An umbrella review of the facilitators and barriers to implementing Artificial Intelligence (AI) solutions within hospital settings: through the lens of the NASSS framework (spread, scale-up and sustainability)
1University College London
2University of East Anglia
* Corresponding author; email: sigrun.clark@ucl.ac.ukAbstract
Advancements in artificial intelligence (AI) are revolutionising the healthcare sector, but challenges exist in AI adoption and its long-term use. This umbrella review aimed to identify the facilitators and barriers of AI implementation within hospitals and was registered on PROSPERO. Five databases (MEDLINE, HMIC, CINAHL Plus, Web of Science and Cochrane Reviews) were searched in January 2025, 763 articles were screened, with 13 included. The inclusion criteria encompassed studies implementing AI that were conducted within the hospital setting. The quality of the data were assessed using the ROBIS checklist and data were extracted using the NASSS (Nonadoption, Abandonment, and challenges to the Scale-up, Spread, and Sustainability) framework, to demonstrate how AI implementation was affected by: whether the AI solution had been technologically validated to ensure generalisability across departments; evidence the AI solution brings measurable gains; a lack of trust or understanding among hospital staff; the budgets and resources available to onboard the AI solution, train staff, and maintain the solution; the need for national policies on funding and regulating AI solutions. These factors affected the adoption, spread, scalability and sustainability of AI implementation and could be considered in future implementation efforts. The study was funded by the NIHR (NIHR205439).
Article notes
Competing Interest Statement
The study was funded by the NIHR (NIHR205439).
Funding Statement
The study was funded by the NIHR (NIHR205439).
Introduction
Advancements in artificial intelligence (AI) are revolutionising the healthcare sector1,2. AI offers potential economic advantages, enhances medical services, and optimises the use of resources within hospitals1,2. AI is an overarching system that processes large bodies of data and is able to learn from the data to solve problems through a subset of analytical techniques such as machine learning2,3. Machine learning is the process in which a computer system is able to identify patterns, learn from those patterns, and then perform actions to produce results3. A machine learning algorithm will automatically adapt its algorithm based on its experience of receiving repetitions of sample data along with desired outcomes, this is a process known as training the machine learning algorithm. As a result, the algorithm will produce the desired outcome from the training sample data and should also be able to generalise its algorithm to produce the desired outcome from new (non-sample) data4. Despite AI’s transformative potential in healthcare, its adoption and long-term use is shaped by various factors, including education, organisational and financial factors within institutions, and regulatory policies5. Identifying the factors that could act as barriers in implementation can help facilitate the implementation of AI solutions across healthcare systems.
Existing research, including systematic reviews by Hassan et al. and Ahmed et al., has examined an array of barriers and facilitators to AI implementation in healthcare2,6. Key hindering factors identified by these publications included insufficient IT infrastructure, high upfront costs, and ambiguous legal frameworks surrounding liability. Conversely, influential facilitators such as transparent governance structures, well-defined protocols for data security, and ongoing professional development have been shown to bolster acceptance and adoption rates. There have also been numerous systematic reviews focussing on the barriers and facilitators of AI implementation within specific healthcare contexts such as acute care7, nursing8 and infection control9. Although these reviews highlight critical insights into how AI integration into healthcare workflows is influenced by various factors in specific clinical areas, no umbrella review has been conducted to explore cross-cutting factors across clinical settings and types of AI solutions. The aim of this umbrella review is to identify common trends in implementation and collate the factors acting as barriers and facilitators to implementing AI in hospital settings. Addressing this gap would enable the field to glean a more cohesive understanding of the multifaceted issues at play and develop tailored strategies for embedding AI solutions more seamlessly into hospital settings.
Methods
This umbrella review was developed in accordance with the Preferred Reporting Items for Overviews of Reviews (PRIOR) guidelines outlined by Gates et al. (2022)10. The protocol for the review was published on PROSPERO (CRD 42025638087)11.
1.Search strategy
A scoping search of the existing literature was conducted on Google Scholar. As a result of this process the combination of the following search terms were employed: ‘Artificial Intelligence’ AND ‘Hospital’ AND ‘Systematic review’ across five databases: MEDLINE, CINAHL Plus, HMIC, Cochrane Reviews and Web of Science in January 2025. The detailed search criteria can be found in Appendix 1.
2.Inclusion and exclusion criteria
The eligibility criteria that were used for the review can be found in Table 1. The types of studies that could be included encompassed systematic reviews, literature reviews, scoping reviews/rapid reviews. There were no limitations on the types of study design that could be included, so quantitative, qualitative and mixed method study designs were all relevant. Hospital settings were the only type of clinical setting that would be included in the review, and AI and its subset technologies were considered. There were no exclusions based on the country of the study. Publications were limited to those that shared the perspectives on the barriers and facilitators to implementing AI in the hospital setting. Studies not mentioning AI, those conducted outside hospital environments (e.g., dental practices, primary care clinics), or articles that were single empirical studies rather than reviews were excluded.
3.Study selection
Once searches were retrieved from the scientific databases, the results were de-duplicated using Endnote and were then imported into the software Rayyan12 for title and abstract screening. Four researchers conducted screening of the same 25% of title and abstracts and underwent team discussions to identify discrepancies and reconfirm the teams understanding of the eligibility criteria. Two researchers then independently screened the remaining publications. A similar process was then initiated with full text screening, with the group of researchers conducting screening of the same 25% of full texts, followed by team discussions to identify discrepancies and build confidence for two researchers to then screen the remaining full texts.
4.Data extraction
A data extraction form was developed using Microsoft Excel based on the Nonadoption, Abandonment, Scale-Up, Spread, and Sustainability (NASSS) framework13. The data extraction form was piloted among the research team who extracted the same 25% of included publications to build confidence and ensure alignment, three researchers then independently extracted data from the remaining publications.
5.Critical appraisal
The ROBIS tool is a critical appraisal tool for assessing the risk of bias in systematic reviews14. ROBIS was utilised to evaluate literature quality included in the review across different study designs which included qualitative, quantitative, and mixed methods studies. The quality assessment was conducted by three researchers, one conducting the initial quality appraisal, and two additional researchers cross-checking half of the decisions each.
6.Synthesis
The content from the data extraction form was synthesised using narrative synthesis to deductively group together content that could fit under the categories from the NASSS framework.
Results
1.Study Characteristics
The search was performed across five databases, yielding 963 records. After duplicates were removed, 763 records remained for title and abstract screening. After the screening of titles and abstracts, 215 articles remained and underwent a full-text review. As a result of the comprehensive review of records sourced from database searches, 9 articles were selected for data extraction. An additional 4 articles were also selected for data extraction; these articles were handpicked from the scoping process conducted prior to implementing the umbrella review. An overview of the results from the screening process can be found in Figure 1.
All studies were ranked with a high risk of bias based on the ROBIS scores, as demonstrated in Table 2. The researchers agreed on the overall high-risk score for each publication; there were more granular discrepancies, which were then discussed between the researchers. Some of the ROBIS criteria were not applicable to the qualitative studies in the umbrella review, as the criteria were mainly relevant to quantitative meta-analyses. This may have affected the overall ROBIS scores and is a limitation of the research.
Most reviews focused on the use of AI or machine learning-based approaches broadly, whilst one review focused specifically on the use of generative AI chatbots in the form of ChatGPT8. After checking the individual studies included in each review, the research team identified that seven of the 13 reviews in our umbrella review, contained some of the same studies.
The following table summarises the 13 articles included in the umbrella review.
2.The reported factors acting as facilitators and barriers aligned with the NASSS framework
The Nonadoption, Abandonment, and challenges to the Scale-up, Spread, and Sustainability (NASSS) framework has been used to show how AI adoption in the hospital setting depends on (1) the condition, (2) the technology, (3) the value proposition, (4) the adopter system, (5) the organisation, (6) the wider context, and (7) how the domains interact with each other over time. Table 3 highlights these seven domains and summarises the key themes within each domain identified from the umbrella review.
Domain 1: The Condition
The first domain of the Nonadoption, Abandonment, and challenges to the Scale-up, Spread, and Sustainability (NASSS) framework concerns the clinical characteristics of the condition for which a technology is introduced. Very few reviews in this umbrella analysis addressed the extent to which characteristics of specific conditions may impede or facilitate AI implementation. The 2023 review by van der Vegt et al. was the only review that discussed how the definition of the outcome of sepsis could affect the algorithm’s performance. Depending on which outcome definition was used (e.g. initial inflammatory process vs multi-organ dysfunction) the algorithm could either predict the likelihood of developing sepsis or just diagnose sepsis16.
Domain 2: The Technology
Lack of robust technological infrastructure emerged as a key impediment to scaling-up AI-based solutions. Outdated networks, limited bandwidth, or poorly maintained servers produced delays in AI output, undermining clinicians’ confidence in the AI solution’s recommendations2,16–21. It was recommended across the literature that AI needs to be able to integrate within existing computational resources. Hassan et al. (2023) stressed the significance of embedding AI solutions into existing IT infrastructure with minimal disruption. Kamel Rahimi et al., emphasised that technology fails when it does not fit smoothly into daily workflows – for example using an external dashboard often demanded that staff switch between screens, leading to fragmentation and fatigue17.
Concerns around the accuracy of outputs from the AI were also raised across the literature. The potential for false positive warnings led to alert fatigue among users, and inaccurate outputs created burden to the user’s workflow, impacting user perspectives of the AI and impeding adoption2,16–19,21,22. Mitigation strategies to address these concerns were raised in the publications and included technical validation, followed by clinical validation. Additional barriers included the concerns around the quality and relevance of the data being fed into AI platforms2,6,8,17,21,22, and that AI systems could perpetuate biases present in the data they were trained on2,6,19,21. One recurring issue across the studies was the “black-box” phenomenon2,6,18. Many advanced AI models lacked interpretability or transparency, making it difficult for clinicians to trust or fully adopt them. Other concerns from potential end users included whether the AI model had been externally validated to assess its generalisability, and whether the outputs produced by the AI could be generalisable across different departments or hospitals16,17,21. This domain revealed that transparent, user-friendly, and well-integrated technology is a key prerequisite for the scaling and spread of AI solutions in hospitals.
Domain 3: The Value Proposition
The third NASSS domain examined whether the innovation filled a recognised need for diverse stakeholders: patients, clinicians, administrators, and technology developers. Multiple reviews observed that the perception of the value of the AI solution can either catalyse or block widespread adoption.
Kamel Rahimi et al. indicated that end users were more inclined to support the AI solution if it delivered measurable gains through cost savings and improved clinical outcomes17. Some examples of gains included reduction in surgery cancelations, reductions in patient’s length of stay in hospitals and improved efficiency of clinician’s workflows19,23. Other reviews highlighted that a lack of empirical evidence on the added value of AI application in certain practices could act as a barrier to its implementation2,6,18,24. The review by Hassan et al. mentioned that an enabler to AI implementation would be to identify a problem that the AI should address, rather than implementing the AI solution without a use case2. Similarly, Kamel Rahimi et al. shared how concerns from end users regarding whether new AI solutions were relevant to their workflow could be a barrier to adoption17.
Domain 4: The Adopter System
The fourth NASSS domain addressed users’ potential and willingness to adopt the technology—specifically, healthcare professionals, hospital administration teams, data scientists, and patients. The reviews consistently found that staff receptiveness to AI was influenced by factors such as awareness, training, perceived autonomy, and trust.
A lack of trust was frequently reported across the publications due to a misunderstanding of outputs or concerns around the accuracy of the AI outputs2,6,16,17,21,23,24. Reviews found that trust improved when users actually had experienced an AI solution making correct predictions16,21. Often discussed as a contributor to the scepticism towards AI was the lack of training and education among staff2,6,7,9,16,18,21,22. The provision of training for the interpretation of AI outputs, and raising awareness among staff in regard to using AI could help to facilitate trust and uptake in AI implementation2,16–21,24.
Some of the reviews identified a fear among clinicians that AI could undermine their professional expertise or lead to job displacement, this finding was similar to reviews that reported that balancing the use of AI with clinician autonomy could be an enabler to its implementation2,6,16,18,19,24. Factors cited in the literature to address these concerns included co-designing the AI solutions with those who were going to implement the solution or be affected by its implementation such as patients, healthcare professionals, and data scientists2,16,17,21,22,24. The literature also cited the importance of developing a multi-disciplinary team including data scientists and clinicians who would be involved in implementing the AI solution together, to improve team understanding of the solution.
Overall, the adopter system domain underscores that successful AI implementation requires a workforce that sees genuine benefits, feels equipped to use the tool, and experiences no existential threat to professional identity. Without these conditions, scale-up is likely to stall.
Domain 5: The Organisation
The fifth domain covered factors that affect an organisation’s ability to adopt AI systems, such as funding decisions, leadership involvement, and readiness to adapt operational processes. Many reviews detailed how organisational culture and resource constraints can impede new technologies17,18. The reviews highlighted that the organisational cost for AI implementation could be a barrier to its spread and scalability, specifically the cost for the AI solution, the costs associated with integrating the AI into existing infrastructure, the cost for training staff and the cost for reserving time for staff or appointing new job roles to implement the AI solution2,6,16–19,21,22. Another point was raised in relation to sustainability and that there would be costs associated with maintaining the AI and maintaining a level of training among staff.
The synergy between senior leadership commitment and AI expansion emerged as a vital enabler in multiple reviews2,16,17,21. Kamel Rahimi et al. highlighted that leadership backing allowed hospital teams to allocate staff time and resources to implement AI solutions. Another structural issue was how AI intersects with existing workflows, the publications cited that a facilitator to AI implementation would be non-disruptive integration or staggered integration to existing workflows2,16,17,21. Other facilitators discussed in the reviews included having an implementation plan to provide guidance on the adoption of the AI solution, including aspects such as who would be responsible for receiving the alerts2,17,21.
Domain 6: The Wider Context
NASSS’s sixth domain addressed political, regulatory, economic, and sociocultural factors. A cross-cutting finding from these reviews was that the external policy environment based on regulatory bodies and government actions can either bolster or undermine local AI initiatives. The authors repeatedly cited that the trajectory of the AI solution in the hospital setting is shaped by external bodies, from government agencies to professional societies and accreditation institutions. Many of the reviews discussed the need for action on: financial and legal endorsement2,6; national regulatory standards that consider ethics around consent, data protection, data ownership, legal liability2,6,17,19; and an agreed standard for the accuracy of AI before it is approved for clinical practice2.
Economic considerations also emerged strongly in the domain of the wider context across the literature. Kamel Rahimi et al. found that financial challenges in health systems could create favourable conditions for AI if the solution was efficient and produced cost saving, but at the same time the high cost of AI could be a barrier for many health systems17. Hassan et al. flagged the need for healthcare systems implementing AI solutions to know all the cost upfront from starting, to scaling, and sustaining AI2. Ahmed et al. raised the risk of a two-tier health system forming, with some hospitals gaining access to supportive AI solutions whilst others may face barriers due to funding constraints6.
Domain 7: Interactions Between Domains and Adaptation Over Time
The final NASSS domain focused on how the previous six domains interlinked with each other, over time, shaping the resilience and long-term viability of AI interventions. The included reviews collectively showed that complex interdependencies determine whether an AI solution is implemented. Most commonly intertwined were the domains technology and the adopter system: the complex outputs from AI solutions could lead to a lack of understanding among hospital staff leading to barriers in adopting the technology17,21,23. Potential mitigation measures to the lack of understanding included conducting user-feedback and pilots that could lead to iterative updates to the AI solution, making the outputs more user-friendly, or providing more context and transparency behind the technology2,16–19,21,24.
Discussion
This umbrella review has offered a broad perspective on the factors acting as barriers and facilitators influencing the adoption, spread and sustainability of AI implementation in hospital settings. Although the included systematic reviews covered diverse clinical contexts—from operating room management, infection prevention, critical care documentation and wearable patient monitoring, certain cross-cutting insights emerged.
The most common factor that affected adoption was hospital staff perspectives on the trust and understanding of the AI solution2,6,9,16–24. Implementation science has long shown that healthcare professionals’ scepticism or poor understanding of a new intervention routinely obstructs roll-out. This is consistent with existing literature on healthcare professionals’ perspectives on e-health interventions, or complex interventions more generally whereby a lack of understanding, lack of training or lack of trust for a specific intervention or solution can affect implementation25,26. National and local organisations seem to recognise the importance in ensuring an understanding of AI systems, NHS England have developed a report on developing healthcare workers’ confidence in AI, which sets outs strategies to develop training curricula on using AI27,28. Local hospitals sites are following suit: Bedfordshire Hospitals NHS Foundation Trust issued a dedicated Artificial Intelligence Policy in 2024 that mandates staff training on AI fundamentals29. Such policies may help to improve the trust and understanding gaps flagged above.
Spread and scalability can be facilitated if AI solutions are externally validated to ensure generalisability across different contexts such as departments and hospitals16,17,21. Generalisability of research findings across different hospitals is a goal across most health research30. However, there is a growing debate that the generalisability of machine learning models across different hospitals is unachievable due to differences in the operational characteristics of hospitals, instead it may be more appropriate to ensure the algorithm works well within one local setting31–33.
To facilitate the sustainability of AI solutions, the cost of maintaining the AI solution in terms of technology and staff training needs to be considered2,6,16–19,21,22. Addressing system level factors such as funding and regulations to sustain AI implementation, is an approach that the UK Government is now pushing forward – with funding and plans for supporting AI implementation within the NHS34,35.
The findings from this umbrella review can support the future implementation of AI solutions in hospital settings. It is evident from the published studies that AI solutions that are co-produced with clinicians, patients and other end-users, tested in the real world and refined in an iterative way are more likely to respond to the needs of users and be implemented successfully. Furthermore, support from hospital senior leaders can facilitate uptake.
The strengths of this review include that it has been guided by the NASSS framework from the outset, ensuring that our data extraction and synthesis systematically captured the domain insights regarding AI implementation and sustainability in hospitals. This research leveraged a large, multidisciplinary research team, allowing us to conduct double screening at both title and abstract and full-text stages on a subsample of the articles, as well as double data extraction for a subsample of publications, and a full double quality appraisal on all included articles. These measures aimed to minimise bias and enhance the reliability of our findings. Although we used the ROBIS tool for critical appraisal, all reviews were judged to have a high risk of bias, reflecting both the wide heterogeneity of included papers and the fact that ROBIS is more tailored to quantitative systematic reviews. It does, however, remain critical to note that many of the reviews included in this umbrella analysis face their own methodological shortcomings, as indicated in the risk-of-bias assessments. We also recognise that as seven of the included 13 reviews contained some of the same included literature, this may have caused limitations in our research by over-representing themes.
In summary, this application of the NASSS framework shows that implementing AI in hospital settings is a complex undertaking, involving individual, organisational, and technical factors. The emergent conclusion is that sustainable AI adoption requires thoughtful alignment across each NASSS domain.
Conclusion
This umbrella review pinpointed both the facilitators and barriers for adopting AI solutions in hospitals, and through the NASSS framework we have been able to emphasize the role of technology, value, the adopter system, the organisation and the wider context. The common factors that affected adoption included a lack of trust in and understanding of the AI solutions, whilst scale-up and spread was affected by the generalisability of the AI models. Sustainability was affected by the availability of funds and resources to maintain the AI solution. Some of the factors that acted as facilitators in implementation included co-design activities, comprehensive training, and gaining buy-in from senior leadership. Implementation could also be facilitated by regulatory bodies who can produce transparent governance frameworks and policies that address funding, data security and legal liability.
Data Availability
All data produced in the present study are available upon reasonable request to the authors
Appendix Group
Appendix
1.Appendix 1. Search strategy
MEDLINE – 13 January 2025
Ovid MEDLINE(R) ALL <1946 to January 10, 2025>
CINAHL Plus – 13 January 2025 (no limits on date of publication or language)
HMIC – 13 January 2025
HMIC Health Management Information Consortium <1979 to November 2024>
Cochrane Reviews – 13 January 2025
”artificial intelligence” OR “ai” OR “a.i.” OR “machine learning” OR “deep learning” in Title Abstract Keyword AND “hospital” OR “tertiary care” OR “secondary care” OR “acute care” OR “inpatient care” OR “intensive care” OR “ward” in Title Abstract Keyword AND “systematic review” OR “meta-analysis” OR “meta-analyses” OR “meta-synthesis” in Title Abstract Keyword
Results: 3
Web of Science – 13 January 2025
”artificial intelligence” OR “ai” OR “a.i.” OR “machine learning” OR “deep learning” in Topic AND “hospital” OR “tertiary care” OR “secondary care” OR “acute care” OR “inpatient care” OR “intensive care” OR “ward” in Topic AND “systematic review” OR “meta-analysis” OR “meta-analyses” OR “meta-synthesis” in Topic
Results: 363
Appendix 2. List of included articles
- Bellini V, Russo M, Domenichetti T, Panizzi M, Allai S, Bignami EG. Artificial Intelligence in Operating Room Management. J Med Syst. 2024;48(1):19. doi:10.1007/s10916-024-02038-2
- Kamel Rahimi A, Pienaar O, Ghadimi M, et al. Implementing AI in Hospitals to Achieve a Learning Health System: Systematic Review of Current Enablers and Barriers. J Med Internet Res. 2024;26:e49655. doi:10.2196/49655
- Kucukkaya A, Arikan E, Goktas P. Unlocking ChatGPT’s potential and challenges in intensive care nursing education and practice: A systematic review with narrative synthesis. Nurs Outlook. 2024;72(6):102287. doi:10.1016/j.outlook.2024.102287
- Lee C, Britto S, Diwan K. Evaluating the Impact of Artificial Intelligence (AI) on Clinical Documentation Efficiency and Accuracy Across Clinical Settings: A Scoping Review. Cureus. Published online November 19, 2024. doi:10.7759/cureus.73994
- Piaggio D, Zarro M, Pagliara S, et al. The use of smart environments and robots for infection prevention control: A systematic literature review. Am J Infect Control. 2023;51(10):1175-1181. doi:10.1016/j.ajic.2023.03.005
- Baig MM, GholamHosseini H, Moqeem AA, Mirza F, Lindén M. A Systematic Review of Wearable Patient Monitoring Systems – Current Challenges and Opportunities for Clinical Adoption. J Med Syst. 2017;41(7):115. doi:10.1007/s10916-017-0760-1
- Tan S, Mills G. Designing Chinese hospital emergency departments to leverage artificial intelligence—a systematic literature review on the challenges and opportunities. Front Med Technol. 2024;6. doi:10.3389/fmedt.2024.1307625
- van der Vegt AH, Campbell V, Mitchell I, et al. Systematic review and longitudinal analysis of implementing Artificial Intelligence to predict clinical deterioration in adult hospitals: what is known and what remains uncertain. Journal of the American Medical Informatics Association. 2024;31(2):509-524. doi:10.1093/jamia/ocad220
- van der Vegt AH, Scott IA, Dermawan K, Schnetler RJ, Kalke VR, Lane PJ. Deployment of machine learning algorithms to predict sepsis: systematic review and application of the SALIENT clinical AI implementation framework. Journal of the American Medical Informatics Association. 2023;30(7):1349-1361. doi:10.1093/jamia/ocad075
- Hassan M, Kushniruk A, Borycki E. Barriers to and Facilitators of Artificial Intelligence Adoption in Health Care: Scoping Review. JMIR Hum Factors. 2024;11:e48633. doi:10.2196/48633
- Ahmed MI, Spooner B, Isherwood J, Lane M, Orrock E, Dennison A. A Systematic Review of the Barriers to the Implementation of Artificial Intelligence in Healthcare. Cureus. Published online October 4, 2023. doi:10.7759/cureus.46454
- Lambert SI, Madi M, Sopka S, et al. An integrative review on the acceptance of artificial intelligence among healthcare professionals in hospitals. NPJ Digit Med. 2023;6(1):111. doi:10.1038/s41746-023-00852-5
- Ouanes K, Farhah N. Effectiveness of Artificial Intelligence (AI) in Clinical Decision Support Systems and Care Delivery. J Med Syst. 2024;48(1):74. doi:10.1007/s10916-024-02098-4