Introduction

With the increasing prevalence of type 2 diabetes1, diabetic retinopathy (DR) is now one of the leading cause of preventable vision loss among working-age adults, especially in low-income and middle-income countries (LMICs)2. Diabetic macular edema (DME) and proliferative diabetic retinopathy (PDR) are the two main causes of vision loss in the population with diabetes. DME, characterized by retinal thickening in the macular area, has now surpassed other DR conditions as the predominant cause of moderate or severe vision loss in patients with diabetes3. It is estimated that DME affects ~2.7% of adults diagnosed with diabetes globally. Prompt detection and intervention of DME could substantially reduce the risk of permanent vision loss and increase the chance of vision gain4,5,6. Thus, DR screening, especially for DME, is widely recognized as important for the prevention of vision loss in the population with diabetes7.

DR screening, mostly utilizing the two-field non-stereoscopic fundus photography (FP) and visual acuity (VA) test, shows good sensitivity and specificity for retinopathy detection7. However, detecting DME reliably with FP remains challenging3, because the appearance of retinal thickening or edema on non-stereoscopic retinal photograph is often subtle. Relying on surrogate markers (e.g. exudates, hemorrhage), FP-based DR screening has been shown to have poor sensitivity and modest specificity for DME, which can lead to unnecessary referrals and missed diagnoses8,9. Although previous studies found incorporating manual spectral-domain optical coherence tomography (SD-OCT) in screening improved detection of DME, widescale implementation is hampered by the high costs of manual OCT and the need for well-trained technicians to acquire and interpret the images8,10. As the prevalence of diabetes mellitus (DM) continues to rise, the burden of regular retinal screening will escalate, further straining medical resources and significantly increasing healthcare costs11. These challenges highlight the urgent need to develop automated screening solutions that can not only perform sensitive and specific DME detection, but also avoid imposing greater workload for healthcare workers.

In recent years, automated self-imaging OCT (SI-OCT) has been developed to enable patients to capture high-resolution OCT images of the macular by themselves12,13,14. The accuracy of measuring retinal central subfield thickness (CST) and diagnosing macular diseases using SI-OCT has been previously validated12,13,14. More importantly, one of the SI-OCT devices (Master OCT, MOPTIM, Shenzhen, China) has recently been approved by the National Medical Products Administration of China (NMPA) as an imaging tool for retinal diseases in clinical practice (Fig. 1)14. With the fully automated process and high-resolution OCT imaging, the SI-OCT has the potential to improve DR screening with better performance on DME detection.

Fig. 1: Illustration of self-imaging optical coherence tomography (SI-OCT) and the typical images of diabetic macular edema (DME).
figure 1

a A participant performing OCT self-imaging, b typical fundus photography centered on the macula in a patient with diabetes demonstrating retinal hemorrhage and exudates, c normal and d DME OCT images collected using the SI-OCT.

Full size image

However, prior to the widescale application, there has been no economic data, which is important for policymakers to determine if the SI-OCT provides good value and should be applied in DR screening. In this study, we aimed to fill these gaps by investigating the performance and cost-effectiveness of incorporating SI-OCT in FP-based screening for DR by building a Markov model.

Results

Demographics

We conducted a cross-sectional and observational study in a rural area of Shaoguan, one of the underdeveloped cities of Guangdong Province in China. A total of 1780 participants with known DM were recruited, with a mean age of 62.6 (8.5) years, and 950 (53.4%) were women. All participants were offered macular self-imaging using the SI-OCT, in addition to standard VA testing and two-field mydriatic FP. Among them, 134 were excluded due to refusal to accept screening, ungradable images, and other reasons (shown in Supplementary Fig. 1). A total of 1646 participants (92.5%) were finally included and analyzed.

Two independent retinal specialists reviewed all screening data (including visual acuity, medical records, intraocular pressure, FP, and OCT images from SI-OCT). They definitively diagnosed each participant as having either non-referable diabetic retinopathy (non-RDR; defined as DR grades R0 or R1) or referable diabetic retinopathy (RDR; defined as DR grades R2 or R3) according to the United Kingdom National Health Service (UK NHS) screening protocol. DME was diagnosed as foveal involvement of abnormal intrarenal and/or subretinal fluid with concurrent thickening affecting the 1 mm diameter CST. Consensus between the two graders was required, and their diagnoses served as the reference standard for this study.

Among the 1646 participants, RDR was present in 80 participants (4.9%), and DME in 65 participants (3.9%). Specifically, 10 participants (0.6%) were found to have RDR and DME, while 70 (4.3%) were shown to have RDR but no DME. There was no diagnostic discrepancy between the two independent retina specialists. Demographic characteristics of the study population were provided in the Supplementary Table 1.

Performance of screening strategy

Three screening strategies for RDR were compared in this study: the conventional FP-based strategy, a combination strategy of FP and SI-OCT (FP + SI-OCT), and a combination strategy of FP and manual SD-OCT (FP + SD-OCT). The screening performance for DME for each strategy are reported in Table 1, using the following metrics: sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). Compared with the traditional FP-based screening strategy, the FP + SI-OCT strategy showed higher sensitivity and specificity for the detection of DME (sensitivity [95%CI]: 87.69% [79.86–95.52%] versus 61.53% [50.04–73.04%], specificity [95% CI]: 98.29% [97.64–98.94%] versus 92.47% [91.09–93.85%]). Eight participants (8 out of 65, 12.31%) with non-center-involving DME had CST lower than our pre-specified DME threshold (305 mm) based on SI-OCT. Twenty-seven out of 1581 participants (1.71%) who had CST larger than 305 mm but were deemed to have no DME after reviewed by the two specialists, mostly due to neovascular age-related macular degeneration (nAMD, 18 out of 27, 66.67%) and segmentation errors (5 out of 27, 18.52%). The positive predictive value (PPV [95% CI], 67.86% [57.69–78.03%] versus 25.16% [18.30–31.91%]) was higher for the FP + SI-OCT strategy than the FP-based screening protocol, while the negative predictive value (NPV [95% CI], 99.49% [99.14–99.83%] versus 98.32% [97.67–98.97%]) remained mostly unchanged, probably due to the low overall prevalence of DME, as expected from screening asymptomatic patients. The performance of FP + SD-OCT strategy for DME screening was derived from the reported best performance (sensitivity 100%, and specificity 98.34%) in the previous studies1.

Table 1 DME screening performance using the different strategies
Full size table

Since the screening criteria for RDR were the same across the three strategies, with all based on biomarkers on FP, the sensitivity, specificity, PPV and NPV for screening RDR were 92.50%, 94.44%, 45.96% and 99.60% for all strategies, respectively.

Cost-effectiveness analysis

We built Markov models to compare the costs and effects of the three different screening strategies, and the primary outcome was incremental cost-effectiveness ratio (ICER). The current screening protocol (FP-based strategy) served as the status quo. The models were based on a hypothetical cohort of 100,000 individuals with DM through a total of 10 1-year Markov cycles. The estimated costs per person, QALYs gained per person, incremental costs and QALYs and ICERs for each strategy are shown in Table 2. The total estimated cost of traditional FP-based strategy (status quo) for a participant with DR was $1273.70, and the total expected QALYs gained were 7.93435 over the 10 years. Notably, although incorporating SI-OCT in the current strategy required extra financial investment in equipment, the FP + SI-OCT strategy resulted in a gain of 1 quality-adjusted life-year (QALY) at a cost of $8016 compared with the traditional FP-based strategy. In contrast, FP + SD-OCT strategy required much more investment in not only equipment but also personnel, thus resulted in a gain of 1 QALY at a cost of $45,753, which is much less cost-effective than the FP + SI-OCT strategy.

Table 2 Base-case cost-effectiveness results from DR screening programs using different strategies
Full size table

Compared with the pre-specified standardized threshold, the FP + SI-OCT strategy, with an ICER less than per capita GDP in Shaoguan ($8408), would be considered as “very cost-effective”. The FP + SD-OCT strategy, with an ICER much larger than three-times per capita GDP ($25,224), would be considered as “not cost-effective”.

Sensitivity analysis

Extensive sensitivity analyses were performed to test the robustness of our model outcomes (shown in Figs. 2 and 3). The tornado diagram shows the 10 parameters that had the greatest impact on the results. In our study, the cost of each strategy, utility for treated or untreated DME, adherence to referral or treatment were common parameters in most scenarios (Fig. 2). The probabilistic sensitivity analysis showed that the base-case ICERs of FP + SI-OCT strategy were robust to randomly distributed parameters (range and assigned distribution for sensitivity analysis are shown in the Supplementary Tables 2 and 3). Under the one-time per capita GDP threshold ($8408) and the current willingness-to-pay (WTP) threshold ($25,224), the FP + SI-OCT strategy was cost-effective compared to the FP-based strategy in 50.27% and 73.62% of the simulations, respectively. In contrast, the FP + SD-OCT strategy was much less likely to be cost-effective (only 18.47% and 1.14% of the simulations under the one-time per capita GDP threshold and the current WTP threshold, respectively). When comparing between the FP + SD-OCT and FP + SI-OCT strategies, the FP + SI-OCT strategy was more cost-effective in 73.62% and 94.68% of the simulations under the one-time per capita GDP threshold and the current WTP threshold, respectively (Fig. 3).

Fig. 2: Deterministic one-way sensitivity analysis.
figure 2

Costs are given in US dollars. The top ten parameters that caused the greatest impact on ICERs are shown. We did 1-way sensitivity analyses for each of the three strategies against current practice (FP-based strategy), respectively (a, b). FP + SI-OCT strategy and FP+manual SD-OCT strategy were further compared (c). The strategy was defined as cost-effective if the ICER was less than three-times per capita gross domestic product (GDP) per QALY gained, which was $25,224 in Shaoguan (per capita GDP $8404). GDP gross domestic product, ICER incremental cost-effectiveness ratio, QALY quality-adjusted life-years, DME diabetic macular edema, RDR referable diabetic retinopathy.

Full size image
Fig. 3: Probabilistic sensitivity analysis of the incremental costs and incremental effectiveness.
figure 3

Costs are given in US dollars. Incremental effectiveness is defined as incremental QALYs. We did probabilistic sensitivity analyses for each of the two strategies against current protocol (FP-based strategy), respectively (a, b). The combined strategy of FP and SI-OCT and the combined strategy of FP and manual SD-OCT were further compared (c). Dashed and solid lines represent one-time and three-times per capita GDP, respectively. The three-times per capita GDP per QALY gained ($25,224) served as the willingness to pay (WTP) per QALY gained, to define the strategy was cost-effective or not. QALY quality-adjusted life-year, GDP gross domestic product, FP fundus photography, SI-OCT self-imaging optical coherence tomography, SD-OCT spectral-domain optical coherence tomography.

Full size image

By taking 10,000 random draws, we got the cost-effectiveness acceptability curves (Supplementary Fig. 2). When all three screening strategies were available, the FP-based strategy was preferred with a lower WTP (below one-time per capita GDP); the FP + SI-OCT strategy became the dominant strategy in 50.53% of simulations at the one-time per capita GDP threshold ($8408). With WTP higher than one-time per capita GDP, the FP + SI-OCT strategy clearly emerged as the dominant strategy. Specifically, at the current WTP threshold ($25,224), the FP + SI-OCT strategy was favored in the majority of simulations (69.36%), whereas the FP + SD-OCT strategy was far less likely to be preferred, within the simulated range of WTP.

Discussion

Our study conducted in rural China showed that incorporating SI-OCT into the current FP-based DR screening practice provided better performance for DME screening. Despite of extra financial investment, the FP + SI-OCT strategy was more cost-effective than the FP-based strategy under the current WTP threshold ($25,224). In contrast, incorporating manual SD-OCT into the current FP-based DR screening practice was not found to be cost-effective.

Fundus photography-based DR screening has been widely implemented and shown to have excellent performance for retinopathy; however, screening for DME continues to pose a substantial challenge3. Previous studies had reported only moderate or even poor performance of FP-based screening for detecting DME, with sensitivity of 40–73% and specificity of 67–79%3,8,9,15. The performance of our FP-based strategy results align with these prior studies. Incorporating OCT did markedly improve DME screening performance compared to FP alone, which could be explained by the 3D visualization of retinal thickness changes in OCT images for detecting DME10. In this study, the FP + SI-OCT strategy showed sensitivity and specificity similar to previous studies using manual SD-OCT to identify DME in clinics, with reported sensitivity of 78–100% and specificity of 79–100%16. We found that 35.38% (23 of 65) participants with DME, who were screened negative by the FP-based strategy, could be correctly referred by SI-OCT, and 7.27% (115 of 1581) participants without DME, who were screened positive by the FP-based strategy, could be kept in surveillance with SI-OCT, avoiding unnecessary referrals and saving costs for more than US$3,000. For the FP + SI-OCT strategy, incorrect referrals occurred in up to 1.71% of cases, with 18.52% due to segmentation errors during the fully automated screening process and 66.67% due to retina thickening caused by other lesions like neovascular age-related macular degeneration (nAMD)17. OCT screening may still advance from future research on improved automated OCT devices or artificial intelligence (AI) models yielding OCT scans with higher levels of automation, resolution and quality. These results confirm that incorporating SI-OCT substantially enhances the accuracy of DME screening and enables automated imaging, analysis, and detection for referrals.

Our study found that incorporating manual SD-OCT into the FP-based DR screening practice is not cost-effective in rural China. Despite the superior screening performance, the expensive manual SD-OCT machines are currently unavailable at most DR screening sites, even in most developed countries. The high cost of incorporating manual SD-OCT in DR screening, including the initial capital, personnel and skill set requirements, may impact government funding decisions, overburden health systems, and impede large-scale implementation. A recent economic study in Hong Kong showed that using manual SD-OCT to screen DME only for patients with type 2 DM was cost-effective at an ICER of $1801.2 per QALY (taking FP-based strategy as the status quo), but with costs up to $86.01 per patient10. However, our results showed that incorporating manual SD-OCT into DR screening was not cost-effective, with an extremely high ICER of $45,753 per QALY (> 3× per capita GDP), and far less likely to be preferred in the sensitivity analysis. The divergence primarily from the differences in the proposed strategies and the cost-effectiveness models. Our study proposed to incorporate OCT into the current FP-based practice rather than deploying OCT solely. Since FP plays an important role in detecting DR and the related complications (e.g., vitreous hemorrhage and tractional retinal detachment), which are difficult to be detected in OCT images, the previously proposed SD-OCT only strategy may not be feasible for real-world application. Moreover, our study considered DR screening as an entirety, with the presence, severity, progression, screening, referral and treatment of DR (non-RDR and RDR) being included in the cost-effectiveness models, rather than DME only.

The FP + SI-OCT screening strategy was shown to be very cost-effective in our study. While this screening strategy required additional investment in equipment and further treatments (US$302,015 per 100,000 individuals screened), it yielded considerable incremental QALYs at the population level, comparable to the more resource-intensive FP + SD-OCT strategy. These would bring lots of benefits to the cost-effectiveness of screening. First, the fully automated SI-OCT have no working time limitation and perform screening with much shorter time when compared with traditional human-dependent screening protocol. The improved efficiency could not only increase the number of patients screened, but also reduce participants’ waiting times and income losses. Moreover, using CST measured with SI-OCT as the criteria not only allows for easy threshold optimization in the referral, but also provides objective and data-driven diagnosis without subjective human biases, which leads to more consistent and unbiased assessments. Screening results and referral recommendations based on CST measurement could be provided in real-time, instead of waiting for weeks after visit. Emerging evidence from other studies suggests that real-time screening results at the screening site improved referral compliance18,19.

Since this study was conducted in Shaoguan, which is a low-income region in China, our findings on automated SI-OCT screening have practical implications for eye care capacity building in developing countries and regions. Burden from diabetic retinopathy, especially DME, is rendered challenging by the disproportionate prevalence and constraints of trained primary eye care providers in developing regions2,20. For example, China had one-third of the world’s population with diabetes estimated at 141 million, but DR screening is not routinely performed, with the active screening rate of DR lower than 20%21. Several challenges impede DR screening. First, there is a limited number of trained primary eye care providers20. Instead of national or large-scale regular DR screening programs driven by the national or regional governments, screening is often conducted by tertiary hospital medics, “Lifeline Express” project or some epidemiologic research programs22,23,24. Second, primary eye care providers have lower accuracy reading FP for determining referable retinal diseases (e.g., DME) versus retina specialists19,25,26. Third, there is a large population of diabetes patients in need of screening in China, which requires huge investments in human resources and equipment27. Fourth, a low rate of engagement and adherence to referral or follow-up care, partly due to the long waiting times for screening results28. These challenges are common in DR screening programs globally, not just in China mainland20. In these settings, screening using automated SI-OCT has the potential to obviate the large workforce of medical care providers for screening and make large-scale screening a reality in developing areas, despite the vast territory, large population, and medical resource constrains.

Emerging evidence suggests that retinal OCT imaging could be useful for detecting individuals at the risk of Referable macular diseases (e.g., nAMD)29, glaucoma30,31, or even cardiovascular disease and cognitive impairment32, which could expand the role of self-imaging OCT beyond the prevention of DR/DME. Rather than screening for a single retinal condition, studies found that AI-based screening for multiple diseases is more cost-effective33, and a recent foundation model for generalizable disease detection offers a potential solution34.

The rapid development of automated devices establishes the foundation to enable fully automated basic eye care, fundamentally transforming existing models for screening, diagnosis, and treatment. Automated devices (e.g., self-administered visual acuity tester, fundus photography and surgical robots) can be deployed in remote and underdeveloped areas to make eye care more accessible, or even at home35. Furthermore, without the extra requirement for human operation, the more individuals screened using the self-administered equipment, the more the direct equipment cost is shared and the lower the screening cost per individual. Although automated eye care devices show promise against global eye care challenges, real-world evidence in the feasibility and cost-effectiveness remains imperative.

The current study has various strengths. Firstly, to the best of our knowledge, this is the first study to evaluate both the performance and cost-effectiveness of automated, regulated digital medical device-based screening for eye disease. The current study provides an example for the feasibility and cost-effectiveness analyses of regulated digital medical prior to deploying at scale in real-world settings. Secondly, all model parameters came from reliable real-world studies and a real rural community screening. Thus, our findings may guide policymakers in China, or other LMICs lacking regular screening programs but with high prevalence of blinding eye diseases. Finally, our automated SI-OCT approach targeted DME detection, the major challenge in DR screening, and demonstrated superior cost-effectiveness.

Nonetheless, several limitations should be noted. Firstly, given literature gaps, some parameters came from non-Chinese studies. Secondly, the performance metrics of FP + SD-OCT strategy for DME was simulated from the published data, and the best screening performance was selected to mimic the ideal screening settings. Thirdly, while the cost of the current SI-OCT is only a fraction of traditional benchtop manual OCT, there is still a need to further reduce manufacturing costs to be more affordable for global implementation. Thirdly, we did not account for the potential improvement in referral rate and earlier follow-up care for OCT screening through its real-time CST measurement, grading and immediate detection of DME18. Fourthly, many screening programs utilize the detection of more-than-mild DR as the threshold for referral, while this study followed the UK NHS screening protocol, wherein patients with DR grades of R2 or R3 were referred. The primary difference between the two screening strategies’ referral criteria is whether to refer patients with isolated cotton wool spots or a venous loop. Given these lesions may not necessarily require prompt treatment, referring such cases could increase referral-related costs, and the anticipated gain in utility may be relatively modest, probably leading to slightly decreasing ICERs. Finally, the current study was conducted in a developing region, with applicability optimally to similar economic status. However, given the higher labor costs in developed nations, SI-OCT screening may be even more cost-effective.

Incorporating the NMPA-approved self-imaging OCT into DR screening shows a promising value in developing regions, by significantly improving both the sensitivity and specificity of DME detection, and the overall cost-effectiveness of DR screening as well.

Methods

Study design, setting and ethics

This was a cross-sectional and observational study conducted in a rural area of Shaoguan, one of the underdeveloped cities of Guangdong Province in China (2021 population: 3.29 million, GDP per capita: US$8,408). This study was approved by the Ethical Review Committee of the Zhongshan Ophthalmic Center (2020KYPJ201) and conducted under the tenets of the Declaration of Helsinki. Written informed consent was obtained from all participants.

Participants

Participants who were patients with known diabetes mellitus (defined by WHO criteria) were recruited from the Shaoguan Diabetic Screening Program between September 13th and December 10th, 2021. All participants were offered macular self-imaging using the SI-OCT, in addition to standard VA testing and two-field mydriatic FP. Participants were excluded from further analysis if the FP or OCT images were ungradable due to poor quality. The flowchart is shown in Supplementary Fig. 1.

Mydriatic fundus photography

Slit-lamp examination was performed prior to pupil dilation. For participants without any contradictions related to the mydriatic drops, one drop of tropicamide 1% was sequentially administered to each eye at 10-min intervals for three times. For participants exhibiting contradictions (e.g. shallow anterior chamber), pupil dilation was not pursued and non-mydriatic FP was performed instead. Participants underwent imaging in a dimly lit room. For each eye, two 45-degree non-stereoscopic fundus photographs were obtained using a non-mydriatic fundus camera (Canon CR‐DGi with a 20D SLR back; Canon, Tokyo, Japan). The first image was centered on the optic disc, while the second was focused on the macular region. Images were obtained by trained photographers, and eyes with inadequate images underwent repeat imaging attempts.

Macular scanning using the self-imaging OCT

All participants underwent the macular self-imaging using the SI-OCT (Master OCT, MOPTIM, Shenzhen, China) with an infrared light source centered at a wavelength of about 840 nm, an axial resolution of 55 µm and scan speed of 80,000 A-scans/second14. Self-imaging was performed in examination rooms at each screening site. Each participant first read printed instruction manual and watched a 2-min video tutorial. They were then left alone to perform self-imaging. To initiate scanning, participants scanned a QR code on the device with their mobile phone to gain access. Guided by automated voice prompts, participants properly positioned their head on the chinrest, fixated on the internal fixation target generated by the self-adjusting sensor, and avoid blinking during imaging acquisition. The default scan protocol obtained 6 radial, concentric B-scans (6-mm length) centered on the fovea. Invalid scans occurred if preparation instructions or fixation were inadequate, prompting the device to notify participants to adjust position and repeat imaging. Up to 3 scan attempts were permitted to pass examination. If needed, participants could access remote assistance by phone or text for any operational issues.

Screening strategies and evaluation

Three screening strategies for RDR and DME were compared in this study: the conventional FP-based strategy, a combined strategy of FP and SI-OCT (FP + SI-OCT strategy), and a combined strategy of FP and manual SD-OCT (FP + SD-OCT strategy). The conventional FP-based strategy represented the current standard U.K. National Health Service (NHS) screening protocol of visual acuity (VA) and fundus photography (FP) assessment, with non-RDR, RDR and DME detected based on VA and surrogate markers on FP36. The FP-based strategy served as the reference and base strategy. The FP + SI-OCT strategy applied VA, FP, and SI-OCT in all participants, defining non-RDR and RDR by the NHS screening protocol, while DME was detected by SI-OCT only, with thickened central subfield thickness (CST) indicating the presence of DME. Finally, to compare the cost-effectiveness of the fully automated approach for DME based on SI-OCT and the technician-dependent approach based on SD-OCT, we simulated the FP + SD-OCT strategy, in which VA, FP, and SD-OCT were applied in all participants. Non-RDR and RDR were detected by the NHS screening protocol, while DME by manual SD-OCT only. The performance of SD-OCT for DME was simulated from the published data, and the best screening performance was selected, with a sensitivity and a specificity of 100% and 98.34%, respectively10.

All fundus photographs were graded for DR by the trained graders at the screening site, according to the standard U.K. NHS screening protocol. DR was graded as R0, R1, R2, R3a, or R3s. R0 was defined as no DR and R1 as background DR with at least one of the following features: microaneurysm(s) or HMa (a term used when it is difficult to distinguish between a microaneurysm and a dot hemorrhage); retinal hemorrhage(s); venous loop(s); any exudate in the presence of other non-referable features of DR; or any number of cotton wool spots in the presence of other non-referable features of DR. R2 was defined as pre-proliferative DR with at least one of the following features: venous beading, venous reduplication, multiple blot hemorrhages, or intraretinal microvascular abnormality. R3 was classified into two groups: R3a was defined as active proliferative retinopathy with at least one of these features: new vessels on the disc, new vessels elsewhere, pre-retinal or vitreous hemorrhage, or pre-retinal fibrosis with or without retinal traction. R3s was defined as stable post-treatment retinopathy, with evidence of previous peripheral retinal laser treatment. For all of the three strategies, RDR was considered to be present if any feature(s) of R2, or R3 were found.

For the FP-based strategy, DME was detected according to the current standard U.K. NHS screening protocol, based on habitual/pinhole visual acuity and the presence of surrogate markers on FP. Specifically, DME was detected with at least one of the following features: exudate within 1 disc diameter (DD) of the center of the fovea, or group of exudates within the macula, or any microaneurysm or hemorrhage within 1DD of the center of the fovea only if associated with a habitual/pinhole visual acuity ≤ 6/12. For the FP + SI-OCT strategy, retinal boundaries on SI-OCT images were automatically segmented by the proprietary software, without manual correction applied. CST was defined as the average thickness, measured from the inner limiting membrane to the retinal pigment epithelium, within a 1 mm diameter circular region centered on the fovea. CST > 305 μm was set as the criterion for DME, based on our published data and the SI-OCT device model37,38. Typical example images of normal macula and DME from SI-OCT are shown in Fig. 1.

Definitive diagnosis and evaluation of strategies

To establish a reference standard to compare the screening performance of each strategy, two experienced retina specialists (S.D.C and L.G.) independently reviewed all screening data to definitively diagnosed each participant as having either non-referable diabetic retinopathy (non-RDR; defined as DR grades R0 or R1) or referable diabetic retinopathy (RDR; defined as DR grades R2 or R3) according to the UK NHS screening protocol. DME was diagnosed as foveal involvement of abnormal intrarenal and/or subretinal fluid with concurrent thickening affecting the 1 mm diameter CST. In cases of diagnostic disagreement between the two specialists, a third retina specialist (X.L.L.) was consulted to make the final determination. The screening performance of the three strategies for DME was compared using the following metrics: sensitivity, specificity, PPV, and NPV. To estimate the 95% confidence intervals (CIs) for performance metrics, we employed a non-parametric bootstrapping technique, by resampling the data and generating 1000 bootstrap samples of equal size to the original dataset and then deriving the 95% CIs as the 2.5th and 97.5th percentiles of the distributions of the respective performance metric values obtained from the 1000 bootstrap samples.

Model design

We built Markov models using a TreeAge Pro 2022 software (TreeAge; Williamstown, MA, USA) to compare the costs and effects of the three different screening strategies, and the primary outcome was incremental cost-effectiveness ratio (ICER). The current screening protocol (FP-based strategy) served as the status quo. The models were based on a hypothetical cohort of 100,000 individuals with DM through a total of ten 1-year Markov cycles (Supplementary Fig. 3). Individuals were allowed to enter the model as either healthy (free of RDR and DME) or unhealthy (affected by RDR and DME) and could transition to death from any health state. In each cycle, individuals either remained in the same health state or transitioned to a more severe state. Those who had neither RDR nor DME were not referred, while those with either condition were referred to tertiary eye centers for definitive examination (the screening and referral pathway is shown in Supplementary Fig. 3). Those who were eventually diagnosed with DME were assumed to receive appropriate anti-VEGF treatment and routine clinical care, which has been proven to improve vision and reduce the risk of blindness5,6, and those with RDR were recommended to have scatter or panretinal photocoagulation and clinical care to reduce blindness risk. Model parameters including transition probabilities, screening and treatment compliance, utilities, and mortality were derived from published studies in China and other Asian countries. As the average age of the population was around 60–64 years of age, a fixed natural mortality rate of 0.85% was incorporated39, with increased mortality odds for those with non-RDR, RDR and DME. Both costs and utilities were discounted at 3% annually in the base-case analysis.

Costs and utilities

All cost, utility and rate values used in the base case and sensitivity analyses are provided in Supplementary Tables 2–5. Costs of screening, referral examination, initial treatment, and follow-up were estimated from a healthcare system’s perspective, so both direct and indirect costs were analyzed. Direct medical costs included the charges for screening, examination, and treatment. Annualized cost of fixed assets was calculated assuming a 5-year lifespan and no salvage value. Staff costs were based on the average time spent per participant and salaries. Direct non-medical costs covered costs of transposition, food, and lodging for hospital visits. Indirect costs included time and wage loss for accompanying family members. As most screened individuals were over age 50, wage loss was not included in screening costs, but the wage loss for one accompanying family member was included. Total economic burden for blindness in year 1 was estimated as $8920, comprising 53.2% direct medical costs, 6.4% direct non-medical costs, and 40.4% indirect costs (family member productivity loss, low vision services, etc.), while follow-up years incurred only the 40.4% indirect costs, which equaled to $36031. Costs were derived from published data and Guangdong provincial fee schedules in Chinese yuan, converted to US dollars at 6.473 yuan/dollar (February 26, 2021).

Quality-adjusted life-years (QALYs) were calculated by assigning utility values derived from published economic analyses in China and other Asian countries, or computed utility values using visual acuity results from clinical trials40. A utility of 1 was assigned for participants without any DR, 0.87 for those with non-RDR, 0.7 for those with treated or untreated RDR, 0.65 for those with untreated DME, 0.72 for those with treated DME41,42,43, and 0.26 for blindness41.

Cost-effectiveness analyses

Taking the traditional FP-based strategy as the control, incremental cost-effectiveness ratios (ICERs) for each strategy were calculated using Eq. (1) as follows:

$${ICERs}=\frac{{\rm{incremental}}\; {\rm{cost}}}{{\rm{quality}}\; {\rm{adjusted}}\; {\rm{life}}\; {\rm{years}}\; {\rm{gained}}\,\left({QALYs}\right)}$$
(1)

Strategies were assessed by comparing the ICERs to a standardized threshold. According to the criteria by the World Health Organization’s (WHO’s) Choosing Interventions that are Cost-Effective (CHOICE) program, strategies that gained one QALY for less than 1 time per capita gross domestic product (GDP) are highly cost-effective, strategies that cost between 1–3 times per capita GDP per QALY gained are cost-effective, and strategies that cost over 3 times per capita GDP per QALY gained are not cost-effective44. The 2021 per capita GDP for Shaoguan was $8,408, so we assumed a willingness to pay (WTP) of $25,224 per QALY gained as the standardized threshold.

Sensitivity analysis

Extensive one-way deterministic and probabilistic sensitivity analyses were performed to account for uncertainties in ICERs. Probability-related data (utilities, compliance, transition probabilities) were largely from published literature, thus 10% variation was used. For costs, most treatment-related costs were from the China’s National Healthcare Security Administration and tertiary hospitals, so 20% variation was applied. A wider variation of 50% was used for the costs of screening program and blindness45. Tornado diagrams show the 10 most influential factors on ICERs. For probabilistic sensitivity analysis, 10,000 random draws from parameter distributions simultaneously varied all variables. Upper and lower bounds from one-way analysis defined distribution parameters. Methods and results reporting adhered to Consolidated Health Economic Evaluation Reporting Standards (Supplementary Table 6).