The Advent of Big Data and Machine Learning in Ophthalmology: The 2026 Cogan Lecture

Abstract

The major objective of this lecture is to summarize the evolution of computational ophthalmology over the past 2 decades through the scientific advances that have transformed ophthalmic research. Topics include the emergence of computational ophthalmology, the role of big data in scientific discovery, the application of machine learning to ophthalmic imaging, and the development of artificial intelligence (AI)-ready data ecosystems. The lecture concludes by exploring the next frontier in computational ophthalmology, highlighting foundation models and agentic AI as emerging approaches with the potential to accelerate scientific discovery, and improve clinical care are discussed. This article is based on the 2026 Cogan Award Lecture presented by Aaron Y. Lee, MD, MSCI, and Cecilia S. Lee, MD, MS, at ARVO 2026.

Over the past 2 decades, ophthalmology has undergone a remarkable transformation driven by electronic health records (EHRs), large-scale clinical registries, advances in imaging, and artificial intelligence (AI). Together, these developments have fundamentally changed how ophthalmic research is performed, allowing investigators to study disease across millions of patients.

Ophthalmology has been uniquely positioned to benefit from this evolution because routine clinical care generates large volumes of quantitative measurements, standardized imaging, and longitudinal data. Combined with advances in computational methods, these resources have transformed the eye into a powerful model for studying not only ocular disease but also systemic health and aging.¹

This review summarizes the evolution of the field through several major themes, including the emergence of computational ophthalmology, the role of big data in scientific discovery, the application of machine learning to ophthalmic imaging, and the development of AI-ready data ecosystems. Finally, the next frontier in computational ophthalmology is examined highlighting foundation models and agentic AI as emerging approaches that may further accelerate scientific discovery and improve clinical care.

The Emergence of Computational Ophthalmology

Before EHRs large-scale clinical databases and AI clinical research advanced one specific question at a time. Investigators manually reviewed paper charts and assembled carefully constructed cohorts, making thoughtful study design and close clinical observation essential. Although these studies were modest in size, every variable collected and every imaging feature analyzed reflected a deliberate scientific question.

The transition from paper charts to EHRs fundamentally changed the clinical research landscape. Rather than spending months identifying patients, retrieving physical charts, and manually extracting clinical information, investigators could begin studying longitudinal data generated through routine clinical care. The challenge shifted from collecting data to organizing, integrating, and interpreting increasingly large and complex datasets, allowing researchers to examine entire health systems and follow patients over time.

In the early 2010s, substantial advances in Big Data research emerged in the United Kingdom through the integration of EHR through the National Health Service (NHS), enabling access to large, longitudinal clinical datasets. Investigators at Moorfields Eye Hospital using regional ophthalmic registries built from NHS EHRs, demonstrated how routinely collected clinical data could be transformed into continuously growing research resources. These registries became comprehensive clinical datasets that supported longitudinal analyses of treatment outcomes, disease progression, surgical complications, and comparative effectiveness in real-world practice. They served as the backbone for studies demonstrating only transient visual benefit from switching anti-VEGF agents in chronic neovascular age-related macular degeneration, an increased risk of treatment-requiring diabetic macular edema during the first year after cataract surgery, particularly among patients with more advanced diabetic retinopathy, and an association between socioeconomic deprivation and later presentation with more severe diabetic eye disease.^2–4

As more of clinical ophthalmology became digitized, a new field of computational ophthalmology emerged as an interdisciplinary field of research at the intersection of computer science, machine learning, informatics, and clinical data. In addition to applying computational tools to ophthalmic data, computational ophthalmology integrates clinical, imaging, molecular, and population-level data from national clinical registries and longitudinal cohorts to enable quantitative, data-driven discovery that addresses questions that cannot be answered using conventional approaches alone.

Utilizing Big Data for Scientific Discovery

The establishment of the Intelligent Research in Sight (IRIS) Registry by the American Academy of Ophthalmology (AAO) further accelerated the growth of computational ophthalmology by making national-scale ophthalmic data available for research.^5,6 Many clinically important questions that had previously required large multicenter collaborations became increasingly tractable.⁷ Studies using the IRIS Registry demonstrated that immediately sequential bilateral cataract surgery is not associated with a higher risk of postoperative endophthalmitis nor significantly worse refractive outcomes than delayed surgery, identified sociodemographic disparities in the diagnosis and surgical management of pediatric strabismus, and revealed racial and ethnic differences in the surgical management of Fuchs endothelial corneal dystrophy.^8–11

Big data also broadened the role of ophthalmology beyond the study of eye disease alone. Longitudinal multi-disciplinary population-based cohorts enabled investigation of the relationship between ophthalmic health and cognitive aging. The finding that cataract extraction is associated with reduced risk of subsequent dementia, together with the establishment of prospective retinal imaging cohorts such as the Adult Changes in Thought (ACT) study, reinforced the concept of the eye as an accessible window into aging and neurodegenerative disease.^12–14

The advent of big data also transformed the interpretation of evidence. As datasets grew larger, statistical significance alone became less informative.¹⁵ Rather than relying primarily on P values, greater emphasis became placed on visualizing raw data, allowing longitudinal trajectories, distributions, and population-level patterns to reveal relationships that were more likely to be clinically and biologically meaningful.

Applying Machine Learning to Ophthalmic Imaging

For many decades, ophthalmic imaging remained primarily qualitative, serving as visual records interpreted by clinicians rather than quantitative sources of biological information. However, as ophthalmic care transitioned to digital workflows, imaging modalities such as optical coherence tomography (OCT), color fundus photography (CFP), and visual field testing became routinely digitized and increasingly centralized, creating enormous imaging datasets that would soon become amenable to computational analysis. Much like how EHRs transformed clinical documentation into research data, machine learning transformed these image-based data into a computational resource, enabling quantitative analysis at an unprecedented scale.

One major application of machine learning is the integration of multimodal imaging and clinical data for ophthalmic disease detection. A multimodal deep learning model combining OCT, CFP, and demographic and clinical variables accurately detected glaucoma while identifying both established risk factors, such as age, intraocular pressure, and optic disc morphology, and previously underappreciated features, such as pulmonary function, associated with disease.¹⁶ These findings illustrated that machine learning cannot only achieve accurate disease prediction, but also generate interpretable insights into disease biology through the integration of multimodal imaging and clinical information.

Another application of machine learning is the automated quantification of metrics routinely evaluated during clinical examinations. Computer vision algorithms have been developed to quantify eyelid position, estimate intraocular pressure so sensitive that ocular pulse amplitude can be visualized from Goldmann applanation tonometry recordings, and characterize eye movements from routine clinical videos, achieving performance comparable to that of expert clinicians.^17,18 Together, these studies demonstrated that machine learning can transform routine clinical examinations into objective, reproducible, and scalable computational biomarkers. By converting subjective clinical observations into standardized quantitative measurements, these approaches have the potential to improve diagnostic precision, longitudinal disease monitoring, and large-scale clinical research.

Beyond disease detection and automated measurement, machine learning has also enabled direct modeling of the relationship between ocular structure and visual function. Traditionally, structural imaging and functional testing were interpreted separately, requiring clinicians to mentally integrate OCT findings with visual field testing or microperimetry. Deep learning enabled the development of models that predict retinal sensitivity directly from OCT in macular telangiectasis type 2.¹⁹ By learning the relationship between retinal microstructure and pointwise visual function, these models generated functional maps directly from volumetric OCT data. Similar approaches were subsequently applied to glaucoma, where multimodal deep learning models predicted pointwise visual field sensitivity directly from OCT and infrared reflectance optic disc imaging, providing new insights into the structure-function relationship between retinal nerve fiber layer anatomy and visual field loss.²⁰ Collectively, these studies demonstrated that machine learning can bridge structural imaging and functional assessment by quantitatively modeling their underlying biological relationships, revealing patterns that were previously difficult to measure and leading to new insights into disease mechanisms.

Building AI-Ready Data Ecosystems

As computational ophthalmology matured, another important lesson emerged. The performance of AI and machine learning models depended not only on increasingly sophisticated algorithms, but also on the quality and diversity of the data used to develop them. Early studies established the promise of deep learning, but they also highlighted the limitations of single-center, disease-specific datasets. Progress would increasingly depend on building larger, more diverse, and multimodal datasets capable of supporting trustworthy and generalizable AI.

As part of the National Institutes of Health Bridge2AI program, Artificial Intelligence Ready and Exploratory Atlas for Diabetes Insights (AI-READI) was created to address this need.²¹ Rather than creating another disease-specific imaging dataset, AI-READI was conceived as a multimodal resource integrating retinal imaging with complementary clinical, physiologic, molecular, behavioral, and environmental data to study both pathogenesis and salutogenesis in type 2 diabetes mellitus (T2DM).²¹ The cohort was intentionally balanced across key demographic characteristics and the full spectrum of T2DM severity, ranging from normal, prediabetes, medication-controlled, and insulin-treated T2DM, thereby reducing the inherent selection bias that exist in registry- or EHR-based datasets. Beyond enabling future scientific discovery, AI-READI was designed to facilitate benchmarking, harmonization, and standardization across imaging platforms and computational methods. The overarching goal was to establish a reusable scientific resource that could support a broad range of future research questions and computational methods.

Equally important, AI-READI emphasized that representative data require representative participants. Recruitment strategies prioritized demographic diversity, broad representation across disease severity, and inclusion of communities historically under-represented in biomedical research.²² Community engagement was embedded throughout the lifecycle of the study through community advisory boards, participant partnerships, and transparent communication, underscoring that trustworthy AI depends as much on public trust as on technical innovation.

Responsible data sharing became another central component of this effort. As multimodal datasets become increasingly rich, protecting participant privacy while enabling scientific discovery requires robust governance rather than a single technical solution. AI-READI implemented a layered governance framework that combined controlled-access mechanisms, data-use agreements, and technical safeguards to support responsible open science.²³ Such frameworks recognize that high-quality datasets should function as shared scientific infrastructure supporting broad collaboration while maintaining the trust of the participants who make such research possible.

The development of AI-READI underscores that progress in AI depends on far more than algorithmic innovation. Thoughtful study design, representative recruitment, standardized data collection, ethical governance, global data sharing, and open collaboration are foundational components of trustworthy biomedical AI. These principles echo those that guided the earliest era of clinical investigation, despite the dramatic increase in the scale and complexity of biomedical datasets.

The Next Frontier in Computational Ophthalmology

The development of AI-ready data ecosystems has established the foundation for the next phase of computational ophthalmology. With increasingly large, diverse, and multimodal datasets now available, attention is shifting toward computational systems that can use these powerful resources to accelerate scientific discovery and improve clinical care (Fig.).

Figure.

Evolution of clinical research in ophthalmology from the era of paper charts, through the digital era and now in the era of AI enabled clinical care.

One of the most promising research developments in this new phase is the emergence of foundation models. Rather than developing algorithms for individual clinical tasks, researchers are increasingly building foundation models that learn generalizable representations from large, multimodal datasets and can be adapted across a wide range of downstream applications.

OCTCube represents one example of this emerging approach using multimodal retinal imaging.²⁴ OCT provides volumetric three-dimensional information that captures retinal structure in remarkable detail. OCTCube was designed to leverage this structure while jointly learning from complementary imaging modalities, including infrared reflectance and fundus autofluorescence. Rather than optimizing performance for a single prediction task, the goal was to learn reusable representations that could support diverse clinical and research applications. Beyond improving predictive performance, foundation models have the potential to accelerate biomarker discovery, improve clinical trial design, and enable more efficient development of future AI systems.

Another promising research direction lies in transforming how biomedical research itself is conducted. Modern clinical research increasingly depends on computational expertise for cohort construction, database queries, statistical analysis, and reproducible workflows. At the same time, concerns about the reproducibility of published research and the growing divide between increasingly sophisticated data science methods and clinical domain expertise have created new barriers to translating important clinical questions into reproducible computational analyses.

These challenges motivated the development of LLM-Assisted Testing of Clinical Hypotheses (LATCH).²⁵ LATCH uses a multi-agent framework to translate natural language clinical questions into reproducible analytical workflows, assisting investigators with database querying, cohort construction, and statistical analysis while preserving human oversight for study design, interpretation, and scientific judgment. Importantly, LATCH operates on database schemas rather than patient-level data, preserving privacy while enabling automated analytical workflows. The system also generates transparent analytical reports that document cohort construction, variable mapping, and statistical code, making each analysis readily reproducible. In validation studies, LATCH successfully reproduced hundreds of published clinical hypotheses within minutes, illustrating its potential to dramatically accelerate biomedical research. By automating many of the technical aspects of data analysis, systems such as LATCH have the potential to improve reproducibility, lower barriers to clinical research, and allow investigators to focus more directly on clinically meaningful questions.

Together, foundation models and agentic AI represent complementary directions for the future of computational ophthalmology. Foundation models seek to learn broadly useful biological representations from increasingly rich biomedical data, whereas agentic systems seek to make scientific investigation itself more efficient, transparent, and reproducible. Rather than replacing human expertise, these technologies are expected to have their great impact by augmenting clinicians and scientists, helping them integrate increasingly complex data, generate new hypotheses, and translate discovery into improved patient care.