Abstract
Background
The rising incidence of diabetes and its complications, particularly diabetic retinopathy (DR), in sub-Saharan Africa presents a significant public health challenge, compounded by a lack of skilled human resources. Artificial intelligence (AI)-based screening offers a promising way to address this burden. Understanding the integration of such complex interventions from research to routine clinical practice is crucial for sustainability.
Methods
A qualitative study design was employed using the Normalization Process Theory (NPT) framework, which supports the evaluation of whether an innovation will be sustainable in everyday practice by analysing how new practices become embedded into routine work. Clinical staff and patients involved in AI-based DR screening at three clinics in Kigali, Rwanda were interviewed using an NPT-based questionnaire through semi-structured interviews and focus groups. Verbal consent was obtained for recording and transcription. Interview data were thematically analysed, with codes generated to align with the four NPT constructs: coherence, cognitive participation, collective action, and reflexive monitoring.
Findings
In total, nine clinical staff members and 67 patient participants were interviewed. Participants reported a coherent understanding of the program's purpose, value, and benefits. They valued the technology for its ability to address the gap between the scarcity of competent healthcare providers and the growing burden of DR. Challenges identified related to workload, division of labour, initial patient distrust of AI, and restrictive organizational policies regarding operator access.
Interpretation
The NPT framework proved valuable for analysing the implementation of this complex intervention, providing insights into user perceptions and generating actionable recommendations for enhancement. While positive adoption was observed, further research is needed to fully understand the intervention's long-term impact on health outcomes such as visual preservation and treatment success.
Funding
Fundus cameras were donated by Topcon. The company had no input into the design and/or analysis of the study.
Keywords: Diabetic retinopathy, Artificial intelligence, Qualitative, Africa
Research in context
Evidence before this study: Several studies have shown that real-time implementation of artificial intelligence (AI) for large-scale diabetic retinopathy screening is feasible including recent research conducted in Rwanda. The results of these studies provide evidence on the benefits of AI screening in promoting adherence to prescribed treatment for diabetic eyecare in sub-Saharan Africa.
Added value of this study: Outside of structured research studies, little is known about the factors that may influence the adoption of AI-based screening into daily clinical practice. To successfully and sustainably implement new technologies, such as AI, in healthcare in the future, it is important to understand the factors that contribute to its success or failure, which this study aims to evaluate.
Implications of all the available evidence: Successful implementation and adoption into daily practice of new technology in global healthcare settings requires recognition and an understanding of the factors that contribute to its success or failure. This analysis can inform future optimization strategies and resource allocation decisions regarding the integration of AI-based diabetic retinopathy screening programs in clinical practice in sub-Saharan Africa and globally.
1. Introduction
The global burden of diabetes mellitus and its complications, particularly diabetic retinopathy (DR), is a growing public health concern, with a notable increase in incidence and prevalence across sub-Saharan Africa [1]. This escalating burden is often unmet by the availability of sufficient and proficient human resources in many low-income countries, highlighting a critical gap in diagnostic and management capabilities for DR, a leading cause of preventable blindness [2].
Technological advancements, such as artificial intelligence (AI)-based screening tools, offer a promising solution to address this deficit. Recent research, including in Zambia [3] and the RAIDERS study in Rwanda [4], [5], have demonstrated the significant benefits of AI in DR screening. Following the RAIDERS study [4], [5], fundus cameras (NW400, Topcon, Japan) and the AI platform, Cybersight AI, (Orbis International, New York, USA) were deployed in three clinics in Kigali, Rwanda, signifying a transition from a research setting to real-world clinical application.
While the efficacy of such interventions may be proven in controlled research environments, their sustainable integration into routine community-based clinical practice presents unique challenges. Understanding the facilitators and barriers to this transition is essential for successful implementation and long-term impact. This is the precise point at which the theoretical framework of the Normalization Process Theory (NPT) is introduced [6]. The NPT is an ideal tool for this investigation, as it was specifically developed to analyse how new practices or technologies become embedded into routine work [7]. This study therefore explores the practical application and acceptance of AI-based DR screening within the Rwandan healthcare system.
2. Materials and methods
This study adopted a qualitative, exploratory design, utilizing the NPT framework and adhering to the Standards for Reporting Qualitative Research (SRQR) guidelines to investigate the processes of implementing and integrating an AI-based diabetic retinopathy screening program into routine clinical practice in Rwanda. The study was conducted in three clinics in Kigali, Rwanda: 1) Rwanda Diabetic Association (RDA) diabetes clinic (a non-governmental organization); 2) Rwanda Military Referral and Teaching Hospital (RMRTH) (a public hospital); and 3) La Frontier Medical Clinic (a private clinic). Fundus cameras and the AI platform designed to screen for DR were deployed in all three locations.
A data triangulation approach was taken by including both clinical staff and patient participants. Inclusion criteria for clinical staff consisted of ensuring the participant was directly involved in operating the fundus cameras and AI platform (n = 9) and consented to participation in the study. Patient participants must have had experience with the intervention through having undergone AI-based DR screening (n = 67) and completed the exam with either a positive or negative outcome. Screening positive was defined as more than mild retinopathy detected by the AI system. A purposive sampling strategy was employed to ensure representation from all three clinics and capture diverse perspectives.
2.1. Data collection
Data were collected through semi-structured interviews and separate focus groups for clinical staff and patients. A pre-designed questionnaire was developed based on the four constructs of the NPT framework (Coherence, Cognitive Participation, Collective Action, and Reflexive Monitoring) guided the interviews. Staff responded to all the questions while only a specific set of questions were asked to patients.
The questionnaire was pretested using a small group of five respondents to ensure the translations were sound and that the choice of questions to ask patients was appropriate and adequate for analyses. The majority of interviews (n = 65) were conducted in the local language, Kinyarwanda, with two respondents interviewed in English. Pretesting resulted in minor modifications to the Kinyarwanda translations to improve clarity. Verbal informed consent was obtained from all participants prior to their interview for audio recording and transcription, with anonymity and confidentiality assured. The study was approved by the Rwanda National Health Research Committee (NHRC/2020/PROT/025) and the Rwanda National Ethics Committee (945/RNEC/2020). The tenets of the Declaration of Helsinki were followed throughout.
2.2. Data analysis
The audio-recorded interviews were transcribed verbatim, and thematic analysis guided by the NPT framework was employed to analyse the qualitative data. A data triangulation approach was applied by comparing themes across clinical staff and patient interviews to improve qualitative rigor. Transcripts were reviewed iteratively by authors WM and OU to identify initial themes and codes, which were then systematically grouped and categorized under the four core constructs of the NPT framework. Authors discussed any differences identified in themes and coding and results were documented once consensus among authors was reached. Participant quotes are in italics with ellipses (…) used to show where words were removed and [text] indicates where they were added for increased readability. To ensure data confidentiality and reduce re-identification risk in a small setting, clinic names and specific job roles were removed from all participant quotes.
3. Results
In total, nine clinical staff members, 77·8% (n = 7) female, ranging in age from 59 to 25 years old, and 67 patient participants, 55·2% (n = 37) female, mean age 48.25 years(range 78-23y) were interviewed. Among the nine clinical staff four were from the non-governmental organization, two from the private clinic, and three from the public clinic. All patients were African and all clinic staff were Rwandans except one. Of the patient participants, 34·3% (n = 23) screened positive for a referral from the AI exam. The NPT framework was utilized to organize and present the study's findings, highlighting the facilitators and barriers to implementation across each of its four constructs. (Table 1).
Table 1.
Summary of perceived facilitators and barriers.
| NPT Construct | Perceived Facilitator | Perceived Barrier |
|---|---|---|
| Coherence |
|
|
| Cognitive Participation |
|
|
| Collective Action |
|
|
| Reflexive Monitoring |
|
|
3.1. Construct 1: coherence (sense-making)
This construct explores how participants understood and made sense of the AI-based screening intervention. The findings indicate a strong sense of coherence, as participants reported a common understanding of the program's purpose, value, and benefits. A key facilitator was the shared recognition that the AI tool provides a crucial bridge between the growing burden of DR and the scarcity of proficient human resources in the region.
A key indicator in the frame work is whether the intervention is easy to describe, which the staff could in their own words, “You use a special camera to take a picture of the retina and then AI tells you if it can see any damage from diabetes in the picture.” Interview responses indicate that clinical staff particularly valued the AI's ability to provide instant interpretation, which led to a faster turnaround time for patients and created an effective triage system for referrals. One clinical staff member noted, “Before the implementation of this AI-based screening, we used to refer all patients to ophthalmologist on a regular basis, which resulted in long queues and decreased compliance to the referral system, now we have a tool to determine and convince those who really need an urgent consultation.”
Patients articulated positive sentiments when asked whether the intervention had a clear purpose as supported by the response, “The computer based system is great as it tells you if you are normal, so that you don't need to go to spend hours at the hospital waiting for the screening of retinopathy. For me the first screening was normal, but when I came six months later the printed paper was in red so they transferred me to the eye hospital, and doctors there told me that diabetes was damaging my eyes and even my kidneys might be affected. These warnings helped me in taking serious measures to control the blood sugar.”
The primary barrier identified from patient interviews was initial patient distrust. A traditional belief that humans are superior to machines and robots led some patients to doubt the AI's results until they were confirmed by a qualified healthcare provider. This finding revealed that the successful use of the intervention was not solely dependent on the AI's technical performance but that patients valued its integration into a hybrid workflow, where the human clinician's role as a validator was essential for building trust. Patients responded, “We trust what our health care provider tells us after they look at the test because they are the ones to take care of our lives. If we start doubting about their knowledge, we might even stop taking the medicine, they prescribe…” and “I think that sometimes they [clinical staff] doubt the response given by the computer because sometimes they [clinical staff] have to repeat the exam multiple times before giving you [the patient] the results, but overall, I think it is in their competence to know if the test worked as expected.”
3.2. Construct 2: cognitive participation (engagement)
This construct assesses the commitment and engagement of participants to continue with the intervention. Responses support the ongoing participation with the technology among both staff and patients. Staff members were prepared to invest their time and energy, viewing new technology as a facilitator of service delivery and quality of care. The added value of the intervention to patient care was seen to outweigh any potential increase in workload or adjustments to their routines. A clinical staff member captured this by stating, “When an unsuspicious patient is found to have significant retinopathy, then the point of having such an intervention that can screen multiple patients to isolate that one who might be lost in the crowd becomes evident.”
An additional finding from several patient interviews was the role of patient-to-patient testimonies in popularizing the intervention. Several patients came for screening after hearing about the technology from other patients, as described in one response, “Yes I got to know about it happening here at RDA from my friend when they used a computer to take pictures of his eyes and provided a response in few minutes. We are glad that they had brought in new technologies”.
While inquiring whether patients would be willing to engage with the technology with commitment and regularity, the key emerging theme was that the technology saved them time. Patients articulated this as follows, “I always prefer a clinic where I don't spend a lot of time because I run a small business and if I close it the whole day then it becomes difficult to survive. For example I came from Kabuga, I pass by many clinics but I know that here at RDA I can find multiple services in short time and still go back to my work earlier…”.
Patients sentiments reflected an acceptance that the new technology would also fit well within the overall goals and activity of their diabetes clinic, with one patient stating, “Yes I think the nurses like the technology and they know what they are doing. They understand the procedure and can give explanation whenever you ask without rushing you”.
3.3. Construct 3: collective action (work done)
This construct examines how the intervention was integrated into existing work practices by the clinical staff and its impact on roles, responsibilities, and resources. This construct was not investigated among patients, but only included in clinical staff interviews (n = 9). The findings revealed a key barrier: an increased workload for the same staff members involved in daily patient care. This necessitated the creation of fixed screening days, leading to patient inconvenience and additional appointments. Another staff member remarked,“…It would be beneficial to all of us nurses and our patients to train all the staffs in the clinic to perform this investigation so that it is done whenever a patient requires it, instead of having fixed days of screening.”
A critical point of contradiction emerged from the data: the skills required to operate the AI camera were reported to be easily acquired, requiring only one day of training. However, the clinics' administrative bodies restricted the number of operators to ensure proper handling and preservation of the valuable, donated equipment. This highlights that the main barrier to scaling and full integration was not the technology's complexity, but an organizational policy driven by asset protection.
3.4. Construct 4: reflexive monitoring (appraisal)
Reflexive monitoring involves the ongoing appraisal and evaluation of the intervention by its users, the clinic staff. Over time users developed a positive perception of the intervention and provided constructive feedback. Users identified issues such as over-screening of the same patients, which led to workflow disruption and unnecessary appointments, and suggested a need for refresher courses on screening intervals and continuous mentoring. One staff remarked, “Sometimes I feel like we have screened everyone. You see this young man I was seeing just now we screened him when you were here for research and again, we screened him this year. Many of the patients coming now have already been screened.”
The feedback went beyond technical adjustments. Staff expressed a desire for additional training not only on the functionality of the fundus camera and AI platform but also on the pathological processes of DR. This was noted as a way to boost their confidence in delivering results and answering patient questions that went beyond the scope of the screening tool. Clinic staff linked increased knowledge and skills to a sense of professional development remarking, “As we screen patients, we are noticing an increase in awareness in this screening method, and more patients becoming increasingly curious about it and ask a lot of questions; we think patients are recommending it to other patients…We would benefit from additional training in diabetic retinopathy pathology to answer some of the patient's questions when we are delivering results to them.”
Furthermore, staff at one clinic expressed a need for a portable fundus camera to enable its use in mobile clinics and diabetes camps, demonstrating an active appraisal of the intervention's limitations and a clear vision for its expansion.
Patients contributed some reflections on the technology when asked whether they felt the staff were well-trained to carry out the intervention. “We believe that to become a health care provider supposes a certain level of integrity, so if the nurses here…have received proper training, they should be able to use it to test our eyes the same way they test us for other problems. After all technology tends to simplify things better than how they used to be done”.
4. Discussion
This research demonstrates the value and flexibility of the NPT framework for analysing the implementation of a complex intervention in a low-income context. The NPT framework has been successfully applied to analyse complex healthcare interventions, including digital health initiatives, in various contexts. A systematic review by May et al. (2018) provides a comprehensive overview of how NPT has been used to study the implementation of healthcare interventions, including in resource-limited settings [8]. Research has demonstrated its value in understanding the factors that enable or inhibit the embedding of new practices [8].
The findings on patient distrust and the subsequent reliance on human confirmation align with the concept of “optimal trust,” in AI, which balances reliance on technology with human oversight, is a well-established theme in the literature [9]. Studies on AI adoption in healthcare in low-resource environments, like those in sub-Saharan Africa, often highlight the crucial role of human validation in building and maintaining patient trust [9]. Patients are more likely to accept and use AI tools when they are integrated into a system where a trusted human healthcare provider confirms the AI's output [9]. This study provides a valuable data point from a sub-Saharan African population on how this trust is developed over time, moving from scepticism to confidence through positive experiences and human validation.
While much of the existing literature focuses on institutional drivers, there is a growing body of work that recognizes the power of social networks and peer-to-peer influence in health technology adoption, particularly in community settings [10]. The findings on cognitive participation in this study highlight a uniquely community-driven approach to technology adoption. The growing confidence in the technology was not just a result of institutional support or positive clinical outcomes but was also driven by informal conversations among patients. Patient-to-patient testimonials and word-of-mouth are powerful mechanisms for popularization and normalization, as they address scepticism and build confidence in a way that formal communication channels cannot [10].
The most critical finding from the collective action construct is the disconnection between the intervention's technical simplicity and its organizational barriers. The core challenge was not the difficulty of using the AI tool but a management decision that restricted its use, creating a bottleneck and hindering its full integration into routine practice. This barrier was relieved by implementing training for all levels of staff on the technology providing greater access and empowerment to staff for the integration of the technology. This underscores a vital point: the sustainability of a technological intervention depends as much on supportive organizational policies and a re-evaluation of workflow as it does on user acceptance. Research consistently shows that the failure of a healthcare technology is often not due to technical flaws but to organizational barriers, such as poor workflow integration, lack of supportive management, and misaligned policies [11], [12]. The most advanced technology can be rendered ineffective if a management decision creates a bottleneck that prevents its optimal utilization.
Finally, the findings from reflexive monitoring reveal, as has been reported by other researchers [13], that the AI is more than just a diagnostic tool; it is a catalyst for system-wide improvement. By bypassing human specialist scarcity and bringing immediate triage capabilities to community clinics that previously lacked eye care, the tool serves as a fundamental healthcare equity strategy. The staff's feedback on fears around over-screening and their request for deeper clinical knowledge suggest that the AI prompted them to critically appraise their own roles and practices. This process of reflexive monitoring is a core component of the NPT framework [6]. Further, reflective responses revealed aspects from the initial research protocols that were not carried over to the sustained clinical implementation that were still seen as beneficial to staff, “We are happy that we now know clearly where to refer patients. It would be good to know whether the patients come to the hospital like with the research before – you told us who came and who did not.” This sort of reflection on the impact of the technology and ensuring the patients adhere to their referral demonstrates the perceived value of the intervention due and also highlights the importance of making sure staff know how their work is creating sustained impact. Closing the loop for referral adherence is a critical aspect for any screening program to prevent vulnerable patients from falling through systemic gaps.
The NPT framework proved valuable for analysing the implementation of this complex intervention, offering insights into user perceptions and generating actionable recommendations for enhancement. The study identified a range of facilitators, including a shared understanding of the tool's value, strong user commitment, and the power of patient advocacy. It also highlighted critical barriers, such as initial patient distrust, organizational workflow disruptions, and restrictive policies that impede full normalization.
To operationalize these findings into practical implementation steps linked to NPT constructs:
(1) For Coherence, programs should establish a hybrid diagnostic workflow where clinicians confirm AI results to alleviate initial patient distrust.
(2) For Cognitive Participation, regular communication on the technology's time-saving benefits should be prioritized.
(3) For Collective Action, management must remove restrictive asset-protection policies and train all staff levels to distribute the workload evenly.
(4) For Reflexive Monitoring, continuous education focusing on DR pathology should be implemented to sustain user confidence and professional empowerment.
While the study provides valuable insights into the adoption process, a key recommendation for future research is the collection of robust follow-up data to fully understand the intervention's long-term impact on health outcomes, such as visual preservation and treatment success.
Our study carries several limitations that require cautious interpretation of the findings. First, the sample size of clinical staff was small, and all three clinics were located in an urban setting, which may not reflect rural experiences. By bypassing human specialist scarcity and bringing immediate triage capabilities to community clinics that previously lacked eye care, the tool serves as a fundamental healthcare equity strategy. Second, there is a risk of social desirability bias, as interviewees may have offered overly positive responses regarding the technology. Finally, the study lacks long-term clinical outcome data to definitively confirm the intervention's sustained clinical efficacy.
Finally, a significant limitation is that the retinal cameras and AI software licensing were provided free of charge through research infrastructure, creating an artificial economic environment. In real-world low-income settings, upfront capital equipment costs and recurring proprietary software licensing fees pose major implementation barriers for resource-constrained health systems [14]. Real-world implementation also incurs hidden costs, such as staff time redirected for training and workflow adjustments to manage scheduling. However, literature suggests that while the initial financial barrier is high, AI-based screening is ultimately more cost-effective than the standard clinical workflow over time [15]. It allows low-cost, mid-level clinic staff to handle screening via automated task-shifting, relieving the financial and labour burden on scarce, highly salaried ophthalmologists at the hospital level [16], [17]. Furthermore, for patients, the immediate point-of-care results remove the significant financial burden of repeat travel and lost wages typically required to collect delayed manual screening results. Because this study bypassed these foundational financial hurdles, further research is required to evaluate the true budgetary impacts and financial sustainability of non-subsidized AI implementation in this context.
While the study in Rwanda identified ‘asset protection’ and geographic/infrastructure limitations such as the inability to use non-portable cameras in mobile clinics as physical constraints to collective action, high-income countries (HICs) face entirely structural and institutional barriers. Research in HICs indicates that implementation is bound by complex regulatory governance such as HIPAA/GDPR data transmission compliance, medical liability concerns, and highly rigid, shifting insurance billing codes (such as CPT codes) that dictate which tiers of clinical staff can legally or financially operate the screening devices [18]. Furthermore, whereas staff in Rwanda see AI as a vital bridge to cover human resource scarcity, clinicians in high-resource settings frequently frame the barrier around role ambiguity, potential workforce competition, and changes to specialist referral pathways [19].
Ultimately, this research contributes to the growing body of evidence-based medicine in Africa by providing a framework for analysing how complex medical technologies are integrated into low-resource settings. These findings can inform policymakers and healthcare providers, ensuring that AI implementation is not just a technical deployment but a thoughtfully managed process that accounts for human factors, organizational policies, health equity and the realities of clinical practice. Lastly, these findings inform the development and implementation of AI-based screening interventions in other parts of Rwanda and globally.
Generative AI statement
AI was not used in the writing or preparation of this manuscript, nor were any figures or artwork generated or modified using generative AI.
CRediT authorship contribution statement
Wanjiku Mathenge: Writing – review & editing, Writing – original draft, Supervision, Formal analysis, Data curation, Conceptualization. Olivier Uwizeye: Writing – review & editing, Investigation, Data curation. Noelle Whitestone: Writing – review & editing, Writing – original draft. David H. Cherwek: Writing – review & editing, Supervision, Funding acquisition, Conceptualization.
Funding
Fundus cameras were donated by Topcon Healthcare, Inc.
Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgements
The authors are grateful to all participants in the study for their time and responses. In addition the authors wish to express their appreciation for the Cybersight AI team including Nicolas Jaccard and Gabriella Lanouette.
References
- 1.Burgess P.I., MacCormick I.J., Harding S.P., Bastawrous A., Beare N.A., Garner P. Epidemiology of diabetic retinopathy and maculopathy in Africa: a systematic review. Diabet Med. 2013;30(4):399–412. doi: 10.1111/j.1464-5491.2012.03756.x. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Vision Loss Expert Group of the Global Burden of Disease Study, GBD 2019 Blindness and Vision Impairment Collaborators Causes of blindness and vision impairment in 2020 and trends over 30 years, and prevalence of avoidable blindness in relation to VISION 2020: the right to sight: an analysis for the Global Burden of Disease Study. Lancet Glob Health. 2021;9(2):e144–e160. doi: 10.1016/S2214-109X(20)30489-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Bellemo V., Lim Z.W., Lim G., et al. Artificial intelligence using deep learning to screen for referable and vision-threatening diabetic retinopathy in Africa: a clinical validation study. Lancet Digit Health. 2019;1(1):e35–e44. doi: 10.1016/S2589-7500(19)30004-4. [DOI] [PubMed] [Google Scholar]
- 4.Mathenge W., Whitestone N., Nkurikiye J., et al. Impact of artificial intelligence assessment of diabetic retinopathy on referral service uptake in a low-resource setting: the RAIDERS randomized trial. Ophthalmol Sci. 2022;2(4) doi: 10.1016/j.xops.2022.100168. Published 2022 Apr 30. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Whitestone N., Nkurikiye J., Patnaik J.L., et al. Feasibility and acceptance of artificial intelligence-based diabetic retinopathy screening in Rwanda. Br J Ophthalmol. 2024;108(6):840–845. doi: 10.1136/bjo-2022-322683. Published 2024 May 21. [DOI] [PubMed] [Google Scholar]
- 6.Murray E, Treweek S, Pope C, et al. Normalisation process theory: a framework for developing, evaluating and implementing complex interventions. BMC Med. 2010;8:63. Published 2010 Oct 20. doi: 10.1186/1741-7015-8-63. [DOI] [PMC free article] [PubMed]
- 7.May CR, Albers B, Bracher M, et al. Translational framework for implementation evaluation and research: a normalisation process theory coding manual for qualitative research and instrument development. Implement Sci. 2022;17(1):19. Published 2022 Feb 22. doi: 10.1186/s13012-022-01191-x. [DOI] [PMC free article] [PubMed]
- 8.May CR, Cummings A, Girling M, et al. Using normalization process theory in feasibility studies and process evaluations of complex healthcare interventions: a systematic review. Implement Sci 2018;13(1):80. Published 2018 Jun 7. doi: 10.1186/s13012-018-0758-1. [DOI] [PMC free article] [PubMed]
- 9.Ghassemi M., et al. Building and maintaining trust in artificial intelligence in clinical care: a systematic review. J Med Internet Res. 2020;22(7) [Google Scholar]
- 10.Vella K., et al. The role of social networks in the adoption of health information technologies. Health Policy Technol. 2021;10(2):1–9. [Google Scholar]
- 11.Grol R., et al. Implementation of evidence-based practice in healthcare: a systematic review of the literature. Int J Qual Health Care. 2011;23(4):452–458. [Google Scholar]
- 12.Ludwick D.A., Doucette J. Adopting electronic medical records in primary care: lessons learned from health information systems implementation experience in seven countries. Int J Med Inform. 2009;78(1):22–31. doi: 10.1016/j.ijmedinf.2008.06.005. [DOI] [PubMed] [Google Scholar]
- 13.Greenhalgh T., et al. How do health information technologies affect clinical work and professional practice? A meta-review of the literature. BMC Med. 2017;15(1):1–13. [Google Scholar]
- 14.Rogers M., et al. Feasibility and equity of AI-driven models for diabetic retinopathy screening in sub-Saharan Africa: a systematic review and meta-analysis. Cureus. 2025;17(2) [Google Scholar]
- 15.Ginting A. Effectiveness of tele-ophthalmology and AI-based retinal imaging for diabetic retinopathy screening in rural settings: a systematic review. Int J Med Sci Health Res. 2026;46(2) [Google Scholar]
- 16.Bascaran C., Mwangi N., D’Esposito F., Gordon Iris, Ulloa J.A.L., Mdala S., et al. Effectiveness of task-shifting for the detection of diabetic retinopathy in low- and middle-income countries: a rapid review protocol. Syst Rev. 2021;10(1):32. doi: 10.1186/s13643-020-01553-w. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Cleland C.R., Rwiza J., Evans J.R., Gordon I., MacLeod D., Burton M.J., et al. Artificial intelligence for diabetic retinopathy in low-income and middle-income countries: a scoping review. BMJ Open Diabetes Res Care. 2023;11(4) doi: 10.1136/bmjdrc-2023-003424. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Chen JS, Lin MC, Yiu G, Thorne C, Kulasa K, Stewart J, Nudleman E, Freeby M, Han MA, Baxter SL. Barriers to implementation of Teleretinal diabetic retinopathy screening programs across the University of California. Telemed J E Health 2023 29(12):1810–1818. doi: 10.1089/tmj.2022.0489. [DOI] [PMC free article] [PubMed]
- 19.Tran J., Estevez J.J., Howard N.J., Kumar S. Barriers and enablers influencing the implementation of artificial intelligence for diabetic retinopathy screening in clinical practice: a scoping review. Clin Exp Ophthalmol. 2025 Sep-Oct;53(7):791–802. doi: 10.1111/ceo.14567. [DOI] [PMC free article] [PubMed] [Google Scholar]