All published studies and gray literature that reported implementation findings related to adoption or reported factors impacting adoption were considered in this review. Therefore, studies with various designs, including quantitative and qualitative studies, literature reviews, thought articles, conference papers, and reports, were included in the initial search and review. “Health care organizations” were defined as organizations that are engaged in providing care to patients or involved in some aspect of providing agency to health care players. Health care players were defined as anyone involved in the process of providing or receiving care, including policy makers; administrative professionals; clinicians; physicians; and, most importantly, patients and their families. All types of AI technologies were considered in this review. Articles were not excluded based on variations in settings (hospital vs community setting) or countries where the research was conducted. Only articles in English were included. Due to the speed at which the landscape for AI is advancing, only articles that were published between January 2011 and December 2023 (when the search was conducted) were included.

This review is intended to synthesize findings from publications that reported on the barriers to and facilitators of the adoption of AI implementations. A search was conducted on MEDLINE, ScienceDirect, and IEEE Xplore in December 2023. Keywords were selected in reference to the question identified to formulate the scope. Keywords included “artificial intelligence”; “healthcare” or “health care”; “hospital,” “health services,” or “health facilities”; “adoption”; “barriers”; “obstacles”; “challenges”; “facilitators”; and “enablers” (Textbox 1).

A total of 2 reviewers independently screened the articles from the initial search by reviewing their titles and abstracts. Articles meeting the inclusion criteria were identified. Articles that did not meet the inclusion criteria were excluded. Any discrepancies were resolved through discussion.

Of the articles identified, the full text of the semifinal set of articles was reviewed to further refine selected articles. This process was iterative, and some exclusions were made during the writing phase, as the findings evolved. An Excel spreadsheet (Microsoft Corp) was used to record the articles identified. Recordings included the following details: the name of the article, authors, journal, whether the article was peer reviewed, type of paper, discipline, country, region, method, population, end users, and type of AI application (if specified). Duplicate studies were identified and removed to ensure there was no overlap.

On the basis of the conventional content analysis approach used (as described in the Methods section), a total of 18 categories of barriers and facilitators were identified (Figure 2). Interestingly, the themes were found to provide perspective on both facilitators and barriers. For example, if the theme explainability was identified as a barrier, the same theme was tabulated as a facilitator to capture what the articles recommended for overcoming challenges with explainability to increase adoption. As such, the reporting of the results for each theme provided perspective on the theme being both a barrier and facilitator, with the exception of governance, which was entirely noted as a key facilitator.

Explainability can be defined as the ability to deconstruct an algorithm to understand the mechanism by which it arrived at the output. Explainability has gained prominence due to the fast-paced growth of ML algorithms such as DL. These algorithms are labeled “black box” due to the difficulty in interpreting and tracing the techniques used by the AI models, thereby impacting trust and demanding the need for transparency [31,32]. According to Holzinger et al [33], “explainable AI deals with the implementation of transparency and traceability of statistical black-box machine learning methods, particularly deep learning.” While transparency has a wider definition and explainability is a component of transparency [34], most studies have noted explainability in the context of algorithmic transparency; therefore, findings from these 2 interrelated concepts have been discussed together.

Several studies noted a lack of explainability as a barrier to adoption. Baxter et al [35] reported concerns from adopters around the lack of explainability regarding the prediction of the AI algorithm embedded in the EHR to predict unplanned readmission; specifically, the lack of explainability regarding what features of the algorithm were driving the output was an impediment to trust among adopters. Other studies noted that the way the data were being used to train algorithms was not clear. The lack of traceability and logical understanding of how the algorithm arrived at a recommendation contradicted a key foundation of evidence-based medicine, which relies on high standards of explainability. Clinicians expressed the need to understand both the scientific and clinical bases of the recommendations provided by the AI to confidently validate and apply the decision [19,20,36-38]. Morrison [36] particularly quoted stakeholders seeking clarity on the extent to which there is a need to provide transparency on AI output to patients and how this is directed by legislation or data protection laws. In a study conducted by Nadarzynski et al [39], users of an AI-enabled chatbot reported hesitancy to use the technology due to a lack of transparency on how the chatbot accurately arrives at responses to health inquiries.

According to Holzinger et al [33], “explainability is an important element for consideration in order to enhance trust of medical professionals.” To facilitate adoption, improving algorithmic transparency will be a key consideration to change attitudes and build the trust of adopters [35,40]. Additional recommendations around facilitating adoption were to have processes in place to support clinicians in case of disagreements on decisions due to a lack of transparency and explainability [41]. Furthermore, revealing the process of how the algorithm was developed, who was involved in the development process, whether clinicians were consulted, and how the data were processed would enable acceptability [42,43].

In health care, causality is especially important when using automated decision systems; therefore, Holzinger et al [33] emphasized that AI systems should support the understanding and explanation of the causal models as opposed to simply solving through pattern recognition. Similarly, Gillner [44] also noted that this opacity of the output is not aligned with the “medical ethos.” Weinert et al [45] recommended that investing into explainable AI that produces a transparent and understandable AI could help address the issue of acceptability. Moorman [38] reported that successful adoption was achieved by publishing evidence on the algorithm’s underpinnings and providing clinicians with details on how data elements interacted within the algorithm to produce the predictive output.

A prominent theme that came through was around the prevention of algorithm bias to ensure equity and fairness and avoid concealed discrimination. Algorithmic bias has been defined by Panch et al [46] as “the instances when the application of an algorithm compounds existing inequities in socioeconomic status, race, ethnic background, religion, gender, disability or sexual orientation to amplify them and adversely impact inequities in health systems.” Such biases have been visibly found in glomerular filtration rate and pulmonary function and have continued to persist despite efforts to address them. Inaccurate and underrepresentative training data sets for AI models can cause bias, misleading predictions, adverse events, and large-scale discrimination, causing barriers to adoption [32,41,47,48]. Baxter et al [35] and Chua et al [49] reported clinical stakeholders’ concerns around the relevance of the AI model’s output, especially because the algorithm did not consider social determinants of health to predict risk outcomes for readmissions. Similar concerns were raised by participants from other studies around the risk of algorithm bias as a challenge for adoption. Others expressed dissatisfaction that the AI algorithm may not be representative of the patient population among whom it is implemented or may have been trained with a biased training data set that has been retrofitted to produce certain results, therefore not providing a representative outcome of interest [36,49-51]. AI not accounting for patients’ health determinants was noted as a “grand challenge” [19].

Inadequate data from representative groups, algorithms designed to represent a majority, and missing variables that impact predictions are components that contribute to bias [21]. This can be addressed by engaging clinicians in the design and development of the algorithm to ensure that appropriate measures are taken to address bias before the AI algorithm is deployed.

One of the major themes that came through was around the value, usefulness, and accuracy of an AI algorithm. Accuracy and quality of the AI algorithm’s output were primary reasons for adoption hesitancy in the context of the functionality of the AI. In some studies, patients reported the need to assess the usefulness of the AI before using it and had concerns around the quality and accuracy of the output, thereby questioning the value of AI as a whole [39,42,52-55]. Clinicians also reported concerns around the usefulness of the output based on the lack of accuracy and inactionable output contributing to a low likelihood of use. This also included dissatisfaction if recommendations were too similar, inappropriate, or not useful [35,37,38,40,56]. Ease of use; complex interfaces; and inconsistent performance, for example, due to false positives and negatives add burden to the workflow, creating more work for clinicians, thereby impeding adoption [42,52,57-59]. Morrison [36] identified a lack of an agreed standard and benchmark for accuracy (how accurate does an AI tool need to be before it is approved for clinical practice) as an impediment to implementation and, subsequently, adoption, as, if a standard existed, it could provide rationale for the accuracy. Temsah et al [60] recommend the application of evidence-based oversight mechanisms to ensure accuracy and dependability. Finally, Choudhury [61] noted that if an algorithm is not performing up to standards or adds more work to the clinicians, this can impact adoption, as clinicians perceive it as a high risk.

Perceived benefit, perceived usefulness, usability, ease of use, usefulness, accuracy, and reliability of the output of the AI are key contributors to adoption [19,45,62-67]. In particular, usability and acceptability should be assessed with the intended user in mind [41]. Perceived benefit is especially important when it contributes to improved efficiency in clinical processes [62].

From a patient’s perspective, there is value if an AI technology can be used from home for minor consultations, such as skin cancer detection using an AI-enabled app [42]. Ease of use of the technology and accessibility to information for minor health concerns [39] are especially seen as valuable, as they negate the need for a visit to the physician; however, in the event that a visit is required based on the AI’s recommendation, then the integration of the technology with the health care system is considered beneficial [42]. It is essential that as the AI system matures, it is designed such that it can “adopt and challenge contradictory rules and behaviours” [43].

Patient safety concerns causing adverse effects were noted by Mlodzinski et al [48] and Vijayakumar et al [59]. The lack of accuracy of AI output was also considered to pose a potential risk of harm by both clinicians and patients, as, in some cases where the algorithm may output false negative results, it may provide an incorrect sense of reassurance and cause a delay in treatment. However, in cases where the algorithm is too sensitive, thereby providing false positive results, it may add work and costs to the treatment process [39,42,57,61,68]. The risk of harm can be lowered if AI algorithms are developed with the 5 rights (in the case of an automated decision system), similar to other clinical decision support tools [40]. In addition, Sangers et al [42] proposed that in some cases, AI applications should provide only risk indication instead of a diagnosis to reduce the risk of harm. Chen et al [68] reported that policies and mechanisms to safeguard professionals could address challenges associated with a potential risk of harm due to a lack of output accuracy.

Lack of accuracy; doubts about unsafe results; privacy breaches; patients’ perceptions and acceptance of automation; and uncertainty about developers’ reliability, availability, usability, and perceived usefulness were found to be obstacles to gaining trust [39,42,47,50,53,54,69,70]. Clinicians expressed fear around having to reframe their professional identity and responsibilities [57]. Fear around what AI really meant was noted as another barrier [36,42].

Facilitators of trust included the endorsement of the technology by experts as well as academically backed clinicians, including regulating bodies such as the government; evidence of output accuracy based on the evaluation of AI; and positive opinions from trusted thought leaders in the respective clinical fields [42,58,60]. Another facilitator of trust is the engagement of patients in the development of AI. This could facilitate trust in the public and address concerns around trust in data sharing [71]. Overall, trust was largely found to be associated with perceived usefulness; however, one study noted that “if peoples’ confidence and beliefs are improved, they will use the product despite its usefulness” [52]. In addition, Fan et al [72] reported that initial trust is a key indicator of the intention to use the AI application and noted that if the confidence to use the AI application is high based on performance expectancy, then this will increase trust in using the AI.

The lack of human intervention was found to be a barrier to adoption for both clinicians and patients. From a clinician standpoint, physicians expressed that they would be less likely to use an AI, given their familiarity with the patient’s condition and the value of intuition in clinical decision-making [32,35]. Mlodzinski et al [48] reported concerns about potential systemic bias present in the AI that could impact the patient-provider relationship. From a patient standpoint, the lack of human presence was seen as a limitation due to a lack of empathy and emotional connectivity with another human or simply not having the presence of a human physician to verbally communicate and discuss, such as when using an AI app or chatbot [51,64]. By contrast, the lack of human presence was, in some cases, seen as a benefit due to the anonymity in sharing intimate or uncomfortable health concerns [39,42]. Hemphill et al [73] reported increased confidence in patients when AI was combined with clinical interpretation.

Several studies particularly highlighted the importance of strategic alignment with initiatives. Baxter et al [35] reported on how a lack of alignment among different organizational initiatives led to varying outcomes and disjointed communication. Strohm et al [57], on the other hand, talked about how strategic alignment could lead to better dispersing of funds across different departments.

Sun and Medaglia [50] pointed out the necessity of outlining a comprehensive “top-down strategy” that would include organizations’ goals and resource distribution for AI implementation. Weinert et al [45] elaborated that to overcome the barrier to including AI initiatives in the organizational strategy, the German government introduced a new law that supported organizations with financial assistance to implement innovative digital technologies such as AI, as there was hesitancy among organizations to include expensive AI implementations as part of their strategy due to a lack of funds.

Several studies noted that to start the journey of implementing AI, there is a need to identify a problem and not merely use data to come up with a solution. Therefore, use cases should be identified based on notable problems that can be addressed by AI solutions. One particular study mentioned how the lack of a use case affected the implementation of the AI model [35,36].

A lack of sufficient buy-in from end users and a lack of endorsement from organizational leadership emerged as barriers, including not engaging stakeholders early in the process. It is critical to incorporate clinicians and other stakeholders, such as patients, in the development life cycle, especially the testing phase with the application of a user-centered design and testing approaches. This may be time intensive but proves to be an effective approach to enable adoption [39,40]. As Ongena et al [69] pointed out in their findings, patients should be involved when developing AI systems, specifically for diagnostic, treatment planning, and prognostic purposes. Pou-Prom et al [74] reported the usefulness of engaging end users in designing, deploying, and refining the AI solution. Moorman [38] recommended maximizing buy-in and engagement at all levels of stakeholders, from leadership to users, and especially engaging a clinician leader from the onset. Finally, Goldstein et al [75] noted the inclusion of champions at the leadership and clinical levels to achieve successful implementation.

This review found that the lack of integration of the AI system into the workflow can be a barrier to successful implementation and adoption. For example, Baxter et al [35] reported impacts on success due to variations in existing workflows for risk assessments and readmission scores across different areas. Similarly, Strohm et al [57] mentioned how the lack of integration and standardization of workflows led to variations in workflows. For other types of AI solutions, such as apps, the data not being integrated into the health care system or workflow was seen as a barrier for patients to adopt the solution [42,53]. Helenason et al [58] and Schepart et al [56] both noted the importance of conformity with the workflow when integrating AI tools into the environment where they would be used.

Recommendations included the following: ensuring that AI applications easily integrate with existing IT systems, integrating data from patients’ use of AI-enabled apps into the health care system and workflow, and considering the integration of the AI into the clinical workflow but maintaining autonomy for clinicians to have the final say [19,20,38,42,49,57,75]. In situations where the AI system is deployed in different areas, using a common model may improve alignment in workflows [35,36]. Chen et al [68] reported that clinicians saw AI integration into the workflow as a positive if the system was seen as potentially eliminating routine work, allowing them to focus on other tasks. Moorman [38] reported that it was helpful to assess existing unit workflows, communications, escalation, and event management processes before implementation to address challenges brought up by clinicians concerned about added work.

Training refers to educating the users on various aspects of the technology, such as the outcomes of the AI model, its benefits, and how it supports the clinical workflow, and providing new skills such as technical and data science skills to staff, especially laggards and champions, to assist with the adoption [36,41,57,62]. Kelly et al [21] particularly noted that to provide clinicians with clarity on how an algorithm could improve patient care, approaches such as using a decision curve analysis that would provide quantified benefits of using a model to inform actions that need to be taken would be helpful. Skepticism and a lack of understanding were seen as barriers to AI adoption [48,51]. Chen et al [68] found that AI adoption increased among radiologists who were more familiar with AI. Sun [76] recommended that implementation teams should consider influencing clinicians by sharing AI knowledge via more informal communications, such as social media communication or in-person communication. Moorman [38] found it helpful to develop educational material with input from clinicians to tailor it to the clinical role and hospital culture. Training and awareness should include building an understanding of the technology; providing clarity around language such as the definition of AI; education around data use in health care; and building awareness on the value of AI, including breaking down concepts that dispel fear [36,37,39,42,50,52,71]. Clinicians’ awareness and knowledge of AI before using it contributed to its successful acceptance [59,73]. In addition, users feel that the technology is “unqualified” based on their perception of the premature nature of the technology [39]. Misunderstanding of the capabilities of AI technologies in the general public and the health care sector is a challenge to adoption, with gaps in awareness around the value, advantages, and high expectations of AI [37,50]. Overall, Alsobhi et al [70] emphasized the urgency of accelerating AI adoption through the dissemination of AI training.

Shortages of personnel with the required skills were reported as barriers, along with the quality of IT infrastructure available for AI implementation [32,35,36,40,45,50]. Hickman et al [71], in particular, noted the lack of technological (infrastructure) maturity to allow for the integration of AI. The presence of AI experts, perhaps a multidisciplinary team, particularly with clinical scientists, data science and subject matter experts with AI skills, an innovation manager, AI experts to provide training, and local champions within departments involved in the end-to-end process, was considered an important element for adoption [32,57,59,62,71]. Goldstein et al [75] noted that for scalability, where the AI application would be deployed in multiple sites, having designated resources from the onset of the project with clear roles and responsibilities was seen with success. Yang et al [77] recommended cultivating talent with both high-level medical and technology knowledge and understanding how the 2 domains can be used to meet patients’ needs. In a different perspective on the shortage of resources, Chen et al [68] noted that radiographers and radiologists held more positive attitudes toward the adoption of AI, as it would help address workforce shortages in the radiology field in the United Kingdom.

The need for evaluation on multiple fronts was noted by various studies. Studies indicated that the technical evaluation of AI is necessary as a first step to validation. Technical evaluation must be followed by clinical validation (based on established methods in clinical research) and economic validation. Wolff et al [78] particularly noted the lack of clinical and economic measurements as a barrier to practical implementation. Evaluations should be tailored toward digital technologies to gather empirical evidence surrounding the value of AI’s use [53,78]. However, the cost of conducting an empirical evaluation and a quantified clinical trial–type evaluation may be a deterrent to the pace of developing the technology. Therefore, this should be considered when selecting the type of evaluation to be conducted. A focus on assessing the effectiveness and accuracy was duly highlighted. Implementers should consider validating or testing the algorithm with synthetic data [40-42,57]. Establishing a standard methodology for the validation of AI algorithms and the overall evaluation of AI will be critical to gaining confidence from adopters [50]. Hickman et al [71] suggested that having structures in place for the continued monitoring of standards that impact AI (eg, regulatory standards) and ensuring that infrastructure is in place to evaluate and monitor algorithms continuously are necessary.

Data quality, security, ownership, and storage were prominent themes in the reviewed studies. In terms of data quality and integrity, several issues were identified as barriers to developing good AI models that provide value to users. These were issues around variability, the nature of unstructured data, incompleteness of data, the data not representing the reality of clinical care, and the absence of data standards (specifically around how and what data are collected). Having metadata standards, terminologies, data quality metrics, and common data models were identified as facilitators in resolving some of these issues [40,43,45,50]. Fragmented access to data and limited sources, such as the availability of data only from EHRs or data silos, were also noted as barriers [78].

Data access, integrity, and provenance are key to the development of models. Institutions that were the most successful in implementing AI were thoughtful about how to guarantee data integrity [40]. Wolff et al [78] noted the challenges with data silos and fragmented access to medical data, including limitations on the availability of data only from the EHR, to enable the development of robust AI applications. In terms of data ownership, the dilemma around who owns the data, whether it would be the government, institution (eg, hospital), or patient, is a barrier to adoption, as it leaves questions around how data would be integrated or accessible for future AI advancement [37,50]. While this may not present a direct adoption challenge, it does indirectly impact the availability of data required for development and to produce meaningful outputs, which is an impediment to adoption, as identified earlier.

Data security was identified as a major contributor toward hesitancy, with concerns around cybersecurity relating to both training and testing data as well as fear of trackers and spyware obtaining unsolicited data [39,41,42,45,48,77]. Furthermore, the ability of deidentified data to be reidentified poses a major risk for the individuals and institutions providing their data. This further contributes to resistance to sharing the data that could help expand data sets for AI training. Several strategies for preventing security breaches have been proposed, and these could be helpful in securing data, especially health data that are accessible over the web [20,42].

Concerns around data processing include the lack of understanding of how data are stored, processed, and accessed; the establishment of protocols; and compliance with existing privacy policies, such as the General Data Protection Regulation and the Health Insurance Portability and Accountability Act [20,36]. A survey of patients conducted by Ongena et al [69] indicated the need for patients to be informed about how and specifically which data are processed. One study explicitly highlighted the issue of data ownership. Concerns over data ownership were particularly evident when patients linked data ownership to trust in the technology [47].

Concerns about privacy and ethics were focused on maintaining confidentiality, ensuring processes are in place to obtain consent, and having informed consent with clarity on how the data will be processed [32,39,47,56]. Having clear consent processes related to how data are generated through the use of AI and how these data flow as well as defining the meaning of consent and transparency on strategies to maintain privacy are seen as facilitators of adoption. This is especially applicable to “clinical and epidemiological use cases of ML in both decision support and automation categories, as data from patients or the public are essential to train algorithms in these areas” [20]. In addition, Sun and Medaglia [50] particularly pointed out the unethical use of data, such as data being used by commercial organizations. Ensuring transparency to end users, especially patients involved in the ethical and legal frameworks that guide the development of AI systems, could be helpful [69,73]. Weinert et al [45] particularly identified ethical issues, specifically as they relate to liability, as a barrier to AI adoption. Wolff et al [78] noted that integrating “privacy-by-design” technologies into AI applications that incorporate advanced data protection features could mitigate such challenges.

Governance was primarily noted as a key facilitating factor, playing the role of enabling the full cycle of AI. It is critical to have a governance structure in place to oversee the development and rollout of AI from conception to implementation, with governance tools providing guidance on various stages of the process. Governance should include diverse professionals with clear articulation of accountability, including nuances in reactions to accountability [35,40,55,58,64]. Isbanner et al [64] noted the importance of articulating accountability. This is especially important in health care because “ethical and governance challenges matter to the public.” Wolff et al [78] recommended outlining specific responsibilities for different stakeholders to delineate accountability-based steps in the process, for example, identifying who would review an x-ray image analysis and identifying liability and culpability (eg, obligatory human check of a decision obtained by an AI application). According to Sunarti et al [47], the governance body should include “developers of software, government officials, health care, medical practitioners and advocacy for patients groups.” The lack of accountability in the decision-making process is a challenge; therefore, framing this in the governance model could be a way to address adoption issues related to accountability [37,50,73]. Other functions of the governance body would be to ensure funding and connectivity to the wider data science community, ensure alignment with strategic initiatives in the organization, and act as a long-term centralized knowledge repository for performance oversight. A governance model should have mechanisms and systems in place to facilitate changes impacting AI technologies in development or use based on cyclical changes in the technology or changes in the external landscape, such as the ones initiated by regulatory bodies [32,41,79]. Formalized analysis of ethical considerations in the development and use of AI should be a key component of governance. Governance should also be linked to the data governance committees for various data processing, quality, and integrity oversights [43]. One particular solution proposed by Morrison [36] was to have national-level governance templates that would facilitate national data protection via the implementation of impact assessments. In contrast, governance can hinder data sharing. Therefore, governance bodies should maintain a rigorous process without becoming a constraint [36].

A lack of regulation and policies from the government, including uncertainty around legal direction or law, was presented as a barrier to the application of AI technologies [37,45,57,60,77,78]. Other researchers noted that there was no clarity in the area of regulatory structures with regard to which regulatory body should be consulted for AI developments and deployments [36,53]. Therefore, it would be essential for governments to establish regulatory bodies and legal frameworks to provide guidance on various aspects of AI development and application [20,41]. In addition, ambiguity surrounding malpractice liability policy as it relates to physicians’ legal responsibilities, for example, in case of diagnostic errors, remains a barrier to AI adoption [49].

The lack of and uncertainty surrounding funding are presented as barriers to implementation [32,51,56]. Researchers have suggested that there is a need to have costs identified from the start-up stage all the way to scalability. Funding can especially be a barrier if an AI technology is lacking in evidence, with little to show for the value it provides. Sun and Medaglia [50] and Xing et al [53] reported that financial barriers in the context of cost and benefits, the lack of a sustainable business model, and insufficient funding to meet public demands should also be considered. Sun and Medaglia [50] additionally noted challenges associated with the adoption of IBM Watson in China due to patients having to pay high fees for the service. Finally, funding should be cohesively considered not only for the development of the technology but also for resources required to implement the technology, such as technical subject matter experts, project managers, and champions [36,40,41,51,57]. Weinert et al [45] noted that to overcome the barrier of lack of resources and meet the financial investment demands of AI implementations, the German government introduced a new law that could help organizations bridge funding gaps; however, they could not conclude whether this would facilitate any progress, as the announcement was just made.