Social Determinants of Health, Vision, and Treatment in Exudative Age-Related Macular Degeneration Authors: Timothy M Janetos, Adovich S Rivera, Yang Li, Chris A Andrews, Paul J Bryar, Manjot K Gill, Jeremy A Lavine, Joshua D Stein, Dustin D French, the SOURCE Consortium Published: 2026 ABSTRACT AbstractPurposeTreatment deviations in exudative age-related macular degeneration (exAMD) are common and lead to worse visual outcomes; reasons remain incompletely understood. We sought to identify distinct treatment patterns in newly diagnosed, treatment-naïve exAMD and to evaluate associations with social determinants of health (SDH) and baseline vision.DesignMulticenter retrospective cohort study.ParticipantsA total of 15 664 newly diagnosed, treatment-naïve exAMD patients receiving at least one intravitreal injection between January 1, 2014 and January 1, 2024 across 17 academic medical centers in the Sight Outcomes Research Collaborative ophthalmology big data repository.MethodsTreatment patterns over the first 12 months were grouped using sequence clustering. Mixed-effects logistic regression evaluated associations between cluster membership and sex, age, race and ethnicity, primary language, urbanicity, primary insurance, distressed communities index, Charlson comorbidity index, and baseline logarithm of the minimum angle of resolution visual acuity.Main Outcome MeasuresSequence-derived treatment cluster membership and its association with SDH and baseline vision.ResultsTwo clusters were identified: “frequent treatment” (67.4%) and “nonpersistent” (32.6%); the latter characterized by a few initial injections followed by early treatment cessation. Nonpersistence was best differentiated by ≥ 7 months without an office visit or treatment (Youden index, 0.83; area under the curve, 0.93). In the adjusted model, factors associated with nonpersistent cluster membership included Black race (adjusted odds ratio [aOR], 1.42; 95% confidence interval [CI], 1.11–1.80), Hispanic ethnicity (aOR, 1.38; 95% CI, 1.00–1.90), older age at index date (aOR, 1.08 per decade; 95% CI, 1.03–1.13), rural residence (aOR, 1.32; 95% CI, 1.11–1.58), increasing distressed communities index (aOR, 1.03; 95% CI, 1.01–1.05), Medicaid insurance (aOR, 1.62; 95% CI, 1.03–2.52), worse baseline logarithm of the minimum angle of resolution acuity (aOR, 1.71 per unit; 95% CI, 1.54–1.91), and lower Charlson comorbidity index (aOR, 0.95 per point; 95% CI, 0.93–0.97).ConclusionsDemographics, SDH, and vision were associated with nonpersistence in newly diagnosed exAMD. Social determinants of health‑informed identification of high-risk individuals may guide qualitative work on modifiable barriers and intervention development to prevent irreversible vision loss.Financial Disclosure(s)Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.Keywords: Exudative AMD, Treatment adherence, Social determinants of health, Health disparities, Intravitreal injection Age-related macular degeneration (AMD) remains a leading cause of vision loss for those aged 65 years and more.1 Global prevalence is predicted to rise to 288 million affected by 2040.2 Within the United States, AMD affects 13.6% of the population aged 40 years or more.3 Approximately 10% to 15% of patients with AMD will develop exudative age-related macular degeneration (exAMD) disease, which contributes to the majority of severe irreversible vision loss.4 Untreated, exAMD will lead to legal blindness in approximately 90% of patients.4,5 The introduction of intravitreal pharmacologic agents targeting VEGF, and more recently angiopoietin-2, has revolutionized management of the disease and can effectively restore or stabilize vision. Treatment, however, is burdensome, often requiring frequent intravitreal therapy, especially in the first year after diagnosis.6 Previous research has demonstrated high rates of treatment nonadherence among patients undergoing intravitreal therapy for exAMD, leading to worse visual outcomes.7,8 Reasons for nonadherence are likely multifactorial, complex, and span across patient, provider, and systemic barriers.9 Social determinants of health (SDH), defined as the conditions in which people are born, learn, work, play, worship, and age, are well-known drivers of health outcomes including treatment adherence and health care access.10 Evaluation of SDH associations with treatment in exAMD is understudied, and identification of potential drivers of treatment disparities is important to identify high-risk individuals and to inform areas for intervention. Importantly, factors leading to nonadherence may differ depending on the quantity of nonadherence, which is an umbrella term encompassing deviation from provider-recommended treatment.11 Prior studies examining treatment adherence in exAMD have used variable definitions, and there is no widely accepted threshold for what determines clinically significant adherence.9 This is important as patients who miss one treatment versus those who discontinue or have large gaps in care are likely different groups, with ultimately different visual outcomes. An expert consensus for exAMD treatment adherence published in 2021 defined nonadherence as missing ≥2 treatment or monitoring visits over 12 months, where a missed appointment was defined as a visit exceeding the recommended date by >2 weeks; nonpersistence was defined by no appointment in the last 6 months.12 This consensus definition has not yet been validated using clinical data. Therefore, the aim of this study was twofold. First, to describe the treatment patterns of exAMD across a large cohort of newly diagnosed patients to identify valid clusters representative of potentially appropriate and inappropriate treatment to better define an operational outcome using a data-driven approach. Second, to explore the association between identified treatment clusters, SDH, and baseline vision to identify potential treatment disparities and future research areas for intervention development. Methods Study Design and Data Source This retrospective cohort study uses data from the Sight Outcomes Research Collaborative (SOURCE) Ophthalmology Big Data Repository (https://www.sourcecollaborative.org/).13 Sight Outcomes Research Collaborative captures structured electronic health record (EHR) data from all patients receiving eye care at participating academic medical centers from the time each site went live on the EHR to the present. This study used data from 17 SOURCE sites, each contributing up to 10 years of data. Sight Outcomes Research Collaborative sites use the same data extraction methods to ensure data harmonization with data tables including demographics, eye examination and testing data, outpatient orders/results, medications, –9/10 diagnosis, Current Procedural Terminology (CPT) 4 codes, charges, and surgeries. Data are updated periodically and include both ocular and nonocular care data. Sight Outcomes Research Collaborative data are fully deidentified before analysis using Datavant software (Datavant). Additionally, SOURCE is linked to other data sets that provide SDH data, including community of residence, among others. Northwestern University Institutional Review Board designated this study as exempt, and informed consent was not obtained because SOURCE uses only deidentified data. This study adheres to the Declaration of Helsinki, and the Strengthening the Reporting of Observational Studies in Epidemiology reporting guidelines were followed. Inclusion Criteria The cohort included newly diagnosed, treatment-naïve exAMD patients defined by first CPT-4 code of an intravitreal injection (67028) in a patient with an International Classification of Diseases (ICD) 9/10 diagnosis code of exAMD (362.52, H35.32xx). There was a 3-year “lookback” period for both ICD-9/10 diagnosis and CPT code to include only first-time or treatment-naive injection patients. The index date was defined as the first intravitreal injection, and patient data were obtained 12 months after index date, with complete site data obtained 12 months post index date for each participant. Exclusion criteria included a concurrent diagnosis of diabetic macular edema or cystoid macular edema during the follow-up period of 12 months post index date. Additionally, SOURCE sites were dropped if there were <100 eligible patients at the site and patients were dropped if they lacked all basic sociodemographic variables or if they died during the follow-up period. The index date inclusion ran from January 1, 2014 to January 1, 2024, with the end of last possible follow-up of January 1, 2025. Analysis was performed at the patient level with bilateral patient data obtained from the first exudative eye or a random eye from patients treated bilaterally at the index date. Sequence Analysis To analyze treatment patterns among patients during the first year of treatment, a sequence analysis was performed.14 This method was chosen as it allows for multiple unique sequences across patients, and we, therefore, did not assume a priori an operational definition of treatment adherence given the heterogeneity of treatment patterns often dependent on temporal treatment paradigms, provider treatment strategy, therapeutic response, and drug of choice. Sequence analysis maintains this heterogeneity in treatment patterns, which cannot be adequately apprehended by prespecified operational definitions. Sequences were created for each patient by assigning each month a mutually exclusive event during the 12-month follow-up. Months were categorized as having (1) an injection (I), (2) an office visit without an injection (V), or (3) neither an office visit nor an injection (N). Next, a cluster analysis was conducted to group similar patient sequences based on order and duration. Optimal matching was used to calculate a dissimilarity matrix among all sequences. Partitioning around medoids method was used to determine cluster grouping. Two-, 3-, and 4-cluster solutions were explored, and the final cluster solution was determined based on cluster validity metrics (average silhouette width, Calinski-Harabasz index), cluster stability (Jaccard similarity index), and visual inspection of representative clusters for clinical validity and usefulness.14 All patient sequences are included in each solution evaluated. Patient demographics by cluster membership were presented as descriptive statistics and compared using t tests or χ2 tests. Given that individual patient cluster assignment is descriptive of the cohort and may shift with addition of new patients, particularly for treatment sequences near the boundary between solutions, we next evaluated whether simple, clinically interpretable operational definitions of treatment adherence could classify sequence clusters to provide a more stable, reproducible classification for a future outcome or prediction task. Metrics evaluated included maximum time gap between injections (or 12 months if only 1 injection), total number of injections, total duration of existing in the “N” state (no injection or office visit), and “N” state at month 4. For each metric, receiver operating characteristic curves were produced to assess discrimination, and the optimal threshold was determined using Youden’s index if relevant. Next, patient cluster membership was regressed against demographics (age at index date, sex, race, and ethnicity), SDH (distressed communities index [DCI], urbanicity of residence, primary insurance, primary language), and clinical covariables (Charlson comorbidity index [CCI] score, baseline best-reported visual acuity). Baseline visual acuity was abstracted at index date. Race and ethnicity are described as a single categorical variable, with Latino or Hispanic individuals grouped regardless of reported race and all remaining categories as non-Hispanic. Charlson comorbidity index is a composite index ranging from 0 to 19 based on comorbidities and age, originally developed to predict risk of 10-year mortality, with higher scores indicating a higher comorbidity burden.15 Distressed communities index ranges from 0 to 100 (low to high distress) and is a composite index of socioeconomic factors based on zip code to measure a community’s overall economic distress. Percent missingness was assessed for each covariable. Multiple unadjusted bivariate mixed-effects logistic regressions with cluster membership as the primary outcome against each covariable, followed by a fully adjusted multivariable mixed effects logistic regression model including all covariables were performed. A random intercept for SOURCE site was included. The multivariable model was assessed for covariable collinearity via variance inflation factors. All models were complete case analyses. All analyses were performed using R version 4.5.1 (R Foundation for Statistical Computing). Results After all exclusion criteria, the cohort consisted of 15 664 patients (Fig 1). A 2-cluster solution was chosen based on cluster validity, stability, and clinical relevance (Supplemental Table 1 and Supplemental Figure 1, available at www.ophthalmologyscience.org). This grouped patient treatment sequences into 2 clearly defined groups as visualized in Figure 2. Cluster 1 (10 556 patients) was defined as “frequent treatment” and comprised sequences typically defined by a series of initial monthly injections, followed by an increased interval between injections. Cluster 2 (5108 patients) was defined as “nonpersistent”, which was largely comprised of sequences demarcated by a few initial injections or visits without future treatment throughout the 12 months. Clusters are visually represented by all treatment patterns and the 10 most frequent representative patterns in Figure 2. Descriptive statistics by cluster membership are shown in Table 1. As expected for this population, overall, the cohort consisted of generally older (average age at index date, 79.0; standard deviation, 10.8), White (82.1%), primary Medicare insurance (70.3%), and primary English-speaking (94.8%). Covariable missingness was <10% for each covariable, except for visual acuity (missingness, 11.3%). Patient inclusion flowchart. AMD = age-related macular degeneration. Cluster membership by frequent treatment cluster (left) and nonpersistent cluster (right). Green = injection; blue = office visit without injection; yellow = no office visit or injection. Top figures indicate all sequences. Bottom figures show the most frequently represented sequences within each cluster, with bar height contributing more to each cluster. Table 1Patient Demographics by Treatment Cluster MembershipCharacteristicOverall (n = 15 664)Frequent Treatment (n = 10 556)Nonpersistent (n = 5108)P ValueMale sex, n (%)5949 (38.0)4006 (37.9)1943 (38.0)0.929Language, n (%)< 0.001 English14 850 (94.8)10 089 (95.6)4761 (93.2) Spanish230 (1.5)107 (1.0)123 (2.4) Other584 (3.7)360 (3.4)224 (4.4)Area, n (%)< 0.001 Urban13 830 (88.3)9446 (89.5)4384 (85.8) Rural1022 (6.5)649 (6.1)373 (7.3) Missing812 (5.2)461 (4.4)351 (6.9)Insurance, n (%)< 0.001 Commercial3103 (19.8)2237 (21.2)866 (17.0) Medicare11 016 (70.3)7362 (69.7)3654 (71.5) Medicaid157 (1.0)73 (0.7)84 (1.6) Missing1388 (8.9)884 (8.4)504 (9.9)Race and ethnicity, n (%)< 0.001 White12 857 (82.1)8854 (83.9)4003 (78.4) Black519 (3.3)305 (2.9)214 (4.2) Latino or Hispanic464 (3.0)247 (2.3)217 (4.2) Asian446 (2.8)286 (2.7)160 (3.1) Other279 (1.8)168 (1.6)111 (2.2) Missing1099 (7.0)696 (6.6)403 (7.9)DCI distress score, mean (SD)31.4 (25.9)30.5 (25.6)33.2 (26.5)< 0.001 Missing (n, %)915, 5.8%521, 4.9%394, 7.7%CCI, mean (SD)5.24 (2.47)5.32 (2.49)5.08 (2.42)< 0.001 Missing (n, %)0, 0%0, 0%0, 0%LogMAR, mean (SD)0.32 (0.39)0.30 (0.34)0.38 (0.47)< 0.001 Missing (n, %)1774, 11.3%1099, 10.4%675, 13.2%Age at index date, mean (SD), yrs79.0 (10.8)78.8 (10.5)79.3 (11.2)0.015 Missing (n, %)0, 0%0, 0%0, 0%CCI = Charlson comorbidity index; DCI = distressed communities index; logMAR = logarithm of the minimum angle of resolution; SD = standard deviation. Patient Demographics by Treatment Cluster Membership CCI = Charlson comorbidity index; DCI = distressed communities index; logMAR = logarithm of the minimum angle of resolution; SD = standard deviation. The evaluated simple operational definitions to correctly classify patients to cluster membership are shown in Table 2. Overall, total duration in the “N” state (no injection or office visit duration) had the highest discrimination with a threshold of greater than or equal to 7 months (Youden’s Index, 0.83; area under the curve, 0.93; sensitivity, 88.3%; specificity, 95.0%) to classify to nonpersistent cluster membership. Total number of injections during the first year also had a high level of discrimination with a threshold of ≤5 injections (Youden’s Index, 0.82; area under the curve, 0.90; sensitivity, 86.3%; specificity, 96.2%). Table 2Performance of Simple Binary Metrics for Classifying Nonpersistent Cluster MembershipMetricMaximum Time between InjectionsTotal Number of InjectionsTotal Duration without Injection or VisitMonth 4 without Injection or VisitOptimal threshold≥3.1 mo≤5 injections≥7 mo—Area under the curve0.640.900.930.70Youden index0.370.820.830.41Sensitivity, %52.586.388.366.9Specificity, %84.696.295.073.8 Performance of Simple Binary Metrics for Classifying Nonpersistent Cluster Membership Results of the unadjusted and adjusted analyses are shown in Table 3. The multivariable model (“adjusted odds ratio [OR]”) included all variables assessed, and all variance inflation factors were <2. In the unadjusted models, there were notable associations with nonpersistent cluster membership including age at index date (OR, 1.05 per decade; 95% confidence interval [CI], 1.02–1.09), Black race (OR, 1.52; 95% CI, 1.26–1.82), Hispanic ethnicity (OR, 1.77; 95% CI, 1.45–2.16), Asian race (OR, 1.27; 95% CI, 1.03–1.56), other race (OR, 1.35; 95% CI, 1.06–1.74), Spanish primary language (OR, 1.98; 95% CI, 1.50–2.59), other primary language (OR, 1.27; 95% CI, 1.07–1.52), rural area of residence (OR, 1.32; 95% CI, 1.14–1.52), increasing DCI (OR, 1.04 per 10-unit; 95% CI, 1.03–1.06), Medicaid primary insurance (OR, 2.17; 95% CI, 1.55–3.03), worse baseline logarithm of the minimum angle of resolution visual acuity (OR, 1.70; 95% CI, 1.55–1.86), and CCI (OR, 0.95 per unit increase; 95% CI, 0.94–0.97). In the adjusted model, nonpersistent cluster membership was significantly associated with Black race (OR, 1.42; 95% CI, 1.11–1.80), Hispanic ethnicity (OR, 1.38; 95% CI, 1.00–1.90), increasing age at index date (OR, 1.08 per decade; 95% CI, 1.03–1.13), rural area of residence (OR, 1.32; 95% CI, 1.11–1.58), increasing DCI (OR, 1.03 per 10-unit; 95% CI, 1.01–1.05), Medicaid primary insurance (OR, 1.62; 95% CI, 1.03–2.52), worse baseline logarithm of the minimum angle of resolution visual acuity (OR, 1.71; 95% CI, 1.54–1.91), and CCI (OR, 0.95 per unit increase; 95% CI, 0.93–0.97). Table 3Mixed-Effects Logistic Regression: Unadjusted and Multivariable-Adjusted Associations with Nonpersistent Cluster MembershipVariableCategoryOR95% CIP ValueaOR95% CIP ValueSexFemale1.00 (ref)——1.00 (ref)——Male0.990.92–1.060.751.000.92–1.100.92Age at index datePer decade increase1.051.02–1.09< 0.011.081.03–1.13< 0.01Race and ethnicityWhite1.00 (ref)——1.00 (ref)——Black1.521.26–1.82< 0.011.421.11–1.80< 0.01Latino or Hispanic1.771.45–2.16< 0.011.381.00–1.900.048Asian1.271.03–1.560.021.100.83–1.450.50Other1.351.06–1.740.021.290.94–1.770.12LanguageEnglish1.00 (ref)——1.00 (ref)——Spanish1.981.50–2.59< 0.011.510.94–2.430.09Other1.271.07–1.52< 0.011.060.81–1.370.69AreaUrban1.00 (ref)——1.00 (ref)——Rural1.321.14–1.52< 0.011.321.11–1.58< 0.01DCIPer 10-unit increase1.041.03–1.06< 0.011.031.01–1.05< 0.01Primary insuranceCommercial1.00 (ref)——1.00 (ref)——Medicare1.020.92–1.130.691.000.89–1.120.99Medicaid2.171.55–3.03< 0.011.621.03–2.520.04Baseline logMARPer logMAR unit1.701.55–1.86< 0.011.711.54–1.91< 0.01CCIPer unit increase0.950.94–0.97< 0.010.950.93–0.97< 0.01aOR = adjusted odds ratio; CCI = Charlson comorbidity index; CI = confidence interval; DCI = distressed communities index; logMAR = logarithm of the minimum angle of resolution; OR = odds ratio; ref = reference. Mixed-Effects Logistic Regression: Unadjusted and Multivariable-Adjusted Associations with Nonpersistent Cluster Membership aOR = adjusted odds ratio; CCI = Charlson comorbidity index; CI = confidence interval; DCI = distressed communities index; logMAR = logarithm of the minimum angle of resolution; OR = odds ratio; ref = reference. Discussion By using a data-driven sequence analysis, this study identifies a distinct group of newly diagnosed treatment-naïve exAMD patients defined primarily by treatment with a few initial intravitreal injections followed by cessation of treatment. In contrast, the alternative cluster was characterized by an initial loading phase followed by variable but consistent treatment. This group was established using sequence clustering rather than assuming a definition for appropriate or inappropriate treatment. Overall, 32.6% of patients belonged to the nonpersistent cluster. An American Academy of Ophthalmology Intelligent Research in Sight registry study of nearly 150 000 treatment-naïve patients defined nonpersistence as no follow-up within 6 months of last injection and lost to follow-up as no follow-up within 12 months with rates of 14.3%, and 11.6%, respectively.16 Similar to our study, odds of nonpersistence were higher among Black patients, older age, and primary Medicaid insurance. Use of the IBM MarketScan database of 8167 patients demonstrated a discontinuation rate closer to our study of 30.4% between 24 and 60 months after treatment initiation.17 A key contribution of our study was the evaluation of several distinct SDH covariables including urbanicity of residence, primary language, DCI, and primary insurance in addition to basic demographics, CCI, and baseline vision. Black race had increased odds of belonging to the nonpersistence cluster. We previously published on total treatment rates among exAMD patients and noted a decreased overall treatment rate among Black patients compared to White patients using the All of Us research data set.18 Similar racial disparities have been identified in time to treatment initiation.19 Distressed Communities Index is a measure of a community’s economic well-being based on multiple indicators, with an increasing DCI indicating higher community distress. Increasing DCI was associated with increased odds of belonging to the nonpersistence cluster; however, the effect was small in the multivariable model (OR, 1.03 per 10-unit increase, P = < 0.01). A survey-based factor comprised of high neighborhood disorder, high loneliness, high discrimination, and high stress levels was also associated with a decreased rate of treatment in the All of Us research study, indicating that lived environment may have structural barriers to receiving care. Rural area was also a significant factor in this study, and travel distance and reliance on a caregiver for travel have all been identified as potential barriers to receiving intravitreal therapy among patients with exAMD.7,18 In unadjusted estimates, Hispanic and Spanish primary language were also significantly associated with nonpersistence cluster membership. Similarly, in the All of Us research study, a survey-based factor comprising poor English proficiency, transportation difficulty, rural care access issues, worse health literacy, and provider concordance (e.g., how much a patient identifies with their provider) was associated with a lower treatment rate.18 Interestingly, increasing CCI was significantly associated with decreased odds of nonpersistence in both the unadjusted and adjusted models. This may be a function of integration into the health care system for chronic disease management, and these patients already may have overcome barriers to receiving care and have a higher level of medical engagement. Alternatively, older age at index date and worse baseline visual acuity were both significantly associated with increased odds of nonpersistence, both of which have been previously associated with nonadherence.7 As this group is already older with little age variation, the effect of age on group membership may be minimal. Prior qualitative work has noted procedural dissatisfaction as a potential reason for discontinuation, and poor baseline vision or end-stage disease may not realize benefits of treatment.9 Ultimately, many of these factors are highly inter-related and likely only associated with the underlying true causes of the identified health disparity. Therefore, qualitative research is necessary and needed in this realm to fully understand the unmeasurable patient barriers to care. Currently, the few qualitative studies examining adherence to intravitreal therapy among exAMD patients that were conducted among patients actively receiving treatment, do not specifically focus on those at a high risk for nonpersistence, and all were conducted outside of the United States.20, 21, 22, 23, 24 Although some barriers leading to treatment nonpersistence may be universal to all patients, future qualitative studies among individuals who meet our operational definition may help identify areas for intervention specific to groups at the highest risk and should be a future area of research. Furthermore, EHR-derived SDH, demographic, and clinical factors may be more useful as a tool for prediction to identify patients at a high risk for nonpersistence to intervene early. Metrics such as CCI and DCI can be calculated or abstracted for any patient presenting for a first-time injection with a new diagnosis of exAMD. Based on the evaluated operational definitions to classify cluster membership, a potential outcome for prediction, or future adherence studies, could be a total of ≥7 months without treatment or office visit or ≤5 injections within the first year after diagnosis, both of which demonstrated good sensitivity and specificity to differentiate the clusters. Interestingly, the former aligns particularly well with the expert consensus definition of “nonpersistence” of 6 or more months without an appointment. Therefore, future steps would be to validate these definitions externally, using outcomes aligned with provider intended treatment, missed treatments, or even functional visual outcomes to demonstrate clinical validity and usefulness. As noted, there are multiple definitions in the literature of nonadherence and nonpersistence, and our identified cluster likely more aligns with nonpersistence or discontinuation/lost to follow-up. This distinction is important, as it has implications on ultimate visual outcomes and reasons behind nonpersistence may be different than those who stay in care but perhaps have difficulty attending all treatments or visits. In this case, the frequent treatment group is most likely a mixed group comprised of a range of adherence patterns. Strengths of this study include its multicenter nature with >15 000 patients spanning across the United States and integrating more granular SDH covariables. Furthermore, we used a sequence clustering algorithm to explore treatment patterns, which allowed for the development of operational definitions to be used in future adherence studies or for prediction. Limitations include inability to assess provider-intended treatment, and therefore, nonpersistence is ultimately assumed. It is possible that this cluster also included patients who had robust response to treatment or who had treatment discontinued due to advanced stage and futility of continued treatment. We would expect these patients to be less frequent in number, but ultimately validation based on provider treatment intent is needed. Additionally, because cluster solutions are unsupervised and developed from the full cohort, individual cluster assignment may shift as additional patients are added in future analyses, particularly for patients with sequences near the cluster boundary. Therefore, the operational definition provided above provides a reproducible way to classify new patients for future studies. Furthermore, disease misclassification leading to measurement error is possible in large data sets based on ICD-9/ICD-10 codes and a single CPT code. Specific medication administration codes (e.g., J codes) are not reliably reported in SOURCE and therefore some intravitreal injections may include other medications such as corticosteroids. However, we attempted to minimize this by excluding patients with other diagnostic codes that may lead to an intravitreal injection. Additionally, although missingness was overall low, it may not have been at random and only complete case model analyses were performed, which may introduce biased estimates. Finally, patients were conditioned on survival during the 12-month follow-up, and no concurrent diagnosis of diabetic macular edema or cystoid macular edema was including, excluding 868 patients, which may lead to selection bias. Ultimately, future research should focus on using SDH-informed prediction to identify high-risk individuals while qualitative studies are needed to develop potential interventions. Conclusions In a large, multicenter cohort of newly diagnosed, treatment-naïve exAMD patients, sequence clustering identified a distinct “nonpersistent” treatment pattern affecting one-third of patients. Demographic, clinical, and SDH factors, including Black race, Hispanic ethnicity, older age, rural residence, Medicaid insurance, distressed community, and worse baseline vision were associated with nonpersistence, whereas higher CCI was associated with reduced odds. Future research should focus on SDH-informed prediction to identify high-risk individuals and on qualitative studies to develop targeted interventions to reduce treatment disparities and prevent irreversible vision loss. Manuscript no. XOPS-D-26-00576. Footnotes Supplemental material available atwww.ophthalmologyscience.org. Disclosure(s): All authors have completed and submitted the ICMJE disclosures form. The authors made the following disclosure(s): J.D.S.: All support for the present manuscript ‑ R01EY034444, R01EY032475, R01EY035016; Grants or contracts from any entity ‑ Dompe, Bausch/Lomb. D.D.F.: All support for the present manuscript ‑ NIH, RPB. J.A.L.: All support for the present manuscript ‑ Therini Bio; Consulting fees ‑ Line 6 biotechnology, Genentech. M.K.G.: All support for the present manuscript ‑ NIH, RPB; Consulting fees ‑ Genentech-Roche, Regeneron, 4DMT, Ocular Therapeutics; Participation on a Data Safety Monitoring Board or Advisory Board ‑ Kriya Therapeutics. P.J.B.: All support for the present manuscript ‑ NIH, RPB. T.M.J.: All support for the present manuscript ‑ NIH, RPB. A.S.R.: All support for the present manuscript ‑ National Eye Institute: R01EY036341 (Co-Investigator); Grants or contracts from any entity ‑ National Heart, Lung, and Blood Institute: R21HL181701 (Principal Investigator). This work was supported by the National Eye Institute (R01EY034444 “Health Differences in Utilization, Quality, and Outcomes for Three Common Ocular Conditions [HealthDOC]”); the National Eye Institute (R01EY036341 “Health Evaluation and Access to Leverage Technology for Improved Diabetic Retinopathy Outcomes [HEAL-DR]”); Research to Prevent Blindness, New York, NY; and the National Center for Advancing Translational Sciences (K12TR005104 to T.M.J.). The funding organizations had no role in the design or conduct of this research. Support for Open Access publication was provided by Northwestern University. HUMAN SUBJECTS: Human subjects were included in this study. Northwestern University Institutional Review Board designated this study as exempt and informed consent was not obtained because SOURCE uses only deidentified data. This study adheres to the Declaration of Helsinki and the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guidelines were followed. No animal subjects were used in this study. Author Contributions: Conception and design: Janetos, Rivera, Andrews, Bryar, Gill, Lavine, Stein, French Analysis and interpretation: Janetos, Rivera, Li, Andrews, Bryar, Gill, Lavine, Stein, French Data collection: Janetos, Rivera, Li, Stein, French Overall responsibility: Janetos, Rivera, Li, Andrews, Bryar, Gill, Lavine, Stein, French Obtained funding: Janetos, French Contributor Information Timothy M. Janetos, Email: timothy.janetos@northwestern.edu. the SOURCE Consortium: Tony N. Pira, Manju L. Subramanian, Sejal Amin, Paul A. Edwards, Divya Srikumaran, Fasika Woreta, Jeffrey S. Schultz, Anurag Shrivastava, Baseer Ahmad, Louis R. Pasquale, Alon Harris, Vaidehi Dedania, Paul J. Bryar, Dustin D. French, Merina Thomas, Suzann Pershing, Sophia Y. Wang, Preethi Ganapathy, Anne E. Coleman, Victoria Tseng, Michael Deiner, Catherine Sun, Jennifer Patnaik, Prem Subramanian, Saleha Munir, Wuqaas Munir, Joshua D. Stein, Lindsey De Lott, Brian L. Vanderbeek, Rithambara Ramachandran, Rajeev Ramchandran, Ore-ofe O. Adesina, Brian C. Stagg, Barbara Wirostko, Arjun Dirghangi, Brian McMillian, Arsham Sheybani, Soshian Sarrapour, and Kristen Harris-Nwanyanwu Supplementary Data 3 (top) and 4 (bottom) cluster solutions. Green = injection; blue = office visit without injection; yellow = no office visit or injection. All patient sequences are included in each cluster solution. The 3-cluster and 4-cluster solution maintains the general sequences of clusters 1 and 2, but partitions out less valid or stable sequences with difficult interpretability. Supplemental Table 1 References Associated Data Supplementary Materials Supplemental Table 1