Key points are not available for this paper at this time.
The prevalence of diabetes is projected to reach 783 million individuals by 2045 1. This rise will inevitably lead to a substantial increase in the prevalence of diabetic retinopathy (DR), which affects ~30% to 40% of people with diabetes and progresses to sight-threatening disease in 10% 2. Timely screening reduces DR-related sight loss and has contributed to a decline in certifiable blindness in the United Kingdom, which has a national DR screening programme 3. Despite the benefits of screening and the rising global burden of DR, an estimated 80% of people with diabetes worldwide lack access to regular screening 2. This gap is particularly pronounced in low and middle-income countries (LMICs), where resource constraints and personnel shortages hinder the establishment and maintenance of effective DR screening programmes 2. Artificial intelligence (AI) has emerged as a promising tool for automating DR grading, offering the potential to efficiently scale DR screening and alleviate a reliance on human graders. IDx-DR (Luminetics Core) was the first autonomous AI system in medicine to receive Food and Drug Administration (FDA) authorisation in 2018 for more-than-mild DR detection 4. Deep learning-based AI systems have consistently demonstrated high diagnostic accuracy for referable DR detection in numerous retrospective and prospective validation studies 4. AI systems are continually advancing, with Daley et al., in this issue of Clinical and Experimental Ophthalmology (CEO), describing the development of a dual-modality system which uses fundus photographs and optical coherence tomography scans 5. In addition to accuracy, AI systems have demonstrated cost-effectiveness, enhanced clinical workflow efficiency and improved screening completion rates 6, 7. Reflecting this progress, a recent review identified multiple approved AI systems for DR detection in the United States (3), European Union (21), United Kingdom (7) and Australia (8), collectively representing over half of all approved ophthalmic image analysis AI systems 8. However, despite these regulatory achievements and promising validation data, the implementation of AI in clinical settings remains limited relative to the vast at-risk population. In the United States, FDA-approved AI systems for DR screening have been deployed in primary care and diagnostic centres, with billing records indicating over 15 000 uses between 2018 and 2021 9. Deployment has predominantly occurred in higher-income urban areas associated with academic institutions 9. The growth of AI-based DR screening in the United States has been modest, increasing by 1% between 2021 and 2023 compared to 185% for traditional fundus imaging used for DR screening in the same period 10. In Singapore, the SELENA+ AI system has been integrated into the national DR screening programme, representing one of the first deployments of a deep learning-based AI system into routine national-scale screening 11. The integration process encompassed regulatory clearance, cost-effectiveness evaluation, and the embedding of the system into clinical workflows 11. A recent real-world prospective validation study demonstrated that the system had an operational sensitivity of 94.7% and specificity of 82.2% for detecting referable DR 11. A large-scale pilot implementation of the automated retinal disease assessment (ARDA) system (Google/Verily) has been conducted in India and Thailand, collectively encompassing over 600,000 DR screening sessions 12. Encouragingly, prospective evaluations of AI systems are ongoing in rural and underserved populations, such as Indigenous communities and LMICs, where the need is greatest 4, 5, 13, 14. Automated retinal image analysis tools have been used in DR screening programmes in Scotland, United Kingdom (iGraderM) and Portugal (RetMarker) since 2011, with RetMarker reducing human grader burden by 48.4% 15. Although not based on deep learning, these systems demonstrate that automated grading can be safely and effectively incorporated into DR screening programmes over long time horizons 7. A recent scoping review by Tran et al., published in this issue of CEO, identified 18 implementation studies between 2018 and 2023 of AI systems being trialled in clinical DR screening 16. The study identified several barriers (e.g., training, infrastructure and cost) and enablers (e.g., efficiency, accessibility and outcomes) that influence the implementation of AI in DR screening 16. Therefore, despite the promising progress to date, AI adoption at the scale required to meet the rapidly rising prevalence of DR globally faces challenges. AI systems must be seamlessly integrated into clinical workflows to realise their clinical utility. Early experiences, such as Google's ARDA pilot deployment in primary care clinics in Thailand, demonstrated that even high-performing AI systems can fail if clinical workflows are disrupted 17. Factors such as image quality, additional administrative steps, and operator experience influence real-world AI system operation 16, 17. Staff must be trained to use AI systems, and in some settings, high staff turnover could present an ongoing challenge 11. Policies are needed to resolve conflicts when differences between AI system outputs and operator judgement emerge, or in cases where AI systems make errors. Successful deployment also requires adequate infrastructure including reliable fundus cameras and network connectivity for cloud-based AI. In LMICs, resource constraints present substantial challenges to acquiring the necessary infrastructure, but handheld cameras and offline AI systems offer a solution. Once established, AI can enable increased screening efficiency and faster DR classification compared to manual grading 16. Most AI systems for DR screening have been developed using datasets predominantly from Western and East Asian populations, raising concerns regarding generalisability to underrepresented populations. Variations in ethnicity, imaging devices and disease prevalence can impact AI system performance. Therefore, efforts to improve data representativeness and conduct rigorous validation across diverse populations are essential. Paradoxically, populations most in need of accessible DR screening, particularly in rural and underserved communities, often face the greatest barriers to access but also hold the greatest potential to benefit 16. Ensuring equitable access will require targeted public health strategies and policy frameworks that promote appropriate resource allocation. Innovative outreach programmes, such as the Lions Outback Vision initiative in remote Western Australia, exemplify approaches to extend AI-enabled DR screening to marginalised populations. Health economic analyses suggest that AI-based DR screening is cost-effective 7. However, economic challenges remain. Setup costs are non-trivial, involving the procurement and maintenance of secure data infrastructure and licensing fees 11. Inference costs, if incurred every time the AI system analyses an image, can accumulate rapidly, particularly in high-volume screening programmes. Reimbursement for AI-enabled screening is not universal, and many health systems lack clear policies covering payment, leaving providers uncertain about returns 11. These financial uncertainties pose significant barriers, especially in resource-constrained settings, and highlight the importance of clear funding pathways to sustain AI-assisted DR screening initiatives. AI systems process sensitive patient data which could pose security concerns, and healthcare users may have fears related to the commercial use of their data by third-party vendors. Whilst in-house AI systems are an attractive option for screening providers, significant regulatory and technical challenges associated with developing, validating, hosting and managing AI systems and data infrastructure need to be overcome. The projected increase in the global burden of DR represents a major and growing public health issue. AI-enabled DR screening offers an accurate, scalable and potentially cost-effective solution to addressing substantial gaps in screening coverage, especially in resource-constrained settings. The technical feasibility and clinical effectiveness of these AI systems are well established, but their broader clinical implementation requires addressing challenges in workflow integration, infrastructure, equity concerns and economic barriers. Addressing these issues will require coordinated efforts between clinicians, researchers, policymakers, industry stakeholders and end users to develop equitable and sustainable adoption of AI in DR screening. The global prevention of avoidable DR-related vision loss using AI is within reach, but realising this potential requires coordinated and strategic efforts from all stakeholders. Dr. Pearse A. Keane is supported by a UK Research & Innovation Future Leaders Fellowship (MR/T019050/1), Moorfields Eye Charity with The Rubin Foundation Charitable Trust (GR001753) and an Alcon Research Institute Senior Investigator Award. Dr. Keane is a cofounder of Cascader Ltd. and has acted as a consultant for insitro, Retina Consultants of America, Roche, Boehringer-Ingelheim, and Bitfount and is an equity owner in Big Picture Medical. He has received speaker fees from Zeiss, Thea, Apellis and Roche, and grant funding from Roche. He has received travel support from Bayer and Roche. He has attended advisory boards for Topcon, Bayer, Boehringer-Ingelheim and Roche. The authors declare no conflicts of interest.
Nderitu et al. (Mon,) studied this question.