Corresponding Author: Kevin Anggakusuma Hendrawan, MD. Department of Ophthalmology, Faculty of Medicine, Widya Mandala Surabaya Catholic University, Jl. Raya Kalisari Selatan No. 1 Pakuwon City, Surabaya, 60112, Indonesia. Diabetic retinopathy (DR) is the most common microvascular complication of diabetes mellitus, and it is the foremost cause of vision loss among adults and older adults [1]. The worldwide incidence of DR is projected to increase significantly, escalating from approximately 103 million individuals in 2020 to an estimated 130 million by 2030, and reaching around 161 million by 2045 [2,3]. Vision-threatening DR is also projected to increase by 26.3% to 36 million in 2030 and by 57.0% to 44.82 million by 2045 [3]. Moreover, the epidemiological forecast for 2030 indicates that DR prevalence is set to increase significantly in middle- and low-income regions. The anticipated rise varies, with projections ranging from 20.6% to as high as 47.2% across regions such as the Western Pacific, South and Central America, Asia, Africa, the Middle East, and North Africa [2,3]. Despite the increasing burden of DR, many low- and middle-income countries (LMICs) lack structured national screening programs. Although some higher middle- and high-income countries have fully developed and structured national screening programs [4], only 8.5% of upper middle-income countries have national DR screening policies, and most low-income countries lack specific policy data. This gap highlights the need for cost-effective strategies to address DR in LMICs [5].
The gold standard for DR diagnosis is a dilated fundus examination performed by an ophthalmologist using either an indirect ophthalmoscope or a slit-lamp biomicroscope. However, numerous barriers exist to DR screening, including limited access to healthcare, lack of time, high out-of-pocket costs, insufficient knowledge and awareness, and poor coordination of care [6]. These challenges are even more pronounced in developing countries. The introduction of artificial intelligence (AI), deep learning (DL), and telemedicine with retinal imaging is a promising and efficacious technology designed to identify patients with DR [6-8].
Screening and prompt treatment before the onset of symptoms can significantly decrease the risk of severe visual impairment and the economic impact associated with vision-threatening DR (VTDR) and diabetic macular edema [9]. Research conducted in high-income countries suggests that implementing AI or teleophthalmology in DR screening programs can significantly reduce screening costs compared to human grading [10-12]. This price gap may be primarily due to the substitution of technology costs for labor costs. However, due to low labor costs in developing countries, conclusions drawn from high-income countries may not be equally applicable. Therefore, this study aims to summarize data on the economic evaluation of DR screening modalities in developing countries, including the use of AI, telemedicine, or other community-based screening methods when compared to standard care.
Materials and Methods
Data sources and search strategy
This systematic review was conducted following the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines [13,14]. This study was registered in PROSPERO (No. CRD420251015270). Five reviewers conducted independent searches for relevant studies published in PubMed, ScienceDirect, the Cochrane database, and Google Scholar. We conducted a search using the following keywords: (diabetic retinopathy) AND (screening) AND (community based OR telemedicine OR teleophthalmology OR artificial intelligence) AND (cost OR incremental cost OR cost-effectiveness OR cost-utility OR economic evaluation) AND (developing countries OR low-income countries OR middle-income countries). The terms from each category were independently cross-referenced with those from other categories.
Selection criteria and selection
Studies were included if they met the following criteria: recruited subjects with either type 1 or type 2 diabetes mellitus, utilized a low- or middle-income population/setting, involved patients undergoing primary screening for DR, compared screening modalities (such as telemedicine, AI, or community-based programs) with standard care, and conducted a health economic analysis or provided adequate data for assessing the cost-effectiveness of the employed technology. The classification of “developing countries” used here follows the World Bank data based on the time the studies were conducted.
This study excluded articles that did not employ economic model analyses and only referenced or briefly addressed the potential economic costs of AI-based DR screening. Studies lacking adequate data, those reporting comorbid eye diseases, as well as case reports, guidelines, editorials, commentaries, opinions, and reviews, were also excluded from the analysis. Five reviewers independently evaluated the titles and abstracts of the selected articles. The researchers conducted a thorough review of the full texts of potentially relevant studies to select the final studies for inclusion. Disagreements were addressed through discussions.
Quality assessment
The quality assessment of all included studies was conducted independently by six investigators utilizing the JBI Critical Appraisal Checklist for Economic Evaluations [15]. It was used to assess the risk of bias of studies evaluating the cost-utility and/or the cost-effectiveness of screening methods. This checklist consist of 11 considerations: (1) Is there a well-defined question? (2) Is there a comprehensive description of alternatives? (3) Are all important and relevant costs and outcomes for each alternative identified? (4) Has clinical effectiveness been established? (5) Are costs and outcomes measured accurately? (6) Are costs and outcomes valued credibly? (7) Are costs and outcomes adjusted for differential timing? (8) Is there an incremental analysis of costs and consequences? (9) Were sensitivity analyses conducted to investigate uncertainty in estimates of cost or consequences? (10) Do the study results include all issues of concern to users? (11) Are the results generalizable to the setting of interest in the review?
Data extraction and analysis
The researchers extracted and synthesized all relevant data, including study characteristics, intervention details, cost measures, and effectiveness outcomes, in a narrative format. We analyzed the studies and presented the findings based on four cost-effectiveness metrics. First, “cost” refers to the expense incurred when adopting a new intervention in comparison to the regular or existing approach. Second, “quality-adjusted life years (QALYs)” is a metric used to quantify the supplementary advantages of quality and quantity of life. Third, the “incremental cost-effectiveness ratio (ICER)” is a metric utilized in health economics to evaluate the relative costs and outcomes of two or more interventions to achieve a unit of effect. Fourth, the “incremental cost-utility ratio (ICUR)” is a particular kind of ICER in which the unit of effect is utility (generally in QALYs) [16]. Additionally, the financial data in the articles was converted to US dollar based on the year of publication to help with comparisons.
Results
Study selection and characteristics
Fig. 1 illustrates the flowchart of the literature search and study selection procedure. The process began by retrieving relevant studies systematically from appropriate databases, yielding a total of 1,451 studies. Consequently, we excluded duplicate studies, meta-analyses, reviews, conference proceedings, studies with inaccessible full texts, and studies with titles and abstracts that were not relevant to our review. After the initial screening, 17 studies were obtained. We then conducted a review of the full text of these 17 studies and excluded papers with unclear methods or incomplete and/or irrelevant targeted outcomes. Ultimately, a total of seven studies were included in the systematic review.
The seven selected studies in this review reported the economic evaluation of DR screening modalities by assessing the cost-effectiveness and cost-utility of AI-based screening and telemedicine in comparison to the standard screening method for DR, which included the option of no screening. Four of the studies were conducted in China, and one each in Brazil, India, and Thailand. Table 1 summarizes the characteristics of these studies [9,17-22].
Quality assessment
Table 2 outlines the procedure used to evaluate the reviewed studies, which applied the JBI Checklist for Economic Evaluation [9,17-22]. This study reviewed seven studies that investigated the economic evaluation of DR screening methods. The outcomes in all studies had negligible reporting bias. All studies included in our systematic review and meta-analysis are of high quality.
Economic evaluation of screening methods modalities
Table 3 summarizes the economic evaluation of each studies for the diabetic retinopathy screening [9,17-22]. Lin et al. [17] assessed the cost-effectiveness and cost-utility of AI-based telemedicine screening compared to human grading for DR diagnoses in urban China. They employed a decision-analytic Markov model with real-world data from the Shanghai Digital Eye Disease Screening Program. The study estimated costs, effectiveness, and QALYs over 30 years. The results showed that the total expected cost per participant for AI-based screening was $3,182.50, compared to $3,265.40 for human grading. The expected years without blindness were 9.80 years for AI-based screening and 9.83 years for human grading, while QALYs were 6.748 for AI-based screening and 6.753 for human grading. Meanwhile, the ICER for AI screening was $2,553.39 per year without blindness, and the ICUR was $15,216.96 per QALY, indicating that AI-based screening was not cost-effective under Shanghai’s cost-effectiveness threshold (set at $22,600 per QALY, equivalent to the local gross domestic product [GDP] per capita). Moreover, the sensitivity analyses suggested that AI-based screening could become cost-effective if referral compliance increased by 7.5%, or if the cost of human grading increased by 50%, or if AI screening costs decreased by 50%. These results indicate that AI-based screening was not cost-effective under Shanghai’s cost-effectiveness threshold (set at $22,600 per QALY, equivalent to the local GDP per capita).
Li et al. [18] evaluated the cost-effectiveness of AI-based DR screening compared to ophthalmologist screenings and no screenings in rural China using a Markov decision model. The study simulated the long-term development of DR in a hypothetical cohort of 10,000 diabetic patients over 50 years and estimated the ICER for each screening strategy. From a health system perspective, AI-based screening was found to be cost-effective compared to no screening, with an incremental cost of $5,182.25 and an incremental utility of 0.33 QALYs, resulting in an ICER of $15,598.72 per QALY, which was below the cost-effectiveness threshold (three times the per capita GDP, $30,230.36 per QALY). AI-based screening was also found to be more cost-effective than ophthalmologist-based screening, as the latter incurred an additional $2,070.63 and was less effective (-0.31 QALYs). Their sensitivity analyses also confirmed the robustness of these findings, with AI-based screening remaining the most cost-effective strategy under various parameter modifications. Probabilistic sensitivity analysis using Monte Carlo simulations revealed that AI screening had a higher probability of cost-effectiveness compared to ophthalmologist screening, particularly as the willingness-to-pay (WTP) thresholds increased.
Next, Srisubat et al. [19] evaluated the cost-utility of DL compared to trained human graders for DR screening in a nationwide program in Thailand. The study assessed lifetime costs and health outcomes from both societal and healthcare provider perspectives by using a decision tree-Markov model. The results indicated that DL-based screening leads to a slight reduction in costs (approximately $2.70 per screened patient) while achieving similar QALYs, resulting in an incremental net monetary benefit of around $24.10. Sensitivity analysis confirmed the cost-effectiveness of DL even if its cost increased to $4.00 per patient. Moreover, incorporating improved compliance rates for treatment referrals with DL suggested potential effectiveness benefits ranging from $20 to $50. Although DL screening detected more referable cases, leading to higher initial treatment costs, it ultimately resulted in fewer cases of bilateral blindness, which translates to greater long-term societal cost savings. The study concluded that DL-based DR screening is an economically viable alternative in middle-income countries, especially where there is a shortage of trained human graders and low compliance with referrals for treatment.
Huang et al. [20] assessed the cost-effectiveness of AI-based DR screening compared to ophthalmologist-based screening and no screening in rural China using a Markov decision model. The analysis, conducted over a 35-year simulation period, evaluated costs and effectiveness from both the health system and societal perspectives. From a health system viewpoint, the AI approach provided an additional 0.16 QALYs per person at a cost increase of $180.19, resulting in an ICER of $1,107.63 per QALY. This figure is well below China’s 2019 cost-effectiveness threshold, typically defined as one to three times the GDP per capita ($10,255.03-$30,765.09). AI screening also outperformed ophthalmologist screening, being both more effective and less costly. In contrast, the latter added $34.86 in costs and resulted in 0.04 fewer QALYs, making it a less favorable option. When broader societal costs, including factors such as productivity loss and travel expenses, were considered, the ICER for AI screening increased to $10,347.12 per QALY, remaining well within the acceptable range. The sensitivity analyses also confirmed that the model’s outcomes were stable across a range of input assumptions, with the most influential variables being the utility assigned to the no-DR health state and the costs related to ophthalmologist follow-ups. Additionally, the probabilistic analysis showed that AI screening remained the most cost-effective strategy in every simulation scenario tested.
Zhang et al. [9] conducted a cost-utility analysis to compare telemedicine-based and community-based DR screening against no screening in both rural and urban settings in China. They used a Markov model to estimate costs and QALYs over a long-term period from a societal perspective. The findings demonstrated that both telemedicine and community screening methods were more cost-effective than no screening at all. In rural areas, community screening yielded an ICUR of $4,179 per QALY gained, while the telemedicine screening had a lower ICUR of $2,323 per QALY, making it the more cost-effective option. Similarly, in urban settings, community screening had an ICUR of $3,812 per QALY, whereas telemedicine screening had an ICUR of $2,437 per QALY, further confirming the economic advantage of telemedicine. Ultimately, telemedicine proved to be the more cost-effective strategy compared to community screening, with an ICUR of $1,212 per QALY in rural areas and $1,141 per QALY in urban areas. The sensitivity analyses also confirmed the robustness of these findings, with ICURs consistently remaining below the cost-effectiveness threshold of three times per capita GDP. Additionally, the probabilistic sensitivity analysis showed that telemedicine screening was more likely to be cost-effective in both settings, particularly as WTP thresholds increased.
Furthermore, Ben et al. [21] conducted a cost-utility analysis comparing opportunistic ophthalmology referral-based screening, systematic ophthalmology referral-based screening, and systematic teleophthalmology-based screening for DR in the context of the Brazilian public healthcare system. The findings showed that systematic teleophthalmology-based screening was the most cost-effective strategy compared to opportunistic screening, with an ICUR of $9,792 per QALYs gained. Systematic ophthalmology referral-based screening was more expensive and less effective than systematic teleophthalmology-based screening, confirming the superior economic value of teleophthalmology. The sensitivity analyses also indicated that the most influential parameters affecting cost-effectiveness were utility values for blindness, discount rate, and treatment costs for VTDR. Meanwhile, the probabilistic sensitivity analysis showed that at the minimum WTP threshold, the probability of systematic teleophthalmology screening being cost-effective was 0.46, and 0.67 at the maximum WTP threshold.
Finally, Rachapelle et al. [22] conducted a cost-utility analysis to evaluate the cost-effectiveness of telemedicine-based DR screening in rural southern India, compared to no screening, and to determine the optimal screening interval. They developed a Markov model using real-world data to estimate the cost per QALY gained from a one-time teleophthalmology screening camp and to assess the cost-effectiveness of different screening intervals. The model incorporated data from patient interviews, provider cost estimates, and diagnostic test accuracy assessments. Cost inputs included health provider costs (personnel, equipment, and screening logistics) and household costs (travel, lost income, and hospital fees). Effectiveness was measured through utility values derived from time trade-off interviews with diabetic patients. The results demonstrated that a one-time telemedicine-based DR screening was cost-effective compared to no screening, with an ICUR of $1,320 per QALY, which falls within the World Health Organization’s (WHO) cost-effectiveness threshold of one to three times India’s GDP per capita ($1,061-$3,183 per QALY). Further analysis revealed that increasing the screening frequency to every 2 years remained cost-effective ($2,435 per QALY), whereas annual screenings ($4,029 per QALY) would exceed the cost-effectiveness threshold. From a societal perspective, which includes indirect patient costs, screening once every 5 years remained cost-effective, while more frequent intervals became less economically viable. The sensitivity analyses also identified key cost drivers, including utility values for blindness, DR progression rates, and screening costs. They confirmed that telemedicine remained the most cost-effective strategy under varying conditions. The probabilistic sensitivity analysis indicated a high probability of telemedicine screening being cost-effective at a WTP threshold of three times GDP per capita.
Discussion
The advancement of digital technologies has opened new pathways for ophthalmic care by offering alternative DR screening models such as telemedicine and AI-based analysis [8,23-26]. Telemedicine facilitates remote assessment via image transmission [27], while AI offers automated retinal grading [28]. These technologies can enhance screening coverage, prioritize referrals, and shift care from opportunistic to systematic delivery [29,30]. Nevertheless, although these modalities provide promising alternatives for screening in remote areas, a health economic assessment of these methods is crucial to ensure cost-effectiveness and guide future policy decisions.
The designs, training, and validation of AI models substantially varied across the reviewed studies. These variations ranged from algorithm type, data sources, validation approaches, and referral thresholds. Some studies used AI models validated within their own screening programs, while others relied on external data or previously published performance metrics. This limits the comparability of cost-effectiveness estimates across the studies and the strength of the pooled conclusions. Because cost-effectiveness outcomes are highly sensitive to diagnostic performance and referral criteria, this heterogeneity weakens the strength of pooled conclusions and illustrates the importance of context-specific evaluation when interpreting or applying results across different settings. Supplementary Table 1 presents a structured summary of AI model features [9,17-22].
Most studies in this review found that AI-based or telemedicine screening strategies were cost-effective compared to standard care or no screening at all. In rural China, AI screening was more effective and less costly than ophthalmologist-based screening, with ICERs below China’s WTP threshold [18,20]. Zhang et al. [9] also demonstrated that telemedicine was more cost-effective than both community-based and no screening in both rural and urban areas. Similarly, systematic teleophthalmology-based screening in Brazil was superior to opportunistic referral-based screening [21]. Meanwhile, from a health provider’s perspective, telemedicine screenings every 2 years have been proven to be cost-effective in rural India [22].
In Thailand, DL-based screening was slightly more cost-effective than trained human graders [19]. In contrast, Lin et al. [17] found that AI-based screening in Shanghai was not cost-effective compared to human graders in a telemedicine model due to minimal cost reduction and lower QALYs gained and years without blindness. They suggested that human graders were more cost-effective due to low labor costs. However, increasing the referral compliance of patients with sight-threatening DR would result in the AI-assisted model becoming more cost-effective.
Furthermore, although AI-based and telemedicine screening were generally cost-effective, their economic feasibility depends on overall affordability and implementation strategies. For example, Lin et al. [17] reported that AI-based screening reduced costs by only 2.5%, suggesting that labor costs and system efficiency influenced economic feasibility. The study also found that the potential economic benefits of AI decreased when patient adherence to referral and treatment was suboptimal. Similarly, Srisubat et al. [19] noted that AI-based screening saved only $2.70 per patient; however, improving compliance with treatment would achieve greater cost savings.
Moreover, screening intervals also influence economic feasibility. Rachapelle et al. [22] reported that while annual screening is recommended in many countries, it has higher additional costs from both health provider and societal perspectives. Evidence beyond this review, such as that of Emamipour et al. [31], further supports the notion that a personalized screening model, where intervals are adjusted based on individual risk factors, is more cost-effective than fixed-interval strategies. However, this approach resulted in a slightly higher proportion of delayed diagnosis of sight-threatening DR, highlighting the trade-off between cost savings and timely detection.
The effectiveness of AI-based and telemedicine screening is primarily assessed through QALYs gained, as they represent improvements in both lifespan and quality of life. Most studies demonstrated incremental QALY improvements from AI-based or telemedicine screening compared to standard care, suggesting that these technologies provide measurable health benefits [9,18-22]. However, the magnitude of gain varied considerably across settings, depending on factors such as screening sensitivity, disease stage at detection, and adherence to referral. Notably, Lin et al. [17] reported AI-based screening had slightly lower QALYs than human graders, showing no clear advantage. This finding suggests that while AI and telemedicine screening strategies generally improve health outcomes, their impact on QALYs and thus their cost-effectiveness may be modest in contexts where clinical efficacy or follow-up systems are suboptimal.
Differences in economic modeling assumptions, including time horizons, discount rates, WTP thresholds, and perspectives substantially affected cost-effectiveness outcomes across the reviewed studies (Supplementary Table 2) [9,17-22]. Time horizons ranged from 25 to 50 years, with most other studies using 30 to 40 years and only one study used a lifetime horizon. Meanwhile, discount rates were generally 3%, though some used a slightly higher rate of 3.5% or even 5%. Longer time horizons and lower discount rates tend to yield more favorable ICERs by capturing extended health benefits [32,33]. Some used explicit country-specific threshold, while others applied the WHO-recommended benchmark of one to three times the country’s GDP per capita. This common reference enables the results to be interpreted within each country’s economic context. In addition, studies varied in their perspective. Most adopted a healthcare system or provider perspective, but only a few included patient or societal costs such as productivity losses, which may underestimate the true economic value of screening [34]. The studies were also mostly Markov-based, although three studies employed a hybrid decision tree-Markov model approaches to model initial screening decisions and long-term outcomes. Therefore, these contextual and methodological differences limit direct comparability of cost-effectiveness ratios and reinforce the need to interpret findings within each country’s economic and health system context.
Nevertheless, although AI and telemedicine-based screening appear cost-effective in LMIC settings, their real-world implementation faces substantial logistical and system-level barriers. Several studies reported that low patient awareness, out-of-pocket costs, and long travel distances reduced participation in DR screening [18,22]. Infrastructure limitations, such as the unequal distribution of ophthalmologists, prolonged wait times, and the absence of referral systems, were also highlighted as key challenges in Brazil and India [21,22]. Other studies have noted that workforce shortages in rural regions limit the feasibility of human grading and make AI a more practical, yet context-dependent, alternative [17,21]. Moreover, while some regions benefited from strong screening infrastructure and low labor costs [19], such conditions are not widely generalizable [17]. Therefore, these barriers underscore the need to tailor implementation strategies to the local health system’s capacities.
Beyond the scope of the reviewed studies, broader literature highlights additional barriers that affect DR screening uptake in LMICs. These include low diabetes-related health literacy, mistrust of nonphysician healthcare providers, and limited awareness of retinopathy risk [35,36]. Infrastructure limitations, such as poor internet connectivity, insufficient system interoperability, and inadequate equipment maintenance, have also been reported [37,38]. Although not directly modeled in economic evaluations, these factors influence patient participation, referral completion, and long-term program sustainability and should be considered in implementation planning.
The methodological quality of the reviewed studies was assessed using the JBI Checklist for Economic Evaluations (Table 2). Most studies met the core criteria, including clearly defined research questions, appropriate comparators, and credible valuation of costs and outcomes. However, some studies demonstrated methodological limitations. Ben et al. [21] received an “unclear” rating for question 4 (clinical effectiveness), as the cited diagnostic accuracy values were not validated within the local context. Zhang et al. [9] also received “unclear” for question 4, as estimates of clinical effectiveness were incorporated from secondary sources without clear reporting or validation. Rachapelle et al. [22] also received an “unclear” rating for question 4, as the utility values and diagnostic performance assumptions were not explicitly reported or justified. In addition, they were rated “no” for question 2 because the intervention strategies were differentiated only by screening frequency, with minimal detail on how the interventions were delivered. Moreover, Lin et al. [17], Ben et al. [21], and Rachapelle et al. [22] were rated “no” for question 11, as the authors explicitly acknowledged that their findings were not generalizable beyond the specific settings studied. These limitations highlight the importance of transparent reporting and local insights in supporting reliable economic evaluations across diverse healthcare contexts.
Limitations
Overall, this research conducted a comprehensive economic evaluation of AI- and telemedicine-based screening strategies for DR across diverse developing countries. By incorporating multiple modeling approaches and care delivery strategies, the findings provide valuable insights for policymakers considering implementation in resource-constrained settings. However, several limitations should be carefully considered when interpreting the results. The reviewed studies varied in key economic parameters, such as time horizons, discount rates, and WTP thresholds, making direct comparisons of cost-effectiveness challenging. Many studies used hypothetical cohorts and assumed referral compliance, reducing external validity. Utility values and transition probabilities were often adapted from prior literature. Despite these limitations, the findings highlight the potential of AI and telemedicine screening to improve DR detection in LMICs, although further longitudinal studies and standardized cost-effectiveness analyses are necessary to evaluate long-term scalability.
Conclusions
This review finds that AI- and telemedicine-based screening strategies for DR are, in most settings, cost-effective alternatives to conventional screening methods or no screening at all in developing countries. However, their economic value is highly context-dependent, influenced by local labor costs, referral adherence, and the choice of screening interval. Moreover, although these technologies have significant potential to improve access and equity in eye care, their successful implementation requires investment in infrastructure, workforce readiness, and patient engagement. Future studies should prioritize real-world evaluations, standardized economic modeling, and longitudinal data to support scalable, evidence-informed screening policies in resource-constrained settings.
PDF Links
PubReader
ePub Link
XML Download
Full text via DOI
Full text via PMC
Download Citation
Supplement1
Print
