The Association of Lapses in Diabetic Retinopathy Care with Vision Impairment

Abstract

Objective

To identify the association of lapses in diabetic retinopathy (DR) care with vision impairment among patients with type 2 diabetes (T2D), stratified by DR severity and treatment status.

Design

Retrospective cohort study.

Subjects

Adults with T2D seen at the Wilmer Eye Institute from 2013 to 2023.

Methods

Propensity score modeling was used to estimate the probability that a patient would experience a lapse in care during a 2-year observation window based on sociodemographic (age, sex, self-reported race or ethnicity, insurance), ocular (baseline visual acuity, ophthalmic comorbidities, history of retinal treatments), and medical factors (Diabetes Complication Severity Index, Charlson Comorbidity Index). Doubly robust logistic regression models using inverse probability weighting of the propensity scores were used to estimate the association of care lapses with vision impairment. We also estimated average marginal effects to examine the probability difference in vision impairment associated with care lapses among patients grouped by DR severity (no DR, nonproliferative DR, proliferative DR) and treatment status (no treatment, any treatment, anti-VEGF only, panretinal photocoagulation only).

Main Outcome Measures

Vision impairment (or vision worse than Snellen equivalent 20/40) at the end of the observation period.

Results

A total of 45 764 patients with 90 440 eyes contributed to the study. Most patients were female (52%), and the average age was 61.7 years. A total of 81% had a lapse in care. Patients who had a lapse in care had 50% higher odds of having vision impairment compared with patients who never lapsed (P < 0.001). This association was seen across all subgroups of DR severity and treatment status (all P < 0.001).

Conclusions

For patients with T2D, a lapse in care was associated with higher odds of having vision impairment, regardless of underlying severity of DR or treatment status. These findings highlight the importance of longitudinal DR care and reducing lapses in care to prevent vision loss.

Financial Disclosure(s)

Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.

Diabetic retinopathy (DR) is one of the most common causes of vision loss globally.1, 2, 3, 4, 5, 6, 7 Screening and regular follow-up for patients with diabetes are critical for detecting the development or progression of DR and initiating appropriate treatment.⁸ Lapses in DR care among patients with diabetes can lead to delays in initiation of these vision salvaging treatments, potentially resulting in vision loss.8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18

Studies have reported that a high proportion of patients have visual acuity (VA) loss or vision-threatening disease progression after an extended lapse in DR care.^{9,11,12,19,20} However, the majority of these studies focus on patients at the highest risk of vision loss, such as those with proliferative DR (PDR) and diabetic macular edema that require treatment. Furthermore, most existing studies rely on either randomized controlled trials with standardized protocols or observational studies that use fixed time intervals to define a lapse in care (eg, not returning in 6 months or 1 year).^10,12,14,19,21, 22, 23 Fixed definitions may not align with provider-recommended follow-up schedules, limiting their accuracy in identifying clinically meaningful lapses. In contrast, using an adaptive definition that incorporates individualized provider recommendations offers greater sensitivity and specificity by accounting for the urgency of follow-up and the individual patient's risk of disease progression.^9,17,24 The objectives of this study were to evaluate the association between lapses in DR care with vision impairment among patients with type 2 diabetes (T2D) that considers providers' individual recommended follow-up intervals, and how this association varies by the severity of DR and DR treatment status.

Methods

Study Design and Patient Selection

This was a retrospective cohort study of adult patients aged ≥18 years with T2D first seen at the Wilmer Eye Institute at Johns Hopkins Hospital for DR screening or treatment from April 4, 2013 to September 7, 2021.²⁴ Patients with T2D were identified if they had a qualifying International Classification of Diseases code at any clinical encounter or a hemoglobin A1c value of ≥6.5% as previously described.²⁴ All DR screening and treatment office visits in a 2-year observation window after the patient's entry into the cohort until 2023 were analyzed to standardize the observation period and opportunity for lapse in care. The study was approved by the Johns Hopkins institutional review board (institutional review board #0026432) where informed consent was waived. The study adhered to the tenets of the Declaration of Helsinki and followed the Strengthening the Reporting of Observational Studies in Epidemiology reporting guidelines.²⁵

Lapses in DR Care

The main explanatory variable was whether the patient ever had a lapse in DR care, as previously defined.²⁴ In brief, the provider's recommended follow-up time frame was compared with the patient's actual office encounter. A lapse in care was defined if the patient returned ≥10 weeks when recommended to return ≤8 weeks, ≥24 weeks when recommended to return >8 to ≤20 weeks, ≥40 weeks when recommended to return >20 to ≤32 weeks, and ≥64 weeks when recommended to return >32 to ≤52 weeks.²⁴

VA Outcomes

The main outcome variable was vision impairment at the end of the 2-year observation period (±6 months). The best recorded VA in the electronic health record from all possible measurement methods was extracted. Possible measurement methods for each eye included at distance or near, with or without habitual correction, with or without pinhole, with manifest refraction at distance or near, cycloplegic refraction at distance, or final dispensed refraction at distance. Measurements were converted to logarithm of the minimum angle of resolution for analysis.26, 27, 28 Vision impairment was defined as vision worse than logarithm of the minimum angle of resolution ≥0.3 or Snellen equivalent of 20/40. For eyes that did not have a VA measurement in this timeframe, intermediate VA measurements were extracted and used to impute the outcome VA (see following section).

Additional Covariates

Sociodemographic, ocular, and medical covariates were extracted from the electronic health record as previously described.²⁴ Sociodemographic covariates included age, sex, self-reported race and ethnicity, and insurance. Ocular covariates included baseline logarithm of the minimum angle of resolution VA; presence of other ophthalmic comorbidities—glaucoma, cataracts, or other retinal diseases; and history of retinal treatments—panretinal photocoagulation (PRP), intravitreal anti-VEGF, focal laser treatment, or other intravitreal injection treatments (Table S1, available at www.ophthalmologyscience.org). Medical characteristics extracted included the Diabetes Complication Severity Index and Charlson Comorbidity Index.29, 30, 31, 32 Patients with missing sociodemographic data were excluded. The diagnostic codes used to identify each clinical condition can be found in our prior publications³³ and Table S1.

Multiple Imputation for Missing Data

Missing data were addressed using multiple imputation by chained equations implemented by the mice R package.³⁴ We generated 50 data sets with imputed missing values of outcome VA at 2 years (50.4% of eyes), DR severity (48% of eyes), baseline VA (1.5% of eyes), Charlson Comorbidity Index (0.04% of patients), and Diabetes Complication Severity Index (0.04% of patients). For each imputed measure, we chose predictor variables and imputation models that respected the clustering of eyes within patients and enforced temporal ordering between predictor and outcome. Where outcome acuity was unobserved, intermediate measurements of VA (in addition to baseline VA) were used to predict outcome VA. Following simulation evidence in the literature on missing imputation, we imputed 50 data sets to protect against loss of power and to ensure convergence of estimates after pooling using Rubin rules.^35,36

Statistical Analysis

Gradient Boosting Machine propensity score modeling was used to estimate the causal effect of lapses in care on vision impairment across the 50 imputed data sets. We first modeled whether patients had a lapse in care as a function of a set of sociodemographic, ocular, and medical characteristics. The causal estimand targeted by these models was the average treatment effect among the treated.

Because different propensity score methods can produce varying treatment effect estimates and model performance,37, 38, 39 we tested 11 approaches using the R packages MatchIt, WeightIt, and twang. Diagnostics such as covariate balance and common support were assessed for each method to ensure the technique was able to generate balanced treatment/comparison groups across the entire range of propensity scores.^40,41

We used doubly robust logistic regression models using inverse probability weighting of the propensity scores to estimate the average treatment effect among the treated of a lapse in care on vision impairment.⁴² In this framework, both the weighting and covariate adjustment contribute to the estimation, providing consistent recovery of the average treatment effect among the treated as long as either the propensity score model or the outcome model is correctly specified. The outcome analysis used the survey package to estimate weighted models and the margins package to estimate marginal effects^43,44 Reported models accounted for clustering at the level of individual patient eyes using a robust sandwich variance estimator implemented in the survey package.⁴³

We chose to report the Gradient Boosting Machine–based estimates because this method demonstrated strong performance in achieving covariate balance across comparison groups. Additionally, its estimates tended to fall in the middle of the range across all methods, making it a representative choice.

Sensitivity Analysis

A sensitivity analysis was performed of only eyes with both baseline VA as well as outcome VA at 2 years of follow-up without multiple imputation for missing data.

Exploratory Subgroup Analysis

After estimating our analysis model, we calculated average marginal effect (AME) estimates of having a lapse in care on vision impairment on the patient level.⁴⁴ The AME represents the average change in the predicted probability for having vision impairment associated with lapses in care. We calculated the AME across the entire distribution of patients in our sample, as well as subsamples grouped by DR severity. The subgroup analysis included patients in the following groups: (1) those with no DR or nonproliferative DR (NPDR) not on treatment, (2) no DR or NPDR on any treatment, (3) all with NPDR not on treatment, (4) all NPDR on any treatment, (5) all with PDR on any treatment, (6) PDR only on anti-VEGF, and (7) PDR with PRP only. Any treatment was defined as receipt of PRP, focal laser, intravitreal anti-VEGF, or other intravitreal injection.

We also tested the significance of a 3-way interaction model between lapses in care, DR severity, and any treatment. Because the 3-way interaction was not found to be statistically significant (P = 0.25) we only report the results using the main model.

E-Value Analysis

The E-value quantifies the minimum strength of association that an unmeasured confounder would need to have with both the exposure and the outcome, on the odds ratio (OR) scale, to explain away the observed association. We used the EValue R package to conduct this calculation.^45,46

All analyses were performed in R, version 4.4.2 (R Project for Statistical Computing). All analysis codes are available on GitHub.⁴⁷ All P values were 2-sided and P < 0.05 was considered statistically significant.

Results

A total of 45 764 patients with 90 440 eyes contributed to the study (Fig S1, available at www.ophthalmologyscience.org). Using the imputed data, the average age was 61.7 years, 52% (N = 23 965) of patients were female, 48% (N = 21 808) were non-Hispanic White, 37% (N = 16 802) were non-Hispanic Black, 4% (N = 1917) were Hispanic, and 11% (N = 5237) were of other races and ethnicities. A total of 43% (N = 19 879) of patients had Medicare insurance, 32% (N = 14 803) had private insurance, 9% (N = 3905) had Medicaid, 7% (N = 3184) had other insurance, and 9% (N = 3993) had no insurance (Table 2). The majority of patients in the cohort (81%; N = 37 105) had a lapse in care during their 2-year observation period (Table 2). Of the eyes included, 70% (N = 63 258) had no DR, 20% (N = 18 468) had NPDR, and 10% (N = 8714) had PDR.

Table 2.

Baseline Characteristics of Patients Stratified by Having a Lapse in Care Using the Imputed Data

	Lapse in Care N (%)	No Lapse in Care N (%)	Total N (%)
Total	37 105 (81%)	8659 (19%)	45 764 (100%)
Demographic characteristics
Sex†
Male	17 486 (80%)	4313 (20%)	21 799 (48%)
Female	19 619 (82%)	4346 (18%)	23 965 (52%)
Age (yrs)†
>18 to <20	82 (71%)	34 (29%)	116 (<1%)
>20 to <45	4359 (88%)	622 (12%)	4981 (11%)
>45 to <65	16 544 (83%)	3419 (17%)	19 963 (44%)
>65	16 120 (78%)	4584 (22%)	20 704 (45%)
Race/ethnicity†
Non-Hispanic White	16 714 (77%)	5094 (23%)	21 808 (48%)
Non-Hispanic Black	14 272 (85%)	2530 (15%)	16 802 (37%)
Hispanic	1710 (89%)	207 (11%)	1917 (4%)
Other∗	4409 (84%)	828 (16%)	5237 (11%)
Insurance†
Private	11 881 (80%)	2922 (20%)	14 803 (32%)
Medicaid	3548 (91%)	357 (9%)	3905 (9%)
Medicare	15 907 (80%)	3972 (20%)	19 879 (43%)
None	3422 (86%)	571 (14%)	3993 (9%)
Other	2347 (74%)	837 (26%)	3184 (7%)
Ocular characteristics
Severity of DR†
No DR	50 030 (79%)	13 227 (21%)	63 258 (70%)
NPDR	15 517 (84%)	2951 (16%)	18 468 (20%)
PDR	7736 (89%)	978 (11%)	8714 (10%)
DR treatments†
PRP	618 (98%)	15 (2%)	633 (<1%)
Focal laser	35 (97%)	1 (3%)	36 (<1%)
Anti-VEGF	2479 (94%)	151 (6%)	2630 (3%)
Other injections	162 (89%)	21 (11%)	183 (<1%)
Ophthalmologic comorbidities†
Glaucoma	11 425 (86%)	1835 (14%)	13 260 (15%)
Prior cataract surgery	2033 (83%)	428 (17%)	2461 (3%)
Other retinal disease	550 (87%)	80 (13%)	630 (<1%)
Baseline visual acuity (logMAR)‡	0.25 (0.51)	0.17 (0.37)	0.24 (0.49)
Medical characteristics
Diabetes Complication Severity Index‡	0.85 (0.93)	0.66 (0.84)	0.81 (0.92)
Charlson Comorbidity Index‡	1.43 (1.04)	1.36 (0.94)	1.42 (1.02)

DR = diabetic retinopathy; logMAR = logarithm of the minimum angle of resolution; NPDR = nonproliferative diabetic retinopathy; PDR = proliferative diabetic retinopathy; PRP = panretinal photocoagulation.

^∗Other race and ethnicity included Asian, American Indian or Alaska Native, Native Hawaiian or Other Pacific Islander, other, unknown, choose not to disclose, unable to obtain, and ≥2 races.

^†Categorical variable.

^‡Continuous variable—reported mean + standard deviation.

Patients who ever had a lapse in care had 50% increased odds of having vision impairment compared with comparable patients who had never lapsed (OR = 1.50, 95% confidence interval [CI] 1.40 to 1.60, P < 0.001). The AME of lapses in care on vision impairment was 0.061, meaning the typical patient who lapsed had a 0.06 higher probability of experiencing vision impairment than an otherwise equal patient who did not lapse (95% CI 0.052 to 0.071, P < 0.001). In the sensitivity analysis of only complete cases, a total of 22 889 patients with 44 824 eyes were included (Table S3, available at www.ophthalmologyscience.org). Results were similar patients who ever had a lapse in care had an 19% higher odds of having vision impairment compared with comparable patients who never lapsed (OR = 1.19, 95% CI 1.11 to 1.27, P < 0.001) with the AME of lapses in care on vision impairment 0.03 (95% CI 0.02 to 0.04, P < 0.001).

To test the robustness of our findings, we evaluated multiple propensity score approaches. Figure S2 (available at www.ophthalmologyscience.org) shows that propensity score-adjusted ORs ranged from 1.13 to 1.56 (all statistically significant). The inverse probability of treatment weighting Gradient Boosting Machine demonstrated favorable balance diagnostics. Diagnostics to assess reduction in postmatching covariate differences showed substantial reductions, with a maximum absolute standardized mean difference of 0.04 (Fig S3 and S4, available at www.ophthalmologyscience.org). Density and box plots of propensity scores across treatment or control groups show substantial overlap, supporting the plausibility of exchangeability across groups.

In the subgroup analysis, patients who had a lapse in care in each DR severity and treatment status group had a higher probability for vision impairment compared with patients who did not lapse (Fig S5, available at www.ophthalmologyscience.org). The AMEs of a lapse in care on vision impairment among patients with no DR or NPDR not on treatment (0.061; 95% CI, 0.052–0.071) or on any treatment (0.071; 95% CI: 0.060–0.083) were both significant (P < 0.001). The AMEs were significant for patients with NPDR not on treatment (0.065; 95% CI, 0.055–0.075) or on any treatment (0.071; 95% CI, 0.059–0.083), P < 0.001 for both. The AMEs for patients with PDR on any treatment (0.058; 95% CI, 0.048–0.068), PDR on anti-VEGF treatment only (0.059; 95% CI, 0.048–0.069), and PDR on PRP only (0.062; 95% CI, 0.050–0.074) were all significant (P < 0.001) (Fig S5).

In the E-value analysis, the point estimate E-value to reduce the observed association to null was 1.75. The CI E-value to shift the CI to include the null was 1.65. An unobserved confounder would thus need to have associations with both the exposure and outcome that are larger than the observed OR in the study, net of covariates included in the outcome model, to lead us to fail to reject the null hypothesis.

Discussion

In this retrospective cohort study, we find that patients with T2D who have had a lapse in DR care have higher odds of experiencing vision impairment at the end of a 2-year observation period compared with similar patients who do not have any lapse. This association is consistent regardless of the severity of DR or whether the patient is receiving treatment such as PRP or intravitreal anti-VEGF.

Our result that patients who have a lapse in care have a higher probability of vision impairment is largely consistent with what is reported in the literature. The increased risk of vision loss among patients with severe NPDR or PDR is well-documented in the literature. Moderate to severe NPDR and PDR are characterized by vascular closure and neovascularization, respectively, which can lead to complications such as macular edema, retinal detachment, and vitreous hemorrhage, all of which can contribute to vision loss.⁴⁸ Previous studies have shown that these patients experience significantly worse VA compared with those without DR.^49,50 Additionally, several studies have reported that the likelihood of vision loss is significantly higher in patients with more severe DR at the time of diagnosis.^5,51,52 Beyond the clinical measure of VA, several studies have also reported on the significantly greater vision related-functional burden in patients with severe NPDR and PDR.53, 54, 55

Prior literature has also suggested that patients with PDR or diabetic macular edema undergoing intravitreal anti-VEGF therapy that have lapses in care are more likely to have poor VA outcomes compared with patients who do not have a lapse in care.^12,16,56,57 Similarly, in our study, patients on any treatment, which included anti-VEGF therapy, PRP, focal laser therapy, and intravitreal or intraocular steroid therapy, had greater odds of vision loss if they experienced a lapse in care. Additionally, some studies have reported worse outcomes in patients with PDR treated with anti-VEGF after lapses in care compared with those treated with PRP.^9,19,58 While we stratified the PDR group by treatment type—anti-VEGF only, PRP only, or any treatment—and found an association between lapses in care and vision impairment across all 3 groups, we did not perform direct comparisons between PRP and anti-VEGF groups. This will be further investigated in future studies.

We also identified increased odds for vision impairment even in patients who had NPDR. The risk of vision loss among those with NPDR is low. Although 16% of patients with mild NPDR will progress to proliferative stages within 4 years, the risk of vision loss in this group is significantly lower than in those with more advanced stages of retinopathy.^5,8 However, systemic risk factors such as poor glycemic control and longer duration of diabetes are associated with increased risk of progression to clinically significant changes even in this low risk group.59, 60, 61 It is unclear why we might see vision impairment at the end of the 2-year period in this low risk group. Patients who do not have vision changes could be less likely to return for care, and the patients who do return, even in this low-risk population, could have other ocular conditions that are driving their symptoms, such as cataracts. Because this was a retrospective study, we are unable to attain the precise etiology of vision impairment. Previous studies by Obeid et al^12,13 reported that for patients receiving treatment for PDR or neovascular age related macular degeneration, a better baseline VA was associated with significantly higher odds of not returning for care within 12 months of their last treatment procedure.

Most published risk factors of lapses in care have modest effect sizes, with marginal ORs ranging from 1.04 to 1.46.^24,62 Common risk factors include age, race and ethnicity, insurance, comorbidity burden (Charlson Comorbidity Index), and other ocular comorbidities, such as glaucoma, as well as social risk factors such as income, housing or food insecurity, and mental health, also influence access to eye care.⁶³ Major predictors of nonrefractive vision impairment include older age (OR, 1.05), low income (OR, 2.23), lower educational attainment (OR, 2.11), and longer duration of diabetes (≥10 years; OR, 2.67),^64,65 with additional risk factors in diabetic populations including worse baseline vision, sex, race, smoking, and greater DR severity at diagnosis.⁵ In this context, our E-value analysis supports the robustness of our observed association between lapses in DR care and vision impairment. The point estimate E-value of 1.75 indicates that an unmeasured confounder would need to be associated with both a lapse in care and vision impairment by an OR of ≥1.75 each, beyond our measured covariates, to fully explain away our observed effect.⁶⁶ The presence of a confounder is possible but not likely given the comprehensive adjustment for many sociodemographic, ocular, and medical covariates in our model. Nonetheless, the moderate size of the E-values highlights the need for cautious causal interpretation, as residual confounding is not implausible in an observational study.

There are several limitations to this study. Because this was a single-center study performed at a tertiary academic institution, we do not know the extent to which these findings are generalizable. Replication of the study in larger multicenter networks will be important to confirm its applicability across varied practice settings and patient populations. Furthermore, we had relatively few patients with vision-threatening DR, including PDR that can be better studied in multicenter studies. The relatively few patients in some of these DR severity categories, particularly patients undergoing treatment, further limits the generalizability of our subgroup analyses, which is only exploratory. There are also limitations of our lapse in care definition. This definition is consistent with our prior use of the variable²⁴ and aimed at future interventions to easily identify lapses in care. Future studies could examine the strength of our findings using composite definitions of lapses in care. Additionally, given that this was a retrospective study, we also do not know the reason why patients might experience a lapse in care, whether it was influenced by clinic characteristics (eg, challenges in making a follow-up appointment) or patient characteristics (eg, the role of social determinants of health). We also do not know the precise etiology of the vision loss experienced by patients (eg, whether the patient experienced a progression of cataracts that could be fixed by cataract surgery). This study was also limited by our inability to track patients' hemoglobin A1c values over the 2-year observation period, which is associated with vision outcomes in these patients. These unmeasured confounders could be contributing to both lapses in care as well as worsening VA, thus potentially overestimating the direct association of a lapse in care on VA. However, based on our E-value analysis, we suspect that it is unlikely for such an unmeasured confounder to be a major issue. Additionally, given the retrospective nature of this study, we can only report an association with lapses in care with vision impairment and cannot definitively report any causality. However, we attempted to estimate the association of a lapse in care on vision impairment using inverse probability weighting modeling to mitigate effects from baseline factors that may have influenced our outcome.

In conclusion, in this retrospective cohort study, we found that lapses in care among patients with T2D are associated with vision impairment at the end of a 2-year observation period regardless of the underlying DR severity or treatment status. This finding underscores the importance of diligent and repeated eye care for this population.

Manuscript no. XOPS-D-25-00476.