Abstract
Diabetic retinopathy (DR) is the leading cause of preventable blindness among working-aged adults. It affects approximately one in three people with diabetes and guidelines recommend screening every 1-2 years depending on individual risk factors. Given the scale of diabetes and the demand for diabetic retinopathy screening, artificial intelligence (AI) is theorised to address the growing public health challenge of effective and accessible DR screening. This paper aims to: (1) synthesise evidence regarding efficacy of AI as a screening tool for DR, (2) understand the challenges in adopting this technology in clinical care pathways, and (3) appraise the health economic evidence for using AI in DR screening. We conducted a narrative review of selected peer-reviewed publications regarding validation studies of AI in DR screening, qualitative studies regarding usability and acceptability, implementation science studies and health economic evaluations. AI has been validated as an accurate screening tool for DR. However, there are challenges to its adoption, scale-up and spread within screening programmes around the world. There is a growing body of qualitative and real-world evidence which identify and seek to understand the implementation challenges, which include image gradability, workflow adjustments and costs of acquiring AI tools. Finally, AI is considered a cost-effectiveness tool for DR screening in most countries, especially high-income countries.
Introduction
Diabetes mellitus is a major global health challenge (1) which affects over 800 million people worldwide, with prevalence tripling since 1990 (2). An estimated one-third of people with diabetes will develop diabetic retinopathy (DR), making it the most common microvascular complication (3) and leading cause of preventable blindness among working-aged adults (4). The economic impact of DR is also substantial. In Australia, for example, the cost of diabetic macular oedema alone was estimated to be $2.07 billion in 2015 (5). To prevent the vision loss and costs associated with DR, current guidelines recommend screening every 1-2 years for people with diabetes, and at least annual screening for anyone with evidence of DR (6).
DR screening involves imaging of the retina and its blood vessels, primarily through fundoscopy or optical coherence tomography (7). The standard approach to DR screening is initial screening by primary eye care providers (optometrists, general practitioners or ophthalmologists via telehealth), with ophthalmology referral when DR is detected. Despite these procedures only requiring a photo of the eye, adherence to screening guidelines remains low around the world (8,9). Another concern is that ophthalmology follow-up among those with referable DR is reported to range from merely 5% within recommended timeframes to 51% within a one-year period (10).
Artificial intelligence (AI) may address these screening challenges of accessibility, cost-effectiveness and adherence to ophthalmology follow-up. Existing cost-effectiveness analyses indicate that AI screening may not only reduce costs but also improve clinical outcomes compared to manual grading by ophthalmologists (11). However, the reliability of models is limited by the availability of evidence regarding AI’s real-world utility (11). Furthermore, there is a need for evidence-based approaches for the use of AI in DR screening which can be transferable to other healthcare settings with their own unique complexities. Acknowledging these uncertainties, our paper aims to: (1) synthesise evidence regarding efficacy of AI as a screening tool for DR, (2) understand the challenges in adopting this technology in clinical care pathways, and (3) appraise the health economic evidence for using AI in DR screening.
I. Quantitative Studies
Diagnostic Accuracy
As the use of artificial intelligence popularises in healthcare, accuracy benchmarks become strong indicators in validating human-screening replaceability.
Specifically, implementing AI systems to evaluate the presence of DR within a patient’s retinal image requires a closer look at the algorithm’s results. Several metrics that are based on the algorithm’s responses are considered in determining its accuracy. These include specificity, sensitivity and gradability, which hold significant influence in making decisions about the algorithm’s quality and reliability.
In this context, specificity refers to the proportion of participants in a sample whom the algorithm correctly identifies as not having DR – referred to as a “true negative”. A higher specificity score for a particular algorithm would indicate that it is less likely to erroneously diagnose a patient with DR, given the patient’s input fundus photograph. Conversely, sensitivity refers to the proportion of participants in the same sample whom the algorithm correctly identifies as having DR – referred to as a “true positive”. In actual medical use, this metric would indicate the algorithm’s greater accuracy in correctly diagnosing a patient with DR (12). This is often prioritised in screening programmes to minimise the number of missed diagnoses, also known as “false negatives”. The implications of a false negative are detrimental in patients with DR. Overlooking patients with disease can result in the patient not receiving adequate treatment or therapy, therefore putting them at risk of vision loss or even blindness.
These accuracy metrics are critical in evaluating an AI algorithm’s reliability and safety for a patient. As algorithms advance, their reliability is increasing with tremendous promise in screening accuracy. A 2021 Australian study on AI screening in DR conducted by Scheetz et al., (12) using AI to screen various Indigenous and non-Indigenous populations, demonstrated the growing robustness of this technology. The AI used in the study had 96.9% and 87.7% sensitivity and specificity, respectively (12). These results overwhelmingly surpassed the United States Food and Drug Administration (FDA) recommendations with strong sensitivity and specificity scores.
Gradability of Retinal Images
Additional criteria in evaluating an AI algorithm’s DR screening accuracy include gradability. A critical factor in conducting accurate screening is image quality. This relates to the overall ability of a given fundus photograph to be visually interpreted by a healthcare professional or algorithm. Thus, sharper image quality may warrant greater diagnostic accuracy (13); however, if an image is unable to be properly assessed due to its quality, it is deemed “ungradable” or unable to have a diagnosis interpreted from it. This results in a patient being referred to an ophthalmologist without a confirmed diagnosis of DR.
While still an important consideration in manual grading, image quality serves an immense purpose in AI grading. Because of the algorithm’s sole dependence on the data that it has been fed, often of a very high image quality, it is crucial that a patient’s photo meets a standard of similar quality for the AI to make a diagnosis (13). Although there are many screening algorithms now, their reliability is not equal; in fact, this variation is often due to a difference in retinal cameras and data quantities. For instance, a 2022 Thai study on deep learning screening for DR conducted by Ruamviboonsuk et al., (13) including a data set of over 7000 participants, corroborated this variation in gradability. It found that certain cameras produced photos that were more or less prone to being classed as “ungradable” due to image quality. The Australian study by Scheetz et al. (12) above supports this notion by comparing a Canon, DRS and Topcon Maestro camera – the Topcon Maestro scored nearly 40% less in specificity than its counterparts. Not only were different cameras used in training the artificial intelligence model, but photos of varying quality and focus were also incorporated to help the model recognise and diagnose photos of lesser quality (12). Ultimately, greater exposure to different scenarios of DR or non-DR photos of the fundus contribute to increasing AI diagnostic accuracy by supplying more information to draw a better conclusion on.
Comparison to Human Screening Accuracy
As per the Thai study (13), the analysis found that the algorithm had 91.4% in sensitivity and 95.4% in specificity. In contrast, the retinal specialist over-readers had a sensitivity of 84.8% and specificity of 95.5%. These differences are far from trivial indicators of AI models’ competence in DR detection and function as promising statistics of a technology that may be more widely implemented. Within the Thai study’s data group, the particular AI model reassured its capabilities among accurate diagnoses (13).
Adherence to Follow-up/Patient Satisfaction
Although not necessarily indicative of an AI algorithm’s raw accuracy in DR screening, tracking the proportion of patients who followed up with their referrals is insightful data in considering real-world implementation. To start, “adherence to follow-ups” refers to compliance in attending a recommended appointment with a specialist after being positively diagnosed with DR. Revisiting the Australian study (12), of the 28 participants who were given a referral, 9 of them attended a specialist appointment. Although the referral follow-up yielded about a 32% success rate, a post-screening questionnaire, completed by 87.7% of the participants, indicated high satisfaction rates of 93.7%. Ultimately, high satisfaction rates have often been indicative of greater referral follow-up adherence, however there is limited data to validate this theory. Satisfaction rates also validate the feasibility of replacing manual grading by AI algorithmic analysis. However, when similar AI algorithms were implemented in Brazilian clinics for DR screening, adherence to referrals showed significantly lower compliance (14).
Medical Professional Satisfaction
Further, implementing AI algorithms in medical screening requires more than patient satisfaction: the healthcare professional’s thoughts are equally important in such decisions. Referring again to the Australian study by Scheetz et al. (12), healthcare professionals generally found the system user-friendly and advantageous. Particularly, the real time reports and diagnoses were admired by the professionals. This study provided its own metric for quantifying staff satisfaction, of which was abundant.
II. Implementation Science
Organisational and health system experiences using artificial intelligence
When considering the implementation of AI in the screening for DR, it is important to consider the impact it has on countries with wealth disparities as well as the organisational structures within the system.
A study that was carried out in the capital of Northern Brazil by Gustavo et al. (15), replaced manual checks for DR with hand-held cameras for screening to research the implementation of AI in Brazilian systems. This showed a poor adherence rate with only a ¼ of the target population screened (3561/15,000), eventually leading to the halt of this scheme placed in all 45 clinics (15). The reason for this was said to be the poor organisational structure of the primary care systems, since there was a restraint in actively reaching out to patients. This led to a lack of clear communication in informing patients of the scheme and the benefits, leading to a low participation rate and visitation to the clinic (15). This is a major disadvantage for the healthcare system overall since it may increase the cost of future spending due to the lack of early detection of retinal diseases and may negatively affect the patient who may experience vision loss (16). Furthermore, the poor organisation by the primary care clinics meant that there was uneven patient attendance ranging from 0 to 40 patients per day making the screening process inefficient due to the requirement of having a specialist present to handle the camera in the event of an attendance (15). This leads to several challenges such as being understaffed or overstaffed, meaning it is operationally ineffective and may lead to longer waiting times for patients. Additionally, specialists may become unmotivated since their expertise isn’t being fully utilised while also expending their time unproductively, which may lead to a reduced quality of care – for example paying less attention to patient needs when screening.
Upon reflection of this study, it was noted that other countries had carried out similar schemes with a higher success rate than Brazil (15). It was thought to be due to differences in health education since countries such as the UK, Denmark and Iceland view health education as a key component of their healthcare systems, with greater investment in educating the public about the consequences of DR. As a result, people in these countries are more likely to attend regular retinal screenings. In contrast, Brazil’s health education system is largely decentralised (17), which makes delivering consistent health education more challenging. Consequently, people may be less aware of the harmful effects of DR and are less likely to visit clinics for regular screenings. UK’s and Denmark’s success was mainly due to the large national database of screening which is beneficial due to a higher feasibility for scaling nation-wide (15). This is because a central database means AI screening can be deployed nation-wide regardless of the size or resources available for individual clinics, ensuring equal access to everyone for screening. In contrast, Brazil relies on referrals and volunteer programmes for screenings (15) hence making standardisation and equal access opportunities challenging due to the lack of similar resources.
Another study that was carried out in the United States (18) aimed to identify the common workflow themes leading to a successful adoption of autonomous AI. This study showed that centres with Executive and Clinical champions, who are essentially administrators and leaders, provided a smoother workflow by bridging clinical and administrative teams, promoting staff engagement and training staff. When follow-up appointments for patients with positive signs of disease were scheduled before leaving the centre they were more successful signs of adoption with higher patient satisfaction.
Disadvantages that come with using AI systems
The use of AI in DR screening presented several disadvantages compared to manual checks. A study carried out in Thailand by Peranut et al. explored how well an AI system could support DR screening in community hospitals, displaying that the screening process took longer compared to manual methods (6:30 minutes vs. 5:38) (19), thereby delaying result delivery.
Although the AI system showed strong concordance (agreement) with specialists, only 64.18% of the cases it identified as positive were confirmed (19), highlighting a significant false positive rate that may result in excessive and potentially unnecessary referrals and increased patient anxiety. Furthermore, the AI flagged more images as ungradable often due to cataracts (an eye condition where the lens of the eye becomes cloudy, leading to blurred vision) which could hinder accurate diagnosis and place additional demands on ophthalmologists for follow-up evaluations.
Furthermore, doctors often induce mydriasis (increased pupil dilation) to increase screening accuracy by reducing ungradable images. AI systems lack flexibility to work with blurred imaging due to its learning with perfect images, unlike manual graders who can alter viewing of the retina through different light angles etc. without inducing pupil dilation. Mydriasis can cause temporary blurred vision, light sensitivity and difficulty focusing (20), which may interfere with daily activities like reading or driving for the patient who has received this. Therefore, the use of AI may increase the need for mydriasis in patients undergoing DR screening which isn’t always convenient for the patient.
In conclusion, when implementing AI systems in healthcare across countries, it is important to consider the resources available in the country and the type of organisation within the centres. It is also important to address the disadvantages that come with using these systems in comparison to manual grading for DR, so that in the future they can be improved.
III. Qualitative Studies
Challenges to implementing artificial intelligence
Implementing AI in real-world settings, especially in healthcare settings like DR screening, often comes with several challenges (21, 22). Such challenges go beyond technical limitations and involve human, ethical and cultural issues. For example, low patient participation was reported in Brazil, ethical concerns around false positive results appeared in Thailand and cultural attitudes affected acceptance in both settings (21, 22). Here are common challenges that show the success of AI depends not just on its technical power, but on how it fits into human context and how people feel about it, including how it affects their work. To address these challenges requires involving end users early in the design process and fostering open communication about the role of AI.
Lack of trust and acceptance – a culture issue
Some people question the accuracy and fairness of AI decisions. For example, in healthcare settings doctors may hesitate to use AI diagnostic tools because they don’t understand how the AI reaches its conclusions. This is often referred to as interpretability or explainability. For instance, in a Thai study by Ruamviboonsuk et al. (21), doctors expressed concerns over transparency in how an AI algorithm arrived at its diagnosis to justify their decision when acting upon the AI’s advice.
Opportunities of using artificial intelligence in DR screening
AI offers many benefits in healthcare, particularly in DR screening, and enables innovative opportunities for advancement by enabling smarter systems which can adapt and improve over time. This potential has been demonstrated in recent studies where AI systems improved accuracy after being trained on large datasets. AI allows healthcare providers and clinics to manage and analyse retinal images more efficiently to handle vast amounts of data quickly, uncovering trends and insights that would be difficult or impossible for humans to detect alone. AI can also support communication between doctors and patients by automatically generating visual explanations of retinal scans, allowing for in-person counselling on DR and its risk of blindness. Furthermore, AI’s ability to work continuously without fatigue makes it ideal for environments that require constant monitoring with high volumes of work, such as diabetic eye clinics and telemedicine screening programmes. As AI becomes more integrated into everyday operations, it offers the potential to transform diabetic eye care by increasing access, reducing screening time and creating new roles focused on managing and improving intelligent systems.
AI has shown great promise in enhancing diabetic retinopathy screening through improved accuracy, speed and accessibility. However, its success depends on thoughtful implementation, clinical training and, most importantly, patient and provider trust. Future research and continued feedback from healthcare professionals will be essential to maximise AI’s potential in eye care, including addressing the challenges of explainability for a given output after DR screening.
IV. Health Economics & Policy
When evaluating the use of AI in DR screening, it is vital to acknowledge the economic standpoint. Specifically, whether these algorithms can be implemented in the long run. It is important to understand that governments will always seek the most cost-effective method when achieving tasks. This may sacrifice the optimal accuracy of the AI algorithms; therefore, cost-effectiveness serves as a balance between the two, allowing for both resources saved and medical safety remaining.
Cost-effectiveness of artificial intelligence
Cost effectiveness refers to the scenario in which the desired outcome or goal is achieved, while allocating the least amount of money/resources possible. Many base-case scenarios surrounding AI-based DR screening tend to be constructed by the status quo, which, while they contain the most accurate results, are not sustainable long-term as ultimately the costs of screening with AI outweigh the benefits (more accurate results leading to less cases of people with diabetes developing blindness). Therefore, cost-effectiveness must be utilised to enhance patient outcomes, while maintaining the balance of cost and benefits. A study on health economics in China by Wang et al. (1) found that while the most accurate AI screening model had a specificity of 87.7% and a sensitivity of 93.3%, the best cost-effective model had a lower aggregate percentage by 4.3%, with a sensitivity of 96.3% and a specificity of 80.4%. However, it is important to note that while these statistics are relevant to China, they are not necessarily applicable to other countries, as cost-effectiveness varies from area to area. This is due to matters such as GDP and willingness-to-pay levels (the financial threshold in which a government is willing to pay for a given healthcare outcome, e.g. costs/QALYs gained), which heavily impact the number of resources that can be allocated towards funding AI-based DR screening.
Is the most accurate algorithm the most cost-effective?
In the study conducted by Wang et al. (23) previously mentioned above, the status quo and cost-effective model appeared to show discrepancies in their costs and benefits. The status quo would cost approximately US$1563 million in a 30-year time span, and provide a diabetic patient with 9.1689 quality-adjusted-life-years (QALYs) (23). The best cost-effective model would cost US$14.8 million and would gain 839 QALYs for the entire population in this time span (23). While the best cost-effective algorithm is important to enhance the longevity of AI-based DR screening programmes in clinical settings, the decrease in specificity and sensitivity come with effects that make the more accurate algorithms appear more appealing. Expanding on this, a decrease in sensitivity in AI-based DR screening leads to a higher percentage of misdiagnosis and a 4.5-fold higher risk of patients progressing to blindness (23). A decrease in specificity would lead to unnecessary referrals, leading to excess costs of medical bills, medication and transportation (this would severely affect those in rural areas as they would be forced to spend more than those in urban areas for transportation) (23). However, despite these potential issues, it is in the best interest of governments to continue using the best cost-effective method to prolong the usage of this screening in the medical field. In the ideal scenario, all countries would use the most effective method of AI-based DR screening. However, the constant changing of the economy does not allow for the sustainability of these programmes, therefore using the cost-effective method is the only way to ensure that these programmes fulfil their function, both in the present and the future.
Cost-saving vs Cost-effectiveness and how they are utilised in different subgroups within a population
Cost-saving methods are those designed to reduce the number of resources and money spent on AI-based DR screening by as much as possible, while cost-effectiveness takes into consideration both the costs and the desired outcomes required. The difference between cost-saving and cost-effective methods is vital knowledge when assessing which AI-based DR screening models would be most effective in different areas. This is due to the difference in those who make up the population of an area as well as the location. A study that highlights how cost-saving and cost-effectiveness can be utilised differently in the same area was conducted by Hu et al. (24) in Australia, where a Markov model simulated DR progression over 40 years in both 65,160 Indigenous and 1,197,818 non-Indigenous Australians over the age of 20. The results were presented in the incremental cost-effectiveness ratio (ICER – evaluates cost-effectiveness of one intervention compared to another). Using the status quo, it was predicted that 96,269 non-Indigenous Australians and 3,396 Indigenous Australians would develop blindness during this time, which would cost the Australian government around AU$13,634.6 million (24). This study used 3 different AI methods and found that when the correct AI–based DR screening strategy was applied, it would prevent 38,347 blindness cases, gain 172,090 QALYs and save AU$595.8 million in the non-Indigenous population, and prevent 1,211 blindness cases, gain 9,800 QALYs and save AU$19.2 million in the Indigenous population (24). Every method would become cost-saving for the Australian government, and when the correct strategy was chosen, the AI DR screening could be implemented cost-effectively in both rural and urban Australia. This research was particularly important for members of the Indigenous population in Australia, or those who live in rural or remote areas, as they have less accessibility to healthcare than those living in urban areas.
In summary, cost-effectiveness plays a key role when designing and applying the use of AI models into clinical DR screening. When applied effectively, it can sustainably improve the lives of many people who have diabetes and may progress to DR.
Conclusion
AI is an accurate, efficacious screening tool for DR. However, there are challenges to its adoption, scale-up and spread within screening programmes around the world. Indeed, there is a growing body of qualitative and real-world evidence which identify challenges to adoption such as patient and healthcare worker trust, including explainability of algorithmic outputs, adjustment of clinical workflows, raising patient awareness of DR screening and ensuring follow-up is completed by patients. Other considerations for implementation include retinal camera quality and image gradability, both of which can derail a successful screening programme. Finally, AI is considered a cost-effective screening tool for DR in most countries, especially high-income countries. However, these health economic modelling studies should be updated and iterated over time as more real-world evidence is generated.
Bibliography
(1) Burton MJ, Ramke J, Marques AP, Bourne RRA, Congdon N, Jones I, et al. (2021) The Lancet Global Health Commission on Global Eye Health: Vision Beyond 2020. The Lancet Global Health. (9:4) pp. 489–551.
(2) Zhou B, Rayner AW, Gregg EW, Sheffer KE, Carrillo-Larco RM, Bennett JE, et al. (2024) Worldwide trends in diabetes prevalence and treatment from 1990 to 2022: a pooled analysis of 1108 population-representative studies with 141 million participants. The Lancet [Internet]. (0:0) Available at: https://www.thelancet.com/journals/lancet/article/PIIS0140-6736(24)02317-1/fulltext.
(3) Yau JWY, Rogers SL, Kawasaki R, Lamoureux EL, Kowalski JW, Bek T, et al. (2012) Global Prevalence and Major Risk Factors of Diabetic Retinopathy. Diabetes Care. (35:3) pp. 556–564.
(4) Bourne RRA, Jonas JB, Bron AM, Cicinelli MV, Das A, Flaxman SR, et al. (2018) Prevalence and causes of vision loss in high-income countries and in Eastern and Central Europe in 2015: magnitude, temporal trends and projections. Br J Ophthalmol. (102:5) pp.575–585.
(5) Deloitte Access Economics. (2015) The economic impact of diabetic macular oedema in Australia. Deloitte Access Economics [Internet]. Available at: https://www.diabetesaustralia.com.au/wp-content/uploads/The-economic-impact-diabetic-macular-oedema.pdf
(6) Optometry Australia. (2018) Examination and management of patients with diabetes. Optometry Australia [Internet]. Available at: https://www.optometry.org.au/wp-content/uploads/Professional_support/Guidelines/clinical_guideline_diabetes_revised_sept_2018_final_designed.pdf
(7) Goh JKH, Cheung CY, Sim SS, Tan PC, Tan GSW, Wong T.Y. (2016) Retinal Imaging Techniques for Diabetic Retinopathy Screening. J Diabetes Sci Technol. (10:2) pp. 282–294.
(8) Rahmati M, Smith L, Boyer L, Fond G, Yon DK, Lee H, et al. (2024) Factors Affecting Global Adherence for the Uptake of Diabetic Retinopathy Screening: A Systematic Review and Meta-Analysis. American Journal of Ophthalmology. (268) pp. 94–107.
(9) Foreman J, Keel S, Xie J, Van Wijngaarden P, Taylor HR, Dirani M. (2017) Adherence to diabetic eye examination guidelines in Australia: the National Eye Health Survey. Med J Aust. (206:9) pp. 402–406.
(10) Bresnick G, Cuadros JA, Khan M, Fleischmann S, Wolff G, Limon A, et al. (2020) Adherence to ophthalmology referral, treatment and follow-up after diabetic retinopathy screening in the primary care setting. BMJ Open Diabetes Res Care. (8:1) e001154.
(11) Hu W, Zhu Z, He M, Zhang L. (2024) Cost-effectiveness analysis of artificial intelligence-based diabetic retinopathy screening in Australian primary care settings. Invest Ophthalmol Vis Sci. (65:7) p. 335
(12) Scheetz J, Koca D, McGuinness M, et al. (2021) Real-world artificial intelligence-based opportunistic screening for diabetic retinopathy in endocrinology and indigenous healthcare settings in Australia. Sci Rep 11. e15808. Available at: https://doi.org/10.1038/s41598-021-94178-5.
(13) Ruamviboonsuk P, Tiwari R, Sayres R, Nganthavee V, et al. (2022) Real-time diabetic retinopathy screening by deep learning in a multisite national screening programme: a prospective interventional cohort study. The Lancet Digital Health. Available at: www.thelancet.com/journals/landig/article/PIIS2589-7500(22)00017-6/fulltext.
(14) Melerbi FK, Melo GB. (2022) Feasibility of screening for diabetic retinopathy using artificial intelligence, Brazil. Bull World Health Organ. Available at: http://dx.doi.org/10.2471/BLT.22.288580.
(15) Melo GB, Silva LH, Santos TR, Costa MA. (2024) A comprehensive strategy of diabetic retinopathy screening: Implementation and early challenges in Aracaju, Brazil. Diabetes Metab Syndr. (18:3) e100192. Available from: https://www.sciencedirect.com/science/article/pii/S2666970623000719.
(16) Malerbi FK, Melo GB. (2022) Feasibility of screening for diabetic retinopathy using artificial intelligence, Brazil. Bull World Health Organ. (100:10) pp. 643–647. Available from: https://pmc.ncbi.nlm.nih.gov/articles/PMC9971534/.
(17) Collins C, Araujo J, Barbosa J. (2000) Decentralising the health sector: issues in Brazil. Health Policy. (52:2) pp. 113–127. Available from: https://www.sciencedirect.com/science/article/abs/pii/S0168851000000695.
(18) Goldstein J, Weitzman D, Lemerond M, Jones A. (2023) Determinants for scalable adoption of autonomous AI in the detection of diabetic eye disease in diverse practice types: key best practices learned through collection of real-world data. Front Digit Health. (5) e1004130. Available from: https://www.frontiersin.org/journals/digital-health/articles/10.3389/fdgth.2023.1004130/full.
(19) Chotcomwongse P, Ruamviboonsuk P, Karavapitayakul C, Thongthong K, Amornpetchsathaporn A, Chainakul M, et al. (2025) Transforming non‑digital, clinical workflows to detect and track vision‑threatening diabetic retinopathy via a digital platform integrating artificial intelligence: implementation research. Ophthalmol Ther. (14:2) pp. 447–460. Available from: https://link.springer.com/article/10.1007/s40123-024-01086-8.
(20) Azhar A, Riaz R, Khan A, Kumari V, Habib A, Akilimali A. (2024) Ryzumvi: pioneering advances in countering drug‑induced mydriasis. Ann Med Surg (Lond). (86:6) pp. 3530–3534. Available from: https://pmc.ncbi.nlm.nih.gov/articles/PMC11152829/.
(21) Ruamviboonsuk P, Krause J, Chotcomwongse P, Sayres R, Raman R, Widner K … & Ratanapakorn, T. (2019) Deep learning versus human graders for classifying diabetic retinopathy severity in a nationwide screening program. NPJ Digital Medicine. (2:1) p. 25. Available at: https://doi.org/10.1038/s41746-019-0099-.
(22) Malerbi FK & Melo GB. (2022) Feasibility of screening for diabetic retinopathy using artificial intelligence in Brazil. International Journal of Retina and Vitreous. (8:1) pp. 1–9. Available at: https://doi.org/10.1186/s40942-022-00365-5.
(23) Wang Y, Liu C, Hu W, Luo L, Shi D, Zhang J, et al. (2024) Economic evaluation for medical artificial intelligence: accuracy vs. cost-effectiveness in a diabetic retinopathy screening case. Nature. Available at: https://www.nature.com/articles/s41746-024-01032-9.
(24) Hu W, Joseph S, Li R, Woods E, Sun J, Shen M et al. () Population impact and cost-effectiveness of artificial intelligence-based diabetic retinopathy screening in people living with diabetes in Australia: a cost effectiveness analysis. The Lancet. (67) Available at: https://www.thelancet.com/journals/eclinm/article/PIIS2589-5370(23)00564-3/fulltext.