The use of large datasets in ophthalmic research has increased exponentially in the last 10 years. Although initial analyses used claims databases as well as the American Academy of Ophthalmology (AAO) Intelligent Research in Sight (IRIS) registry, additional datasets have become available more recently, such as the National Institutes of Health All of Us registry, UK Biobank, Epic Cosmos (Epic Systems, Verona, WI), and TriNetX (TriNetX, Cambridge, MA). Some publications using these datasets are thought provoking; however, many others are unlikely to affect clinical practice. Why?
It is essential to understand the types of questions that “big data” datasets are most appropriate to answer. In our view, these applications fall into several broad categories:
-
•
Identification of novel associations . Examples include associations between medication exposure (eg, calcium channel blockers, glucagon-like peptide–1 agonists) and a disease outcome (eg, glaucoma). , It is important to note that such analyses are not conclusive on their own, given their retrospective nature; rather, they should serve as the foundation for further prospective studies to further elucidate these associations.
-
•
Clinical care patterns at the population level . Examples include analyses of adherence to AAO preferred practice patterns, testing frequency, cost analyses, and trends in diagnosis and treatment patterns , —all which can be highly valuable. Such analyses can guide future interventions to optimize screening efforts or to improve care quality at the practice level.
-
•
Real-world evidence (RWE) analyses . If an intervention is explicitly coded (eg, the use of ranibizumab or aflibercept) and the outcome is unambiguous and observable in these datasets (eg, visual acuity), “big data” can provide valuable RWE that often complement data from prospective cohort studies or randomized clinical trials. , Such analyses are particularly useful for post-marketing safety surveillance and pharmacovigilance, as randomized trials are often underpowered to detect uncommon adverse events. In addition, RWE studies have the advantage of including certain patient populations (eg, specific minorities) that may not have been well represented in prior prospective studies. ,
-
•
Evaluation of rare events . Analysis of rare events (eg, rare diagnoses, rare postoperative complications, risk factor analysis of rare reoperations) at individual institutions is typically challenging because of their low incidence. However, larger datasets provide larger sample sizes, thereby increasing the statistical power of such studies. These analyses can help clinicians assess the risk of such events or can provide guidance regarding the clinical management of such events. ,
-
•
Evaluation of health equity and structural determinants of vision outcomes. Large-scale datasets are uniquely suited to evaluate disparities in access to care, time to diagnosis, treatment initiation, and follow-up patterns across diverse populations. By linking self-reported demographic variables, insurance status, and geographic indicators to care use patterns, big data studies can identify structural barriers to optimal ophthalmic care. Findings from such studies may help inform policy interventions, targeted screening programs, and resource allocation strategies aimed at reducing preventable vision loss at the population level.
In contrast, important limitations of large registry-based datasets must be acknowledged:
-
•
Lack of granular longitudinal imaging and functional data. Most multicenter datasets do not include imaging and functional data, such as optical coherence tomography or standard automated perimetry. This limitation reflects both the absence of widely adopted ophthalmic imaging standards and the technical challenges of extracting structured imaging data across institutions. As a result, attempts to infer disease progression using proxy measures, such as changes in International Classification of Diseases ( ICD ) codes, are inherently unreliable. Coding practices are inconsistent, often influenced by billing requirements, and are not sufficiently sensitive to capture subtle or longitudinal changes in disease severity.
-
•
Residual and unmeasured confounding. Although statistical methods can adjust for measured confounders, residual and unmeasured confounding remain unavoidable in large observational datasets. Selection bias, treatment bias, and clinician-driven decision-making can further complicate interpretation. For example, a surgeon may preferentially select one procedure over another based on nuanced clinical features not captured in the structured data fields typically evaluated in big data studies. Even with advanced analytic techniques, such datasets are rarely capable of establishing causation and, at best, generate hypotheses that require prospective validation.
-
•
Reliance on billing codes for phenotype definition . Diagnostic codes in ophthalmology are frequently assigned for billing rather than precision. As a result, ICD -based phenotype definitions may lack specificity and fail to capture important clinical distinctions. For instance, differentiating confirmed glaucoma from glaucoma suspect, or distinguishing disease severity stages, is often unfeasible without imaging, functional testing, or detailed chart review. Reliance on coding alone risks substantial misclassification bias.
-
•
Informative missingness and non-random follow-up. In electronic health record datasets, testing frequency and follow-up intervals are driven by clinical judgment rather than standardized protocols. Patients perceived to be at higher risk may undergo more frequent testing, whereas others may be seen less often or may be lost to follow-up. As a result, missing data are often non-random. Such informative follow-up patterns can introduce bias that is difficult to correct with statistical techniques. A recent publication highlighted factors associated with socioeconomic data missingness in the IRIS Registry.
-
•
Incomplete capture of care across systems. Many registries represent care delivered within specific health care systems or contributing networks. However, patients may receive testing, procedures, or medications outside the captured system, potentially for systematic reasons. This issue leads to incomplete outcome ascertainment or exposure misclassification. This fragmentation can distort estimates of treatment effects and clinical outcomes.
Interpreting Big Data: Key Principles
We urge both researchers and peer reviewers to apply several guiding principles when interpreting results and preparing or reviewing manuscripts that use large-scale datasets.
Multiplicity and Spurious Associations
Stay updated, free articles. Join our Telegram channel
Full access? Get Clinical Tree
