The Role of Urban Residence, Race and Ethnicity, and Glycemic Control in Receiving Standards of Care and Progression to Vision-Threatening Diabetic Retinopathy

In 2021, 9.6 million Americans (or 26% of those with diabetes) had diabetic retinopathy, of whom 1.8 million (or 5% of those with diabetes) had vision-threatening diabetic retinopathy (VTDR) (1,2). To prevent the development and progression of diabetic retinopathy, the American Diabetes Association and other professional organizations recommend monitoring of glycemic (HbA_1c, continuous glucose monitoring), blood pressure, and serum lipid control and adequate ophthalmologic screening, including at least annual dilated retinal examination once retinopathy is detected (3–5). Despite these guidelines, diabetic retinopathy constitutes one of the most common causes of blindness in the U.S., with prevalence rising over time, especially among younger individuals and racial and ethnic minority individuals (6,7).

Mild or moderate nonproliferative diabetic retinopathy (NPDR) without diabetic macular edema (DME) is characterized by early signs of retinal damage (e.g., microaneurysms, small retinal hemorrhages) and requires screening with funduscopic examination or photography, since people with diabetes are often asymptomatic (8). Progression to severe NPDR and proliferative diabetic retinopathy (PDR) is characterized by anatomical changes in blood vessels such as intraretinal microvascular abnormalities, neovascularization (for PDR), vitreous hemorrhage, and retinal detachment. DME involves swelling of the macula due to fluid leakage from damaged blood vessels and exudative changes that distort vision. Together, severe NPDR, PDR, and DME are referred to as VTDR, and individuals may progress from mild-to-moderate NPDR to VTDR over time and with poor glycemic control (9).

Fragmentation and lapses in diabetes care that hinder receiving care are highly pervasive and deleterious to prognosis after retinopathy diagnosis (10–13). However, the burden of lapses in care is likely not shared uniformly across the population, with socioeconomically disadvantaged and non-Hispanic (NH) Black, Hispanic, and Asian populations disproportionately affected (10,14,15). In most studies with characterization of disparities in management and progression of diabetic retinopathy, investigators relied on regional samples (13,16) or the studies were based on self-reported data (17,18). Therefore, the extent of these disparities in meeting standards of care and progression of diabetic retinopathy to VTDR in a diverse real-world setting is currently unknown. Whether these differences persist beyond glycemic control before detection of retinopathy is also unknown. In addition, adults living in small towns are disproportionately impacted by most complications of diabetes; however, the association with diabetes retinopathy seems less clear (19).

To address these gaps in knowledge, we used a large integrated database of electronic health records, from all 50 states and Washington, DC, to examine differences in meeting standards of diabetes care for people with diabetic retinopathy and their progression to VTDR. These data can be used to understand the contributions of place of residence, race and ethnicity, and historical glycemic control in management of diabetic retinopathy and the need for interventions to address risk factors at community and individual levels.

In this real-world analysis of electronic health records from the U.S. we used data of adult patients newly detected with NPDR without DME from the Epic Cosmos research platform from January 2015 to December 2023. The Cosmos platform integrates Health Insurance Portability and Accountability Act of 1996 (HIPAA)-complaint, deidentified data from >250 million patients seen at health care organizations from all 50 states and DC that use the Epic electronic health records (20). Patient geographic resolution is limited to the state level and encounters are date shifted to protect privacy.

The data for this study comprise all adults (aged ≥18 years) with type 1 or type 2 diabetes who had at least one in-person encounter (inpatient, outpatient, or telehealth) in each of the 2 years before diagnosis and had no recorded date of death 2 years from detection of NPDR without DME. The index date was the first encounter where an ICD-10, Clinical Modification (ICD-10-CM), code for retinopathy was recorded by a health care provider on Epic Cosmos (Supplementary Table 1). Type 1 diabetes was identified with use of diagnosis codes (ICD-10-CM E10). Type 2 diabetes was identified with use of a combination of diagnosis codes, medications, and laboratory parameters as specified by the Surveillance, Prevention, and Management of Diabetes Mellitus (SUPREME-DM) computable phenotype (21). We excluded patients with other forms of diabetes (ICD-10-CM E08, E09, E13) and did not consider other ophthalmic complications associated with diabetes such as cataract and glaucoma. A flowchart of analytic sample selection can be found in Supplementary Fig. 1.

Current residential location was categorized as urban or rural/small town based on U.S. Department of Agriculture Rural-Urban Commuting Area 2010 primary codes (22). Zip codes were classified according to measures of population density, urbanization, and daily commuting flows into metropolitan, micropolitan, small town, and rural commuting areas.

Glycemic control in the year prior to detection of NPDR was defined on the basis of the value of HbA_1c closest to date of detection. We defined three categories of glycemic control: HbA_1c <7.0%, 7.0%–8.9%, and ≥9.0%. Patients without a recorded HbA_1c in the year prior to detection of NPDR without DME were additionally categorized as having unknown glycemic control.

Standard of care in diabetes management after diagnosis was defined as meeting at least two of the following three criteria: at least one visit per year to an ophthalmologist, at least one visit per year to a primary care provider, and at least one screening each of HbA_1c, blood pressure, and lipids. We assessed whether each patient received standard of care in each of the first 2 years as well as both years after detection of NPDR.

Time to outcome was measured in no. of months from the start of the follow-up period, which began 1 month after detection of mild-to-moderate NPDR. VTDR was defined according to the earliest diagnosis of any severe NPDR, PDR, and DME with use of ICD-10-CM codes. Patients were censored on death or at last in-person encounter in a participating health care system.

We used investigator-specified covariates to account for confounding. Epic Cosmos harmonizes sociodemographic and clinical variables across participating health care systems. We extracted investigator-specified sociodemographic characteristics, namely, year of birth, biological sex (male, female), and race and ethnicity (Hispanic, NH White, NH Black, NH other). We extracted the most recent anthropometric parameters (BMI, blood pressure) and laboratory parameters (HbA_1c, LDL cholesterol, HDL cholesterol) in the year before diagnosis, insurance status (Medicare, Medicaid, Private, uninsured) for month of diagnosis, history of comorbidities (hypertension, hyperlipidemia, cardiovascular disease, cerebrovascular disease, obesity) in the 2 years before diagnosis, and prescriptions (insulin use, other hypoglycemic medications, antihypertensive and antihyperlipidemic medications, antidepressants, antipsychotics) in the year before diagnosis. We computed the number of in-person health care encounters in each of the 2 years after diagnosis as a measure of health care use. We estimated the longitudinal continuity of care within a health care system using the Bice-Boxerman index (BBI), an indicator of dispersion of care sought by the patient, in the first 2 years after detection of NPDR (23). The BBI is a measure of patient-level care continuity that ranges from 0 to 1 such that 0 reflects completely disjointed care and does not require knowledge of the main care provider or providing health system. We define high care continuity according to BBI ≥0.7 (24).

Areal socioeconomic position for the zip code of residence for patients is available through linked Social Vulnerability Index (SVI) 2020 data from the Centers for Disease Control and Prevention (25). Vulnerability according to the SVI was assessed as national percentile ranking (0–100, with higher values indicating greater vulnerability) based on a composite of 16 American Community Survey 2016–2020 indicators for four themes, namely, socioeconomic status, household characteristics, racial and ethnic minority status, and housing type and transportation.

Sociodemographic and clinical characteristics are summarized as mean and SD for normally distributed continuous variables, median and interquartile range (IQR) for nonnormally distributed continuous variables, and percentages for categorical or nominal variables. We used multiple imputation by chained equations (10 data sets, 30 iterations) to account for missingness in key covariates (Supplementary Table 2). For all analysis we used Rubin rules to pool estimates (26).

To describe the proportion of patients who received standards of care (ophthalmology visits, primary care visits, and measurement of their ABCs [HbA_1c, blood pressure, and LDL cholesterol]) we stratified the population by race and ethnicity and by urban/rural setting (our highest level of stratification), in addition to distributions according to categories of glycemic control (HbA_1c <7.0%, 7.0%–8.9%, ≥9.0%, or unknown/unavailable).

To estimate the independent associations (risk ratios and robust 95% CIs) of glycemic control, race and ethnicity, and urban residence with meeting standards of care, we used Poisson regression with robust SEs. We estimated the associations for each of the 2 years after detection of mild or moderate NPDR. We reported associations before and after adjusting for prespecified covariates and state fixed effects.

To estimate the relative hazards (hazard ratios [HRs]) of glycemic control, race and ethnicity, and urban residence we used Cox proportional hazards regression without additional adjustment for covariates. We subsequently adjusted for other prespecified covariates and state fixed effects. We included statistical interactions between categories of glycemic control and race and ethnicity to characterize differences of NH Black, Hispanic, and NH other, relative to NH White. As a sensitivity analysis, we treated all-cause mortality in the period starting from 1 month after start of follow-up as a competing event to estimate the cause-specific HR (27,28).

All analysis was carried out with R 4.2.3 using arrow 11.0.0.3, mice 3.14.0, survival 3.5.3, and other packages.

This secondary data analysis study was declared exempt from an ethics approval requirement by the Institutional Review Board of Emory University.

The code for the analysis is available at https://github.com/jvargh7/cosmos_soc_vtdr. Epic Cosmos access is available through institutional representatives of participating institutions and the Epic Cosmos team after completion of certification requirements.

The analytic sample consisted of 102,919 adults with newly detected mild-to-moderate NPDR, of whom 7,122 (6.9%) lived in rural areas (Table 1), comprising individuals from all 50 states and DC (Supplementary Fig. 2). Average (SD) age of the analytic sample was 62.6 (13.5) years, and 51% were female, 59% NH White, 21% NH Black, and 11% Hispanic. Of the analytic sample, 28% were patients diagnosed with type 1 diabetes and 93% were diagnosed with mild or unspecified NPDR without DME in the earliest available patient record. Only 3.9% and 15.0% of patients did not visit a health care system that is part of the Epic Cosmos network in the first or second year after detection of NPDR, respectively. A similar proportion of participants achieved care continuity in the first year after diagnosis in both urban and rural areas. Most patients were privately insured in both rural (78%) and urban (80%) areas. A higher proportion of patients had Medicare in rural (50%) than urban (40%) areas. The proportion of NH White adults was higher in rural (85%) relative to urban (56%) areas. Rates of glycemic control in the year prior to detection of NPDR were also similar between urban and rural areas.

Of the analytic sample, 16%, 88%, and 55% visited an ophthalmologist, visited a primary care practitioner, and had their ABCs measured in the first year after diagnosis (Table 1). Of the analytic sample, 14%, 78%, and 50% met the same targets in the second year after diagnosis (Table 1). Proportions visiting an ophthalmologist or primary care provider were similar between urban and rural areas in both years after diagnosis. More participants in urban areas had their ABCs measured in both years after diagnosis, relative to rural areas. Receipt of standards of care in both years was similar by race and ethnicity in urban and rural areas (Fig. 1).

Relative to those with good glycemic control in the year before detection of NPDR without DME, patients with HbA_1c between 7.0% and 8.9% had similar likelihood of achieving at least two of three standards of care in the first and second year after diagnosis (Table 2). However, those with HbA_1c ≥9.0% or without a historical measure of HbA_1c available in the year before were less likely to receive standards of care in the first year (HbA_1c ≥9.0%, risk ratio [RR] 0.95 [95% CI 0.94, 0.96]; HbA_1c unavailable, 0.56 [0.55, 0.58]), second year (≥9.0%, 0.94 [0.92, 0.95]; unavailable, 0.65 [0.63, 0.66]), and both years (≥9.0%, 0.89 [0.87, 0.92]; unavailable, 0.51 [0.50, 0.53]) after detection. Relative to NH White adults, NH Black and NH other adults did not differ in receiving standards of care. However, Hispanic adults were less likely to receive standards of care in the first year (0.97 [0.95, 0.99]), second year (0.95 [0.93, 0.98]), and both years (0.93 [0.90, 0.96]) after detection. Relative to rural areas, patients in urban areas were 1.10 (1.08, 1.13) and 1.12 (1.09, 1.15) times more likely to receive at least two of three standards of care in the first and second year after diagnosis, respectively (Table 2). After adjustment for covariates, the observed differences by urban residences persisted such that patients in urban areas were 1.05 (1.03, 1.08), 1.06 (1.04, 1.09), and 1.08 (1.05, 1.12) times more likely to receive at least two of three standards of care in the first year, second year, and both these years after diagnosis.

Of the analytic sample, 14% progressed to VTDR after detection of mild-to-moderate NPDR (Table 1). The median duration of follow-up was 35.5 months (IQR 17.6, 63.0). Of 12,966 patients subsequently diagnosed with VTDR, 54.0% were diagnosed in the first 24 months. The median time for progression among cases was 21.4 months [IQR 9.5–41.2]. However, the median time to event did not differ between urban (21.4 months [IQR 9.5–41.2]) and rural (21.4 months [IQR 9.1–42.0]) areas. Most participants who progressed from NPDR without DME progressed to NPDR with DME (53%) or to PDR (40%).

Progression to VTDR was higher among individuals with poor (HbA_1c 7.0%–8.9%, ≥9.0%) and unknown historical glycemic control in both urban and rural areas (Fig. 1). After adjustment for covariates, hazards of VTDR (HbA_1c 7.0%–8.9%, HR 1.23 [95% CI 1.17, 1.29]; ≥9.0%, HR 1.63 [95% CI 1.54, 1.72]) continued to be higher for individuals with poor historical glycemic control relative to individuals with good glycemic control (Table 2). Moreover, individuals for whom HbA_1c was unavailable in the year prior to detection had higher relative hazards of VTDR (HR 1.43 [95% CI 1.35, 1.51]). Relative to NH White adults, NH Black (HR 1.17 [95% CI 1.12, 1.22]) and Hispanic (HR 1.23 [95% CI 1.16, 1.30]) adults had higher hazards of VTDR, while hazards for NH other adults did not differ (HR 1.03 [95% CI 0.98, 1.12]). Urban residence was also associated (HR 1.07 [95% CI 0.99, 1.15]) with progression to VTDR (Table 2). Magnitude of associations between glycemic control and VTDR was similar across race and ethnicity groups in the case of HbA_1c 7.0%–8.9% and ≥9.0%. NH Black and Hispanic adults had higher relative hazards in the case of historical HbA_1c <7.0% and unavailable HbA_1c (Fig. 2). Associations were similar to those of the main analysis with all-cause mortality treated as a competing event (Supplementary Table 3).

In this analysis of electronic health record data of patients with NPDR living in the U.S., the estimates of receiving standards of care and progression to VTDR suggest opportunities for improvement in clinical practice and access to services. The present analysis suggests that urban residence, race and ethnicity, and historical glycemic control were independently associated with receiving standards of care and prognosis of diabetic retinopathy. We highlight three policy and practice-relevant findings applicable to those diagnosed with NPDR. First, in the analytic sample only one of six visited an ophthalmologist and only one in three had their ABCs measured in the 2 years after detection of NPDR. Second, both poor glycemic control and unavailability of HbA_1c in the year prior to detection of NPDR are risk factors for lapses in care and progression to VTDR. Third, racial and ethnic minority individuals, although similar to NH White individuals in receiving standards of care, were more likely to progress to VTDR.

The low rates of receipt of standards of care in the years after detection of NPDR present opportunities and challenges. For instance, although urbanicity conferred an advantage to rurality in receiving standards of care, patients living in urban areas were also more likely to progress to VTDR. We believe this association is potentially due to unmeasured and unadjusted confounding by socioeconomic characteristics such as poverty and housing insecurity (17,29), given similar rates of glycemic control and receipt of standards of care seen for urban populations above. The higher hazards of progression to VTDR in urban areas may also have contributed to the relative risk of receiving standards of care, suggesting confounding by indication.

The rate of follow-up with an ophthalmologist was low in all groups within the analytic sample, at 16% and 14% during the first and second years after diagnosis with NPDR, respectively. Although it is possible that care was provided by providers and systems other than those participating in Epic Cosmos, these results are consistent with those of a previous study that suggest that of 36,497 patients, 63% experienced a lapse in care of >2 years within a 9-year period (10). Potential teleretinal screening programs may provide a cost-effective solution in both urban and rural settings to identify, refer, and monitor patients with diabetic retinopathy (30). However, our findings emphasize that screening alone may not be sufficient. Comprehensive and multidisciplinary care is essential to retain the most vulnerable populations in care. As expected and shown in this analysis, patients at higher risk for disease progression include minority populations who often have worse glycemic control. Of interest, Hispanic/Latino individuals were the least likely to receive standards of care and had the highest risk of progressing to VTDR. These findings are consistent with a recent analysis of diabetic retinopathy in the U.S. that shows the highest prevalence of type 2 diabetic retinopathy among Hispanic/Latino individuals (7,31).

Despite 88% and 78% of participants following up with a primary care provider in the first and second years after diagnosis, respectively, only 55% and 50% received the recommended ABC measurements in those same years. This represents a disconnect between guidelines for required follow-up screening after diagnosis and actual primary care practice. Nevertheless, opportunities exist for health care providers to reduce the burden of VTDR through management of retinopathy within the first critical years after diagnosis (18). Moreover, use of artificial intelligence–enabled approaches to monitor progression to VTDR can close the gaps in availability of ophthalmologists, although their performance needs to be continually monitored (32,33).

Consistent with randomized trials and prospective studies, poor glycemic control was associated with progression to VTDR (34,35). However, we did not observe heterogeneity in the associations between poor glycemic control and VTDR across race and ethnicity groups. Additionally, relative hazards were higher for NH Black and Hispanic adults if they did not have a prior HbA_1c measurement or in the case of HbA_1c <7.0%, suggesting lapses in long-term diabetes management as well as postdetection care for retinopathy—both indicative of structural barriers to access (1,10,16).

Our analysis has several strengths such as the large sample size, duration of follow-up, and comprehensive information on care-seeking behavior. However, there are some limitations. Although Epic Cosmos incorporates information from >33,000 clinics and 1,412 hospitals, it does not include all hospitals that use Epic systems, since participation in the database is voluntary. Several of the institutions participating in Cosmos are also well-resourced academic medical centers, which may not be representative of health care access for the U.S. population. Moreover, fewer than one-half of the health care systems in the U.S. use Epic, suggesting that our estimates of care continuity, achievement of standards of care, and VTDR diagnosis may be downwardly biased (36). We believe that VTDR is potentially underdiagnosed in our analytic sample, given that only 28.6% of the analytic sample visited an ophthalmologist in the follow-up period. We also did not have information on residence at time of retinopathy diagnosis, although rates of migration from urban to rural areas and vice versa are expected to be low (37,38). Nevertheless, the analytic sample consisted only of individuals who sought care in a Cosmos-participating health system at least once in the 2 years prior to inclusion. Second, we did not have fundus images to verify the diagnosis and progression of retinopathy and, instead, relied on diagnostic codes. We also did not have information on clinical notes, or other individual and social determinants of health, given the nature of the data set.

In conclusion, our results indicate significant disparities in diabetic retinopathy care including greater adherence to standards of care in urban populations relative to rural populations, as well as minority race and ethnicity and poor or unknown glycemic control being risk factors for not receiving standards of care and progression to VTDR. The findings from this study provide a better understanding of the complex interplay of social factors affecting the management of diabetic retinopathy. Future research should be conducted to characterize disparities in management across socio-ecological levels, including, but not limited to, the role of patient adherence, provider attitudes, health system performance, and structural barriers to health (39).

Acknowledgments. The authors thank Brendan Joyce and Kayla Yates of the Epic Cosmos research team for technical input.

Funding. The work of F.J.P. is partly funded by the National Institutes of Health (P30DK111024).

Duality of Interest. F.J.P. has received research support (to Emory University) from Dexcom, Insulet, Tandem, Novo Nordisk, and Ideal Medical Technologies and has received consulting fees from Dexcom and Medscape. No other potential conflicts of interest relevant to this article were reported.

Author Contributions. J.S.V. and F.J.P. conceptualized the study with input from A.M.H. V.R.K. led the data extraction, supported by J.S.V. V.R.K. had full access to the data in the study. J.S.V. and J.B. wrote the first draft of the manuscript with input from A.M.H. and F.J.P. All authors reviewed and edited the subsequent drafts. J.S.V. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.

Prior Presentation. Parts of this study were presented at the 84th Scientific Sessions of the American Diabetes Association, Orlando, FL, 21–24 June 2024. A video presentation can be found in the online version of the article at https://doi.org/10.2337/dci24-0024.

Handling Editors. The journal editor responsible for overseeing the review of the manuscript was Steven E. Kahn.