Skip to main content
  • Augmented Intelligence: Harnessing AI for Glaucoma

    By Annie Stuart, Contributing Writer

    Download PDF

    While technology is providing more valuable health information than ever before, health care efficiency is under increasing pressure from a multitude of factors, including the inability to analyze the never-ending flood of information that threatens to engulf researchers and clinicians alike.

    Can artificial intelligence (AI) come to the rescue? “We need AI—especially given the aging population, increase in health care costs, and shortage of ophthalmologists,” said Lama A. Al-Aswad, MD, MPH, at Envision Health Tech­nologies and Visi Health Technologies in New York City.

    While some ophthalmologists worry about technology and AI taking their jobs, Dr. Al-Aswad said, “the future is an augmented intelligence where we and machines work together to increase efficien­cy, enable personalized medicine, and improve patient outcomes.”

    Benefits for Patients and Physicians

    Among its many benefits for the field of glaucoma, AI has the potential to help in screening, diagno­sis, detection of progression, and prediction, said Felipe A. Medeiros, MD, PhD, at Duke University in Durham, North Carolina.

    Screening in primary care settings. Louis R. Pasquale, MD, FARVO, considers the use of AI for glaucoma screening in a primary care setting to hold considerable potential. “This would involve acquiring and subjecting a nonmydriatic optic nerve photograph to an AI algorithm,” said Dr. Pasquale, at Mount Sinai in New York City. In retina, for instance, some clinicians are already screening for diabetic retinopathy (DR) by apply­ing the FDA-approved algorithms for the detection of DR, Dr. Pasquale noted.1 Looking ahead, he said, “It seems natural that AI-guided glaucoma screening will be an early application to the field of glaucoma.”

    Many researchers are working on developing highly accurate AI models for this application, said Dr. Medeiros. “It’s essential to have very high specificity in screening to avoid many false positives, which could overburden the health care system.”

    For instance, Dr. Medeiros and his colleagues trained an AI model to predict spectral-domain OCT measurements of retinal nerve fiber layer (RNFL) thickness using just simple fundus photo­graphs.2 “This results in a much more objective assessment of the photographs for detecting dam­age from glaucoma,” he said.

    Predicting function and progression. One area in which AI excels is in the analysis of complex imaging data, which could provide insights into a patient’s visual function, said Dr. Medeiros. This could help identify patients with the fastest rates of functional loss, Dr. Pasquale said, and it might predict who might need glaucoma surgery, for example.

    Using data collected from nearly 6,500 patients in the United Kingdom, David P. Crabb, PhD, and colleagues explored whether it was possible to use structural measures from OCTs and infrared reflectance optic disc imaging to predict what a person’s visual fields would look like.3

    “We developed a policy-driven, deep learning model to decide what inputs were best to predict visual function,” said Prof. Crabb, at City, Uni­versity of London. He added that AI models can learn by consensus without being given any prior information. “We found we could quite accurately predict what people could see, which could be particularly helpful for patients who can’t undergo visual function tests.”

    Increasing access, reducing costs. Joel S. Schuman, MD, FACS, at NYU Langone Health in New York City, also cited AI’s potential to provide patient care before actual damage occurs and to reduce the cost of care. Using AI to identify early on those patients who are most likely to progress quickly will allow health care dollars and more in­tensive care to be directed where it’s most needed, he said. “Using AI in this way helps address issues of equity, making excellent care more accessible to all,” Dr. Schuman added.

    Analyzing whole systems. In the future, AI may help take personalized medicine to the next level by looking at the interplay between the eye and the whole system—the body—in order to track disease progression, uncover how interven­tion affects disease, and aid with prognosis, said Dr. Al-Aswad.

    Previous studies hint at the veracity of this approach, she noted. For instance, Dr. Al-Aswad said, “Using deep-learning models trained on data from nearly 285,000 patients, a Google Research study astonished the world by predicting from a retinal fundus photo the patient’s gender, smoking status, age within two to three years, and systolic blood pressure within about 11 mm Hg.” It even came close to the accuracy of the Framingham Risk Score, which estimates the 10-year risk of heart attack.4

    Advancing research. AI can help the field of glaucoma not just through clinical decision-mak­ing but also through research, said Dr. Pasquale. He pointed to a study by Han et al., in which AI was used to classify a large fundus image reposi­tory to grade cup-disc ratio.5 The researchers then compared these measurements across different genetically defined ancestry groups. The results add to our understanding of the genetic archi­tecture for cup-disc ratio, which is “an important endophenotype for glaucoma,” said Dr. Pasquale. He added that applying AI to other big data resources like metabolomics and proteomics data­sets may lead to the discovery of new biomarkers for glaucoma.

    Working with colleagues in neuro-ophthalmol­ogy, Dr. Pasquale and colleagues have also learned that it is possible to use AI algorithms to detect structural and functional endpoints in random­ized clinical trials, using data from previously published studies.6 Given the accuracy of their machine learning approach, AI could reduce the need for expert graders, the researchers wrote. AI also could “supplant the need for designated read­ing centers in future glaucoma neuroprotection randomized clinical trials,” Dr. Pasquale said. This would reduce costs and allow for more resources to be diverted to other investigations, he noted.

    Saving vision. What if an algorithm was able to identify glaucoma patients who need drastic IOP lowering at the outset of clinical care? “A clinical trial that randomizes such patients to an aggressive IOP target—perhaps, an IOP lowering of 30% from untreated IOP—and ultimately shows that [such an] aggressive intervention saved vision would be impressive,” Dr. Pasquale said.

    Images of optic rims from study of RimNet AI system.
    RIM MEASUREMENT. Researchers built an automated system known as RimNet for the purposes of direct rim identification in glaucomatous eyes and optic disc damage grading. Adapted from Rasheed HA et al. Ophthalmol Sci. 2023;3(1):10024.
    Optic nerve with severe glaucomatous damage.
    ADVANCED. Striations of the lamina cribrosa in an optic nerve with severe glaucomatous damage.
    Findings of AI assessment of shape patterns of optic nerve and peripapillary RNFL.
    ASSESSING SHAPE. In one study, Dr. Pasquale and colleagues used AI to assess shape patterns of the optic nerve and peripapillary RNFL and correlate those findings with VF loss. Adapted from Saini C et al. Ophthalmol Sci. 2022;2(3):100161.
    An acquired optic disc pit with no evident rim tissue.
    MONITORING. An acquired optic disc pit located in the inferior temporal region. Note the absence of rim tissue.

    Special AI Challenges in Glaucoma

    Despite many potential applications, the use of AI in glaucoma poses certain challenges, and much work is needed to bring its full potential to fruition.

    No gold standard. “We are not ready for arti­ficial intelligence in glaucoma, partly because we have not created consensus on the definition of glaucoma,” said Dr. Al-Aswad. “For DR, we have a classic definition because we can easily diagnose it by looking at the fundus. But for glaucoma, there is no pathognomonic diagnostic test or symptoms. Instead, diagnosis is based on a constellation of symptoms, diagnostic tests, and longitudinal data.” Further complicating matters is the heterogeneity of the disease among different populations.

    “When ophthalmologists look at fundus photographs, they often can’t confirm whether or not the patient has glaucoma,” said Dr. Medeiros, noting that this can lead to overestimation or underestimation of disease. He added, “Most AI models are developed through a process called su­pervised learning, which requires a labeled dataset to diagnose the disease or detect change over time. In other words, training a deep learning model to detect glaucoma from photographs or identify progression from a series of tests is only as good as the gold standard used to train it.” If such a gold standard is faulty, the deep learning model will also be, he said.

    Working to create these standards is exceeding­ly important, agreed Dr. Schuman, and identify­ing and reaching consensus on features of early glaucoma remains problematic. “However, when looking at the thickness of the RNFL and corre­sponding visual field loss, we don’t have such a hard time agreeing on what constitutes manifest glaucoma. Therefore, I think we could agree on the utility of AI algorithms for this population.”

    A slow-moving disease. Another challenge with glaucoma is that it is usually a slow-moving, disease, making it difficult to observe over a short period of time, said Prof. Crabb. However, he said, if researchers can improve the measurements taken in trials using AI techniques, “I think that could make a big difference by speeding up clinical trials and potentially delivering evidence on new treatments.” In fact, Prof. Crabb and his fellow researchers found that an AI model could identify patients at high risk of progression, thus reducing the number of participants or the length of the study needed to meet clinical trial endpoints.7

    Data availability and quality. “We don’t yet have large datasets that are diverse enough to cover all populations with glaucoma,” said Dr. Al-Aswad. “That can create health disparities at scale.”

    Once you develop a model, added Dr. Medeiros, it must be tested and validated externally on a separate population—one that is representative of the target population and doesn’t introduce racial bias, for example.

    In addition, data sharing is a key issue due to concerns about patient privacy, said Dr. Al-Aswad (see “Ethical and Legal Considerations”). Along with other researchers, she is working on ways to share the algorithm instead. Under this approach, known as federated AI, “A lead institution creates an algorithm, which they share with participating institutions,” she said. “The algorithm can learn from each institution and come back to the lead institution for further adjustment, perpetuating until the most robust model is developed.”

    Data transparency and interoperability. Standardizing image acquisition across different platforms remains a significant challenge, said Dr.

    Pasquale. Acquiring the raw information from the images also is very difficult and time consuming, as there is no Digital Imaging and Communications in Medicine (DICOM) standard for reporting images in ophthalmology, said Dr. Al-Aswad.

    “In cardiology, an echocardiogram and all accompanying data, including volume, is transmitted in the report,” she said. In contrast, Dr. Al-Aswad noted, “in ophthalmology, you need a special license to access the raw data from OCTs or visual fields.” Going forward, it will be important for researchers to push for DICOM standard reporting to make any AI process much easier for training and validation, she said.

    Workflow integration. Another challenge relates to how AI models will be used within routine clinical practice, said Dr. Medeiros. “Can AI help us with the interpretation of images right in our clinics? Potentially, yes. But there has been very little work done on creating clinical decision support systems that would allow these models to actually be integrated into practice and the EHR.” If that doesn’t happen, he said, clinicians may find AI too cumbersome to use and will prefer the simpler tools that are more readily available.

    Moreover, clinicians may need algorithms tailored to the specific populations they serve, said Dr. Pasquale. “We need to allow practitioners to fine-tune algorithms they may use in their offices,” he said, explaining that methods that allow for code-free machine learning and automated machine learning would make this possible.

    What’s Needed to Move Forward

    It will take some time, perhaps five years or more, before we start seeing real results with AI in the field of glaucoma, Dr. Medeiros said. He added that data collaboration, clear image standards, and clinical support systems are essential.

    Defining glaucoma. Members of an FDA initiative called the Collaborative Community on Ophthalmic Imaging (CCOI) have published a number of papers relevant to glaucoma. In a recent paper, Drs. Al-Aswad, Medeiros, Schuman, and other participants in the CCOI glaucoma workgroup delved into the need for objective reference standards for glaucomatous optic neuropathy and its progression.8 Dr. Medeiros emphasized, “We definitely want to develop AI models for screening that target people with moderate or advanced stages of disease. We don’t really need to go after people with very early damage because it becomes too hard to distinguish them from those who simply have some normal variation in the appearance of the optic nerve. Then the specificity becomes very low.”

    Integrating data. In addition, Dr. Medeiros said, much work remains if comprehensive AI models are to be developed. Such models would integrate all available information about a patient—not only results of glaucoma testing but also details on their coexisting systemic disorders, their medications, and any relevant socioeconom­ic conditions. “This would allow us to create rec­ommendation systems to provide guidance about the best treatment option for a specific patient, given everything that is known,” he said.

    For instance, Dr. Medeiros said, a systemic collection of data might reveal that it’s best to pro­ceed with surgery instead of eyedrops if a patient’s overall health profile indicates that it’s unlikely that the person will be able to adhere to an eye­drop regimen. Moreover, the merger of environ­mental exposure data and genetic information with imaging data will make for even more robust AI models of disease, said Dr. Pasquale.

    Training. “We need to think more broadly and train the next generation” in the field of AI, said Dr. Al-Aswad. Efforts to do so are underway at multiple ophthalmology programs across the United States, including at NYU Langone, where Dr. Al-Aswad launched an innovations fellowship. And the NIH is providing significant support and guidance to researchers via its Bridge2AI program, which aims to expand the use of AI in biomedical and behavioral research—and will develop ethical practices for data generation and use.9


    1 The DR algorithms approved at time of press are IDx-DR (Digital Diagnostics), EyeArt (Eyenuk), and AEYE-DS (AEYE Health).

    2 Medeiros FA et al. Ophthalmology. 2019;126(4):513-521.

    3 Kihara Y et al. Ophthalmology. 2022;129(7):781-791.

    4 Poplin R et al. Nat Biomed Eng. 2018;2(3):158-164.

    5 Han X et al. Am J Hum Genet. 2021;108(7):1204-1216.

    6 Doshi H et al. Ophthalmology. 2022;129(8):903-911.

    7 Chen A et al. Am J Ophthalmol. 2022;243:118-124.

    8 Medeiros FA et al. Ophthalmol Glaucoma. Published online Jan. 30, 2023.

    9 Accessed March 9, 2023.

    Meet the Experts

    Lama A. Al-Aswad, MD, MPH Chief executive officer of Envi­sion Health Technologies and Visi Health Technologies, both in New York, N.Y. Relevant financial disclo­sures: AI Optics: C; Envision Health Technolo­gies: E; Visi Health Technologies: E.

    David P. Crabb, PhD Professor of statistics and vision research in the department of optometry and visual sciences at City, University of London in the United Kingdom. Relevant financial disclosures: None.

    Felipe A. Medeiros, MD, PhD Professor of ophthalmology and of biostatistics and bioinformatics, and vice chair for technology at Duke University School of Medi­cine in Durham, N.C. Relevant financial disclo­sures: Google: S.

    Louis R. Pasquale, MD, FARVO Director of the New York Eye and Ear Infirmary of Mount Sinai, pro­fessor of ophthalmology, and site chair of the department of oph­thalmology in New York, N.Y. Relevant financial disclosures: Glaucoma Foundation: S; NEI: S; Research to Prevent Blindness: S.

    Joel S. Schuman, MD, FACS Pro­fessor of ophthalmology at NYU Langone Health, NYU Grossman School of Medicine, in New York, N.Y. Relevant financial disclosures: None.

    Full Financial Disclosures

    Dr. Al-Aswad AI Optics: C; Alcon: C; Bausch + Lomb Americas: C; En­vision Healthcare Technologies: E; GlobeChek: O; Mother Cabrini Health Foundation: S; Save Vision Foundation: C,S; Topcon: C,S; Virtual Field: C; Visi Health Technolo­gies: E; WorldCare Clinical: C.

    Prof. Crabb AbbVie/Allergan: L,S; Apellis: C,S; Bayer: L; Glaukos: L; Roche: S; Santen: C,L,S; Théa Pharmaceuticals: L.

    Dr. Medeiros Aerie: C; AbbVie/Allergan: C; Annexon: C; Apple: L; Bio­gen: C; Galimedix: C; Google: S,P; Heidelberg Engineer­ing: S; Novartis: C; Reichert: C,S; Stealth Bio Therapeutics: C; Stuart Therapeutics: C; Zeiss: C.

    Dr. Pasquale Bioscience: C; Character Bio: C; Eyenovia: C; Glaucoma Foundation: S; NEI: S; Research to Prevent Blindness: S; Skye Twenty Twenty: C.

    Dr. Schuman Aerie: C,US; Alcon: C; Boehringer Ingelheim: C; Bright­Focus Foundation: S; NEI: S; New York University School of Medicine: P; Ocugenix: C,SO,PS,P; Ocular Therapeutix: C,US; Opticien: C,SO,PS; Perfuse: S; Regeneron: C; Slack: C,P; Tufts University School of Medicine: P; University of Pittsburgh Medical Center: P; Zeiss: C,P.

    Disclosure Category



    Consultant/Advisor C Consultant fee, paid advisory boards, or fees for attending a meeting.
    Employee E Hired to work for compensation or received a W2 from a company.
    Employee, executive role EE Hired to work in an executive role for compensation or received a W2 from a company.
    Owner of company EO Ownership or controlling interest in a company, other than stock.
    Independent contractor I Contracted work, including contracted research.
    Lecture fees/Speakers bureau L Lecture fees or honoraria, travel fees or reimbursements when speaking at the invitation of a commercial company.
    Patents/Royalty P Beneficiary of patents and/or royalties for intellectual property.
    Equity/Stock/Stock options holder, private corporation PS Equity ownership, stock and/or stock options in privately owned firms, excluding mutual funds.
    Grant support S Grant support or other financial support from all sources, including research support from government agencies (e.g., NIH), foundations, device manufacturers, and\or pharmaceutical companies. Research funding should be disclosed by the principal or named investigator even if your institution receives the grant and manages the funds.
    Stock options, public or private corporation SO Stock options in a public or private company.
    Equity/Stock holder, public corporation US Equity ownership or stock in publicly traded firms, excluding mutual funds (listed on the stock exchange).