Social Risk Groups in Patients With Diabetes With Differing Eye Care Utilization and Vision Outcomes

Abstract

Purpose: To evaluate whether latent class analysis on social determinants of health (SDoH) data can identify social risk groups that differ by adverse SDoH and vision outcomes in patients with diabetes.

Methods: This was a prospective cohort study of adults ≥18 years with diabetes who completed a SDoH survey. Latent class analysis was used to cluster patients into social risk groups. The association of social risk group and severity of diabetic retinopathy, history of lapses in diabetic retinopathy care, and visual acuity was evaluated.

Results: A total of 1006 participants were included. The three social risk groups differed by sociodemographic characteristics. The average age was 65, 60, and 54 in Groups 1, 2, and 3 respectively. Most (51%) patients in group 1 were non-Hispanic White, 66% in group 2 were non-Hispanic Black, and 80% in group 3 were Hispanic. Group 1 had the lowest burden of adverse SDoH per person (average 3.6), group 2 had 8.2, and group 3 had 10.5. In general, group 1 lacked diabetic retinopathy knowledge, group 2 had financial insecurity and difficulties with transportation, and group 3 had financial insecurity and did not have health insurance. Social risk group was associated with a history of lapses in diabetic retinopathy care, and presenting with worse vision.

Conclusions and Translational Relevance: We identified distinct social risk groups among patients seeking care for diabetic retinopathy that differed by social needs, eye care utilization, and vision. Identifying these groups and their specific needs can help guide interventions to effectively address adverse SDoH and improve eye care utilization and vision outcomes among patients with diabetes.

Introduction

Social determinants of health (SDoH) are the conditions in which people live, work, and play and the wider set of social structures and economic systems that affect the conditions of daily living.¹ SDoH have a major impact on health outcomes, perhaps to a greater degree than medical care, and this is certainly true for diabetic retinopathy.^2,3 Adverse determinants, for example, not graduating from high school, unemployment, low household income, and food insecurity, have been associated with a greater prevalence of diabetic retinopathy.^4,5 Having poor housing and a greater burden of adverse SDoH have been associated with underutilization of eye care.^6–8 Adverse determinants have been linked with an increased risk of vision impairment.^9,10 One of the challenges of identifying the influence of specific determinates on health outcomes is that SDoH are inter-related and influence each other. The World Health Organization suggests that structural determinants, for example, racism, drive downstream intermediary determinants, for example, food availability.¹ Frameworks such as Healthy People 2030 group related determinants into domains, including economic stability, education access and quality, neighborhood and built environment, social and community context, and health care access and quality.¹¹ 

Identifying the most impactful SDoH could provide targets for intervention to address social needs and improve vision outcomes. However, the interconnectedness of SDoH poses considerable challenges in analyzing the effects of adverse determinants on health outcomes. Examining specific SDoH in isolation ignores the impact of other determinants. Summing SDoH in aggregate assumes that each factor has equal importance. Neither method provides the insights needed to guide interventions, specifically understanding which adverse determinants are most impactful on health in which groups of individuals.¹² Latent class analysis (LCA) is a data-driven methodology that attempts to detect the presence of unobserved latent groups among categorical data.¹³ Applying LCA to SDoH data can help to identify distinct social risk groups.¹⁴ We hypothesized that applying LCA to diabetic retinopathy care could highlight distinct social risk groups that differ in eye care utilization, severity of eye disease, and visual acuity. Identifying these different social risk groups can provide insights into the types of interventions needed to address adverse SDoH to improve diabetic retinopathy care and outcomes. 

Methods

Study Design

This was a prospective cohort study of adult patients ≥18 years with type 2 diabetes mellitus seen from July 2, 2021, to March 3, 2023, at the Wilmer Eye Institute at Johns Hopkins Hospital. Patients under the care of providers who agreed to participate in the study (n = 63) were screened for eligibility (Supplementary Fig. S1). Patients were excluded if they had another significant eye disease unrelated to diabetes requiring ongoing monitoring or treatment (e.g., acute retinal necrosis) (Supplementary Table S1). These conditions were identified by manual chart review. Patients were also excluded if were unable to complete the survey or did not speak English or Spanish. All participants provided oral consent. Trained research assistants completed the oral survey with eligible patients either by phone or in clinic (resident clinic or retina clinic). This study was approved by the Johns Hopkins University School of Medicine Institutional Review Board (IRB00279180) and adhered to the tenets of the Declaration of Helsinki. 

The SDoH survey was based on the World Health Organization and Healthy People 2030 framework, and leveraged questions from existing surveys.^1,11 The survey was divided into questions about sociodemographic characteristics, SDoH domains (education access and quality, economic stability, social and community context, neighborhood and built environment, and health care access and quality), and receipt of monetary and nonmonetary services (Supplementary Table S2). The set of questions was developed through literature review of determinants that have been associated with diabetic retinopathy care. The SDoH survey contained 37 core questions and 17 possible subquestions. 

We linked the SDoH survey results with retrospective data in the electronic health record. Ophthalmic diagnoses were extracted, including the severity of diabetic retinopathy (no diabetic retinopathy, nonproliferative diabetic retinopathy, and proliferative diabetic retinopathy), other retinal disorders (including age-related macular degeneration, retinal vein occlusion, and retinal artery occlusion), and glaucoma.⁸ For patients that were seen more than once at the Wilmer Eye Institute, we identified whether they had a history of a lapse in diabetic retinopathy care in the first 2 years at Johns Hopkins Hospital as previously described by our group.⁸ In brief, a lapse in care was defined as not returning for a follow-up visit as recommended by the treating provider. We also extracted the best recorded logarithm of the minimum angle of resolution (logMAR) visual acuity from the office visit closest to the SDoH survey completion date. 

Statistical Analyses

The analyses were divided into three main components. First, we quantified the relationship between each SDoH variable. All items in the SDoH survey were treated as categorical variables and dichotomized with the more adverse SDoH (e.g., uninsured as compared with insured) assigned the higher value (Supplementary Table S2). Missing SDoH data were minimal, ranging from 0% to 16%. For each pair of SDoH variables, we computed a log odds ratio. 

Second, we used LCA to identify distinct social risk groups. LCA is a statistical clustering procedure that identifies different subgroups within populations based on outward characteristics.^15,16 The underlying assumption is that membership in an unobserved social risk group can explain patterns in adverse SDoH identified by the survey. We excluded questions on upstream sociodemographic characteristics, monetary and nonmonetary services, and health outcomes. We used the Akaike information criterion and Bayesian information criteria to select the optimal model with the smallest number of latent classes.^15,17 We evaluated other diagnostic statistics including smallest class size (minimum of 5%), entropy (>0.8), and average latent class posterior probability (>0.9).¹⁵ Ultimately, we chose a model with three latent classes based on a priori model fit and diagnostic criteria (Supplemental Table S3). 

Third, we examined the relationship between the social risk groups and various outcomes using χ² tests of homogeneity. We used the maximum-probability from the LCA model to assign each patient to a social risk group.¹⁸ The outcomes included the severity of diabetic retinopathy (whether the patient had proliferative diabetic retinopathy), history of lapses in diabetic retinopathy care, and visual acuity (dichotomized to worse than 20/40 or a logMAR >0.3, or better than 20/40 or a logMAR ≤0.3). When visual acuities from both eyes were available, an average of the two eyes was used. All statistical analyses were performed in Python (Python Software Foundation, Wilmington, DE, USA; Python Language Reference, version 3.8.9) and Stata statistical software (Version 17.0 for Windows; StataCorp LLC, College Station, TX, USA). 

Results

During the 20-month enrollment period, we identified 3311 eligible patients by phone (n = 2681) and in clinic (n = 630) (Supplementary Fig. S1). Of those contacted, 45% consented to the study, and 1006 patients (n = 647 by phone, and n = 359 in clinic) completed the survey (Supplementary Fig. S1). Baseline sociodemographic characteristics of the study participants are shown in Table 1. Most patients were ≥45 years, approximately one-half were female, 34% were non-Hispanic White, 43% were non-Hispanic Black, and 17% were Hispanic individuals. 

Table 1.

 

Sociodemographic Characteristics and Survey Methodology of Participants by Social Risk Group

 

There were strong associations between the SDoH variables (Fig. 1). For example, in the domain of education access and quality, the highest level of school was associated with the highest degree earned. In the domain of economic stability, having difficulties paying for basic necessities was associated with not having enough food to eat. In the domain of health care access and quality, not having insurance in the past 12 months was associated with the type of main insurance. There were also strong associations between SDoH variables across domains. For example, not having a high school degree was associated with difficulties paying for basic necessities, worrying about losing housing, and not having health insurance. 

Figure 1.

 

Heat map of the association between social determinants of health (SDoH) variables on the log odds scale. SDoH variables are grouped by domain (education access and quality, economic stability, social and community context, neighborhood and built environment, health care access and quality). Darker colors indicate stronger associations in the positive (red) or negative (blue) directions. Noncolored cells without a log odds value represent nonstatistically significant (p > 0.05) associations.

Using LCA, the best performing model included three distinct groups. The three-group model had the lowest Bayesian information criteria and passed other model fit and diagnostic criteria (Supplementary Table S3). Most participants could be strongly placed into one of the three groups as demonstrated by the ternary plot except for one individual who had a 43%, 36%, and 21% probability of being placed into group 1, 2, and 3 respectively (Supplementary Fig. S2). 

The three social risk groups identified by the LCA analysis differed by sociodemographic characteristics (Table 1). The average age was 65, 60, and 54 in groups 1, 2, and 3, respectively. There were more female patients in group 2 (66%) compared with groups 1 and 3. Most of the patients in group 1 were non-Hispanic White (50%), whereas most (66%) in group 2 were non-Hispanic Black, and most (80%) in group 3 were Hispanic individuals. 

The three social risk groups had different patterns of adverse SDoH across the various domains. Group 1 had the least burden of adverse SDoH per person (average, 3.6 ± 1.7), group 2 had an average of 8.2 ± 2.2, and group 3 had the highest burden with an average of 10.5 ± 2.8. Group 3 included 16% of the cohort (n = 164) and had the highest proportion of patients without insurance, without a high school degree, and with financial insecurity (Fig. 2 and Table 2). Group 2 included 30% of the cohort (n = 305), and these patients also had financial insecurity but all were insured. Compared with the other groups, group 2 had more difficulties with transportation to the eye doctor (e.g., either had to take a taxi, bus, or train, walk, or other, or it took them a long time to get to the eye doctor). A greater proportion in group 2 were also not married or living with a partner. Group 1 included 53% of the cohort (n = 537) and had the fewest adverse SDoH. Group 1 also all had insurance and was largely financially secure. The main adverse SDoH in group 1 was not having knowledge about diabetic retinopathy. 

Figure 2.

 

Radar graph displaying the proportion of adverse social determinants of health (SDoH) in each social risk group. Group 1 (green line) has the overall least adverse SDoH burden and group 3 (purple line) has the most. Group 3 has the highest proportion of patients without insurance, without a high school degree, and with financial insecurity. Group 2 (orange line) has financial insecurity and difficulties with transportation to the eye doctor. Group 1 had insurance and financial security, but many lacked knowledge about diabetic retinopathy.

Table 2.

 

Adverse Social Determinants of Health (SDoH) by Social Risk Group

 

The three social risk groups were associated with some of the health outcomes. There were no differences in distribution of proliferative diabetic retinopathy by social risk group (P = .10) (Table 3). There was a greater proportion of patients who had a history of a lapse in diabetic retinopathy care in groups 2 and 3 (P < 0.001), and who presented with visual acuity of <20/40 (P < 0.001). 

Table 3.

 

Percent Distribution of the Outcomes (Severity of Diabetic Retinopathy, History of Lapses in Diabetic Retinopathy Care, and Visual Acuity) by Social Risk Group

 

Discussion

In this prospective cohort study, we evaluated a multidimensional and comprehensive set of SDoH across multiple domains among a group of patients with diabetes seeking eye care at an academic medical center. We confirmed the interconnectedness of SDoH and identified associations between SDoH within and across different domains. Using LCA, we were able to group individuals into distinct social risk groups that differed by social need and health outcomes. These social risk groups were associated with having a history of lapse in diabetic retinopathy care, presenting with vision worse than 20/40, but were not associated with having proliferative diabetic retinopathy. 

This work further emphasizes the interconnectedness among SDoH. Associations between SDoH variables in the same domain were expected. Many questions within domains relate to the same idea. For example, the highest level of school was related to the highest degree that a patient obtained. Finding these associations bolsters the internal validity of our survey and reliability of the study design. The associations found between SDoH across different domains are consistent with what others have reported. For example, financial insecurity as identified by the Protocol for Responding to & Assessing Patients’ Assets, Risks & Experiences questions was related to housing instability and concerns about losing housing. This finding is consistent with the notion that poor housing is an indicator of poverty.^19,20 

Using LCA, we were able to identify three distinct social risk groups with differing adverse SDoH among a heterogeneous group of patients with diabetes seeking eye care. The distinct social risk groups identified by LCA suggests different interventions might be needed for each subgroup to address adverse SDoH. Overall, group 1 had the fewest adverse SDoH and primarily had trouble with diabetic retinopathy knowledge. This finding suggests that improving diabetic retinopathy education could be effective for those in group 1. Many in group 2 experienced financial insecurity and trouble with transportation to the eye doctor, suggesting interventions to improve transportation could be effective for those in group 2. And many in group 3 had financial insecurity and did not have insurance, suggesting interventions to improve health care access could be beneficial for those in group 3. These data also suggest that a single intervention that has been the focus of many attempts to improve health care outcomes is likely not to be effective for all members of a given population. For example, those with difficulties with transportation and accessing health care in groups 2 and 3 are unlikely to benefit from diabetic retinopathy education alone, but may benefit from a combination of interventions. Interventions should be driven by patient's individual social needs, but LCA could be a means by which to identify which ones are most important within subgroups and to prioritize interventions within a given health system. 

The social risk groups identified in the LCA of this diabetic population have several similarities to ones identified in a group of Washington DC Medicaid beneficiaries.¹⁴ McCarthy et al.¹² identified four social risk groups using their data. Group 1 similarly had the least social risks. Group 2 had economic instability and housing instability, group 3 had difficulties with transportation to health care, and group 4 had the most adverse SDoH with economic instability, housing instability, difficulty with transportation, and spending time in jail/prison. A similarity between the risk groups identified both in the present study and the McCarthy study is that economic instability is a feature common to all the groups other than group 1. The other groups further distinguished themselves with additional adverse SDoH. This finding is consistent with frameworks and studies that suggest that economic instability might operate slightly upstream of other domains.^1,7 

The social risk groups had different health outcomes. Social risk group was associated with a history of lapses in diabetic retinopathy care, whereby patients with more adverse SDoH were more likely to have lapses in care. This finding is consistent with prior studies suggesting health disparities in the risk for lapses in diabetic retinopathy care.⁸ Social risk groups were associated with vision of <20/40, again consistent with prior studies showing health disparities in the risk of vision impairment.²¹ Although we chose to use the average visual acuity when bilateral measurements were available, sensitivity analyses (data not shown) suggested similar results when we chose the better visual acuity, a randomly selected eye, as well as only right eyes. Last, we did not find an association between social risk group and the most severe form of diabetic retinopathy, proliferative diabetic retinopathy. This lack of association could be a result of various factors. It could be that we did not have sufficient power to detect differences. Proliferative diabetic retinopathy was rare, only 5% in group 1 and higher in groups 2 and 3 (9% and 6%), but this difference did not reach statistical significance. It could also be that proliferative diabetic retinopathy was not detected in the groups who were more likely to have lapses in diabetic retinopathy care thus lower eye care use and less likely to be diagnosed by an eye care provider. 

LCA is a powerful analytic approach for categorizing groups or classes within heterogeneous populations.²² The technique grew out of quantitative psychology but has since been applied widely in other disciplines including health sciences.²³ LCA has been leveraged in other aspects of medicine, for example in identifying a subphenotype of acute respiratory distress syndrome characterized by more severe inflammation, shock, metabolic acidosis, and worse clinical outcomes.²⁴ In technical terms, LCA is a special kind of finite mixture statistical model that uses a probabilistic approach to identify clusters of subjects using observations on many discrete variables.²² LCA is designed for categorical observations and latent variables, whereas other latent variable measurement models, for example, factor analysis, are more appropriate with continuous indicators and continuous latent variables.²⁵ Because it uses a probabilistic model, LCA offers advantages over other subgrouping approaches, such as cluster analysis.²² For example, probabilistic-based classification offers a more nuanced understanding of class membership uncertainty. Furthermore, LCA can handle multiple categorical variables simultaneously, and there are robust model selection indices to aid in model selection.^15–18 

Our study has several limitations. First, LCA is a data-driven approach and as such requires significant data to generate meaningfully distinct groups. The application of LCA in routine clinical care could be limited because SDoH data are not typically collected as part of usual patient care. This study further bolsters the growing need to collect SDoH at the point of care to improve health. Second, assignment of the maximum probability of each patient into a social risk group does not consider the uncertainty in the assignment and might result in some estimation error. However, because most of the participants in this study could be strongly placed into one of the three classes, this estimation error is not expected to be large or change the primary results. Furthermore, there is evidence to suggest that cut models, as in the one used here, are more robust to model misspecification as compared with joint models.²⁶ Third, many standardized multidimensional SDoH collection tools exist, but none that include all relevant determinants for diabetic retinopathy care. Although the questions used as part of this multidimensional survey were taken from other standardized questionnaires, there could be some measurement error. We tried to get around this issue by building redundancy into the survey, for example, asking about the highest level of schooling in addition to degree earned. Fourth, the generalizability of these results is unknown. Our sample represents English and Spanish speaking populations of patients with types 2 diabetes who sought care at a tertiary academic ophthalmic practice in an urban environment. We chose to specifically include patients with type 2 diabetes because there are slightly different ophthalmic screening guidelines for patients with type 1 and type 2 diabetes.²⁷ It is likely that the adverse SDoH that impact health outcomes and vision care are different depending on the population. The specific findings and social risk groups identified in this study might not be applicable to others, but the methodology of applying LCA to SDoH data is generalizable. Despite these limitations, we evaluated a comprehensive multidimensional set of SDoH in a prospective cohort of patients with diabetes seeking eye care and were able to identify distinct social risk groups with differing social needs and vision outcomes. 

In conclusion, patients with diabetes can be clustered into distinct social risk groups that have different eye disease outcomes. Groups with greater adverse SDoH burden have lower eye care use and worse vision. The unique adverse SDoH in each social risk group can help guide interventions to address social need to potentially improve vision and health outcomes. 

Acknowledgments

Supported by a Career Development Award from the Research to Prevent Blindness (CXC), K23 award from the NIH/NEI (award number K23EY033440) (CXC), K24 award from the NIH/NHLBI (award number K24HL148181) (DCC), and an unrestricted grant from Research to Prevent Blindness (Wilmer Eye Institute). Cai is the Jonathan and Marcia Javitt Rising Professor of Ophthalmology. 

Disclosure: C.X. Cai, None; D. Han, None; D. Tran, None; J.A. Moreno, None; S.L. Zeger, None; D.C. Crews, None