DEPARTMENT OF APPLIED IT FOSTERING AI INNOVATION IN SWEDISH HEALTHCARE: Essential Factors for an Enabling Policy Environment Adrian Sjödal Upul Sanjeewa Thesis: 30 hp Program: Digital Leadership Master’s Programme Level: Second Cycle Year: 2024 Supervisor: Jan Canbäck Ljungberg Examiner: Olgerta Tona Abstract Amidst the rapid technological advancements in AI, a growing number of proponents emphasize its significant potential in revolutionizing healthcare. However, the highly regulated nature of the healthcare sector poses numerous regulatory challenges at various stages of AI development and deployment. Despite regulations traditionally serving as governing and controlling mechanisms across various sectors including healthcare, they also possess the potential to play a crucial enabling role in fostering innovation. The aim of this study is to identify the essential areas of such a conducive regulatory landscape which enables the AI development within Swedish healthcare. Following a thorough review of contemporary literature in the subject area, we have identified five initial areas of policies which provided the base for the creation of the conceptual model of the study. The study followed the qualitative analysis realm of research and data were collected through a set of systematic interviews with expert stakeholders in the healthcare sector. The data collected through the interviews was analyzed using Braun and Clarke’s six step thematic analysis, a process which was instrumental in deriving the findings. As for the outcome of the study, we have iden- tified six essential policy areas for a conducive policy environment in Swedish healthcare for enabling AI innovation. These six areas are Data sharing and Open data initiatives, Data Ownership and Access Control, Data Quality, Standards and Interoperability, Regulatory Experimentation and Knowledge. The thesis urges leg- islators to carefully consider these identified policy areas, thereby introducing new reforms and modifying existing regulations to promote AI innovation in Swedish healthcare. Keywords Digital Innovation, Healthcare, Artificial Intelligence, AI Innovation, Policy envi- ronment, Healthcare Regulations, Innovation facilitation, Policy areas. i Foreword The authors wish to extend heartfelt gratitude to Prof. Jan Canbäck Ljungberg for his invaluable guidance, unwavering support, and insightful feedback throughout the course of this thesis. The authors remain deeply grateful to all the participants of the interviews, whose expertise and contributions as resource personnel have enriched this study immeasurably. Their willingness to share knowledge and experiences has been instrumental in shaping the findings and conclusions of this research. ii Table of contents 1 Introduction ........................................................................................................ 1 1.1 Background ................................................................................................ 1 1.2 Purpose and Research Question ................................................................. 2 1.3 Disposition ................................................................................................. 3 2 Literature Review ............................................................................................... 4 2.1 Digital innovation in healthcare ................................................................. 4 2.2 Policies for digital innovation .................................................................... 5 2.3 Artificial Intelligence (AI) ......................................................................... 5 2.4 AI utilization in Healthcare ....................................................................... 6 2.5 Digitalization and AI utilization in Swedish Healthcare ........................... 7 2.6 Current Regulation of Health Data in Sweden .......................................... 8 2.7 Regulation of AI in healthcare ................................................................... 9 2.8 Conducive policy environment for AI in healthcare ................................. 9 3 Conceptual Framework .................................................................................... 11 3.1 Conceptual framework for a conducive policy environment facilitating AI innovation within healthcare ........................................................................... 11 3.1.1 Data Sharing and Open Data ............................................................... 13 3.1.2 Data Ownership and Access Control ................................................... 14 3.1.3 Data Quality ......................................................................................... 15 3.1.4 Data Standardization and Interoperability ........................................... 16 3.1.5 Regulatory experimentation ................................................................. 17 4 Method .............................................................................................................. 19 4.1 Interviews ................................................................................................ 19 4.1.1 Selection of interview participants ...................................................... 19 4.1.2 Interview guide .................................................................................... 21 4.2 Transcription ............................................................................................ 21 4.3 Data analysis ............................................................................................ 23 5 Findings ............................................................................................................ 24 iii 5.1 Thematic Analysis ................................................................................... 24 5.1.1 Data Sharing and Open Data Initiatives .............................................. 25 5.1.2 Data Ownership and Access Control ................................................... 26 5.1.3 Data Quality ......................................................................................... 28 5.1.4 Standards & Interoperability ................................................................ 30 5.1.5 Regulatory experimentation ................................................................. 31 5.1.6 Knowledge ........................................................................................... 32 6 Discussion ........................................................................................................ 34 6.1 Areas of Policy Initiatives ....................................................................... 34 6.2 Interrelation of policy areas ..................................................................... 36 7 Implications and Limitations ............................................................................ 40 8 Conclusion ........................................................................................................ 42 9 References ........................................................................................................ 43 10 Appendices ....................................................................................................... 54 10.1 Appendix I - Final Interview Guide ......................................................... 54 iv List of Tables Table 1. List of identified policy areas and references ........................................ 11 Table 2. List of interviewed participants ............................................................. 20 Table 3. Transcription codes ............................................................................... 22 v List of Figures Figure 1. Conceptual framework ......................................................................... 13 Figure 2. Six policy areas extracted from the interview transcript data .............. 24 Figure 3. Dynamic interrelationships among policy areas .................................. 38 vi 1 Introduction This chapter provides the background for this master’s thesis, delineating the re- search area and the central research question. Additionally, it offers an overview of the structure, presenting the chapters and sections comprising the study. 1.1 Background Digital innovation, characterized as innovation leveraging digital technologies (Nambisan et al., 2020), plays a crucial role in the contemporary digital landscape, facilitating the creation of new markets (OECD, 2019) while transforming old ones to become more up to date (Paunov & Planes-Satorra, 2019). A specific digital tech- nology which has become a major influencer in digital innovation is Artificial Intel- ligence or AI (Trocin et al., 2021; Mariani et al., 2023; May et al., 2020). AI is widely seen as a beacon of potential, having the capability to influence and improve various fields of society. One particular field which could benefit from the affordances of- fered by AI is the healthcare sector (Wang & Lin, 2022; Rajpurkar et al., 2022: Noor- bakhsh-Sabet et al., 2019; Rosenberg et al., 2024), a domain filled with challenges of today and tomorrow. The Swedish healthcare, being no exception of facing these challenges, has a huge potential in reaping the benefits offered by AI, due to the significant level of digitalization already attained (Tucker, 2023). AI has shown great potential of being useful in improving healthcare administration and economics, as well as in many other clinical aspects including manufacturing of new medicines, diagnosis, treatment, as well as in medical data management and analysis (Davenport & Kalakota, 2019, Habehh & Gohel 2021). While the AI utili- zation have found broad success in image-based diagnosis (Esteva et al., 2017; Raj- purkar et al., 2022), it has since expanded its usability in further areas (Rajpurkar et al., 2022; Lee & Yoon, 2021) (e.g., Capio S:t Görans Sjukhus, 2023; Karolinska Universitetssjukhus, 2023; Björnheden, 2022; Holm & Benjaminsson 2023). Even though Sweden has a relatively high rate of healthcare workers compared to other countries in Europe, patients still suffer from long waiting times for diagnosis and treatment. A contributing factor to this issue is the increased administrative workloads for medical personnel (Anskär et al., 2019), as the amount of total admin- istrative work for nurses is now reaching the equivalency of two days per work week (Ivarsson Westerberg et al., 2021). On top of this, from a study conducted in 2015, more than 80 percent of Swedish primary care physicians do not regard the 1 healthcare system to work at a good level, and one third of the physicians perceives there has been a decline in the system’s quality during the last three years (Osborn et al., 2015). The future of Swedish healthcare faces further obstacles, such as an increased demand of human resources. In the coming decade, people in the age span of 80 years and older will increase by almost 50%, necessitating a significant expan- sion in the healthcare workforce to meet their growing needs. To be able to meet this future demand auxiliary nurses and elderly care workers needs to increase by almost a third, nurses by more than 20% and doctors by close to 10% (Sveriges Kommuner och Regioner, 2022). AI might be the solution to help solve this situation as the tech- nology already has shown strong results. AI has shown success in the administrative area of healthcare, being able produce medical journaling ten times faster than phy- sicians, which could drastically decrease the administrative workload, showing the tangible potential of what the technology affords (Rosenberg et al., 2024). 1.2 Purpose and Research Question Despite the huge potential to play a key role in enhancing overall care for people, AI faces several hurdles from different perspectives at various stages of its development and deployment. Those challenges include regulatory barriers, health economics, ethical considerations, and lack of knowledge among healthcare professionals (An- dersson et al, 2021). Given the highly regulated nature of the healthcare domain, new technology integrations such as AI in healthcare faces numerous regulatory chal- lenges (Petersson et al, 2022, Liu et al, 2022). Although the regulations in general are known for imposing governance and control, they can also function in an enabling capacity, specifically in AI innovation by insti- tuting appropriate regulatory frameworks in both over-regulated and under-regulated domains within the healthcare sector. For example, implementing regulations per- taining to privacy and data protection can enhance the confidence among individuals and institutions in permitting access to patient data for secondary purposes such as for AI development activities. Studies on AI development and incorporation of it across various sectors have em- phasized the significance of regulatory landscapes in varying degrees (Petersson et al., 2022; Liu et al., 2022). Moreover, research conducted on the integration of AI within healthcare has highlighted the necessity of supportive and robust policy frameworks as critical elements (Lee & Yoon., 2021; Hashiguchi et al., 2022). The absence of such supportive policy environments has been identified as a significant impediment to the development and integration of AI within the healthcare domain. With this study, we aim to identify the essential areas of a conducive policy environ- ment for promoting AI innovation within healthcare. We believe that this study’s theoretical contribution will go towards how digital innovation could be facilitated 2 in large scale sectors with the help of policies. And the generic contribution would be to assist policy makers in understanding in how to create a conducive policy en- vironment towards AI integration within Swedish healthcare. Therefore, we aim to answer the following research question, What are the essential policy areas needed for creating a conducive policy environment in facilitating AI innovation within the Swedish healthcare sector? 1.3 Disposition This study begins with a review of previous literature and the conceptual framework which together offers the essential information to comprehend the study. The litera- ture review first displays an overview of digital innovation and its relation to healthcare and regulations, followed by an explanation of AI, the implementation of it in Swedish healthcare and how the current regulatory field looks like. The legisla- tive environment is further detailed in section three of the thesis, the conceptual framework, where the initial five researched areas are presented. Each area is de- scribed with general information, followed by a discussion of the Swedish context and the regulations surrounding it. Specifics of how the study was conducted, including the interviews, the list of inter- viewees, transcription, and data analysis, are available to read in the methods section. Section five, findings, presents the interview data corresponding to each area identi- fied in the thematic data analysis. Additionally, the analysis of these data, along with prior literature, can be found in section six, the discussion. Finaly, the conclusion section summarizes the key takeaways of the study, the theoretical and generic con- tributions, suggestions for future research, and identified limitations. This section is followed by references and appendices. 3 2 Literature Review In this chapter, the synthesis of previous research related to the subject area of this master thesis is presented. Given the recent rapid advancements, AI in healthcare remains a relatively new area of research. Consequently, the articles referenced in this section mostly cover recent years, highlighting the evolving nature of research within this field. Recognizing the need to encompass a broad and contemporary re- search coverage within the specific research niche, we researched across various scholarly databases, including the University of Gothenburg's internet library and Google Scholar. 2.1 Digital innovation in healthcare Like the name suggests, digital innovation is defined as the process of innovating with the utilization of digital technologies (Nambisan et al., 2020), therefore, inher- iting the affordances of digital means, meaning there are certain abilities which are unique to digital innovation (Nambisan et al., 2017). With digital affordances such as openness, editability, interactivity, and distributability, (Kallinikos et al, 2010) parties utilizing digital innovation not only have the capacity to create digital prod- ucts, services and platforms that transforms the business landscape, but they can also enact in novel approaches as hindrances such as scaling no longer are limited to physical constraints (Ciriello et al., 2018). Digital innovation also possesses a num- ber of unique traits compared to its traditional counterpart, the first one being the importance of data. Data lays the foundation of understanding a certain area and to- wards enhancing different types of products, processes and services. The second trait of digital innovation is the increased focus on services as digital technology has the capability to possess such means. The final two characteristic elements are that dig- ital innovation typically operates at a faster pace and that innovation more commonly is shared among multiple actors. Although, it should be noted that these traits differ depending on the area of utilization. (OECD, 2019). Digital means have increasingly been integrated into healthcare due to its positive impact across various areas. It has empowered physicians and patients in treatment and diagnosis (Čartolovni et al., 2022; Awad et al., 2022; Petersson et al., 2022), increased the public access to healthcare (Awad et al., 2022; Hashigushi et al.,2022) and reduced economic expenses (Petersson et al., 2022; Awad et al., 2022), to name a few. Looking ahead, AI is seen to further revolutionize the healthcare sector, lev- eraging new possibilities in the digital health domain (Johnson et al., 2021). 4 2.2 Policies for digital innovation The notion that innovation can generate societal benefits, including improved well- being, environmental sustainability, and job creation, underscores the necessity of nurturing it. Policies are widely recognized as vital facilitators of digital innovation, serving to accelerate its progress and broaden its accessibility across diverse seg- ments of society. This perception is widely held, evidenced by the multitude of pol- icy initiatives from governments worldwide to promote digital innovation (Planes- Satorra & Paunov, 2019). The OECD Digital Innovation report on policymaking (2019) outlines numerous do- mains suitable for policymaking aimed at facilitating digital innovation. Suggestions range from introducing policies in new areas such as data access and a more agile policy responsiveness, to revising older policy areas, such as increasing the use of regulatory sandboxes and strengthen innovation collaborations. However, the OECD (2019) stresses the fact that there is no one-fits-all solution covering every sector. Instead, they advise policymakers to tailor each sectors’ policy environment to its own needs as each sector operates distinctively and thereby is affected differently. For instance, accessing data in the healthcare sector poses greater complexity due to the higher sensitivity of health-related information. 2.3 Artificial Intelligence (AI) During recent years, Artificial Intelligence, or more commonly referred to as its ab- breviation AI, has grown exponentially (Johnson et al., 2021). AI brings hope as it has the ability to positively contribute to large parts of society, both in terms of the public sphere and the private economy. While predominantly utilized in the digital sphere, the technology also presents opportunities in physical spaces, including sec- tors like agriculture and healthcare. While the anticipated benefits of AI are signifi- cant, there are also concerns surrounding its capabilities and potential, posing risks particularly in terms of privacy and other established human rights. (Madiega, 2024). Defining AI can be seen as a complicated matter due to the absence of a universally accepted definition within the scientific community. (Madiega, 2024). While some organizations offer concise definitions of AI (e.g. Swedish Medical Product Agency, 2023), others provide broad definitions covering more basic functions in the AI do- main (e.g. AI HELG, 2018). What could be considered as a more universal standard is the definition made by the OECD (Organization for Economic Co-operation and Development) which also lays the foundation for the European Union definition (Madiega, 2024). The OECD definition is as follows: 5 An AI system is a machine-based system that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. Different AI systems vary in their levels of autonomy and adap- tiveness after deployment. (OECD, 2024) This definition is accepted and encouraged by the Swedish national AI group, AI Sweden (AI Sweden, 2022). As this thesis research is conducted within the Swedish context, the OECD definition is what is referred to when AI is mentioned. While the term AI originally revolved around the concept of machines attaining hu- man-level intelligence, today, AI's utilization often centers on machine learning (ML). In basic terms, ML runs on an algorithm which registers inputs of data and creates predictions based on what that data contains (Helm et al., 2020). DL (Deep Learning) is an increasingly utilized subgroup of ML which instead analyzes in sev- eral layers, thus being able to register patterns at a more complex level (LeCun et al., 2015; Helm et al., 2020). Within healthcare, this specific type of AI has proven great potential, encompassing the capability to outperform physicians in imaging such as cancer detection (Holzinger et al., 2019). The AI lifecycle encompasses a series of stages that guide the development and de- ployment of artificial intelligence systems. De Silva and Alahakoon (2022) intro- duced the "CDAC AI Life Cycle," a comprehensive framework encompassing dis- tinct phases; design, develop, and deploy and 19 constituent stages across those phases that span from conception to production, applicable to any AI initiative. Fur- ther, Saltz (2024) introduced six main stages of the AI life cycle, which include Prob- lem Definition, Data Acquisition and Preparation, Model Development and Training, Model Evaluation and Refinement, Deployment, and Machine Learning Operations. From initial problem definition and data collection to model development, deploy- ment, and continuous monitoring, each phase ensures that AI solutions are robust, efficient, and aligned with ethical standards. 2.4 AI utilization in healthcare AI is already utilized in a variety of fields in healthcare, improving efficiency and treatment of patients. Diagnosis of diseases and cancer is one area AI has shown high potential, especially within radiology where AI has proven to be more accurate than today’s physicians (Esteva et al., 2017; Lee & Yoon, 2021; Johnson et al., 2021). This has led to future predictions that the radiologist profession will face major chal- lenges, with some foreseeing a complete replacement by AI. Other than being uti- lized in diagnosis, AI can further offer personal assistance for patients, unloading 6 work for nurses and improving operational management of the hospital and its work- ers (Lee & Yoon, 2021). While AI provides opportunities to improve healthcare in various ways, it also, in its current state, introduces new ethical and social challenges (Mohammad Amini et al., 2023; Čartolovni et al., 2022; Lee & Yoon, 2021). One problem is the lack of trans- parency of AI (Pawar et al., 2020), a phenomenon which often is referred to as the black box problem (von Eschenbach, 2021). The term black box refers to the inner workings of AI being opaque for outside observers (von Eschenbach, 2021), which could inflict harm for its reliability and trust (Peterson et al., 2022; Čartolovni et al., 2022), complexify the accountability for its actions (Swedish Medical Product Agency, 2023; Čartolovni et al., 2022) and risk providing unequal and discrimina- tory services (Swedish medical Product Agency, 2023). One solution for this issue has been presented as explainable AI (XAI), which provides, like the name denotes, explanations for the outputs (Swedish Medical Product Agency, 2023; Pawar et al., 2020). XAI also possesses the capability to trace what data acts as basis for its pre- diction which enhances the ease of identifying both solutions and errors (Pawar et al., 2020). 2.5 Digitalization and AI utilization in Swedish Healthcare The general digitalization of the Swedish healthcare sector is expanding. In 2022 almost all documentation done within the healthcare sector is digitized. In 2016, healthcare actors were allowed to offer consultation digitally, and recently, with Covid-19 as a strong enabler, physical health centers have also begun to provide these e-services (Janlöv et al., 2023). In 2016, a nationwide strategy was also put in place called 'eHälsa 2025’, with the goal of digitizing the healthcare sector to become the best in the world in providing e-health towards equal health outcomes for all citizens. (eHälsa 2025, 2016; Nordic Innovation, 2022: Janlöv et al., 2023). Even though the digitization and digitalization of healthcare are improving, several underserved areas exist such as the lack of data transferability between regions, the absence of data standards and interoperability (Janlöv et al., 2023; Nordic Innovation, 2022) and a deficient regulatory environment surrounding e-health (Ekman et al., 2022). And as interoperability proves valuable for AI development (Williams et al., 2023 Lehne et al., 2019), and a lack of regulations can restrict AI development (Dav- enport & Kalakota, 2019) these challenges also impact the integration and utilization of AI. Although hurdles exist, AI is currently being integrated into various fields of Swe- dish healthcare in forms of projects and small-scale utilizations. For example, AI is 7 used to combat sepsis in an emergency care unit (Capio S:t Görans Sjukhus, 2023), detecting lung cancer (Karolinska Universitetssjukhus, 2023), locating healthy or- gans when conducting radiation therapy (Björnheden, 2022) and preventing fall re- lated injuries (Holm & Benjaminsson 2023). But as the future presents significant challenges for Swedish healthcare, including a growing elderly population and an increased demand for medical professionals, the necessity for AI is further pro- nounced (AI Sweden, n.d.). 2.6 Current Regulation of Health Data in Sweden Sweden is known for maintaining a robust legislative environment governing the collection, processing, and dissemination of health data, encompassing various na- tional and European Union (EU) regulations (Andersson et al, 2021). The General Data Protection Regulation (GDPR) (European Parliament & Council, 2016), a cor- nerstone regulation of data protection in the EU including Sweden, plays a crucial role in providing guidelines and governing the protection and handling of personal data, including health information (Mohammad Amini et al., 2023). The Medical Device Regulation (MDR), is another significant EU level regulation affecting Swe- dish healthcare which ensures the safety and performance of medical devices, in- cluding AI-based technologies. MDR mandates stringent requirements for clinical evaluation, post-market surveillance, and transparency in medical device use, aiming to enhance patient safety and trust in healthcare innovations (Essén et al., 2022). The regulation's relevance to AI development is significant, as it establishes a structured framework for the safe integration of AI technologies in medical devices by helping to manage the inherent risks associated with AI, ensuring that AI-driven health solu- tions adhere to high standards of reliability and effectiveness (Fraser et al., 2023). In addition, Sweden has a number of national laws regarding healthcare data man- agement. The Patient Data Act (SFS 2008:355) (Socialdepartementet, 2008) pro- vides comprehensive guidelines on handling patient data in healthcare systems, em- phasizing core principles of confidentiality, patient autonomy, and data security, en- suring appropriate handling of sensitive medical information. The Health and Medi- cal Care Act (SFS 2017:30) (Socialdepartementet, 2017), provides guidelines and governance for the organization of healthcare services such as the management of health data for patient care purposes. Furthermore, the Act on Ethical Review of Research Involving Humans (SFS 2003:460) (Utbildningsdepartementet, 2003) mainly governs the use of patient data in the research purposes. It covers the ethical aspects of using human subjects in research ensuring the protection of participants' rights and welfare while participating in medical research contexts. These regula- tions collectively work on safeguarding the patient data in the Swedish healthcare 8 environment while facilitating lawful use for healthcare delivery, research, and pub- lic health initiatives. 2.7 Regulation of AI in healthcare Although AI already has proven to show success in healthcare and is perceived to have a promising future, what AI affords also brings concerns as it could inflict harm on people’s rights and data privacy. These concerns have resulted in the European Union proposing an AI Act as an attempt to govern the currently underregulated field. The act offers a risk-based solution, which categorizes AI’s different areas of usage in different levels of risks ranging from unacceptable risks down to low or minimal risks (European Union, 2023; Mezsaros et al., 2022). The AI Act, alongside other regulations like EU's GDPR, influences the relationship between AI and healthcare. While on one hand it offers rights and protection for the public (Mezsaros et al., 2022), it also can complicate the regulatory landscape, restricting the innovative ca- pabilities of AI (Konttila & Väyrynen, 2022). At the same time, under-regulation poses a significant challenge for AI within the healthcare sector, both in terms of patient safety (Čartolovni et al., 2022), and also for the progression of AI itself (Davenport & Kalakota, 2019). In Sweden, there are currently no laws covering the secondary usage of health data towards innovation, nor an agency to distribute health data (Nordic Innovation, 2022; Socialdeparte- mentet, 2024a). Furthermore, Swedish regulation does not cover any need of data standards nor even the requirement of digitizing health data in the first place (Nordic Innovation, 2022). The regulatory landscape of digital care is seen to be “largely weak”, which could severely affect future implementations of digital means (Ekman et al., 2022). 2.8 Conducive policy environment for AI in healthcare In fostering the development of AI within healthcare, a conducive policy environ- ment plays a critical role in many aspects by ensuring that AI technologies are effec- tively integrated into healthcare systems to enhance patient care, streamline pro- cesses, and advance medical research (Topol, 2019; Gasser et al., 2020; Flahault et al., 2017). A supportive policy environment is capable of providing clarity and guid- ance on crucial aspects such as data privacy and protection, regulatory compliance and standards and ethical considerations, thus boosting the confidence among stake- holders by mitigating risks associated with the AI development and deployment within healthcare settings (Srivastava al., 2023; Gasser et al., 2020). Furthermore, 9 supportive policy environments have the potential to incentivize significant invest- ments in AI research and development through mechanisms such as funding oppor- tunities, tax incentives, and regulatory exemptions. These incentives enable the col- laboration between public and private sectors, facilitating the translation of AI inno- vations from academic research settings to real-world clinical applications (Topol, 2019; Sittig & Singh, 2010). Furthermore, a conducive policy environment fosters trust among healthcare provid- ers, patients, and other stakeholders by promoting transparency and accountability in AI applications. This is achieved by mandating rigorous evaluation and validation processes, as well as mechanisms for monitoring and addressing biases and errors inherent in AI systems, which ultimately contributes towards widespread acceptance and adoption of AI technologies in healthcare (Gasser et al., 2020; Floridi et al., 2018). Therefore, a supportive policy environment is a crucial enabler for the devel- opment of AI in healthcare (Flahault et al., 2017). 10 3 Conceptual Framework This chapter begins by explaining the conceptual model we developed and is fol- lowed by theory for each of the identified initial policy areas. 3.1 Conceptual framework for a conducive policy en- vironment facilitating AI innovation within healthcare A theoretical framework serves as the cornerstone of a study, providing a structured foundation upon which the research is built. Furthermore, it facilitates the data col- lection, data analysis, interpretation of research findings and ultimately guides re- searchers to contribute meaningfully to the field of study (Rocco & Plakhotnik, 2009). The authors were keen to look for a suitable theoretical framework in under- taking the research about the areas of a conducive policy environment for AI devel- opment in healthcare. However, due to the unique nature of the study, none of the pre-existing theoretical frameworks tailored to the specific area of the study. There- fore, the authors opted to develop a conceptual framework drawing onto the already researched areas of AI development, digital innovation and regulations in the healthcare sector. Following an extensive review of literature, we found five distinct areas of policies that we concluded were crucial to ensure the facilitation of AI innovation in healthcare. These five initial policy areas are: Data sharing and Open data, Data qual- ity, Data ownership and Access control, Data Standardization and Interoperability and Regulatory Experimentation. Table 1 below outlines the literature references which served as candidates for selecting the areas of a conducive policy. Table 1. List of identified policy areas and references Policy Attribute References Data Sharing and Open Data Nordic Innovation, 2022 Hashiguchi et al, 2022 Lee & Yoon, 2021 11 Alhassan et al., 2018 Data Quality Sambasivan et al., 2021 Ng et al., 2023 Isgut et al., 2023 eHälsomyndigheten, 2020 Janlöv et al., 2023 Data Ownership and Access Control Sundblad. 2018 Nordic Innovation, 2022 Forcier et al., 2019 European Commission Report (2018) Data Standardization and Interoperability Williams et al, 2023 Lehne et al, 2019 Isgut et al, 2023 Akinola & Telukdarie, 2023 E-Hälsomyndigheten, 2023 Nordic Innovation, 2022 Socialdepartementet, 2020 eHälsa 2025, 2016 Regulatory Experimentation. Attrey et al., 2020 Truby et al., 2022 Buckleyet al., 2020 EU Council, 2020 Ranchordás, 2021 Smuha, 2021 OECD, 2023 IMY (2024) 12 This conceptual framework (see figure 1) of policy areas of a conducive policy en- vironment will act as a guideline during the data collection and data analysis phases of the research. In the following sections these areas are presented in detail. Figure 1. Conceptual framework with policy areas of a conducive policy environ- ment facilitating AI innovation within healthcare. 3.1.1 Data Sharing and Open Data In order to harness the full potential of AI-based solutions such as decision support systems within the healthcare sector, there's a recognized need for a growing volume of diverse data throughout their development and deployment stages (Down et al., 2018). These data encompass various types, including electronic health records con- taining both non-personal and personal information, which significantly influences the efficacy of technologies like AI systems (Miotto et al., 2018). However, given the sensitivity, it is known that access to health data is not straight- forward specifically for secondary usages such as being used in AI (Alhassan et al., 2018). The access and utilization of sensitive health and medical data concerning patients are heavily regulated and protected by diverse national and international pri- vacy and protection laws and regulations (Gulbrandsen et al., 2016). Consequently, a dilemma emerges, wherein the controls established to protect patients may impede opportunities to enhance the healthcare landscape with cutting-edge technologies such as AI (Kemppainen et al., 2019). Therefore, scholars and practitioners highlight the necessity for legislative frameworks that facilitate health data access. In addition, 13 ongoing deliberations are exploring the feasibility of implementing open data initia- tives, potentially through the development of dedicated health data platforms such as European Health Data Space (Directorate-General for Health and Food Safety, 2022) and Federated-Health: A Nordic Federated Health Data Network (Nordic In- novation, n.d.). The Swedish government has recently invested in improving the health data infra- structure with the aim that health data can be accessed no matter region, municipality or field of healthcare. While the primary goal of this initiative is to improve safety by ensuring patients can access their health data wherever they are, the infrastructure also improves the ability to transfer data towards innovation (Socialdepartementet, 2023a). However, as of right now, gathering health data towards innovation lacks support, as it is not covered by law (Nordic Innovation, 2022), no agencies are re- sponsible for allocating or distributing health data, and there is no national digital health data infrastructure. Although, the Swedish government have started to explore this area, as the European union have put forth a proposal regarding establishing a health data agency to facilitate transfers of health data, with innovation as a second- ary usage (Socialdepartementet, 2024a). 3.1.2 Data Ownership and Access Control AI technologies in healthcare are heavily dependent on large quantities of personal data sourced from medical records in health information systems (Miotto et al., 2018). However, considering the sensitivity and privacy protections required, individuals tend to show substantial resistance in sharing those data, making it hard for AI tech- nologies to get access to essential data in the medical and clinical contexts (Sundblad. 2018). In order to formalize data ownership and access provision to health data, legislation could play a vital role in two fundamental forms. Firstly, regulations are necessary to guarantee the security and protection of personal data, thereby ensuring individu- als that they can safely consent to share their data for the further advancement of healthcare systems. Secondly, regulations are also needed to facilitate secure and traceable access to data sources within healthcare information systems (Forcier et al., 2019). Legislators in the European region are leveraging regulations to tackle the complex- ities of personal data protection and formalize access, particularly in areas of sec- ondary data usage such as AI (Forcier et al., 2019). Given the sensitivity of the data in the medical realm, the regulations act in two main forms on the secondary use of such data. Firstly, the regulations aim at strictly preventing unconsented secondary use of personal data by both private and public sector organizations. Secondly, the regulations formalize the access to sensitive personal data while carefully ensuring 14 privacy, opening up the avenues for data access for research (European Commission, Directorate-General for Justice and Consumers, 2018). In Sweden, healthcare providers are made responsible for collecting, processing and storing patient data by the Swedish Patient Data Act (Patientdatalagen), which reg- ulates the handling of patient data to ensure confidentiality, integrity, and availability (Socialdepartementet, 2008). However, provisions have been made to make healthcare data available for research and innovation purposes through regulations and frameworks such as Swedish Ethical Review Act (Lag om etikprövning av for- skning som avser människor), which governs the ethical review of research involv- ing human subjects, including the use of their data (Janlöv et al., 2023). Additionally, Sweden has the Swedish Data Protection Act (Dataskyddslagen) in effect which pro- vides guidelines regarding the processing and sharing of personal data, including healthcare data (Riksdagen, 2018). 3.1.3 Data Quality When using data for decision making, the quality of that data is an essential factor for more accurate decisions. Data quality is not solely dependent on one type of met- ric but a combination of several dimensions. Although research includes a variety of theories on what dimensions data quality consist of, a few consistently highlighted key dimensions are relevance, consistency, accuracy, timeliness and completeness. This means data deemed to be of high quality should not only be accurate, but it should also contain the essential contexts, not been altered or changed, be up to date, and be relevant for the intended area of implementation (Wang et al., 2023). Data quality matters for AI in healthcare as inputs of data lacking the proper contex- tual information could impact the learning process of AI, resulting in incorrect pre- dictions and bias (Sambasivan et al., 2021; Ng et al., 2023; Isgut et al., 2022). These incorrect outputs could result in harm for patients, with even greater risks for people in high-risk environments where an accurate diagnosis could mean the difference of life and death (Sambasivan et al., 2021). Poor quality data could stem from the inadequacies found within the majority of today's available datasets (Ng et al., 2023), as well as from inadequate data practices. Research has shown that lacking data deriving from practices undermining the im- portance of data has a high commonality and could be close to impossible to find in hindsight. Although it is avoidable if you enact in early preventive procedures (Sam- basivan et al., 2021). Healthcare in Sweden does not have a national health data system. This leads to the possibility of different systems being utilized between regions, but also within re- gions as both public and private actors coexist (Janlöv et al., 2023). Neither is there 15 any national data standard in healthcare (E-Hälsomyndigheten, 2023), meaning there exists a potential inconsistency in medical terminology, especially as segments of patient records are simply entered as free text without structure (Janlöv et al., 2023). This absence of standardization could impact data quality which would potentially compromise the accuracy of AI predictions developed using such data (Isgut et al., 2022; Lehne et al., 2019). At the same time, The Swedish e-health agency, eHäl- somyndigheten (2020), states that structured data is favorable for developing AI, and stresses the necessity of a large amount of high-quality data for AI to be successful. 3.1.4 Data Standardization and Interoperability Although data interoperability could vary in different levels of complexity, it can be summed as when a type of data can be exchanged through different systems (Lehne et al., 2019). Standardization of data is a way to reach this interoperability as it ena- bles transferal between datasets (Dzale Yeumo et al., 2017). For AI in healthcare, data interoperability is especially important (Williams et al., 2023 Lehne et al., 2019), as data is a necessity for the development of AI (Borgogno & Colangelo, 2018; Lehne et al., 2019). Utilizing data interoperability in healthcare would mean fully coherent and seamless data interchanges between different healthcare institutions (Akinola & Telukdarie, 2023), which further would enable fueling more data to- wards AI advancement (Williams et al., 2023; Lehne et al., 2019). However, because of a variety of reasons such as a poor infrastructure, lack of data quality (Williams et al., 2023) and non-standardized efforts (Lehne et al., 2019), the current healthcare landscape lacks the abilities to support a such an endeavor (Williams et al., 2023). Without a standardization of data, AI also faces the risk of running on data of poor quality, potentially missing contextual information leading to low quality predictions as a consequence (Isgut et al., 2022; Lehne et al., 2019). Although the Swedish government have underlined the importance of data interop- erability in healthcare to enable data transfers, as of right now, there are no obliga- tions in using standards for data in healthcare, nor even digitizing it. So, when a transferal of a patient’s health data occurs across different healthcare institutions, it is either goes through the same health data system (if the receiving end shares the same system), or it has to go through another specific system made just for transfers (Nordic Innovation, 2022). The Swedish e-health agency, E-Hälsomyndigheten, acknowledges this problem and is currently in the works of adding standards (E- Hälsomyndigheten, 2023). The eHälsa 2025 strategy, a national vision formed in 2016 aiming to improve health and social care by digitization (eHälsa 2025, 2016; Nordic Innovation, 2022), also stresses the importance towards of standards and in- teroperability of information within healthcare (Nordic Innovation, 2022; eHälsa 2025, 2016; Socialdepartementet, 2020), although health data is not specifically mentioned (eHälsa 2025; Socialdepartementet 2020). 16 3.1.5 Regulatory experimentation It is known that regulatory barriers are one of the major impediments in fostering innovation in most industries with no exception for AI in healthcare. The rapid ad- vancement and intricacies of AI technologies have outpaced traditional regulatory frameworks, making it challenging for them to stay current (Truby et al., 2022). Reg- ulatory experimentation has emerged as a promising approach towards fostering in- novation. It involves establishing a controlled environment where researchers, poli- cymakers, and innovators test and observe new innovative products and services at a small scale before rolling out to a wider audience (Attrey et al., 2020; Truby et al., 2022). The concept of ‘regulatory sandboxes’ has gained special attention of the authorities and analysts as one of the regulatory experimentation tools for legally enabling real- world experiments (Attrey et al., 2020; Buckleyet al., 2020). Additionally, the Coun- cil of the European Union has underscored the potential of regulatory sandboxes as a legislative instrument, emphasizing their flexibility and capacity for experimenta- tion to address the future regulatory challenges associated with innovation (EU Council, 2020). In light of the successful deployment of regulatory sandboxes within the financial sector, there is growing anticipation among experts that similar frame- works could yield positive outcomes in fields such as healthcare and AI, and they highlight the need for adaptive measures that facilitate innovation while ensuring legal certainty and regulatory flexibility (Ranchordás, 2021; Scherer, 2016). Nevertheless, academic scholars have highlighted several potential risks associated with the implementation of regulatory sandbox approaches in the context of AI (Ran- chordás, 2021; Smuha, 2021). These concerns encompass the potential for innova- tors to evade accountability both during and subsequent to the testing phase, the risk of compromising safety standards in real-world applications, and the potential expo- sure of participants' personal data (Truby et al., 2022). Concurrently, the European Union's current proposal for AI regulation underscores the significance of incorpo- rating regulatory oversight of sandbox environments for AI technologies (Smuha, 2021; Truby et al., 2022). Due to the potential of playing a vital role in enabling new technologies to be tested in a controlled and constantly monitored environment, regulatory sandboxes have been identified as a promising tool for conquering legislative hurdles specifically in the areas of healthcare and AI (Attrey et al., 2020; Buckleyet al., 2020). European countries such as Norway and Germany are already introduced regulatory sandbox environments with the aim of promoting ethical, privacy-friendly and responsible innovation within AI (OECD, 2023). 17 The Swedish Authority for Privacy Protection (IMY) recently completed a regula- tory sandbox pilot project titled "Safety Measurements in Public Environments using IoT Technology," showcasing promising results. The initiative aimed to assess leg- islative implications associated with data collection using LiDAR sensors in public places for safety monitoring purposes. Within the sandbox framework, two key legal considerations were examined: firstly, whether data gathered through LiDAR sen- sors falls under the purview of data processing regulations outlined in the General Data Protection Regulation (GDPR); and secondly, how this data collection aligns with provisions outlined in the Swedish Camera Surveillance Act. As noted by IMY (2024), IMY's experience with the regulatory sandbox underscored the importance of interdisciplinary collaboration towards helping regulatory authorities in under- standing and implementing legal frameworks within the context of innovative initi- atives (IMY, 2024). However, we were unable find references for regulatory sand- boxes in Sweden for either healthcare regulation or AI regulation. 18 4 Method The method chapter states how the study was conducted. This includes the formation of the interviews, the background experiences of the interviewees, the selection pro- cess, the transcription, and finally, the data analysis. This study was conducted upon qualitative research, which compared to quantitative research, focuses on individuals' experiences rather than countable measurements. Qualitative research is seen to be especially productive within complex areas, such as the healthcare setting (Fossey et al., 2022). 4.1 Interviews The interviews used a semi structured approach following an interview guide, as it is considered suitable when doing qualitative research (Fossey et al., 2002). A semi structured interview lets the interviewer explore outside of the interview guide (Fos- sey et al, 2002; Grossoehme, 2014; DiCicco-Bloom & Crabtree, 2006), which we considered was a crucial aspect because of both the complex nature of healthcare and also the diversity of interviewee backgrounds. Follow up questions and adaptability were deemed necessary to capture the variety of insights from each interview. The interviews lasted from around 20 to 40 minutes and were all conducted and rec- orded via Zoom, an online videocall software which differs from the traditional face to face type of interview. The online way of conducting interviews, which has heav- ily increased since the Covid-19 outbreak, brings both opportunities and challenges. While face-to-face interviews could help catch nonverbal ques and other location- based context, online interviews increase availability by being not bounded to a cer- tain location (de Villiers et al., 2021). For our study we considered online interviews as more practical as our interviewees were located in different parts of Sweden. We also consider our research topic to be relatively independent of the location of the interviewees. Additionally, to capture nonverbal cues effectively, we consistently utilized video calls to observe facial expressions and body language. 4.1.1 Selection of interview participants As healthcare is an environment with many different stakeholders, we decided to interview candidates with wide ranges of backgrounds to capture a broad spectrum of perspectives on the matter. This scope included a variety of parties such as physi- cians, health-tech employees, researchers, examiners and professors in academia. 19 The majority of interviewees were found on the internet and the rest were found either as referrals from interviewees or professional contacts. When using internet to find participants, keywords such as “AI” and “Sjukvård" (healthcare in swedish) were used, which led to finding prominent people and organizations in the field. The selection of possible candidates depended on the factor of having experience or ex- pertise surrounding the utilization of AI in healthcare. Table 2. List of interviewed participants Participant Experience/Background Date Interviewee 1 AI Senior Advisor. Over 20 years in eHealth. Board 03/18 member at several organizations regarding AI and medical informatics Interviewee 2 Consultant with expertise in AI and digitization. 03/18 Project leader for a prior project regarding integrat- ing AI in healthcare. Interviewee 3 Chief Innovation Officer at an eHealth firm. 03/19 Interviewee 4 Senior Innovation Advisor within Health Tech 03/22 Interviewee 5 Senior Researcher, previously data scientist 04/02 Interviewee 6 Chief Physician, chief strategy officer at government 04/09 region level and board member at national AI strat- egy group Interviewee 7 Prior examiner at government agency regarding AI in 04/09 healthcare. Interviewee 8 Head of regulatory compliance at an eHealth firm 04/15 Interviewee 9 Project leader of patient journal system at eHealth 04/19 firm Interviewee 10 Senior AI researcher scientist at health tech firm 04/22 20 4.1.2 Interview guide The first question in the interview guide should preferably be a broad question on the core matter, to help set up an easy mood and make the interviewee feel comfort- able talking (DiCicco-Bloom & Crabtree, 2006). We then decided on the ques- tion ”What do you think are the main key points for AI to succeed in healthcare?”, which we believe is a broad question the participants could answer based on what- ever experience they had. Although, to prevent going too far from the research con- text, before asking our questions, we have included an introduction to the subject to help the interviewees understand better the area we tried to research. This prior con- text was something we further emphasized following the second interview and for- ward, after the first interviewee’s first answer had disjointed from the topic. The first questions also helped us to assess the significance of our initial five policy areas without directly asking and thereby potentially influencing the interviewees. Often the interviewees touched upon one or several of these areas, which meant we could truly understand their importance. This also helped to avoid repetition in the later stage of the interview as the later questions relate to the five initial areas. The interview guide started with an introduction of the study, research context and some broader starting questions, which were followed by questions regarding our initial five policy areas. If there was time over, the interview ended with a more open question asking if they had any further areas of policies they saw as especially im- portant. Although, as it was semi structured interviews, this guide acted as a guide- line and was not strictly followed. 4.2 Transcription To create the transcriptions, interview audio files were processed via an AI transcrip- tion tool called 'Revoldiv’. AI transcription tools can save time, although they are not perfect and can generate errors (McMullin, 2021), so the transcript drafts were then overlooked, revised, and confirmed by the authors. Transcribing is a process where spoken words are interpreted and brought into writ- ten form. The detail of transcription could vary in depth from simply capturing au- dible words to including detailed visual context (Baliey, 2007). Naturalized and de- naturalized transcriptions are two types of this detail level, whereas one focuses on capturing as much as possible, and one limited to what was essentially spoken (Mero- Jaffe, 2011; McMullin, 2021; Nascimento & Steinbruch, 2019). Although, research- ers are not aligned on what type represents which side of this spectrum (Nascimento & Steinbruch, 2019). Nonetheless, the level of detail is a crucial factor in making discoveries which otherwise would get lost. But as transcriptions are interpretations depending on the context of the research, it falls on the authors behind the study to 21 decide what level is deemed necessary to capture the full data (Baliey, 2007; McMullin, 2021). The level of transcribing detail in this study will be mostly limited to audio as visual cues did not give important context to the provided statements by the interviewees. Filler sounds such as “uhm” and “eh” were neither included in the transcription. Repetitions of single words such as “it, it...” or “The, the...” were included if they were considered to give context in the tone of the statements brought by the inter- viewees. Although, for clarity and readability purposes these repetitions were re- moved when presented as quotes in the findings section, unless it altered the meaning. Lastly, unrelated conversations were excluded. Coding for nonverbal and contextual information could be shaped in variety of ways (Bailey, 2007; Azevedo et al., 2017, Oliver et al., 2006). This thesis has based its transcription codes on codebook presented by Azevedo et al. (2017), although the layout and some codes were revised to be more defined and concise. The transcrip- tion codes (Table 2) include codes for pauses, abrupted sentences, inaudible parts, unclear words, and translations as these were found important for transcribing the interviews of the study. Table 3. Transcription codes based upon a codebook presented in the paper by Azevedo et al (2017) although revised to be clearer and more fit for this paper. Situation Code Example You cannot understand (inaudible) I went to (inaudible) and bought what the person is saying food. You are unsure about ?(unclear word)? I went to ?(a restaurant)? And what the person is saying bought food. There is a silence or pause ... I don’t know... Maybe you are correct. There is an interruption - I always wanted to- by someone else. There is a word in Swe- Swedish word We use AI in the Swedish dish (English transla- sjukvård (healthcare) tion) 22 4.3 Data analysis When analyzing the interview data for the findings section, we followed the thematic analysis procedure by Braun and Clarke (2006). They describe the thematic analysis to be conducted in six steps, whereas the initial stage is reading and getting familiar- ized with the interview transcripts. The authors stress the importance of reading the content several times to avoid missing important information. In the second step, coding the transcripts, it is advantageous not to be conservative when assigning codes to prevent missing out on any potential theme, or in our case, any policy area. It’s further important to cover enough context in the quotes to prevent missing any essential information (Braun & Clarke, 2006). To reduce the chance of individual bias when coding the interview data, it is advised that coding is conducted by more than one person and at separate times (Grossoehme, 2014). In our study, the coding was facilitated in the program ATLAS.ti as it was regarded as a functional coding program with the ability for cloud-based usage, meaning multiple people can work on the same dataset. ATLAS.ti further provides the ability to collect all quotes related to specific codes, which will give a clearer picture of finding certain themes when analyzing the interview transcripts. This made it easier to conduct the third step of Braun and Clarke’s (2006) process, to construct themes by grouping codes together. In step four, these themes are reviewed and compared with the initial coding and data to both dispose of deemed unnecessary and unrelated themes, as well as revising the themes which reflected potential. At step five the themes are given names, and in the final step the findings are included into the report (Braun & Clarke, 2006). During the process of searching and establishing themes, we followed both inductive and deductive approaches, meaning that themes were both found by the data itself (inductive) and from already existing data (deductive) (Braun & Clarke, 2006). The existing data were derived from the initial conceptual framework of five policy areas. It should be noted that these areas act as guidance in the analysis phase rather than final conclusions, meaning that the importance of these areas might not necessarily be shared by the interviewees. To fully capture the insights from the interviews, the inductive approach was utilized to find new policy areas outside of the initial ones. 23 5 Findings This chapter presents the key findings resulting from the systematic analysis of the interview data collected. It outlines the process of deriving six policy areas through thematic analysis, providing insights into the methodology employed. Additionally, it offers a detailed examination of each policy area, presenting in-depth analysis and interpretation. 5.1 Thematic Analysis Following Braun and Clarke’s (2006) six step thematic analysis we extracted a num- ber of themes from the interview data. Following step four and five, the found themes were reviewed and compared to our research question and desired study topic, re- sulting in six policy areas considered essential in a conducive policy environment for facilitating AI innovation within Swedish healthcare (see figure 2). These policy areas are Data Sharing and Open Data Initiatives, Data Ownership and Access Con- trol, Data Quality, Standards and Interoperability, Regulatory Experimentation and finally, Knowledge. For each of the six areas we further noted specifics of what the interviewees discussed in relation to the areas. Figure 2. Six policy areas extracted from the interview transcript data including dis- cussion topics. Following Braun and Clarke’s (2006) six step thematic analysis. 24 5.1.1 Data Sharing and Open Data Initiatives All interviewees favored the concept of ‘open data’ initiatives as well as the require- ment for having initiatives to enable data sharing among various organizations and individuals towards promoting the development of AI in healthcare. However, at the same time many of them raised the challenges of practically implementing open data initiatives. Interviewee 3 specifically discussed the challenges of using open data principles in the healthcare domain due to the sensitivity of data and mentioned her view on the practical meaning of open data. Interviewee 3 envisioned it as: So, for me, actually right now, we don’t talk about open platforms, you get the wrong associations. What we mean, all of us, is that it is data that is not locked in. We don't use the data with proprietary models. As I said, we use open standards as much as possible to make more people be able to manage the data and new information. That is what we mean by open data. Establishing health data platforms as a form of open data initiative, as well as their potential of being useful in the AI innovation process, was also discussed by the interviewees 3, 4 and 9. Interviewee 3 highlighted the importance of such data plat- forms for start-ups to be able to scale their innovative ideas: But perhaps we should talk about health data platforms that can be used for innovation, that can be really emphasizing the potential with incu- bators, small, small companies that have great, great ideas that could really help out the healthcare situation. We need to emphasize that eco- system much more in our ways of working. The concept of ‘data lakes’ has also been discussed by many participants as a way of establishing a large pool of commonly accessible data which could be used by heterogeneous stakeholders with varying use cases. Many participants took Euro- pean Health Data Space as an example project of establishing a common data plat- form or a data lake for health data in the EU and highlighted the potential usefulness not only for the AI development, but also for many other use cases. Interviewee 6 and 8 highlighted the potential usefulness of European Health Data Space for inno- vation as below respectively: And we also have the European Health Data Space that will be ratified soon enough with the aim to enable small companies and researchers and others to access data to not only develop Al or innovate in Al ap- plications, but to make use of health data in general. (Interviewee 6) I mean, the European health data space was going to be the one I would focus on there, because it has as one of its benefits is being able to 25 create a sort of a data lake that medical device manufacturers and other deployers or manufacturers of health tech systems could use. (Inter- viewee 8) Further, interviewee 4 discussed an ongoing project on establishing a common data lake within the Nordic countries, initiated by an organization called Nordic Innova- tion. In addition, Interviewee 2 discussed the huge potential of the large collection of healthcare data stored in the Swedish Quality Registers being valuable for various applications, including the development of AI algorithms. Interviewee 2 further stated that data in the Swedish Quality Registers could act as a foundation for data sharing such ‘open data’ initiatives while ensuring the proper data governance and controls. While discussing the open data initiatives within the healthcare domain, participants commonly highlighted the role of regulators in establishing security of such large open collections of sensitive data, as well as the need for proper governance of access to data. Further, interviewee 6 highlighted the practical importance of having com- mand data standards when sharing data and establishing data platforms. 5.1.2 Data Ownership and Access Control Nearly all interviewees emphasized the requirement of provisioning access to health data stored in the systems of various healthcare institutions for further research and development activities aiming to enhance the overall patient care. According to the participants, neither authorities nor the general public has recognized the true poten- tial of the locked health data kept in healthcare systems in Sweden. Due to the chal- lenges of accessing required health data for developing new technologies including AI, some suggested the generation of synthetic data as a viable alternative. Inter- viewee 1 stated it as; Would it be a path to go through synthetic data? Could we make syn- thetic data from our quality registers on the fly or from other purposes from the healthcare regions and so on? Because if we could do that it might be that we could give the researchers and the companies access to more data to train their models and perhaps to be more specific what type of data do we need from you because as today you know we give everything and they spend all their time and all their money on curating the data, making them work for the algorithm [...]. Interviewee 4 discussed an alternative approach to address the issue of transporting data outside of the healthcare systems while still being able to train AI algorithms. Interviewee 4 proposed to instead relocate the algorithms into each institute and train the algorithm there: 26 But to me, I think the question of moving the data around is increas- ingly becoming... I mean, a hopeless thing to do. I think we need to move the Al algorithm. So that the Al algorithm can train on data, it can look at the data, but it can't bring the data with it out... of the silo. Another interesting idea emerging from the interviews was the potential of monetiz- ing the value of health data by introducing economic models for provisioning access to data. Interviewee 2 stated it as: Or should you have some kind of, like, economic model to give? Be- cause it's hugely expensive to develop this kind of data amounts. So should you have a payment scheme or stuff like that? Along the same topic, Interviewee 4 discussed a study conducted by a private organ- ization in order to highlight the economic potential of patient data held in the healthcare systems. And then there's the business case, which is done by Ernst & Young, which at least gives you some indication of what sort of money we are talking about if we actually manage to share health data. According to the interviewees, the reluctance to share data or to provide access to data can either be a consequence of overregulation or lack of awareness about the regulation. While it is generally known that health data can be accessed for research purposes, Interviewee 6 stated it is a cumbersome process and that the data is not allowed to be utilized for commercial purposes. In addition, many of the participants highlighted the lack of clarity in who owns the health data (either individuals or healthcare institutions) needs to be addressed in order to enhance the willingness and clarity on sharing and receiving access to data. The complex structure of the Swedish healthcare environment makes it difficult for institutions in need of health data access for different use cases including AI utilization. This notion was highlighted by Inter- viewee 2; But then you have the bigger challenge of the Swedish government model where you have a distributed organization where the municipal- ities have some tasks that they're responsible for, the regions have some, then the state level is like regulating but it's also controlling. They were, we built like quite a complex framework. And this is like a bigger, big- ger question than developing of Al applications, but, but it also comes into effect when you want to develop Al applications. 27 Interviewee 9 highlighted that current legislation requires modification to make the patients comfortable in sharing their data as well as include provisions for secondary use of health data. Interviewee 9 states that as: And the Patient Data Act is aimed towards care providers and it has a lot of different provisions about what data you're meant to collect and, you know, in terms of quality and so on. And the main, you know, the factor that I would highlight there is that we as medical device manu- facturers, we will often be dependent on that data in its secondary use. However, it is collected for a different use. I mean, it's collected in or- der to give the best quality care to the individual patient. Overall, all the participants highlighted the utmost importance of having access to health data held in various healthcare systems and the role of legislation to facilitate the extraction of the value of such data masses for enhancing overall patient care. 5.1.3 Data Quality Data quality was an aspect which was considered important by the interviewees to provide safe and unbiased predictions. Although, while the interviewees underscored the necessity of good quality, interviewees also clarified what is deemed high quality in health data solely depends on the objective, as stakeholders of health data have different goals when utilizing it. This also leads to different ways of gathering data and thus different types of data sets. Interviewee 3 brought further information to this topic stating that data deriving from one setting might not translate well into another. Interviewee 6 further added that health data which has been gathered as secondary use, could have originally been collected with another purpose rather than the development of an AI. This primary data purpose is important to be aware of, as it allows fully grasping the context and quality of the data. Interviewee 6 then further shared: Yeah, so the first step is to actually know your data and have access to the rawest form of the data because every time you move the data or tweak it in some way, it's, you either lose information or you lose meta information about the data like the amounts of missing or the intended purpose for collecting it in the first place. [...] And if the primary in- tended use for the data was not to, to develop a new Al tool, then you need to understand what the primary collection was for so that you can take that into account when you develop your model. Another part of data quality the participants referred to was perfect data. This type of data fails to contain the imperfections of the real world, and thereby is unable to accurately represent it. Interviewee 2 referred that overcoming this problem is a “key 28 point” towards AI implementation, with interviewee 6 stated to fix this problem was to use “real world data”. Regarding Sweden, both interviewee 2 and 4 considered the Swedish health data to be of good quality, and 4 and 5 further stated Sweden has data which spans through- out a long history. Interviewee 2 stated this mass of data, referred to as Swedish quality register, could be a “gold mine for potential AI developers”. Interviewee 1 further saw potential in these quality registers in the shape of synthetic data, which could fuel companies in producing AI tools for healthcare. Although accessing this data is limited as interviewee 2 stated: But in the same time, it is data that basically is owned by Sweden or like the Swedish healthcare system. And like, then you have to reflect upon the strategy from the people. It's mainly Socialstyrelsen (National board of health and welfare) and that owns these data masses. But it's mainly owned by the Swedish people in that case. And should you let innovation free and let private companies exploit this data to develop services? Or should you have some kind of, like, economic model to give? Because it's hugely expensive to develop this kind of data amounts. [...] So that's the second thing I would say that's like the key point is like access to good quality data. Interviewee 9 discussed the lifespan of the data as a further aspect, debating how long the quality of the data could last. As the healthcare practices and knowledge progress, the interviewee could see that training AI on this type of data would result in poor predictions. Lastly, the case of bias was a further topic participants shared their thoughts on. There were concerns regarding AI running on too homogenous data, failing to de- liver an equal healthcare for all parts of the population, leaving more marginalized groups at a higher risk of being exposed to poor AI predictions. Interviewee 5 ex- plained: And the perhaps, one of the obvious things [regarding ethics] is like it has to be unbiased towards minority groups, towards different genders, different ethnicities, different, I mean, everything because specifically gender and eth- nicity play a role in healthcare. I mean, we know that more, I mean, some genders or some races or ethnicities are more susceptible to certain diseases. And... therefore, you cannot ignore that part of information, genetics play a role and everything, everything matters. But at the same time, you don't, you need to make sure that you are not using this information in the wrong way. 29 Interviewee 3 pointed out the dangers in potential bias for AI, even if it is only re- garded as a medical counseling tool for medical professionals. The interviewee stated that “it [...] takes quite a lot for an individual or a person to say no and say that the algorithm was wrong”, meaning that to disregard the advice given could be a big hurdle. 5.1.4 Standards & Interoperability Among the participants, there was a cohesive stand for the necessity of standards and interoperability. A reoccurring theme when standards was discussed was the lack of data transferability between the Swedish regions. The independency of the regions have resulted in different IT systems, which as a result compromises the ability to seamlessly send and receive information across regions. Interviewee 2 and 10 ex- plained respectively: You don't have even standardized terms and names for things which makes it much more complex to develop normal digitized tools, but when it comes to Al applications, it becomes even more complex. So l would say that a cen- tral area is working with like the frameworks to get, like, standardized terms and make the region start talking the same language. (Interviewee 2) There's a lot of different systems and they are very different. And sometimes you can communicate very, very little with them. It's like it's kind of hard. You have to do plug-in. So if it was a standardized language, I think that would help a lot [...] (Interviewee 10) While some participants focused on standards in relation to data, others referred to the lack of standards in systems used among the regions. Interviewee 10 saw the lack of system standards as an obstacle, hindering the idea of large-scale collective ac- tions towards data gathering for AI development, stating: So they often have like the different computer system in different part of Sweden and that is of course the big problem when you want to share data. I think that is the biggest problem. If you want this like open Al platform for data, you need to have the regions to have the same system, or at least you need a big region. Interviewee 10 further noted the variety of systems in regions also causes complica- tions for software implementations as each system then needs specialized adaptation for successful integration. A different interoperability related problem was pin- pointed by interviewee 9, which also could occur by the result of having different systems in regions and municipalities. Interviewee 9 explained that if a patient visits healthcare in several regions or municipalities, multiple data of a single patient will 30 exist, causing extra work to configure when data is extracted towards AI develop- ment. The movement towards improving interoperability was explained by the interview- ees not as a consequence of recent AI development, but as a long-time issue. Partic- ipants mentioned several upcoming measures to facilitate standards. On the Euro- pean level, there is a project with the goal of streamlining data, resulting in larger datasets and higher quality, as well as the European Health Data space which aims to standardize health data on a European level. In Sweden, interviewee 6 mentioned the number of EHR's (Electronic Health Record) vendors in the regions will decrease to three or less, further contributing towards the facilitation of health data standards, something that is currently in discussion. 5.1.5 Regulatory experimentation It is known that regulatory frameworks often struggle to keep pace with the rapid advancements in technology and require more creative and systematic ways to align with the rapidly changing tech ecosystems. AI is one of the fastest growing technol- ogies across various domains, including healthcare, leaving the governments around the world in the struggle of effectively attempting to regulate. One commonly highlighted insight from the participants was that Swedish legisla- tion requires updates to match recent developments in technology. Interviewee 2 specifically mentioned it as: But when it comes to other regulatory areas or laws basically, most of them are written for an analog world. And that is something that's still a problem within all of the public sector that you have an enormous challenge when it comes to using data. And it has partly to do with the integrity aspect, but also that the lawmaker hasn't really kept up and rewritten laws even to a digitized world, now we're facing like the next big step possibly that would be like the artificial intelligence entering on a broad scale. And the legislature is far, far behind. Even though no participants had any active experience with regulatory sandboxes, the participants saw the benefits of it, recognizing it as a tool to allow experimenta- tion within a controlled environment where different stakeholders work collabora- tively to evaluate new products and services before being implemented in the real world. Below, interviewee 8 mentioned about the regulatory sandboxes: I like the idea. I'm not very familiar with any ongoing initiatives. But one sort of recurring pain point on this area is that when you are dealing with new technology and you want to bring a product to markets that has never been on the market before, you often have a very pressing 31 need to, well, to, you know, A, to obtain legal advice on, you know, how this, this new product might be regulated and what regulations it will trigger and so on. But you also have quite a pressing need in many cases to discuss it with the public authority that will regulate you, which is, you know, Läkemedelsverket (Swedish Medical Products Agency), the NPA in, in Sweden and many, and it will be the health and social care inspectorate, IVO, if it's in the care provider part of the, of the value chain. And they are notoriously difficult to get. And I mean, I can very much understand them on that, but it's very difficult to get the sort of green light ahead of time from them Referring from the success it made in many other industry sectors, participants be- lieved that regulatory sandboxes as a regulatory experimentation tool could play an important role in enabling the AI innovation within healthcare and discussed the im- portance of the active involvement from regulators in order to successfully imple- menting it. 5.1.6 Knowledge A further obstacle participants mentioned hindered AI development in Swedish healthcare was the gap of knowledge. This problem wasn’t exclusive to one party, but to different stakeholders related to healthcare and AI. The first one was the lack of proper knowledge of AI in the medical setting. Interviewee 6 and 7 both stated that medical personnel such as doctors and nurses could have a faulty perception of AI, misrepresenting what it actually is. This lack of understanding hinders the effec- tive integration of AI into the healthcare sector. Interviewee 7 recalls it as being “a matter of attitude […] among the practitioners [in] how ready [...] they [were] to use the Al tools and to accept them and to put them in practice.” Interviewee 6 said: […] One major drawback thus far is the competence about what Al really is in healthcare among coworkers in healthcare. They were trained as medical doctors and nurses and engineers, but few are enough knowledgeable about what Al really is. So they think Al is what they read in the general papers, which is very far from the practice. So that's one obstacle. Other gaps of knowledge exist with the AI developing party which often was referred by the participants as problems most exclusively for “startups” or “smaller players”. These gaps of knowledge existed either as a lack of medical competence or regula- tory competence. Interviewee 10 stated insufficient knowledge of the medical field could result in hinders of communication between the AI developing party and the healthcare professionals, both as verbal miscommunication, but also related to mis- understanding the data or the quality of data. Interviewee 6 further expanded on this 32 topic explaining that for startups to develop their AI products in a certain medical field, they need both knowledge of the particular field and the medical context of it. Participants also argued the importance for the AI innovating party to understand the legislative landscape surrounding the healthcare sector and AI, as it is a quite exten- sive and complex area. As AI technology is continuously progressing, regulations are also moving to catch up, to both replace older legislations and to cover new areas. Interviewee 8 stated a “degree of predictability” was important for startups to navi- gate the future of legislations. The Medical Device Regulation (MDR) was a partic- ularly highlighted when discussing the regulatory landscape, which both interviewee 6 and 8 commented on: So I met lots of companies, but also researchers and internal innovators that have great ideas. They start with like developing the solution and then it grows down the ditch for one of the other reasons that they didn't think of. And medical device regulations is one… regulatory frame- work that you need to know something about[…] (Interviewee 6) And what we're seeing right now is that to, you know, in my opinion, too many parts of this, you know, of this entire aggregated landscape is being replaced at the same time, which is quite, quite unfortunate, because many players are still in implementing the MDR and many players haven't really gotten that far with that. (Interviewee 8) 33 6 Discussion AI has the potential to significantly enhance healthcare by improving medical diag- nosis and treatment, as well as reducing administrative workloads (Davenport & Ka- lakota, 2019; Habehh & Gohel 2021; Rosenberg et al., 2024). Despite the urgent need for improvements in Swedish healthcare (Sveriges Kommuner och Regioner, 2022; Anskär et al., 2019), AI innovation faces significant hurdles, as highlighted by the literature and insights from industry professionals. The current regulatory land- scape lacks the supportive elements to facilitate AI innovation, thereby restricting its ability to enhance healthcare. In this section we discuss the six found essential areas of policies for constructing a conducive policy environment for AI innovation in Swedish healthcare. After each area is discussed, the section concludes by discussing the identified interconnections between the areas and how these relations could take shape. 6.1 Areas of Policy Initiatives Data Quality is a critical part for the development of stable and productive AI sys- tems. In contrast to previous research indicating that the majority of available da- tasets suffer from poor quality (Ng et al., 2023), health data in Sweden is perceived to be of good quality, primarily due to the Swedish quality registers; however, these datasets face challenges related to limited accessibility. Bias, a consequence of poor data, is a potential danger for marginalized groups, even if it only acts as a counseling tool as it might be hard to disregard for medical personnel, implying that AI incor- porates a strong agency. The five attributes of data quality identified by Wang et al. (2023) — Relevancy, Timeliness, Consistency, Completeness, and Accuracy — were deemed important not only in general use but also in the context of Swedish healthcare. With these dimensions mapped out we can further conclude that while standards may facilitate data quality (Isgut et al., 2022; Lehne et al., 2019), they do not cover all aspects of data quality, as the element of relevancy is inherently unique. The absence of Standardization and Interoperability efforts in Swedish healthcare is apparent. The decentralization of Swedish healthcare (Janlöv et al, 2023), more spe- cifically the lack of cross regional health system standardization initiatives, have re- sulted in poor outcomes for fluent data transfers and data access, but also for scaling endeavors and data platforms. Standardization efforts is seen to be the most signifi- cant reason why the healthcare lacks national interoperability in Sweden (Lehne et 34 al., 2019), rather than poor infrastructure (Williams et al., 2023). Furthermore, stand- ardization of IT systems is also a crucial part for enabling AI innovation, which dif- fers from our related research solely acknowledging the standardization of data. Data Sharing and Open Data Initiatives towards secondary use of health data are challenging due to the sensitivity of health data in Swedish healthcare. Kostkova et al., (2016) emphasizes the need for robust governance frameworks and stakeholder engagement strategies to support a balanced agenda that safeguards personal infor- mation while enabling the use of data to improve healthcare. However, there are proposals for implementing health data platforms within Sweden as well as in the EU to make certain health data available for common use (Nordic Innovation, 2022). In a report by EIT Health Scandinavia and Swedish Medtech (EIT Health Scandina- via and Swedish Medtech, 2023) evaluating Sweden’s readiness for the European Health Data Space across various dimensions, it was identified that governance and legislative aspects in Swedish healthcare require changes. While Sweden is advanc- ing in building the necessary infrastructure, the report highlighted that its legal read- iness still lags behind its technical progress. Furthermore, poor data quality and het- erogeneity of data formats have been highlighted as main hurdles in implementing health data platforms, underscoring the need of high-quality data and health data standards. Martin et al. (2017) and González et al. (2018) emphasize the significance of data quality in effective implementations of healthcare data platforms. Their stud- ies emphasize the adverse impacts of data with poor quality and data format hetero- geneity on the usability and reliability of such platforms. Moreover, as identified during our study, ‘open data’ initiatives could theoretically take a different meaning in the healthcare domain. These hypothetical ideas involve granting access to healthcare systems through specific interfaces without the need to transfer data to centralized platforms. Regarding Data Ownership and Access Control, regulatory frameworks have the po- tential to provide third parties with proper access to health data for innovative pur- poses which would fuel AI development. Right now, there is a reluctance of individ- uals and healthcare institutions to share their data. The reluctance could be caused by fear of losing competitive advantage, the lack of knowledge about current legis- lation, or a lack of proper regulation regarding data privacy and protection. On one hand, these types of protective regulations are important as it would establish confi- dence of the individuals and institutions to consent access to their data, but on the other hand, as highlighted by Christensen & Næss, (2020), it is also a dilemma be- tween providing access to health data for innovation and adhering to data protection regulations like GDPR. It centers on balancing the benefits of AI advancements with the need to safeguard patient privacy. Therefore, ensuring robust data protection while enabling data-driven innovation remains a critical and a complex challenge 35 which could be effectively addressed through appropriate regulation. Furthermore, a solution to mitigate the reluctance of transferring data outside healthcare institutions would be to relocate AI algorithms to various healthcare institutions, allowing the algorithms to be trained on-site with their respective datasets. An area which wasn’t recognized to be in major need of policy initiatives in the initial related research of the conceptual framework was Knowledge of cross-func- tional domains. Two knowledge gaps were found to ultimately hampering AI inno- vation in healthcare, the first one being the lack of sufficient knowledge in AI for medical professionals and the second one being AI developers missing proper knowledge either for the medical field or the regulatory field. Regulatory experimen- tation mechanisms such as regulatory sandboxes could be an answer to tackle poten- tial knowledge gaps. With sandboxes being a tool of bringing several stakeholders together, these knowledge gaps could be effectively bridged. To solve the knowledge gap of AI knowledge more specifically for medical professionals, regulatory initia- tives towards increased competence could be a solution. Facilitating knowledge and competence in the healthcare is not a new concept in Swedish healthcare. In 2023 the government invested 3,2 billion SEK towards competence provisions and devel- opment in Swedish healthcare (Socialdepartementet, 2023b). There is also a knowledge steering system (Kunskapsstyrning) in Sweden with the aim of providing medical professionals with the correct proper knowledge through knowledge sharing and implementation (Socialdepartementet, 2024b). Regarding Regulatory Experimentation, there is a clear absence of regulatory sand- boxes initiatives in Swedish healthcare and AI contexts. However, regulatory sand- boxes offer significant potential for AI innovation in healthcare by providing a con- trolled environment where innovators can test new technologies under regulatory supervision. As highlighted by Attrey et al., (2020), these sandboxes facilitate real- world experimentation, promote collaboration between stakeholders, and help iden- tify regulatory gaps early on. On the same time, potential drawbacks do exist high- lighted by authors like Ranchordás, S. (2021), such as insufficient data protection, potential biases in AI algorithms, and the lack of standardized evaluation metrics in the regulatory sandboxes. To mitigate the previously mentioned challenges, robust regulations are needed to ensure data privacy, promote ethical AI practices, and maintain consistent performance standards. Effective governance frameworks should also be established to oversee sandbox activities and ensure compliance with existing healthcare regulations. 6.2 Interrelation of policy areas During the analysis of the collected data, a notable observation emerged regarding the interconnection between the identified policy areas within the context of AI in 36 healthcare. This realization highlights the interconnected nature of various policy areas and emphasizes the importance of understanding their relationships for effec- tive policy making and implementation. The interconnection sometimes appears as an enabler, where the successful imple- mentation of one policy attribute leads to enabling an area or hindrance which an- other policy attribute is trying to resolve. For instance, implementing policies regard- ing establishing data standards could enable the regulators to resolve some of the hindrances surrounding data quality. Consequently, an argument can be made for the merging of these interrelated policy areas, thereby concurrently addressing the hin- drances of multiple domains. Moreover, certain interconnections appear as explicit dependencies, where the real- ization of one policy area necessitates the successful implementation of another. For instance, the effectiveness of policy initiatives regarding the establishment of open data initiatives or health data platforms relies heavily on the prior implementation of policies governing data standards and data quality. Consequently, an argument can be made for the assessment of dependent areas should precede the implementation of high-level policy areas. To visually depict those interrelationships among policy areas and the hindrance they are trying to address, a diagram (Figure 3) has been developed by the authors, which encapsulates the dynamic interactions and dependencies between these key compo- nents. While developing the diagram, the authors considered four common stages of the AI lifecycle mentioned by Saltz (2024) and De Silva & Alahakoon (2022): Data Access (Data Acquisition), Data Pre-processing, AI Algorithm Development and Training, and AI Deployment. These stages are primarily related to the application of policies. 37 Figure 3. Dynamic interrelationships among policy areas and the hindrances they are trying to address within the AI innovation eco systems. Furthermore, the relationships between the selected stages of the AI lifecycle and the policy dimensions are also depicted in Figure 3. Effective regulation for data owner- ship and access control, along with open data initiatives, can facilitate easier data access for AI developers. Implementing regulations for data standards can signifi- cantly reduce the resources required for data pre-processing during AI development. Additionally, the AI algorithm development and training stage can benefit from proper regulations for data ownership, access control, and the use of regulatory sand- boxes. Ensuring data quality through robust regulations can mitigate potential biases during AI development and deployment. While regulatory sandboxes enhance the 38 knowledge dimension, the overall knowledge aspects play a crucial enabling role throughout the AI development and implementation stages. Given the findings of the relationships among various policy areas, it necessitates the policymakers to take more holistic strategies while implementing policies related to AI development in healthcare. Furthermore, this notion calls for researchers to delve deeper into the interrelationships between the given policy areas and contribute both academically and practically. 39 7 Implications and Limitations Our research offers a theoretical contribution by illustrating how policies can serve as enablers for digital innovation in healthcare. The findings indicate that policies are not only crucial for addressing obstacles that hinder AI innovation and develop- ment but also for promoting and facilitating it. While prior research has predominantly focused on the necessity of policies for im- posing control and governance in the integration and development of AI in healthcare and digital innovation at large, our study adopts a different perspective which inves- tigates the role of policies and regulations in actively promoting and enabling AI development within the healthcare sector. This study makes a significant generic contribution towards understanding how AI innovation can be further facilitated in Swedish healthcare. It showcases the essential areas of policies that policymakers need to focus on if the increase of innovative AI measures is of interest. Collectively, these areas map out how a conducive policy environment for AI innovation would take shape within the Swedish healthcare con- text. There are several areas of interest for future studies. While our study tried to grasp a broader picture of the landscape, a more profound understanding of stakeholders' perceptions can be achieved by focusing on specific groups, such as physicians, reg- ulators, or AI start-ups in healthcare. Additionally, future research should explore how conducive policy environments can be developed in other sectors of society. Understanding the differences and similarities across sectors will provide further in- sights into how policies impact digital innovation. Moreover, from a legislative per- spective, it is essential to delve deeper into each proposed area of policymaking iden- tified in this study, and to determine the specific policy initiatives needed to eliminate barriers and foster AI innovation. Finally, although our study identified six essential policy areas for fostering a conducive policy environment for AI development in healthcare, future research should aim to uncover additional areas or refine the ex- isting ones. This endeavor is crucial for developing a comprehensive framework for policymaking in AI healthcare. 40 Our study faces several limitations due to the extensive and complex nature of the healthcare sector, which involves numerous stakeholders. The sample size of ten in- terviewees, while providing valuable insights, is relatively small given the sector's complexity. Increasing the number of interviews could enhance the robustness and generalizability of our findings. A larger sample size would allow for a more com- prehensive understanding of the diverse perspectives within the sector, thereby strengthening the study's conclusions. Future research should aim to include a broader range of participants to validate and extend our results, ensuring a more rep- resentative and reliable framework for policymaking in AI healthcare. Furthermore, we could see potential for unintentional bias introduced by discussing the five initially identified policy areas during the interviews. Focusing on these ar- eas could potentially steer the direction of the interview away from areas the inter- viewees deem more important. Acknowledging this, we tried to mitigate this limita- tion by allocating time at the beginning and end of the interview for the interviewees to provide their own opinions and reflections. The beginning of the interview was considered especially important as the interviewees yet aren’t aware of what these topics were, mitigating any influence. 41 8 Conclusion The integration of Artificial Intelligence (AI) within healthcare holds immense promise for revolutionizing patient care, diagnostic accuracy, and administrative ef- ficiency. However, realizing this potential necessitates a conducive policy environ- ment that addresses the multifaceted challenges inherent in AI innovation. Our study explored the essential attributes of policy frameworks needed to foster AI develop- ment within the Swedish healthcare. Through in-depth interviews with ten profes- sionals spanning diverse backgrounds in AI and healthcare, we identified six pivotal areas requiring policy interventions to facilitate AI innovation. These areas include Data Sharing and Open Data Initiatives, Data Ownership and Access Control, Data Quality, Standards and Interoperability, Regulatory Experimentation, and Knowledge. Each of these areas plays a crucial role in catalyzing AI innovation within the Swedish healthcare sector. In conclusion, our study underscores the imperative of holistic policy frameworks in stimulating AI innovation within the Swedish healthcare landscape. By delineating the multifaceted dimensions of policy intervention required for fostering AI innova- tion, our findings provide a roadmap for policymakers, stakeholders, and researchers alike to navigate the complex terrain of AI integration in healthcare. It's crucial that future research endeavors and actions taken to address the limitations highlighted in this report contribute to advancing our understanding of the intricate interplay be- tween policy interventions and AI innovation within healthcare. Ultimately, the re- alization of AI's transformative potential in healthcare hinges on the concerted efforts of policymakers, stakeholders, and researchers in navigating the dynamic landscape of AI innovation. 42 9 References AI HLEG. (2023). A Definition of AI: Main Capabilities and Disciplines. European Union. https://digital-strategy.ec.europa.eu/en/library/definition-artificial- intelligence-main-capabilities-and-scientific-disciplines AI Sweden (n.d.) Healthcare. https://www.ai.se/en/sector-initiatives/healthcare AI Sweden. (2022). AI Sweden position paper – AI Act. https://www.ai.se/en/news/ai-swedens-position-ai-act Akinola, S., & Telukdarie, A. (2023). Sustainable Digital Transformation in Healthcare: Advancing a Digital Vascular Health Innovation Solu- tion. Sustainability, 15(13), 1-23. https://doi.org/10.3390/su151310417 Alhassan, I., Sammon, D., & Daly, M. (2018). Data governance activities: a com- parison between scientific and practice-oriented literature. Journal of En- terprise Information Management, 31(2), 300-316. https://doi.org/10.1108/jeim-01-2017-0007 Andersson, J., Nyholm, T., Ceberg, C., Almén, A., Bernhardt, P., Fransson, A., & Olsson, L. E. (2021). Artificial intelligence and the medical physics pro- fession-A Swedish perspective. Physica Medica, 88, 218–225. https://doi.org/10.1016/j.ejmp.2021.07.009 Anskär, E., Lindberg, M., Falk, M., & Andersson, A. (2019). Legitimacy of work tasks, psychosocial work environment, and time utilization among primary care staff in Sweden. Scandinavian Journal of Primary Health Care, 37(4), 476–483. https://doi.org/10.1080/02813432.2019.1684014 Attrey, A., Lesher, M., & Lomax, C. (2020). The role of sandboxes in promoting flexibility and innovation in the digital age. OECD Going Digital Toolkit Notes, 2, 1–26. https://doi.org/10.1787/cdf5ed45-en Azevedo, V., Carvalho, M., Costa, F., Mesquita, S., Soares, J., Teixeira, F., & Maia, Â. (2017). Interview transcription: conceptual issues, practical guidelines, and challenges. Revista de Enfermagem Referência, 4(14), 159–168. https://doi.org/10.12707/riv17018 B Braun, V., & Clarke, V. (2006). Using Thematic Analysis in Psychology. Quali- tative Research in Psychology, 3(2), 77–101. https://doi.org/10.1191/1478088706qp063oa Bailey, J. (2008). First steps in qualitative data analysis: transcribing. Family Prac- tice, 25(2), 127–131. https://doi.org/10.1093/fampra/cmn003 43 Björnheden, H. (2022, June 2). AI hittar friska organ som ska skyddas från strål- ning. Sahlgrenskaliv. https://sahlgrenskaliv.se/ai-hittar-friska-organ-som- ska-skyddas-fran-stralning/ Borgogno, O., & Colangelo, G. (2018). Data Sharing and Interoperability Through APIs: Insights from European Regulatory Strategy. SSRN Electronic Jour- nal, 1–42. https://doi.org/10.2139/ssrn.3288460 Buckley, R. P., Arner, D. W., Veidt, R., & Zetzsche, D. A. (2020). Building FinTech Ecosystems: Regulatory Sandboxes, Innovation Hubs and Be- yond. Washington University Journal of Law & Policy, 61(1), 55–98. https://doi.org/10.2139/ssrn.3455872 Capio S:t Görans Sjukhus. (2023, October 24). Nytt AI-stöd förebygger sepsis på IVA [Press Release]. https://capiostgoran.se/nyheter-pressrum/ny- hetsarkiv/2023/nytt-ai-stod-forebygger-sepsis-pa-iva/ Čartolovni, A., Tomičić, A., & Lazić Mosler, E. (2022). Ethical, legal, and social considerations of AI-based medical decision-support tools: A scoping re- view. International Journal of Medical Informatics, 161, 1-9. https://doi.org/10.1016/j.ijmedinf.2022.104738 Christensen, F., & Næss, P. S. (2020). How the General Data Protection Regula- tion affects Health Information System innovation. University of Adger. https://uia.brage.unit.no/uia-xmlui/bitstream/handle/11250/2677186/Fred- rik%20Christensen.pdf?sequence=1 Ciriello, R. F., Richter, A., & Schwabe, G. (2018). Digital Innovation. Business & Information Systems Engineering, 60(6), 563–569. https://doi.org/10.1007/s12599-018-0559-8 Davenport, T., & Kalakota, R. (2019). The Potential for Artificial Intelligence in Healthcare. Future Healthcare Journal, 6(2), 94–98. https://doi.org/10.7861/futurehosp.6-2-94 De Silva, D., & Alahakoon, D. (2022). An artificial intelligence life cycle: From conception to production. Patterns, 3(6), 1–13. https://doi:10.1016/j.pat- ter.2022.100489 de Villiers, C., Farooq, M. B., & Molinari, M. (2021). Qualitative research inter- views using online video technology – challenges and opportunities. Medi- tari Accountancy Research, 30(6), 1764–1782. https://doi.org/10.1108/medar-03-2021-1252 DiCicco-Bloom, B., & Crabtree, B. F. (2006). The Qualitative Research Interview. Medical Education, 40(4), 314–321. https://doi.org/10.1111/j.1365- 2929.2006.02418.x Directorate-General for Health and Food Safety (2022). Proposal for a regulation - The European Health Data Space. European Commision. 44 https://health.ec.europa.eu/publications/proposal-regulation-european- health-data-space_en Dowd, W. N., Cowell, A. J., Regan, D., Moran, K., Slevin, P., Doyle, G., & Bray, J. W. (2018). An Exploratory Cost-Effectiveness Analysis of the Con- nected Health Intervention to Improve Care for People with Dementia: A Simulation Analysis. Health Services and Outcomes Research Methodol- ogy, 18(1), 47–62. https://doi.org/10.1007/s10742-017-0175-y Dzale Yeumo, E., Alaux, M., Arnaud, E., Aubin, S., Baumann, U., Buche, P., Cooper, L., Ćwiek-Kupczyńska, H., Davey, R. P., Fulss, R. A., Jonquet, C., Laporte, M.-A., Larmande, P., Pommier, C., Protonotarios, V., Re- verte, C., Shrestha, R., Subirats, I., Venkatesan, A., & Whan, A. (2017). Developing data interoperability using standards: A wheat community use case. F1000Research, 6(1843), 1–15. https://doi.org/10.12688/f1000re- search.12234.2 eHälsa 2025. (2016) Vision for eHealth 2025. https://ehalsa2025.se/wp-con- tent/uploads/2021/02/vision-for-ehealth-2025.pdf E-Hälsomyndigheten. (2020) Focusrapport – Artificiell intelligens och ehälsa. https://www.ehalsomyndigheten.se/globalassets/ehm/3_om-oss/rappor- ter/fokusrapport_ai_och_e-halsa_20201124.pdf E-hälsomyndigheten. (2023, March 1). Hälsodata. eHälsomyndigheten. https://www.ehalsomyndigheten.se/om-oss/halsodata/ EIT Health Scandinavia and Swedish Medtech (2023). Implementing the European Health Data Space in Sweden. Retrieved from https://eithealth.eu/wp-con- tent/uploads/2023/11/Implementing-the-European-Health-Data-Space-in- Sweden.pdf Ekman, B., Thulesius, H., Wilkens, J., & Arvidsson, E. (2022). Digitalization of Health Care: Findings From Key Informant Interviews in Sweden on Technical, Regulatory, and Patient Safety Aspects. Journal of medical In- ternet research, 24(9), 1–3. https://doi.org/10.2196/38746 Esteva, A., Kuprel, B., Novoa, R. A., Ko, J., Swetter, S. M., Blau, H. M., & Thrun, S. (2017). Dermatologist-level classification of skin cancer with deep neu- ral networks. Nature, 542(7639), 115–118. https://doi.org/10.1038/na- ture21056 EU Council. (2020). Council Conclusions on Regulatory sandboxes and experi- mentation clauses as tools for an innovation-friendly, future-proof and re- silient regulatory framework that masters disruptive challenges in the dig- ital age. https://data.consilium.europa.eu/doc/document/ST-13026-2020- INIT/en/pdf European Commission, Directorate-General for Justice and Consumers. (2018). New opportunities, new obligations : what every business needs to know 45 about the EU’s General Data Protection Regulation https://data.eu- ropa.eu/doi/10.2838/97649 European Parliament & Council. (2016). Regulation (EU) 2016/679 of the Euro- pean Parliament and of the Council of 27 April 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data, and repealing Directive 95/46/EC (General Data Protection Regulation). Official Journal of the European Union. https://eur-lex.europa.eu/eli/reg/2016/679/oj Essén, A., Stern, A. D., Haase, C. B., Car, J., Greaves, F., Paparova, D., Vandeput, S., Wehrens, R., & Bates, D. W. (2022). Health app policy: international comparison of nine countries’ approaches. Npj Digital Medicine, 5(1). https://doi.org/10.1038/s41746-022-00573-1 Flahault, A., Geissbuhler, A., Guessous, I., Guérin, P., Bolon, I., Salathé, M., & Smith, G. (2017). Precision global health in the digital age. Swiss Medical Weekly, 147, 1-3. https://doi.org/10.4414/smw.2017.14423 Floridi, L., Cowls, J., Beltrametti, M., Chatila, R., Chazerand, P., Dignum, V., Lu- etge, C., Madelin, R., Pagallo, U., Rossi, F., Schafer, B., Valcke, P., & Vayena, E. (2018). AI4People—An Ethical Framework for a Good AI So- ciety: Opportunities, Risks, Principles, and Recommendations. Minds and Machines, 28(4), 689–707. https://doi.org/10.1007/s11023-018-9482-5 Forcier, M. B., Gallois, H., Mullan, S., & Joly, Y. (2019). Integrating artificial in- telligence into health care through data access: can the GDPR act as a bea- con for policymakers? Journal of Law and the Biosciences, 6(1), 317–335. https://doi.org/10.1093/jlb/lsz013 Fossey, E., Harvey, C., Mcdermott, F., & Davidson, L. (2002). Understanding and Evaluating Qualitative Research. Australian and New Zealand Journal of Psychiatry, 36(6), 717–732. https://doi.org/10.1046/j.1440- 1614.2002.01100.x Fraser, A. G., Biasin, E., Bijnens, B., Bruining, N., Caiani, E. G., Cobbaert, K., … Rademakers, F. E. (2023). Artificial Intelligence in medical device soft- ware and high-risk medical devices – a review of definitions, expert rec- ommendations and regulatory initiatives. Expert Review of Medical De- vices, 20(6), 467–491. http://do.org/10.1080/17434440.2023.2184685 Gasser, U., Ienca, M., Scheibner, J., & Sleigh, J. (2020). Digital tools against COVID-19: Framing the ethical challenges and how to address them. https://doi.org/10.48550/arXiv.2004.10236 González, L., Pidre, S., Ruggia, R., Serra, F., Mendoza, R., Piñatares, M., & Granja, N. (2018). Data quality management in e-health integration plat- forms: The case of Uruguay. CLEI Electronic Journal, 21(2). http://doi:10.19153/cleiej.21.2.5 46 Grossoehme, D. H. (2014). Overview of Qualitative Research. Journal of Health Care Chaplaincy, 20(3), 109–122. https://doi.org/10.1080/08854726.2014.925660 Gulbrandsen, M., Hopkins, M., Thune, T., & Valentin, F. (2016). Hospitals and in- novation: Introduction to the special section. Research Policy, 45(8): 1493–1498. https://doi.org/10.1016/j.respol.2016.05.010 Habehh, H., & Gohel, S. (2021). Machine Learning in Healthcare. Current Ge- nomics, 22(4), 291-300. https://doi.org/10.2174/1389202922666210705124359 Hashiguchi, T. C. O., Oderkirk, J., & Slawomirski, L. (2022). Fulfilling the Prom- ise of Artificial Intelligence in the Health Sector: Let’s Get Real. Value in Health, 25(3), 368–373. https://doi.org/10.1016/j.jval.2021.11.1369 Helm, J. M., Swiergosz, A. M., Haeberle, H. S., Karnuta, J. M., Schaffer, J. L., Krebs, V. E., Spitzer, A. I., & Ramkumar, P. N. (2020). Machine Learning and Artificial Intelligence: Definitions, Applications, and Future Direc- tions. Current Reviews in Musculoskeletal Medicine, 13(1), 69–76. https://doi.org/10.1007/s12178-020-09600-8 Holm, K., & Benjaminsson, C. (2023, September 21) Färre fallolyckor på vårdav- delningar med hjälp av AI. Sahlgrenskaliv. https://sahlgrenskaliv.se/farre- fallolyckor-pa-vardavdelningar-med-hjalp-av-ai/ Holzinger, A., Langs, G., Denk, H., Zatloukal, K., & Müller, H. (2019). Causability and explainability of artificial intelligence in medicine. WIREs Data Min- ing and Knowledge Discovery, 9(4), 1–13. https://doi.org/10.1002/widm.1312 IMY. (2024). English summary: the Swedish Authority for Privacy Protection, IMY, finishes its second regulatory sandbox pilot. IMY. https://www.imy.se/globalassets/dokument/rapporter/english-summary--- the-swedish-authority-for-privacy-protection-imy-finishes-its-second-reg- ulatory-sandbox-pilot.pdf Isgut, M., Gloster, L., Choi, K., Venugopalan, J., & Wang, M. D. (2022). System- atic Review of Advanced AI Methods for Improving Healthcare Data Quality In Post COVID-19 Era. IEEE Reviews in Biomedical Engineering, 1–17. https://doi.org/10.1109/RBME.2022.3216531 Ivarsson Westerberg, A., Andersson, A., Anskär, E., Castillo, D., Falk, M., & Fors- sell, A. (2021). Papper, pengar och patienter: primärvården i administrat- ionssamhället (1st ed.). Studentlitteratur AB. https://www.studentlittera- tur.se/kurslitteratur/organisation-och-ledarskap/organisation/papper- pengar-och-patienter/ Janlöv, N., Blume, S., Glenngård, A., Hanspers, K., Anell, A., & Merkur, S. (2023). Health system review Sweden. Health Systems in Transition, 47 25(4), i–198. https://iris.who.int/bitstream/han- dle/10665/372708/9789289059473-eng.pdf?sequence=8 Johnson, K. B., Wei, W., Weeraratne, D., Frisse, M. E., Misulis, K., Rhee, K., Zhao, J., & Snowdon, J. L. (2020). Precision Medicine, AI, and the Future of Personalized Health Care. Clinical and Translational Science, 14(1), 86–93. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7877825/ Kallinikos, J., Aaltonen, A., & Marton, A. (2010). A theory of digital objects. First Monday, 15(6). https://doi.org/10.5210/fm.v15i6.3033 Karolinska Universitetssjukhus. (2024). Innovation för livet Årsrapport 2023. https://www.karolinska.se/490761/globalassets/global/4-gamla-kata- loger/utveckling-och-innovation/publikationer/innovationkarolinska- 2023.pdf Kemppainen, L., Pikkarainen,M., Hurmelinna-Laukkanen, P., Reponen, J. (2019). Data Access in Connected Health Innovation: Managerial Orchestration Challenges and Solutions. Technology Innovation Management Review, 9(12): 43-55. http://doi.org/10.22215/timreview/1291 Kostkova, P., Brewer, H., de Lusignan, S., Fottrell, E., Goldacre, B., Hart, G., … Tooke, J. (2016). Who owns the data? Open Data for Healthcare. Fron- tiers in Public Health, 4. http://doi:10.3389/fpubh.2016.00007 LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep Learning. Nature, 521(7553), 436–444. https://doi.org/10.1038/nature14539 Lee, D., & Yoon, S. N. (2021). Application of Artificial Intelligence-Based Tech- nologies in the Healthcare Industry: Opportunities and Challenges. Inter- national Journal of Environmental Research and Public Health, 18(1), 1– 18. https://doi.org/10.3390/ijerph18010271 Lehne, M., Sass, J., Essenwanger, A., Schepers, J., & Thun, S. (2019). Why digital medicine depends on interoperability. Npj Digital Medicine, 2(1), 1–5. https://doi.org/10.1038/s41746-019-0158-1 Liu, J., Wang, M., Kang, X., Zhang, X., & Chen, X. (2022). Seizing the oppor- tunity window of artificial intelligence in China: Towards an innovation policy mix framework for emerging technologies from an evolution per- spective. Systems Research and Behavioral Science, 39(3), 397–414. https://doi.org/10.1002/sres.2875 Madiega, T. (2024). Artificial intelligence act. In European Parliament. European Union. https://www.europarl.europa.eu/Reg- Data/etudes/BRIE/2021/698792/EPRS_BRI(2021)698792_EN.pdf Mariani, M. M., Machado, I., & Nambisan, S. (2023). Types of innovation and arti- ficial intelligence: A systematic quantitative literature review and research agenda. Journal of Business Research, 155, 1–14. https://doi.org/10.1016/j.jbusres.2022.113364 48 Martin, E. G., Law, J., Ran, W., Helbig, N., & Birkhead, G. S. (2017). Evaluating the quality and usability of open data for Public Health Research: A sys- tematic review of data offerings on 3 open data platforms. Journal of Pub- lic Health Management and Practice, 23(4). https://doi.org/10.1097/phh.0000000000000388 Martin, E. G., Law, J., Ran, W., Helbig, N., & Birkhead, G. S. (2017). Evaluating the Quality and Usability of Open Data for Public Health Research. Journal of Public Health Management and Practice, 23(4), e5–e13. https://doi.org/10.1097/phh.0000000000000388 May, A. D., Sagodi, A., Dremel, C., & van Giffen, B. (2020). Realizing Digital In- novation from Artificial Intelligence. ICIS 2020 Proceedings, 6, 1–17. https://aisel.aisnet.org/icis2020/digital_innovation/digital_innovation/6 McMullin, C. (2021). Transcription and Qualitative Methods: Implications for Third Sector Research. VOLUNTAS: International Journal of Voluntary and Nonprofit Organizations, 34(1), 140–153. https://doi.org/10.1007/s11266-021-00400-3 Mero-Jaffe, I. (2011). “Is that what I Said?” Interview Transcript Approval by Par- ticipants: An Aspect of Ethics in Qualitative Research. International Jour- nal of Qualitative Methods, 10(3), 231–247. https://doi.org/10.1177/160940691101000304 Miotto, R., Wang, F., Wang, S., Jiang, X., & Dudley, J. T. (2018). Deep Learning for healthcare: review, Opportunities and Challenges. Briefings in Bioin- formatics, 19(6), 1236–1246. https://doi.org/10.1093/bib/bbx044 Mohammad Amini, M., Jesus, M., Fanaei Sheikholeslami, D., Alves, P., Hassanza- deh Benam, A., & Hariri, F. (2023). Artificial Intelligence Ethics and Challenges in Healthcare Applications: A Comprehensive Review in the Context of the European Mandate. Machine Learning and Knowledge Ex- traction, 5(3), 1023–1035. https://doi.org/10.3390/make5030053 Nambisan, S., Lyytinen, K., & Yoo, Y. (2020). Handbook of Digital Innovation. Edward Elgar Publishing. Nambisan, S., Lyytinen, K., Majchrzak, A., & Song, M. (2017). Digital Innovation Management: Reinventing Innovation Management Research in a Digital World. MIS Quarterly, 41(1), 223–238. https://doi.org/10.25300/misq/2017/41:1.03 Nascimento, L. da S., & Steinbruch, F. K. (2019). “The interviews were tran- scribed”, but how? Reflections on management research. RAUSP Manage- ment Journal, 54(4), 413–429. https://doi.org/10.1108/rausp-05-2019- 0092 49 Ng, M., Youssef, A., Miner, A. S., Sarellano, D., Long, J., Larson, D. B., Hernan- dez‐Boussard, T., & Langlotz, C. P. (2023). Perceptions of Data Set Ex- perts on Important Characteristics of Health Data Sets Ready for Machine Learning. JAMA Network Open, 6(12), 1–14. https://doi.org/10.1001/jamanetworkopen.2023.45892 Noorbakhsh-Sabet, N., Zand, R., Zhang, Y., & Abedi, V. (2019). Artificial Intelli- gence Transforms the Future of Health Care. The American Journal of Medicine, 132(7), 795–801. https://doi.org/10.1016/j.amjmed.2019.01.017 Nordic Innovation. (2022). Bridging Nordic Data. https://www.nordicinnova- tion.org/2022/bridging-nordic-data Nordic Innovation. (n.d.). FederatedHealth: A Nordic Federated Health Data Net- work. Nordic Innovation. https://www.nordicinnovation.org/programs/fed- eratedhealth-nordic-federated-health-data-network OECD. (2019). Digital Innovation: Seizing Policy Opportunities. https://doi- org./10.1787/a298dc87-en. OECD. (2023). Regulatory sandboxes in artificial intelligence. OECD Digital Economy Papers, 356, https://doi.org/10.1787/8f80a0e6-en. OECD. (2024). Explanatory memorandum on the updated OECD definition of an AI system. https://doi.org/10.1787/623da898-en Oliver, D. G., Serovich, J. M., & Mason, T. L. (2005). Constraints and Opportuni- ties with Interview Transcription: Towards Reflection in Qualitative Re- search. Social Forces, 84(2), 1273–1289. https://doi.org/10.1353/sof.2006.0023 Osborn, R., Moulds, D., Schneider, E. C., Doty, M. M., Squires, D., & Sarnak, D. O. (2015). Primary Care Physicians In Ten Countries Report Challenges Caring For Patients With Complex Health Needs. Health Affairs, 34(12), 2104-2112G. https://doi.org/10.1377/hlthaff.2015.1018 Parliament of Sweden. (2008). Patientdatalag (2008:355). https://www.riksda- gen.se/sv/dokument-lagar/dokument/svensk-forfattningssamling/patientda- talag-2008355_sfs-2008-355 Paunov, C., & Planes-Satorra, S. (2019). How are digital technologies changing in- novation? OECD Science, Technology and Industry Policy Papers, 74, 1– 53. https://doi.org/10.1787/67bbcafe-en Pawar, U., O’Shea, D., Rea, S., & O’Reilly, R. (2020). Explainable AI in Healthcare. 2020 International Conference on Cyber Situational Aware- ness, Data Analytics and Assessment (CyberSA), 1–2 https://doi.org/10.1109/cybersa49311.2020.9139655 Petersson, L., Larsson, I., Nygren, J. M., Nilsen, P., Neher, M., Reed, J. E., Tyskbo, D., & Svedberg, P. (2022). Challenges to implementing artificial intelli- gence in healthcare: a qualitative interview study with healthcare leaders 50 in Sweden. BMC Health Services Research, 22(1), 1–16. https://doi.org/10.1186/s12913-022-08215-8 Planes-Satorra, S., & Paunov, C. (2019). The digital innovation policy landscape in 2019. OECD Science, Technology and Industry Policy Papers, 71. 1–55 https://doi.org/10.1787/6171f649-en Rajpurkar, P., Chen, E., Banerjee, O., & Topol, E. J. (2022). AI in health and medi- cine. Nature Medicine, 28(1), 31–38. https://doi.org/10.1038/s41591-021- 01614-0 Ranchordas, S. (2021). Experimental Regulations for AI: Sandboxes for Morals and Mores. Morals & Machines, 1(1), 86–100. https://doi.org/10.5771/2747-5174-2021-1-86 Rocco, T. S., & Plakhotnik, M. S. (2009). Literature Reviews, Conceptual Frame- works, and Theoretical Frameworks: Terms, Functions, and Distinctions. Human Resource Development Review, 8(1), 120–130. https://doi.org/10.1177/1534484309332617 Rosenberg, G. S., Magnéli, M., Barle, N., Kontakis, M. G., Müller, A. M., Wit- tauer, M., Gordon, M., & Brodén, C. (2024). ChatGPT-4 generates ortho- pedic discharge documents faster than humans maintaining comparable quality: a pilot study of 6 cases. Acta orthopaedica, 95, 152–156. https://doi.org/10.2340/17453674.2024.40182 Sambasivan, N., Kapania, S., Highfill, H., Akrong, D., Paritosh, P., & Aroyo, L. (2021). “Everyone wants to do the model work, not the data work”: Data Cascades in High-Stakes AI. CHI ’21: Proceedings of the 2021 CHI Con- ference on Human Factors in Computing Systems, 39, 1–15. https://doi.org/10.1145/3411764.3445518 Saltz, J. (2024, Mars 31). What is the AI life cycle?. Data Science Process Alliance. https://www.datascience-pm.com/ai-lifecycle/ Scherer, M. U. (2016). Regulating Artificial Intelligence Systems: Risks, Chal- lenges, Competencies, and Strategies. Harvard Journal of Law & Technol- ogy, 29(2). https://doi.org/10.2139/ssrn.2609777 Sittig, D. F., & Singh, H. (2010). A New Socio-technical Model for Studying Health Information Technology in Complex Adaptive Healthcare Systems. Quality & Safety in Health Care, 19(3), 68-74. https://doi.org/10.1136/qshc.2010.042085 Smuha, N. (2021). From a ‘Race to AI’ to a ‘Race to AI Regulation’: Regulatory Competition for Artificial Intelligence. Law, Innovation & Technology, 13(1), http://dx.doi.org/10.2139/ssrn.3501410 Socialdepartementet. (2024a) Uppdrag att möljiggöra en nationel digital infra- struktur för hälsodata. Regeringskansliet. https://www.regeringen.se/glo- 51 balassets/regeringen/dokument/socialdepartementet/fokhalsa-och-sjuk- vard/uppdrag-att-mojliggora-en-nationell-digital-infrastruktur-for- halsodata.pdf Socialdepartementet. (2024b, Mars 21). Regeringen och SKR överens om inriktning för kunskapsstyrningen i hälso- och sjukvården. Regeringskansliet https://www.regeringen.se/pressmeddelanden/2024/03/regeringen-och-skr- overens-om-inriktning-for-kunskapsstyrningen-i-halso--och-sjukvarden/ Socialdepartementet. (2023a, September 6). Regeringen satsar på en nationell digi- tal infrastruktur för hälso- och sjukvården. Regeringskanseliet. https://www.regeringen.se/pressmeddelanden/2023/09/regeringen-satsar- pa-en-nationell-digital-infrastruktur-for-halso--och-sjukvarden/ Socialdepartmentet. (2023b, January 19). Regeringens arbete med stärkt kompe- tensförsörjning inom hälso- och sjukvården. Regeringskanseliet. https://www.regeringen.se/artiklar/2023/01/regeringens-arbete-med-starkt- kompetensforsorjning-inom-halso--och-sjukvarden/ Socialdepartementet. (2020). En strategi för genomförande av Vision e-hälsa 2025. Regeringen. https://www.regeringen.se/contentassets/3f5ccc23708542b6971b8a7d3eaa 4379/godkannande-av-overenskommelse-om-en-strategi-for-genoforande- av-vision-ehalsa-2025.pdf Socialdepartementet. (2017). Hälso- och sjukvårdslag (2017:30). Sveriges Riks- dag. https://www.riksdagen.se/sv/dokument-lagar/dokument/svensk-for- fattningssamling/halso--och-sjukvardslag_sfs-2017-30 Socialdepartementet. (2008). Patientdatalag (SFS 2008:355). Sveriges Riksdag. https://www.riksdagen.se/sv/dokument-lagar/dokument/svensk-forfatt- ningssamling/patientdatalag-2008355_sfs-2008-355 Srivastava, D., van Kessel, R., Delgrange, M., Cherla, A., Sood, H., & Mossialos, E. (2023). A Framework for Digital Health Policy: Insights from Virtual Primary Care Systems Across Five Nations. PLOS Digital Health, 2(11), 1–13. https://doi.org/10.1371/journal.pdig.0000382 Sundblad, W. (2018, October 18). Data Is The Foundation For Artificial Intelli- gence And Machine Learning. Forbes. https://www.forbes.com/si- tes/willemsundbladeurope/2018/10/18/data-is-the-foundation-for-artifi- cial-intelligence-and-machine-learning/?sh=566a9b3e51b4 Sveriges Kommuner och Regioner. (2022) Välfärdens Kompetensförsörjning. https://skr.se/skr/tjanster/rapporterochskrifter/publikationer/valfardens- kompetensforsorjning.68208.html Sveriges Riksdag. (2018). Dataskyddslagen (2018:218). https://www.riksda- gen.se/sv/dokument-och-lagar/dokument/svensk-forfattningssamling/lag- 2018218-med-kompletterande-bestammelser_sfs-2018-218/ 52 Swedish Medical Product Agency. (2023). Vägledning rörande användning av ar- tificiell intelligens i svensk sjukvård. https://www.lakemedelsver- ket.se/4a5f16/globalassets/dokument/medicinteknik/artificiell-intelligens- ai/vagledning-anvandning-av-artificiell-intelligens-i-svensk-sjukvard.pdf Topol, E. J. (2019). High-performance medicine: the convergence of human and ar- tificial intelligence. Nature medicine, 25(1), 44-56, https://doi.org/10.1038/s41591-018-0300-7 Trocin, C., Hovland, I. V., Mikalef, P., & Dremel, C. (2021). How Artificial Intelli- gence affords digital innovation: A cross-case analysis of Scandinavian companies. Technological Forecasting and Social Change, 173, 1–12. https://doi.org/10.1016/j.techfore.2021.121081 Truby, J., Brown, R. D., Ibrahim, I. A., & Parellada, O. C. (2022). A Sandbox Ap- proach to Regulating High-Risk Artificial Intelligence Applications. Euro- pean Journal of Risk Regulation, 13(2), 270–294. https://doi.org/10.1017/err.2021.52 Tucker, J. (2023). The future vision(s) of AI health in the Nordics: Comparing the national AI strategies. Futures, 149, 1-10. https://doi.org/10.1016/j.futu- res.2023.103154 Utbildningsdepartementet (2003). Lag (2003:460) om etikprövning av forskning som avser människor. Sveriges Riksdag. https://www.riks- dagen.se/sv/dokument-lagar/dokument/svensk-forfattningssamling/lag- 2003460-om-etikprovning-av-forskning-som_sfs-2003-460 von Eschenbach, W. J. (2021). Transparency and the Black Box Problem: Why We Do Not Trust AI. Philosophy & Technology, 34, 1607–1622. https://doi.org/10.1007/s13347-021-00477-0 Wang, J., Liu, Y., Li, P., Lin, Z., Sindakis, S., & Aggarwal, S. (2023). Overview of Data Quality: Examining the Dimensions, Antecedents, and Impacts of Data Quality. Journal of the Knowledge Economy, 1(20). https://doi.org/10.1007/s13132-022-01096-6 Wang, Y.-H., & Lin, G.-Y. (2022). Exploring AI-healthcare innovation: natural language processing-based patents analysis for technology-driven roadmapping. Kybernetes, 52(4), 1173–1189. https://doi.org/10.1108/k- 03-2021-0170 53 10 Appendices 10.1 Appendix I - Final Interview Guide Introduction First of all, we are very grateful and thankful you agreed to do this interview with us. Our names are… We are currently master students at the Digital Leadership Program at Uni- versity of Gothenburg. This will be our final thesis before graduation. During the interview you can of course ask us any questions. This interview is not a test, so don’t worry if there is a question you don’t know. Just say what you think and what you know. It will be in English so if there are Swedish terms you don’t know in English, you can just say the Swedish ones. Before we start, we want to let you know that this interview will be recorded to help us with transcribing what you say. Although the recording will be kept by ourself and you will be anon- ymous in the sense that we won’t refer by your name in the master thesis. Does it sound good? Any questions before we start recording? [Start the recording] Background information To give a little background regarding our work: We are with this research trying to find certain key themes or areas of policies or future policy-making crucial to help encouraging innovation of AI within Swedish healthcare. The outcome of our work then is that these policy themes could help legislators to understand what parts are important for AI innovation and develop- ment to flourish in the Swedish healthcare. For an example, the theme “open data” might be considered a crucial area for policy making to enable AI innovation. Now, to put us in the context, let us pretend that we (Adrian and Upul) want to form a startup that we want to create an AI tool for healthcare. Because of the current legislative and regu- latory landscape around healthcare, there are going to be hinders such as data access and modes of moving into production. So, our aim is to research about the areas of a conducive policy environment which supports the AI innovation to flourish. Opening questions: 1. What do you think are the main key points for AI to succeed in healthcare, in Sweden? 54 2. What do you think is the biggest hinder preventing AI from innovating in healthcare? 3. What is your perception about the role of legislative environment in promoting AI in healthcare? Are you aware of any regulation on AI in Sweden? 4. Do you think the regulation changes are necessary for the promotion of AI in the cur- rent state of Swedish healthcare? a. If so in what areas? Data Sharing / Open Data Initiatives: 1. What is your perception of Data sharing / Open Data initiatives regarding AI devel- opment in healthcare? And, your opinion on the feasibility of implementing them in Swedish context? 2. What is your opinion about the legislative environment which needs to be in place for such initiatives? Would the current legislation supportive enough or required to make changes to the existing laws or introduce new ones? 3. Are you aware of any Open Data Initiatives or data sharing related initiatives for healthcare at present in Sweden? Data Quality and Integrity Assurance: 1. What is your perception of Data quality and accuracy assurance requirements of data for AI development in healthcare? Do you think additional measures are required in the Swedish healthcare to ensure data quality and feasibility of implementing such changes? 2. What is your opinion about the legislative environment which needs to be in place for such initiatives? Would the current legislation supportive enough or required to make changes to the existing laws or introduce new ones? 3. Are you aware of any Data quality and accuracy assurance related initiatives for healthcare at present in Sweden? Data Ownership and access control: 1. What is your perception of Data Ownership and access control related requirements regarding AI development in healthcare? Do you think additional measures are re- quired in the Swedish healthcare to ensure proper data access and feasibility of imple- menting such initiatives? 2. What is your opinion about the legislative environment which needs to be in place for such initiatives? Would the current legislation supportive enough or required to make changes to the existing laws or introduce new ones? 3. Are you aware of any Data Ownership and access control related initiatives for healthcare at present in Sweden? Data Interoperability / Uniformity / Data standards: 55 1. What is your perception of Data Interoperability / Uniformity related requirements re- garding AI development in healthcare? Do you think additional measures are required in the Swedish healthcare to ensure proper data standards and feasibility of imple- menting such initiatives ? 2. What is your opinion about the legislative environment which needs to be in place for such initiatives? Would the current legislation supportive enough or required to make changes to the existing laws or introduce new ones? 3. Are you aware of any Data Interoperability / Uniformity related initiatives for healthcare at present in Sweden? Facilitation for regulatory experimentation: 1. What is your perception of regulatory experimentation tools such as regulatory sand- boxes on AI development in healthcare? Do you think additional measures are required in the Swedish healthcare to ensure proper use of regulatory experimentation tools and feasibility of implementing such initiatives? 2. What is your opinion about the legislative environment which needs to be in place for such initiatives? Is the current legislation supportive enough, or is it necessary to revise existing laws or introduce new ones? 3. Are you aware of any regulatory experimentation tools such as regulatory sandboxes for healthcare being active in Sweden? Open Ended Questions: What other areas should be considered when creating regulations for enabling AI develop- ment in healthcare according to your experience and perception? 56