At the 2024 American Glaucoma Society annual meeting symposium “AI in Glaucoma: What’s Real, What’s Possible, and What’s Science Fiction,” 6 panelists explored the current status of artificial intelligence (AI) for glaucoma diagnosis and management as well as visions for its future use.
AI Is Perfect for Glaucoma
“Whether we choose to acknowledge it or not, we are in the midst of a glaucoma crisis,” said Dr. Benjamin Xu in his talk “AI Implementation: More Than Meets the Eye.” The prevalence of glaucoma is rising, and the dwindling supply of ophthalmologists is not poised to keep up with future demands.
Additionally, glaucoma is challengingly complex, a common theme among panelists. There is no single test that can pinpoint the disease or predict its progression, diagnosis is subjective, monitoring methods are inherently “noisy,” and there is little to no ability to personalize treatment. However, as Dr. Nazlee Zebardast explored in her talk “Oh, the Places You’ll Go (With AI)!”, that is where AI comes in. The great value of AI lies in its ability to process large amounts of data, and so its potential ability to help in a data-driven field like glaucoma is huge.
Glaucoma-Specific Challenges for AI Exist
“AI is perfect for glaucoma,” agreed Dr. Siamek Yousefi. In his talk “Glaucoma Research: The Next Generation,” Dr. Yousefi walked the audience through a brief evolutionary history of glaucoma, noting rapid advancement over the last 50 years and an exponential increase in studies about AI and glaucoma over the last decade. However, the use of AI in glaucoma remains behind the curve relative to what is being seen in many other areas of medicine, including other areas of ophthalmology.
In his talk “Glaucoma and AI: The World of Tomorrow, Today,” Dr. Filipe Medeiros discussed some reasons for this lag in applying AI to glaucoma diagnosis and management. Citing AI and diabetic retinopathy as a hallmark example of the value of AI in screening, Dr. Medeiros discussed how the retina is easily imaged for patterns, but “it is NOT as easy to assess an optic nerve [for glaucoma].” He described one example in which researchers attempting to regrade a set of nearly 500,000 retinal images to detect referable glaucoma were met with a 72% false positive rate.
In addition to imaging challenges, the many different types of information used in glaucoma diagnosis and monitoring make it especially tricky to integrate with AI. Panelists largely agreed that the recent emergence of generative AI (rather than the application of AI only as a classification tool) along with large-language and foundational models represent a turning point. As these models become better at incorporating multiple types of data, they will be better able to simulate a real ophthalmologist’s diagnostic processes.
AI’s Future May Be Closer Than We Think
Dr. Yousefi described a small study (9 cases) in which ChatGPT-4 was able to combine clinical, ocular imaging, and visual field factors to diagnose different glaucoma phenotypes better than senior ophthalmology residents. Dr. Robert Chang provided other examples in which AI shows ophthalmic potential in his talk “Should We Fear or Embrace Our Future Machine Overlords?” He discussed how generative AI models are performing comparably to or better than humans on the OKAP and USMLE exams, showing good reasoning and rapid improvement, even on open-ended questions; 61% of the time, graders couldn’t tell the difference between an AI-generated answer and a real ophthalmologist’s response. Further, in another study, patients showed a preference for AI- over human-generated answers, specifically rating the AI responses as being more empathetic.
“In terms of coexistence with AI, we are still a bit on the fence as to whether it is friend or foe,” said Dr. Xu. However, panelists agreed that AI has strong potential to help ease clinician burden and to improve patient care. For instance, AI trained to replicate expert assessment of OCT images could elevate general ophthalmologists to a level more on a par with specialists, said Dr. Medeiros, helping to alleviate clinician burden. In addition, Dr. Zebardast suggested that better risk calculation and individualized treatment plans could result if machine learning systems are able to incorporate genetic information with imaging and lifestyle data.
“There are some low-hanging fruits where AI performs really well,” said Dr. Medeiros, citing a few examples where AI could be implemented right away. One instance was AI assessment of OCT image quality to detect artifacts or correct segmentation errors. Another option, added Dr. Chang, is virtual scribing, an application that is currently being researched at Stanford University.
Several Barriers to AI Implementation Must Be Addressed
- Lack of a standard, objective definition of glaucoma
AI works best with objective classification. However, there is no objective, consensus definition for glaucoma. Perhaps agreeing on a working definition would be a good place to start, proposed Dr. Xu.
- Need for reliable ophthalmic datasets
Reliable AI requires reliable datasets, but “big data might mean big biases” cautioned Dr. Medeiros. Many patients are being treated based on their baseline risk factors, which decreases their risk of progression and confounds the resulting data. In addition, all AI systems have blind spots and the potential to manufacture factual errors (hallucinations). Improving dataset quality will help minimize these issues.
- Difficulty integrating multiple types of data
Simultaneously evaluating multiple kinds of data requires all data to be written in the same language. “[Standardization] is better than it was when we started,” acknowledged Dr. Michael Boland in his talk “AI Ethics: Do Androids Dream of Electric Eyeballs?” However, more buy-in is necessary from technology vendors to comply with data standards.
- Need for more rigorous validation of AI tools
AI screening tools need to be validated in a real-world setting and in much more diverse populations. This will ensure fairness and generalizability and improve positive predictive value, especially within specific group parameters.
- Lack of sufficient governance and regulations for AI models
Dr. Boland stressed that for implementation to be successful, governance is essential. “If we are going to depend on these systems to serve patients, we have to make sure they are sustainable and do their job.” He described some regulatory frameworks that are already in place, including new WHO ethics guidelines for the use of AI in medicine and a comprehensive model pipeline from Mass General Brigham for successfully implementing and sustaining an AI system. Dr. Xu also pointed out the crucial need for security regulations to protect patient privacy—one hack could be devastating with so much patient data in one place.
It Is Time to Embrace AI
In the Q&A after the presentations, panelists were asked for their perspectives on the single biggest issue facing AI and glaucoma moving forward. For Dr. Chang, it was that “we’re paralyzed by this definition of glaucoma.” Dr. Medeiros agreed, postulating whether focusing on amassing data and then trying to define things may offer a better approach. Others felt that data standards and governance protocols should be a top priority. “We have a duty to do better for patients,” said Dr. Boland.
Despite these challenges, all panelists felt that AI is poised to be a strong factor for change in the diagnosis and management of glaucoma, but AI needs to become more available for it to be used and so it can improve. As Dr. Chang said, “It is time to embrace AI.”