The Association of Vertically Integrated Delivery Networks on Lapses in Diabetic Retinopathy Care and Presenting Diabetic Retinopathy Severity

Introduction

Diabetic retinopathy (DR) is one of the leading causes of preventable blindness in the adult population, affecting approximately 22% of patients with diabetes.¹ Early detection of DR through timely eye exams is critical to preventing irreversible vision impairment.² In fact, current guidelines from the American Diabetes Association (ADA) and International Council of Ophthalmology (ICO) recommend prompt, early screening after a diabetes diagnosis, followed by regular annual appointments with more frequent monitoring for those with higher risk or active disease.^3,4 Despite this, adherence to ophthalmic follow-up remains a major challenge with nearly 75% of patients having a lapse in DR care and not returning as recommended.^5–7 Missed or delayed ophthalmic care have been shown to be major drivers of poor vision outcomes.^8–11 Therefore, optimizing adherence to ophthalmic follow-up among patients with diabetes is critical.

Higher rates of non-adherence in diabetic patients exemplifies the challenge of managing chronic disease. In addition to maintaining glycemic control, diabetes management depends on regular surveillance of multi-system complications.¹² Fragmented care between primary care providers (PCPs) and specialists likely contributes to lapses in follow-up and worse clinical outcomes. This has led to a growing interest in care integration, which aims to coordinate services across specialties.

Existing literature suggests that integrated PCP and specialty models are associated with increased patient and provider satisfaction, reduced costs, and better diabetes-related outcomes.^13–17 Enhanced communication between PCPs and ophthalmologists may further improve adherence to DR care.¹⁸ However, current research on diabetes care coordination has notable gaps. First, integration efforts have largely centered around primary care and endocrinology, with limited focus on the impact of integrated ophthalmology services. Second, prior studies have emphasized screening uptake rather than long-term adherence and follow-up. Third, most have examined the effects of co-location or integration in community-based settings, such as patient-centered medical homes, rather than within academic health systems. Furthermore, the role of geographic distance to clinic has not been thoroughly explored as a factor that may modify integration benefits.

Our study aims to address these gaps by examining whether having a PCP and ophthalmologist within the same vertically integrated academic health system is associated with reduced lapses in DR care and early detection of DR complications. By focusing on lapses in care, we capture adherence across a longitudinal time scale that extends beyond single-point screenings. Investigating care coordination in an academic vertically integrated context while accounting for distance allows us to understand the benefits of integrated care beyond a community-based setting. Our findings have the potential to identify patients at higher risk for lapses in care and may inform strategies to support adherence and improve outcomes in DR management.

Methods

Study Design and Population

We conducted a retrospective cohort study of adult patients (over 18 years) with diabetes mellitus seen at the Wilmer Eye Institute at Johns Hopkins Hospital between April 4th, 2013 to September 7th, 2023. Patients were included if they had completed ≥ 1 office visits for the screening or treatment of DR at Wilmer.⁷ To ensure all patients had a full two-year follow up, we excluded patients whose initial encounters were after September 7th, 2021, and those whose providers recommended follow-up timeframe was missing or >52 weeks (Supplemental Figure 1). Patients were also excluded for missing sociodemographic, visual acuity, or primary care provider data. Only data from the initial office encounter for each patient was included.

Primary Exposure: PCP Location

We defined a patient’s PCP as the individual serving as the primary point of contact for routine medical care, including physicians, nurse practitioners, and physician assistants. Patients were classified as having an internal PCP if that provider was also located within the Johns Hopkins Health System Corporation,¹⁹ based on structured provider location data extracted from the Johns Hopkins Health data warehouse (Precision Medicine Analytics Platform (PMAP)).²⁰ External PCPs were defined as practicing outside the Johns Hopkins Hospital System Corporation. To validate the data extraction process, provider location was manually verified through chart review of randomly selected patient records (N=42, reviewed by first author S.Y).

Primary Outcome: Lapse in Care

Lapses in care were identified using a previously described methodology.⁷ In brief, office visits relevant to the treatment or screening of DR were first identified. Then, providers’ recommended follow-up timeline was extracted from structured electronic health record fields or unstructured clinical notes. The recommended interval was compared with the actual time to the next visit. A lapse was defined as returning later than the recommended interval, exceeding a pre-set threshold (around 25%) based on literature and expert recommendation. For example, an 8-week recommended follow-up would have a threshold of 2 weeks, and a 20-week follow-up would have a threshold of 4 weeks. Lapses in care were identified in a two-year observation period beginning after the initial visit.

Secondary Outcomes

Secondary outcomes included vision impairment (visual acuity (VA) worse than or equal to 20/40), blindness (VA worse than or equal to 20/200), and VA in logMAR,²¹ all measured in the better-seeing eye at the initial office encounter. Severity of DR was also included and categorized as no DR, unspecified DR, non-proliferative DR (NPDR), and proliferative DR (PDR) as determined at the baseline visit by natural language processing (NLP).^22,23 A sensitivity analysis was performed using the worse-seeing eye.

Covariates

Additional covariates were extracted from the baseline visit and included sex, age, race (Hispanic, non-Hispanic Black, non-Hispanic White, and other), insurance (private, Medicare, Medicaid, other, none) and distance from patient’s home to their ophthalmology clinic.⁷ Distance to clinic was calculated using the GeoPy²⁴ package in Python; patient addresses and ophthalmology clinics were geocoded to a latitude and longitude based on PMAP data from 2018. We categorized patients as living within or beyond the median distance – 7.6 miles – of their clinic. Clinical covariates included the Charlson Comorbidity Index (CCI), Diabetes Complications Severity Index (DCSI), presence of glaucoma, and other retinal comorbidities as previously described.^25,26

Statistical Analysis

All analyses were performed using Stata, version 18 (StataCorp) and Python version 3.8.5 (Python Software Foundation). Descriptive statistics summarized demographic and clinical characteristics of the study population, stratified by PCP location. Continuous variables (age, visual acuity, CCI, DCSI) were compared between internal and external PCP groups using 2-sample t-tests, and categorical or binary variables were compared using Pearson’s chi-squared tests. Multivariate logistic or linear regression models were used to assess the association between PCP type with an interaction term of PCP type and distance from clinic and each primary or secondary outcome, adjusting for age, race, insurance, comorbidities, and ocular disease.

We assessed the interaction between primary care provider (PCP) type and distance to clinic using a likelihood ratio (LR) test that compared a full model including the interaction term with a reduced model without it. Statistical significance was set at α of 0.05. When the LR test suggested a potential interaction, we examined pairwise comparisons between PCP type (internal vs external) and distance category (≤7.6 vs >7.6 miles) using post-estimation contrasts with Bonferroni correction for multiple comparisons. To visualize the combined effects, predicted probabilities for each DR outcome category—no DR, unspecified DR, NPDR, and PDR—were derived from the fitted model using marginal estimates.

Before fitting the final model, we tested the proportional odds assumption with the omodel test, which indicated a violation (P <0.001). Therefore, we used a generalized ordered logistic regression (gologit2) model. Both autofit and partially constrained versions were evaluated; the constrained model, which held the gender effect constant across outcome thresholds, provided the best overall fit (LR P =0.804 for gender; all other variables P <0.05 for the unconstrained model).

Results

A total of 41,581 patients were included in this study (Table 1). The mean age was 61.8 years, with a roughly even distribution of male (48%, N=19,750) and female (53%, N=21,831) participants. Overall, 48% (N=20,030) of patients were non-Hispanic White, 37% (N=15,209) were non-Hispanic Black, 4% (N=1697) were Hispanic, and 11% (N=4645) identified as other. The majority of patients, 81% (N=33,494), had at least one lapse in care during the two year follow-up period.

Table 1 Baseline Characteristics of Patients Included in the Study Stratified by Whether They Have an Internal Primary Care Provider (PCP), Within the Johns Hopkins Health System, or External PCP

Of all patients, 48% (N=19,998) had an internal primary care provider (PCP) and 52% (N=21,583) had an external PCP. Patients with an external PCP were, on average, older (mean age 63 vs. 60 years, P <0.001) (Table 1). Non-Hispanic White patients were more likely to have an external PCP (60% vs. 40%), whereas non-Hispanic Black and Hispanic patients were more likely to have an internal PCP (56% and 58%, respectively; all P <0.001).

For the primary outcome of lapses in care, the interaction between PCP type and distance to clinic was statistically significant (likelihood ratio test, P =0.002), indicating that the relationship between PCP type and lapses differed by distance. Among patients living ≤7.6 miles from the clinic, predicted probabilities of lapse were nearly identical for those with external versus internal PCPs (0.79 vs.80; P = 1.00) (Figure 1). In contrast, among patients living >7.6 miles away, those with an external PCP had a higher predicted probability of lapse compared with those with an internal PCP (0.83 vs.80; P <0.001). In pairwise comparisons of patients living >7.6 miles from clinic, having an external PCP was associated with 19% higher odds of a lapse in care compared with those with an internal PCP (OR 1.19, 95% CI [1.08, 1.32], P <0.001).

Figure 1 The predicted probability for having lapses in diabetic retinopathy care are shown for patients with internal compared to external primary care providers (PCPs), and stratified by the distance the patient lives from the Johns Hopkins health system. The odds ratios (OR), standard error (SE), and P-value for each comparison from the pairwise comparisons are shown below.

For the secondary outcome of vision impairment, the interaction between PCP type and distance was not statistically significant (P =0.65). However, within each distance category, patients with external PCPs had higher predicted probabilities of vision impairment than those with internal PCPs—0.16 vs 0.11 (P <0.001) for ≤7.6 miles and 0.18 vs 0.13 (P <0.001) for >7.6 miles (Figure 2). On pairwise comparisons, having an external PCP was associated with 60% higher odds of vision impairment for patients living within ≤7.6 miles of clinic (OR 1.60, 95% CI [1.43,1.80], P <0.001) and 56% higher odds at >7.6 miles (OR 1.56, 95% CI [1.40, 1.74], P <0.001).

Figure 2 The predicted probability for having vision impairment (visual acuity worse than or equal to 20/40) at initial presentation for patients with internal compared to external primary care providers (PCPs), and stratified by the distance the patient lives from the Johns Hopkins health system. The odds ratios (OR), standard error (SE), and P-value for each comparison from the pairwise comparisons are shown below.

Findings were similar for the outcome of blindness. The interaction between PCP type and distance was not significant (P =0.64). However, there was a significantly higher predicted probability of blindness for patients comparing patients with external versus internal PCP at both distances: ≤7.6 miles: 0.02 vs 0.01, P <0.001; >7.6 miles: 0.03 vs 0.01, P <0.001 (Figure 3). Pairwise comparisons corresponded to an 113% (OR 2.13, 95% CI [1.47, 3.08], P <0.001) and 131% (OR 2.31, 95% CI [1.75, 3.05], P <0.001) increased odds for those with external versus internal PCP for those living closer and further, respectively.

Figure 3 The predicted probability for having blindness (visual acuity worse than or equal to 20/200) at initial presentation for patients with internal compared to external primary care providers (PCPs), and stratified by the distance the patient lives from the Johns Hopkins health system. The odds ratios (OR), standard error (SE), and P-value for each comparison from the pairwise comparisons are shown below.

For baseline visual acuity (logMAR), the interaction between PCP type and distance was significant (P =0.02). At both distances (≤7.6 miles or >7.6 miles), patients with external PCPs had higher predicted baseline visual acuity in logMAR (worse vision) than those with internal PCPs (≤7.6 miles: 0.14 vs 0.10, P <0.001; >7.6: 0.17 vs 0.12, P <0.001) (Figure 4). In pairwise comparisons, all contrasts were statistically significant after Bonferroni adjustment (P <0.001), though absolute differences were small.

Figure 4 The predicted probability for visual acuity (in logMAR) at initial presentation for patients with internal compared to external primary care providers (PCPs), and stratified by the distance the patient lives from the Johns Hopkins health system. The odds ratios (OR), standard error (SE), and P-value for each comparison from the pairwise comparisons are shown below.

Lastly, for baseline DR severity, the overall interaction term was not statistically significant (P =0.62). Nevertheless, within each distance category, PCP type was associated with higher severity (P <0.001) (Figure 5). Among patients living ≤7.6 miles, those with external PCPs had a lower predicted probability of having no DR (0.74 vs 0.85) and a higher probability of having PDR (0.08 vs 0.03) compared to those with internal PCPs (Figure 5A–D). A similar but more pronounced pattern was observed among patients living >7.6 miles away, where the predicted probability of no DR was 0.68 for external PCP versus 0.81 for internal PCP, and the probability of PDR was 0.11 versus 0.05, respectively (Figure 5A–D). In pairwise comparisons, for patients living ≤7.6 miles, the external PCP group had 144% increased odds of presenting with any DR, versus no DR, compared to the internal PCP group (OR 2.44, 95% CI [2.19, 2.73], P <0.001) (Figure 5A). For patients living >7.6 miles, the external PCP group had 138% increased odds of presenting with any DR, versus no DR, compared to the internal PCP group (OR 2.38, 95% CI [2.15, 2.63], P <0.001) (Figure 5A). This trend persisted across all levels of DR severity, with those with external PCPs showing significantly higher odds of more severe disease compared to those with internal PCPs at both distance categories (P <0.001 for all comparisons) (Figure 5B and C).

Figure 5 The predicted probability of no diabetic retinopathy (DR) (Panel A), unspecified DR (Panel B), non-proliferative diabetic retinopathy (NPDR) (Panel C), and proliferative diabetic retinopathy (PDR) (Panel D). The bottom panel depicts the odds ratios (OR), standard error (SE), and P-value from the pairwise comparison for DR severity comparing chances of having unspecified DR, NPDR, or PDR vs no DR (Panel A); NPDR or PDR vs Unspecified DR or No DR (Panel B); PDR vs No DR, Unspecified DR, or NPDR (Panel C).

Results were similar when using the visual acuity of the worse seeing eye at baseline on sensitivity analysis (Supplemental Table 1).

Discussion

In this retrospective cohort study of adults with diabetes at a vertically integrated delivery network, we found that patients whose PCP was outside of the hospital system generally had worse DR care than those with in-system PCPs. Patients with external PCPs had more lapses in DR care, but only when patients lived further away. And, patients with external PCPs were more likely to have vision impairment or blindness and greater DR severity at the initial visit. The effect of external PCP on presenting disease severity was more pronounced among patients living farther away. Our findings suggest that patients with external PCPs, especially those living farther away, may benefit from additional support to reduce disease progression.

Our findings suggest that staying within a vertically integrated system is associated with better follow-up adherence, particularly for patients who live further away. This aligns with prior studies in other subspecialties showing that having a PCP located within the same hospital as specialty care is associated with higher appointment attendance.^27,28 Several mechanisms may underlie this association. First, shared electronic health records and referral systems may streamline communication between PCPs and ophthalmologists, which is linked to better follow up adherence,¹⁸ and enhance interdisciplinary collaboration.^29,30 Second, vertical integration may ease the logistical burden on patients; familiarity with the office site, provider,³¹ and the ability to schedule ophthalmology appointments consecutively with PCP visits may increase the likelihood of adherence. Third, vertically integrated systems may enhance patient and provider education regarding the importance of ophthalmic care. Within an integrated network, PCPs and ophthalmologists can coordinate on clinical guidelines and updates, ensuring the ophthalmology is presented as an essential part of routine diabetes care.

Distance to the hospital also appeared to modify the relationship between PCP type and lapses in care, and to a lesser degree, disease severity. Distance mitigated the association between external PCP and lapses in care–such that patients living closer to the hospital had similar amounts of lapses in care regardless of PCP type. However, distance did not attenuate differences in presenting ophthalmic disease severity. Among both distance categories, patients with an external PCP presented with worse vision and more severe DR. This pattern suggests that lapses in care may be more sensitive to logistical and access-related factors, whereas disease severity could be more driven by longer-term systemic differences in care coordination and referral pathways that geographic proximity alone cannot overcome.

Our findings have several implications for the management of DR care. We show that patients who have external PCPs and live farther from the hospital appear to be at higher risk for lapses in care, and that patients with external PCPs present with worse disease, regardless of distance. Therefore, it may be beneficial for ophthalmologists to encourage patients to remain within the same medical system for their primary care services. Alternatively, for patients living far away with well-established external PCPs, targeted adherence support–such as appointment reminders, transportation assistance, or coordinated scheduling–could help mitigate lapses in care.^32–35 From the PCP perspective, attention to the location of ophthalmology referrals may be important. For instance, patients with external PCPs may have achieved better disease control if their ophthalmology care was more local or better coordinated with their local provider. While emerging artificial intelligence-based screening tools may improve access to DR screenings, patients who screen positive will still need in-person ophthalmic evaluation, underscoring the continued importance of our findings.

Finally, our study had several strengths and limitations. Strengths included a large sample size, the ability to assess lapse in care–providing a more accurate assessment of follow-up adherence–and adjustment for multiple relevant covariates. Limitations include potential misclassification of PCP location, residual referral bias, single-institution design, and retrospective nature, all of which limit causal inference and the generalizability of the findings to other health systems. We also simplified patient distance to clinic as a binary variable; future studies could consider more nuanced measures. Additional work should also examine the association of vertical integration with follow-up adherence and disease severity in other diseases and specialties, as well as evaluate interventions to improve care for patients receiving care across multiple healthcare systems. Studies monitoring other variables, such as lab parameters, glycemic control, and treatment adherence, across DR follow-ups would also help characterize high-risk groups and inform subsequent interventions.

Conclusion

In conclusion, our study demonstrates that having an internal PCP is associated with improved DR care compared to external PCPs within a vertically integrated hospital. Among patients living farther away from the hospital, having an external PCP is associated with higher odds of a lapse in DR care. Having an external PCP is associated with more severe baseline ophthalmic disease compared to patients with internal PCPs. Ophthalmologists and PCPs should consider the potential beneficial role of vertically integrated care delivery services when referring patients with DR, as they may reduce lapses in care and mitigate disease severity.