Artificial Intelligence Applications in Health Care Practice: A Scoping Review

Background: Artificial Intelligence (AI) is often heralded as a potential disruptor that will transform the way we do medicine. The amount of data collected and available in health care, coupled with advances in computational power, have contributed to advances in AI and an exponential growth of publications. However, the development of AI applications does not guarantee their adoption into routine practice. There is a risk that despite the resources invested, benefits for patients, staff and society will not be realized if AI implementation is not better understood. Objective: The aim of this study was to explore how the implementation of AI in healthcare practice has been described and researched in the literature by answering three questions: 1. What are the characteristics of research on implementation of AI in practice? 2. What types and applications of AI systems are described? 3. What characteristics of the implementation process for the AI systems are discernable? Methods: A scoping review was conducted of Medline (PubMed), Scopus, Web of Science, CINAHL, and PsychInfo databases to identify empirical studies of AI implementation in healthcare since 2011 in addition to snowball sampling of selected reference lists. Titles and abstracts were screened and full-text articles using Rayyan.ai software. Data from the included articles was charted and summarized. Results: Of the 9218 records retrieved, forty-five articles were included. Most articles were published recently, from high-income countries, cover diverse clinical settings and disciplines, and intended for care-providers. AI systems are predominantly intended for clinical care, particularly clinical care pertaining to patient-provider encounters. Most possess no action autonomy, but rather support human decision-making. The focus of most research is on establishing the effectiveness of interventions, or related to technical and computational aspects of AI systems. Focus on the specifics of implementation processes


Table of Contents
Only make the preprint title and abstract visible.No, I do not wish to publish my submitted manuscript as a preprint.2) If accepted for publication in a JMIR journal, would you like the PDF to be visible to the public?
Yes, please make my accepted manuscript PDF available to anyone at any time (Recommended).
Yes, but please make my accepted manuscript PDF available only to logged-in users; I understand that the title and abstract will remain v Yes, but only make the title and abstract visible (see Important note, above).I understand that if I later pay to participate in <a href="http

Introduction
Artificial Intelligence (AI) is often heralded as a potential disruptor that will transform the way we do medicine (1,2).The promise of AI lies in its ability to process and learn from large volumes of data and capture patterns otherwise difficult for humans to identify.This ability has raised questions and worries about liability and risks, in particular related to the level of autonomy granted to AI applications (3).Others see a role complementary to humans, e.g.decision support or decision augmentation, where humans (in the roles of clinicians or programmers) provide oversight and collaborate (4)(5)(6)(7).The latter approach has been demonstrated to yield superior performance when compared to experts alone (8).Other benefits include improved patient outcomes, error reduction, health system optimization, cost reductions, and increased value (6).
The amount of data collected and available in health care, coupled with advances in computational power, have contributed to advances in AI applications (9), and an exponential growth of publications on AI in health care, with more than ten thousand records on PubMed in 2021 alone.Included in this are multiple reviews across medical specialties that explore the potential roles of AI to augment health care delivery (10)(11)(12)(13)(14).These include diagnostic (e.g.early cancer diagnosis, diabetes retinopathy screening, or COVID-19 based on CT images), therapeutic (e.g.precision medicine in chemotherapy and for combination drug therapy), and regulatory or administrative applications (e.g.coding of records or economic evaluations), and for population health management (e.g.public health surveillance or predictive epidemiological modelling) (15)(16)(17)(18)(19)(20)(21) However, the development of AI applications does not guarantee their adoption into routine health care practice.Research has identified a number of factors influencing adoption of innovations.These include context (e.g. economic and political context, laws and regulations, socio-cultural factors), organization (e.g.organization structure, resources, processes), group (e.g.professional values and cultures), individual (e.g.attitudes, motivation, user satisfaction, trust), and technology (e.g.usability, design, accuracy, explainability) (22,23).This suggests a need to know more about how AI can be implemented in health care, not only as an innovation, but also with respect to its unique potential and associated concerns.
Earlier reviews have tended to focus only on some aspects of the implementation process of AI in healthcare, for example, regulation and legal issues (24,25), trust and ethics (24)(25)(26)(27)(28)(29), clinical and patient outcomes (30)(31)(32) and economic impact (33).Others have focused their studies on specific AI applications for healthcare, such as predictive medicine, diagnostics and clinical decision making (30,(34)(35)(36).A few reviews have been more overarching, focusing on coproduction processes (37), implementation frameworks (38) and critical implementation barriers or success factors (39) that could inform development of relevant implementation strategies of AI technology.Generally, it is argued that the implementation of AI in healthcare could significantly improve patient and healthcare outcomes, but none of these reviews have actually explored the knowledge base of realworld implementation in everyday clinical practice.
Given the resources invested in developing AI applications, and the risk that effective applications to support, augment, and perhaps even transform healthcare for patients, staff and society will be unnecessarily studied, we sought to explore the research literature on how the implementation of AI in health care practice has been empirically investigated.

Study design
We chose a scoping review methodology in line with the Arksey and O'Malley framework (43) and reported according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews (PRISMA-ScR) checklist (See Figure 1) (44).A previous review suggested that implementation of AI in health care was not well studied (38).A scoping review would thus enable a mapping of the "extent, range and nature of research activity" in this emerging area of research (43).

Identifying the Research Question
To address our aim, we formulated three research questions.
1. What are the characteristics of research on implementation of AI in practice?
2. What types and applications of AI systems are described?
3. What characteristics of the implementation process for the AI systems are discernable?

Identifying Relevant Studies
We focused our search, with support from a university librarian, by iteratively testing synonyms for three concepts: Artificial Intelligence, Healthcare, and Implementation (Table 1).For the purposes of clarity, we differentiated between AI algorithms/models (the actual code), AI applications (the innovation "package"), and AI systems (the application in its context) and used standardized MeSHterms and Subject headings describing AI and its subcategories provided by the databases used for our searches (45).Implementation was defined as, "An intentional effort designed to change or adapt or uptake interventions into routines," based on a review of frameworks for the translation of AI into health care practice (38) Synonyms were joined by the Boolean operator "OR" and then we combined the search strings for each concept with the Boolean operator "AND" (Supplementary File).We used MESH-terms, Headings, and Thesaurus, for the appropriate databases.To cover content in both general and health and healthcare specific sources, five electronic databases were searched: Medline (PubMed), Scopus, Web of Science, CINAHL, and PsychInfo.In addition, we employed snowball sampling by manually reviewing reference lists of review articles we had identified during the screening that might contain relevant references given the topic of the review.

Eligibility Criteria/Identify relevant studies
We included peer-reviewed, empirical studies published in English between December 2011 and February 2022 as preliminary searches suggest AI applications in health care are a more recent phenomenon (Table 2).Given the rapid pace of development of technology and changing datasets, solutions developed before the last decade are likely to be obsolete English-language Practical consideration given the investigators' language proficiency Exclusion criteria Non-empirical designs, including editorials, commentaries, opinion articles, and reports Empirical studies improve the ability to answer the research questions compared to conceptual commentaries or viewpoints Proof-of-concept, feasibility or validation studies not related to implementation of AI As the aim is to explore implementation in practice, studies that stop short of that, e.g.proof-of-concept, validity, or feasibility studies, should be excluded.

Study selection
All identified records were imported to the open-access software Rayyan.ai.Duplicates were removed and the remaining titles and abstracts of the remaining records were screened for eligibility by at least one of the authors.Any uncertainty or conflict was discussed at regular check-ins until consensus was reached among all the authors.These discussions were informed by the multidisciplinary backgrounds of the authors.We also continually reviewed our interpretations of the screening criteria, and when questions were raised, we went back to ensure that the criteria had been applied correctly and in a universal fashion, independent of who screened.We used the AI screening and highlight function of Rayyan.ai,but still screened each record.We also erred on the side of inclusion.Full-text articles were then screened independently by at least two researchers.Conflicts and uncertainty were again resolved through discussion until consensus was reached among all.As we followed the original framework, a quality appraisal of the included studies was not conducted.

Charting the data
We developed a data extraction template to chart data for each of the research questions.To define these conceptual areas, we adopted the WHO Guidance on Ethics and Governance of Artificial Intelligence for Health definition of AI (based on a recommendation of the Council on Artificial Intelligence of the OECD states) (46,47).
An AI system is a machine-based system that can, for a given set of humandefined objectives, make predictions, recommendations, or decisions influencing real or virtual environments.AI systems are designed to operate with varying levels of autonomy (46).

Collating, summarizing, and reporting the results
The extracted data related to research questions 1 and 2 was mapped and summarized.A qualitative thematic analysis (48) was used to analyze data associated with research question 3 to summarize the motives for implementation and elements in the implementation process.Articles were read and reread, with initial ideas sorted into either the domain Motives behind the implementation or Elements in the implementations process.Then, initial codes were identified in each article.The codes were compared based on similarities and differences, collated into potential themes, which were compared to generate a thematic map that was then used to generate clear definitions and names for each theme in the respective domains.Coding and data analysis was performed in pairs and any uncertainties were discussed among all authors until consensus was achieved.

Search results
We identified 9218 records, of which 9179 were identified through database searches and 39 through the snowball search of reference lists in 36 review articles.After removal of duplicates, 5666 records remained and we screened titles and abstracts.In this screening, 5553 records were excluded.The remaining 113 records were assessed for eligibility through full-text review.Sixty-eight articles were excluded for reasons highlighted in Figure 1.A total of 45 articles were included in the scoping review.

RQ 1. Study characteristics
The reviewed body of literature was fairly recent, with majority of the studies (n=32) published between 2020-2022 (49-80  3).

RQ 3. Implementation process characteristics
The research focus in about a third of the studies was to present the effectiveness of the implemented intervention (n = 16) (58,62,(64)(65)(66)69,70,75,(77)(78)(79)81,85,86,89,92).Other research foci included: User experiences (55,63,67,68,73,90), AI usage metrics (52,56,84,88,93), and identification of barriers or facilitators (54,57,59,61,71,91) (Table 3 and Figure 2).Most studies described the implementation process as successful and most AI-based applications were also stated to be in use after the study was completed.In nearly half of the reviewed studies (n=23) the motives behind the implementation were not described.For those studies that did, we identified six types of motives, with to Improve healthcare quality and Achieve better patient outcomes being the two most common.Studies in the first theme described AI systems to improve quality of services (50,75,79,91,92), reduce diagnostic errors (70), length of stay at hospital (77), or unplanned readmissions (54,58), and for the latter to achieve better patient survival (63,74).Another theme, Improve efficiency, focused on healthcare cost reduction, increased service production, and optimization of public services (49,76,78,80,81).Respond to the COVID-19 pandemic was stated as a motive necessitated by the need for access to the most up-todate information (52), the sudden surge in demand for healthcare services (56), prioritization of limited resources (76), and reorganization of service delivery in response to local guidelines for prevention of infection transmission (66).Improve provider satisfaction focused on workload reduction for healthcare professionals (59,73).Empower patients by using AI to support interpretations of laboratory investigations, rather than just the test results, was another motive for implementing AI (90).Three studies had an explicit focus on implementation processes (50,53,72) In the other studies, characteristics common to implementation processes were identified: Co-creation, Contextualization, Non-disruptive work-flow design, Communication, Learning focus, Training, Incentives, and Organizational strategies.Both barriers and facilitators were described.Several implementation efforts involved Co-creation with multidisciplinary stakeholders starting from an ideation phase that included problem identification, requirement collection, and (re)design of clinical workflows to facilitate AI-system integration (49,50,53,56,59,63,72,82). Co-creation also involved end-users in the design of user interfaces (50,72).Contextualization of AI-systems related to the local context and target population was highlighted as important in development and implementation (56,58).Non-disruptive work-flow design was emphasized, where efforts were made to design AI-systems around existing roles and functions of the intended user to avoid radical modification of current practice to fit the AI system (50,53,55).Communication efforts were seen as central to building trust and promoting use by sharing evidence of AI effectiveness with clinicians and describing overall benefits of the technology (50,53,63), appointing champions to promote AI among peers (50,57,79), and encouraging informal communication between clinicians and IT developers to cultivate relationships and build trust in the AI (60).However, one study encouraged the separation of developers and clinicians and made conscious efforts to shift focus away from technical aspects of AI (50).A learning focus could begin in the ideation phase, to understand and assess the problem to be addressed by AI before coding, through development and implementation, by iteratively testing and adjusting workflows (50).After implementation, learning continued through the continuous capture of user feedback to enable improvement (72).Training involved both informal and formal sessions to enable AI use (60,93).After implementation, training could continue in formal peer-group meetings to share best practices and individual training and support for more reluctant users (88).Incentives were used to promote or enforce AI use.More controlling approaches included periodic monitoring and audits (60,88), or removing alternative ways of performing the task altogether to necessitate AI use (88).Gamification was used to promote a feeling of reward and competition (65,69).Organizational efforts involved including hospital's top leadership as essential members of the project team and the design and implementation of the AI system to promote uptake (53,59).One organization formed a special governance committee as a formal mechanism to monitor AI use among healthcare providers (50).Another organization's innovation strategy included innovation managers as part of the organizational structure to promote AI (57).The use of implementation frameworks was mentioned in three studies (54,61,72): the Reach, Effectiveness, Adoption, Implementation, and Maintenance (RE-AIM) framework (94), the Nonadoption, Abandonment, Scale-up, Spread, and Sustainability (NASSS) framework (95), and the Socialisation, Externalisation, Combination, and Internalisation (SECI) model of knowledge dimensions (96).Four studies proposed new frameworks, principles, or recommendations based on their presented findings and implementation experiences (53,59,60,84).Moorman (53) proposed six principles for implementation of AI, including elements of trust/transparency, minimal impact on workflows, stakeholder buy-in, relevant education, actionability of AI outputs, and sustainability through follow-up interactions.Reis et al. (59) proposed a framework for overcoming cognitive and affective resistance to AI implementation centered around concerns of users (physicians), such as transparency/understandability of the AI, involvement of users in the AI training, and trust in the AI system.Sun (60) proposed a power strategy matrix for AI adoption, suggesting that a "boss strategy" or "expert strategy" can influence adoption.Wen et al. (84) presented three desiderata for developing an AI-based platform, where the second focused on improving adoption.

Discussion
Our aim with this study was to explore how the implementation of AI in health care practice has been empirically investigated in the research literature.We found that research on AI systems' implementation is mostly published in high-income countries, covers many different clinical settings and disciplines, and predominantly focuses on care providers as users.The AI models are primarily symbolic or knowledge-based, employ automation or optimisation technologies, and are mainly used to perform tasks related to recognition.AI systems are predominantly intended for clinical care, particularly clinical care pertaining to patient-provider encounters.Most possess no action autonomy, but rather support human decision-making.The focus of most research is on establishing the effectiveness of interventions, or related to technical and computational aspects of AI systems.Focus on the specifics of implementation processes does not yet seem to be a priority in research and the use of frameworks to guide implementation is rare.

Study characteristics
Most studies were published in the last two years (2020-2022), which is unsurprising given the temporal distribution of AI health care studies.Research on AI implementation in healthcare is predominantly conceptual in nature, dominated by commentaries, perspectives, opinion articles, and conceptual frameworks, which raise important questions and issues, but without much-needed empirical evidence (97)(98)(99)(100).Since the empirical evidence base for the implementation of AI solutions in routine healthcare is still narrow and premature, it limits possibilities for generalization both for practice and for the advancement of methodological approaches.Most articles were published in high-income countries, particularly the USA.This finding is consistent with the more developed digital health infrastructure, routine use of electronic health records, and big data initiatives in the North American and European countries and aligns with other reviews of AI applications in various fields of healthcare (101)(102)(103).The many different clinical settings and disciplines could corroborate the data-driven nature of health care, that AI is highly applicable, or that due to its nascent state, AI is still being tried in many different contexts.Given the focus on clinical care, it is not surprising that the intended users were mostly healthcare providers, particularly doctors.A recent scoping review on the use of AI in primary care found a similar predominance of doctors as target end-users (104).This suggests a view of AI systems as tools to support decisionmaking by doctors rather than other health professionals.It was surprising to find a scarcity of implementations of AI applications to handle infectious diseases (except ( 52)) given the overwhelming attention and funding to the management of COVID-19 pandemic in 2020-2022.Another underrepresented area where AI holds a strong promise is mental health (except (81,105)).

Types and applications of AI technology
About half of the AI models were symbolic or knowledge-based.They used human-generated logical representations, rules, and ontologies to infer conclusions and have greater explainability compared to models that are based on pure data-driven or statistical approaches.However, they might not live up to the full potential of AI as they are 'hard-coded, expert cookbooks' that are limited by the knowledge that is encoded into them (106).Data-driven, statistical approaches, such as machine learning, learn predictive functions based on the inputted data.However, these methods are opaque and have implications for healthcare in relation to patient or provider trust, accountability and quality assurance, and patient safety (3,107).The WHO guidance on ethics and governance of artificial intelligence recognises the potential trade-off between transparency and accuracy, but encourages AI explainability and transparency over black-box approaches (47).The predominance of knowledgebased or symbolic models, whose greater transparency and longer existence may ease acceptance among care providers, is in line with previous reviews (108).However, the majority of recently published AI models utilize data-driven or hybrid technologies and knowledge-based models comprised only a minority of the applications (109).The current study found that automation or optimisation technologies were by far the most common, followed by human language technologies.More than half of the AI systems implemented had no action autonomy.Instead, they were human decision support systems where the AI system cannot act on its recommendation or output, but depend on the human operating the system to use or disregard the recommendation made by the AI system.This finding indicates that the decision support systems are the type of AI systems that have achieved adoption the earliest, likely since they enhance human actions and minimally disrupt the clinical workflows (110)

Implementation process
This study found that the way the implementation process of AI systems in healthcare is researched is varied and builds on many types of study designs and methodologies.Nearly half of the included studies did not provide a clear motivation for implementing an AI system, which is a key factor for successful adoption of AI in healthcare (110).The lack of a clear motivation indicates poor alignment with well-defined needs from clinical practice and risks reinforcing a technology focused logic regarding AI implementation AI in healthcare.This observation might reflect the lack of consistent understanding of what is meant by implementation of AI in daily practice and a lack of methodological consistency in how such implementation should be researched and reported.Most of the studies either had a technical or computational understanding of implementation or viewed implementation in terms of effectiveness of the intervention.There was not much focus on the actual process of implementation studies but more on presenting cases of implementation.This indicates the relatively nascent nature of evidence in this field and is similar to other studies which highlight that many of the publications on AI in healthcare focus on the methods and technical aspects of applying the AI model to clinical scenarios but very little on the actual process of its implementation in practice (55,104).Despite the limited focus in the studies on researching the implementation process, our inductive analysis identified the implementation elements: Co-creation, designing non-disruptive workflows, maintaining a learning focus, communication, contextualisation, leadership and conducive organisational structure, trainings, and enforcement or incentivisation of AI use.These aspects are not unique to AI but have been highlighted as important interventions for the adoption of all digital technologies, including AI.For example, involvement of end users in the design and implementation of IT services and applications form the basis of user-centred design, which is seen as an important driver of uptake of digital technologies (111).The commitment, involvement, and accountability of leaders is also a well-known factor for successful implementation in practice (112).Seamless integration with existing workflows was another factor highlighted as central for adoption of AI systems.This finding is consistent with the fact that most studied cases of AI systems' implementation were based on decision support systems that have no action autonomy and can be conveniently incorporated into the routine workflows.However, it is challenging to draw generalized conclusions on the AI implementation strategies from such systems, as they introduce incremental improvements in the workflows and do not represent more disruptive types of AI systems, for example, those with high action autonomy.The findings in this study corroborate the recent work by Gama et al. regarding the uncertainty of what should be considered to be AI and that our understanding of implementation is still in the early stages of development (38).We would add that this understanding is made even more complex by the lack of agreement on what is meant by the term implementation.We rejected numerous studies in the screening because the term implementation was used in a computational sense.For example, the product concept or requirements were implemented as a code, or the coded algorithm was implemented using an existing data set.Even in studies involving real-world settings, the term was used in the meaning of execution of a plan without reflection on the process of execution.The focus of implementation as an intentional effort designed to change, adapt, or the uptake of interventions into routine was scarce in the published literature.

Limitations/Methodological considerations
The strengths of this study include the substantial number of records reviewed and the rigour observed during the screening process.The search strategy was comprehensive and broad, and covered five different electronic databases.However, we did not include a broader search of the grey literature that would have undoubtedly captured additional cases and potentially identified more cases representing ongoing or completed implementation projects not yet published in the research literature.As we aimed to investigate the experiences from implementation in clinical practice, clinical trials, case reports, pilots, feasibility studies and other forms of limited and controlled introduction of AI applications in practice were removed during screening.We expect there to be a lag between the work of technology companies and care providers and subsequent academic publications.However, due to the number of records we identified and the previously found extensive availability of opinion-based articles in the literature in the form of perspectives, insights and narrative reviews (38), we made a conscious choice to focus on peer-reviewed articles.While this procedure might risk excluding relevant knowledge from smaller or unsuccessful implementation attempts or other research adjacent to implementation processes, we delimited the results to the literature based on actual experiences from implementation in everyday clinical practice.Our initial screening of title and abstracts did not require decisions by two reviewers, but all decisions in the full-text screening were confirmed in pairs.We deliberately worked to maintain consistency and mitigate individual variation through bi-weekly meetings where we worked to establish a psychologically safe environment that encouraged all authors to raise or flag doubts, discuss the application of exclusion criteria, or differing interpretations.When in doubt, we would backtrack or repeat without blame, and all conflicts and uncertainties were resolved through discussion until consensus was reached.Additional meetings were held with other experts in the domain to ensure methodological rigour.While the Arksey and O'Malley framework for scoping reviews does not include a quality appraisal (43), we would recommend that future authors consider doing so as the number of articles that carefully consider implementation increases.

Conclusion
The current body of empirical evidence demonstrates a dissonance between the research and practice needs.On one hand, the conceptual and methodological AI research builds on large promises of AI to revolutionize healthcare and problematizes its slow uptake into practice.On the other hand, the current empirically supported knowledge derives mostly from implementations of AI systems with low action autonomy and highlights lessons on the implementation process that are typical to the implementations of other types of information systems.Further research is needed on the more disruptive types of AI systems being implemented into routine care, in order to identify those aspects of implementation unique to AI.This highlights the need for future research to advance in two main streams: (1) to empirically study the implementation processes of various types of AI systems into healthcare practice, and (2) to support the empirical research and practical implementations by developing and disseminating an AI-specific implementation framework that would take into account some of the unique aspects related to uptake of AI in healthcare such as building trust, addressing transparency issues, developing explainable and interpretable solutions and addressing ethical concerns around privacy and data protection.
Click here to enter text.Table 3: Overview of articles included in the scoping review (n=45)

1 .
General information: author/s, publication year, country, clinical setting, study aim, study design 2. Types and applications of AI: AI technology used, type of AI model, type of task performed by AI, level of action autonomy, intended use of AI, intended user of AI 3. Implementation process: Research focus, motives for implementation, elements in the implementation process, and frameworks used. Figures

Table 1 :
Concept areas and synonyms used to develop the search strategy

Table 2 :
Eligibility criteria and their rationaleCriteria Rationale Inclusion criteriaPeer-reviewed Greater credibility since they have been reviewed by peer experts in the field Empirical study design Empirical studies improve the ability to answer the

Year Country Clinical setting Study aim Study design AI technology Task performed Intended use of AI Research focus Motives for implementation AI model Levels of autonomy Intended user of AI Elements in the implementation process
Click here to enter text.Click here to enter text.Click here to enter text.Click here to enter text.Click here to enter text.Click here to enter text.Click here to enter text.