementation of an AI manufacturing platform. Traditional business models may not be sufficient to capture the rapidly evolving AI technologies, which often have value-driven outcomes. Participants expressed concerns about treating medicine as a revenue-generating market, a sentiment echoed by the broader public, especially given the high prices of many drugs. They emphasized that new business models should prioritize ensuring patient access to innovative treatments.

Participants discussed exploring alternative business models in AI medicine manufacturing, such as subscription models, which reduce upfront costs by offering recurring licensing fees. Ethical considerations about data ownership and commercialization were also highlighted. Questions arise about who holds the authority to sell or transfer data for use in AI tools and their training. Additionally, scalability must be considered, requiring preparation and infrastructure development to ensure business models can handle increased demand and maintain effectiveness on a larger scale.

Research.

Further investigation is recommended on factor ‘management of digital healthcare supply chains’ (factor 3) to understand why it was considered irrelevant by participating experts. Additionally, it is important to include the perspectives of underrepresented stakeholder groups that were not part of the current research, such as experts from regulatory affairs, nurses and other professional groups, and technology operators, to ensure a more comprehensive perspective in the societal guide. As the success of an AI technology also depends on its ability to enhance the patient experience and prioritize patient-centricity, future research should explore how AI technologies impact various dimensions of patient experience. Considering that underserved healthcare systems may encounter substantial challenges with digital integration and ethical frameworks, further research should explore the specific needs and constraints of these settings in AI implementation. To deepen insight and enhance the robustness of future findings, a mixed-methods approach is recommended. While this study offers qualitative insights into stakeholder perspectives, quantitative methods could provide broader generalizability and highlight patterns across different settings. For instance, surveys could be conducted to assess the prevalence of identified barriers and facilitators across stakeholder groups or regions. Structured questionnaires could also measure levels of AI readiness, digital literacy, or trust in AI among healthcare professionals and patients. Such efforts will help create a more inclusive and effective societal guide that reflects a wide range of healthcare environments.

Practice.

Several actions were recommended by the FGIs participants to ensure the successful implementation of AI technologies. One of them was a comprehensive approach that integrates various elements – such as regulatory frameworks, ethical considerations, digital infrastructure, human factors, and financial resources – is essential. Equally important is fostering a culture of knowledge sharing within the stakeholder community. This involves not only sharing good practices, but also learning from failures and challenges. By openly sharing both successful examples and lessons learned, stakeholders can gain valuable insights into what works well and what pitfalls to avoid. Additionally, transparency in sharing both successful examples and lessons learned promotes collective learning, which in turn can enhance the effectiveness and sustainability of AI technology integration in healthcare. When using the societal guide, it is important to consider the specific socio-cultural, economic, and regulatory context of each country. Tailoring the guide enables stakeholders to effectively address local barriers and enhance the impact of AI technology integration in medicine. For example, Europe faces challenges in accessing venture capital for funding new business models compared to the US and Asia, which hinders innovation and development in the medical technology sector. To address these funding challenges, innovation-friendly environments should be fostered in Europe. A tool that can guide development and scale-up of AI within existing laws and regulations is, for example, the’innovation funnel for valuable AI in healthcare’, developed by the Dutch Healthcare institute. This funnel considers five domains (value, application, ethics, technology, responsibility), which are key in the innovation process [37].

Recommendations for patient education on AI include organizing webinars covering basic AI concepts and applications in healthcare, such as smart hospitals, diagnosis, prognosis, and personalized medicine. These webinars could feature expert panels and include interactive Q&A sessions. Additionally, providing information in diverse formats, such as patient leaflets and animated videos, ensures broader accessibility and understanding among patients of different backgrounds and age groups.

When introducing new AI systems, it is recommended to proceed in small steps. For example, initially AI can serve as a second pair of eyes for HCPs, which gradually transitions to interpreting AI generated judgments, and ultimately builds towards trust in the technology for decision-making. Furthermore, HCPs will need time to reflect on the changing roles and responsibilities resulting from the integration of AI into healthcare practice. Another recommendation is to implement visualizing tools that illustrate how decisions are derived from AI models. Measures to verify the quality of AI generated results could be incorporated, such as confidence numbers or scales that represent the confidence of an AI in its decisions.

Lastly, establishing a framework to guide both developers and users of AI systems was strongly recommended. This framework should incorporate diverse perspectives to ensure stakeholder consensus and provide a roadmap for daily activities. Currently, such a framework does not exist, indicating room for development in this area.