Summary

Glaucoma doesn't just affect your eyes, it rewires the way your entire body manages stress, inflammation, and cellular strain. That means supporting your long-term eye health calls for more than drops and pressure checks.

Doctor, I Have a Question: How is AI Changing Glaucoma Care?

 

Corresponding authors:

Alon Harris, MS, PhD, FARVO

Professor and Vice Chair Dept. of Ophthalmology

Professor of Artificial Intelligence and Human Health

Interim Director of the Barry Family Center for Ophthalmic Artificial Intelligence & Human Health at Mount Sinai Hospital

Icahn School of Medicine at Mount Sinai

Department of Ophthalmology

New York, NY

 

 

and

 

Gal Jacob Cohen, MD

Artificial Intelligence Fellow

New York Eye and Ear Infirmary of

Mount Sinai

 

Glaucoma is often called the “silent thief of sight” because it frequently remains undetected until significant and irreversible vision loss has occurred. While traditional tools like eye pressure checks and visual field tests remain the bedrock of care, researchers are developing Artificial Intelligence (AI) and Machine Learning (ML) tools to make these assessments more objective, personalized, and accessible.

Below are common questions regarding how these emerging technologies are being studied for future use in the clinic.

1. I am a “glaucoma suspect.” Can AI help determine if I actually have the disease?

SHORT ANSWER: This is a major focus of current AI research. In a study of over 1,000 “glaucoma suspect” eyes followed for nearly six years, an AI model was able to identify which patients were most likely to develop the disease using fundus photographs. Researchers have used AI to find how small decreases in the eye’s nerve layer thickness may increase the risk for fully converting to the disease.”

LONG ANSWER: This is one of the areas where AI is already delivering meaningful results. The term “glaucoma suspect” usually means your eye pressure is higher than average, your optic nerve is borderline, or you have a family history of the disease, but disease has not been confirmed yet. Deciding whether to initiate treatment in these patients is one of the more challenging clinical judgments in ophthalmology, and AI may help in two key ways.

First, it can extract more information from the scans and photographs your doctor already takes. An AI model tested on over 1,000 glaucoma suspect eyes, followed for an average of nearly six years, was able to predict which patients would eventually convert to confirmed glaucoma by analysing serial fundus photographs. Crucially, it did this by detecting small, consistent losses in the retinal nerve fiber layer, the thin sheet of nerve tissue that connects your retinal cells to your optic nerve, that were too subtle for clinicians to recognise on visual inspection alone.

Second, AI can process multiple sources of information simultaneously in a way that is difficult for human clinicians to do consistently. A study using ChatGPT with visual field printouts and optic nerve measurements achieved 96% sensitivity for detecting confirmed glaucoma, and its negative predictive value, meaning how reliably a “no glaucoma” result was actually correct, was 99.3%. In practical terms, if such a tool says a suspect is unlikely to have glaucoma, it is almost always right about that.

What AI cannot do is replace the clinical examination and your doctor still needs to see you and confirm the diagnosis. AI provides additional analysis for the clinical decision, not a substitute for it.

 

2. Can AI tell if my glaucoma is getting worse before anyone notices a change?

SHORT ANSWER: AI is being developed to detect disease progression by finding subtle patterns in your data that might be missed by standard analysis. One study used a form of AI called unsupervised machine learning to identify 18 distinct patterns of vision loss; one specific pattern was found to be a strong predictor that a patient might experience more rapid vision loss in the future.

LONG ANSWER: This is where AI is making some of its most important contributions, because detecting early progression is genuinely hard with standard clinical tools. This conventional method requires multiple tests over several years before a trend becomes statistically reliable. By that point in time, meaningful damage may have already occurred.

AI approaches this differently. Rather than reducing your visual field to one number, an AI technique called deep archetypal analysis was applied to over 2,000 visual fields from patients in a long-running clinical trial, without being given any prior knowledge about glaucoma. It identified 18 distinct spatial patterns of vision loss based purely on the shape and location of the defects. Three were new patterns, and one specific pattern, labelled P15, turned out to be the single strongest predictor of future rapid vision loss, independent of eye pressure or any other clinical measurements. A patient whose current visual field carries this pattern would probably be monitored more closely and maybe treated more aggressively. Ultimately, this will be a decision to be made by the attending glaucoma specialist.

Beyond single-test patterns, AI can also track structural changes in your optic nerve over time from serial photographs. In one study, an AI model detected progressive nerve fiber thinning on fundus photographs, in patients whose disc appearance looked unchanged to examining ophthalmologists, that correlated strongly with actual OCT measurements taken at the same visits. In other words, the algorithm found progression that the doctor’s eye missed.

 

3. Is it true that ChatGPT can now diagnose glaucoma?

SHORT ANSWER: In a recent study by Huang et al. (2026), an advanced version of ChatGPT was able to diagnose glaucoma with 96% sensitivity when provided with visual field and OCT data. While these results match the accuracy of specialists, this was a research demonstration; in a real clinic, a doctor must still weigh these results alongside the clinical examination and the patient’s medical history.

LONG ANSWER: In a research setting, with some limitations, the results were more impressive than most clinicians would have expected. A study published in 2026 tested an advanced version of ChatGPT on diverse patient groups, and the AI was given only the de-identified visual field and structural nerve fiber layer data. No images, photographs or access to the patient’s history.

With the ground truth set as a reference diagnosis made by two glaucoma specialists who reviewed the complete medical record, ChatGPT achieved 96% sensitivity and 83.7% specificity. Its negative predictive value was 99.3%. To translate that last figure: if ChatGPT said a patient did not have glaucoma, it was correct 99.3% of the time.

That said, two caveats are important. First, this was a research study with carefully prepared, de-identified data, not a live clinical workflow. Second, even the researchers noted that ChatGPT’s underlying knowledge base is unknown, meaning it is not always clear what it has and has not been trained on. A novel method like this would need formal validation, regulatory review, and integration into a clinical workflow before it could be used as anything other than a research demonstration.

 

4. Does AI work the same for everyone, regardless of their ethnic background?

SHORT ANSWER: That is definitely the goal and the unmet need and there is a lot of work that needs to be done in this regard. Not all AI methodologies have been validated and proven reliable yet. Ensuring “health equity” is a priority. AI models are trained on datasets which sometimes don’t include representation from all racial groups. The good news is that new research has shown that certain AI models can perform equitably across different ancestries (including African, Asian, and European), which makes it reasonale to expect that these tools will eventually benefit all patient populations and ethnic groups.

LONG ANSWER: Not reliably, yet, and this is one of the most important unsolved problems in using all medical AI. Many of the large AI training datasets used in ophthalmology have been built from populations that skew heavily toward high-income countries and toward patients of European or East Asian ancestry. When a model trained on those populations is applied to patients from other backgrounds, performance often drops. This is not a flaw in the algorithm itself; it reflects what the algorithm was shown during training.

For example, average cup-to-disc ratios and nerve fiber layer thickness measurements differ across populations. An AI trained predominantly on one group may interpret a normal variant in another group as a sign of disease, or miss genuine pathology because it looks different from what the model expects.

A ChatGPT diagnostic study was notable partly because it addressed this directly. The 204 participants included 32% of African ancestry, 24% mixed American, and smaller South Asian and East Asian cohorts. Performance did not differ significantly between participants. This was a promising finding but came from a single study with modest numbers. A larger, independently conducted study across multiple hospital systems also found that careful AI algorithm designs could achieve equitable performance across sex, race, and ethnic backgrounds, though the researchers noted this required deliberate choices in how the model was trained and evaluated.

 

5. Can AI predict if my glaucoma treatment or surgery will be successful?

SHORT ANSWER: This is one of the future goals for the use of AI in clinical practice. Researchers are using AI to analyze past therapeutic outcomes and electronic health records to help predict the likelihood of success for individual treatments including procedures like trabeculectomies or tube shunts. This could eventually help your doctor choose the most effective treatment for you, once validated and approved by the regulatory agencies.

LONG ANSWER: This is an active and genuinely promising area of research, although the tools are not yet in routine clinical use. Predicting whether a surgical procedure will succeed for an individual patient is difficult because the outcome depends on a complex variety of factors: the type of glaucoma, severity and progression rate, prior treatments, the patient’s healing response among others. Historically, surgeons have relied on experience and population-level statistics to counsel patients, without a reliable way to individualise that estimate.

AI is beginning to change that. A Stanford study applied machine learning to electronic health records from 1,571 patients who underwent 2,398 glaucoma procedures, including trabeculectomy, tube shunts, minimally invasive glaucoma surgery, and laser procedures. Using only information available before the operation, the best model identified patients at high risk of pressure control failure with high accuracy. It also revealed something clinically important: failure rates varied dramatically by procedure type, from 57% for tube shunts to 85% for minimally invasive procedures.  A subsequent study validated similar models across 10 academic medical centres and nearly 9,400 patients, confirming that the predictions held up outside the centre where the model was originally trained. That kind of external validation is an important step toward real-world use.

On the medical treatment side, a study trained an AI model on 14,871 eyes where the target eye pressure had been individually set by glaucoma specialists. The AI model’s pressure targets performed as well as the specialists’ targets in reducing visual field deterioration, and both outperformed generic, guideline-based targets. For patients managed by general ophthalmologists without access to subspecialists, this kind of AI-guided target-setting could improve outcomes. It is important to note that none of these tools are available to your doctor today as approved clinical software, this may happen in the coming years after validation testing and regulatory approval.

 

6. Why aren’t these AI tools already in my doctor’s office?

SHORT ANSWER: The future looks bright. but moving from AI research to implementing in a doctor’s office requires rigorous testing, regulatory approval, and ensuring the AI tools work for patients in the real world. AI is currently intended as a supportive tool, not a standalone decision-maker. We must also ensure the AI is “interpretable,” so the doctor understands exactly why the computer made a certain suggestion and allows them to be better informed yet still in control of what decisions are being made.

LONG ANSWER:

The future looks bright, however the gap between the research lab to a real-world clinic is wider than it appears, and for good reasons. The path from a published study to an approved clinical device involves several distinct hurdles.

First, the science needs to hold up at scale. A study involving a specific group of patients is just a starting point. Before a tool is ready for routine use, it must perform reliably across thousands of people at multiple centers using equipment from different manufacturers. Many AI tools that looked promising in initial research have failed at this stage.

Second, regulatory approval is required In the U.S., clinical AI is regulated by the FDA as a medical device. For a tool to be cleared for diagnosis or treatment decisions, it must undergo a rigorous review process to demonstrate that it is both safe and effective for patient use. Currently, no autonomous AI tool for glaucoma diagnosis has received this clearance.

Ultimately, the journey toward widespread clinical use is a long one that requires more than just high-performance algorithms in a lab. To bridge the final gap, the field must transition toward large-scale prospective studies, where the AI is tested to prove its true diagnostic value. Only through this rigorous validation and the subsequent regulatory approval can we ensure that these tools, whether assistive or autonomous, provide a safe and reliable standard of care for every patient.