Artificial Intelligence and the Hunt for Neuroprotection in Glaucoma

Optic nerve, section of the brain seen from the side. Nerve injury. Poor eyesight. Optic neuritis. Inflammation
Credit: Getty Images
Machine learning models can detect optic neuropathy in imaging, discern visual field loss, and now, help researchers to estimate smaller sample sizes for neuroprotective drug trials.

Artificial intelligence (AI) can be applied to a striking array of applications in ophthalmology. In glaucoma alone, researchers have paired its data-crunching powers with technologies such as sensor-laden contact lenses that provide continuous intraocular pressure (IOP) readings. This can help predict the course of a disease and potential visual field loss.1 But the applications of machine learning can go beyond the scope of individualized patient care. Research now shows that machine learning can contribute to the development of new glaucoma drugs. In particular, studies show 2 unique AI models that can assist researchers by reducing the population sample size necessary to show a medication’s effect.

In glaucoma, the only modifiable risk factor is IOP, so most glaucoma drug trials are focused on therapies that lower pressure. But the Holy Grail of glaucoma research, the discovery that will rewrite glaucoma treatment manuals as we know them, is the development of a neuroprotective agent. So far, more than a dozen candidates are in contention, but none of them have yet shown success in evidence-based trials. The reason for that may be as much about the trials themselves as the complexity of the disease. That’s where, according to experts involved in neuroprotection trials, artificial intelligence could play a key role.

“This helps push forward what we’re able to do in medicine and how we’re able to take care of patients,” says Andrew Chen, MD, an assistant professor in the department of ophthalmology at the University of Washington (UW). Dr Chen is lead author of an investigation published in the American Journal of Ophthalmology that employs an AI model to identify eyes that meet the US Food and Drug Administration (FDA) criteria for progression. The idea is to reduce the necessary sample size, or shorten the duration of clinical trials for developing neuroprotective agents.2

In short, artificial intelligence may be able to fast-track drug investigations in a few different ways. One technique is to fine-tune recruitment of participants who would likely meet an established neuroprotective endpoint. Another is to estimate macula ganglion cell-inner plexiform layer (GCIPL) and optic nerve head (ONH) hemiretinal thicknesses in participants to establish an internal control group for trials.2,3  

The Cost of Neuroprotective Agent Trials

Several of the potential neuroprotective therapies include calcium channel blockers, gene therapy, neurotrophins, along with antioxidants and vitamins.4-8 Chiefly, neuroprotective trials are more expensive because the current FDA-approved endpoints for validation (visual field loss criteria) typically progress slowly and chronically in patients with glaucoma, Dr Chen explains. An important challenge is that different progression rates occur between individuals.

Harry Quigley, MD, professor of ophthalmology at Johns Hopkins School of Medicine, agrees. He is a founding member of the American Glaucoma Society, and served 5 year terms as CEO of the Association for Research in Vision and Ophthalmology (ARVO) and editor-in-chief of Investigative Ophthalmology and Visual Science

Clinical trials for IOP lowering agents may require smaller samples and shorter trials for reasons including drug effect and endpoints — IOP-reducing drugs simply need to show that eye pressure decreased and remained lower for a year, while visual field (VF) and optical coherence tomography (OCT) tests showing change may take 2 or 3 years, he explained, adding that IOP measuring is a less complex process than functional vision testing.

“In a neuroprotection trial, you’re having to measure something using the more complicated field and OCT tests that change relatively slowly over time, whereas the IOP change happens pretty much 2 or 3 days after you start taking the new eye drops,” Dr Quigley adds.

Pharmaceutical firms have sponsored many clinical trials for anti-vascular endothelial growth factor (anti-VEGF) injectable medicines to treat retinal diseases. Injectables for glaucoma treatment are less feasible and safe — unless it is a single injection of gene therapy with a viral vector.9 Until individuals with glaucoma reach the moderate stage of the disease, they are able to do many of their normal everyday activities without problems and may not fully realize that their diagnosis could indicate much greater future loss in functional vision, Dr Quigley explains.

“So giving multiple intravitreal injections is almost surely going to cause temporary blurring, and may, in fact, subject the patient to side effects that are negative, such as developing cataracts, such as bleeding inside the eye,” according to Dr Quigley. “So, while we could, I think envision that it would be practical to give 1 very long-lasting injection, the idea that you could do multiple injections every 3 months, or every 6 months, or even every month like they do for macular degeneration in eye disease is not practical for glaucoma.”

AI Can Reduce Sample Size

The dataset for the analysis by Dr Chen and his colleagues includes 7428 eyes of 3871 patients who visited UW and underwent Humphrey® Field Analyzer (HFA) II (Carl Zeiss Meditec), 24-2 visual fields from 1998 to 2018. Investigators reviewed charts to find VF loss in 5 or more locations at more than 7 dB change from baseline in 2 successive tests. They trained the novel AI model with data such as baseline mean total deviation (MD) and basic patient characteristics.2

At 5 years, 13% of the total sample met VF loss progression benchmarks. Variables of MD and age were significant risk factors (P <.0001) — the subgroup with highest odds of progressing exhibited a baseline MD of less than -5.0, and ranged between 60 to 80 years of age.

A training set included 3614 eyes of 1880 patients, and a test set included 3814 eyes for 1991 individuals. The AI model was trained to select its own set of eyes that would most likely demonstrate progression spanning a specified amount of time based on pointwise threshold deviation measurements from the first test, baseline MD, laterality, sex, and age. A separate training was performed for example trial durations of 1 year, 1.5 years, 2 years, and 6-month intervals until 5 years. The model computed risk probability, set at the upper 75th quartile.

Three calculations — not part of the AI training parameters or goals — were run for the test set: first, how many trial participants based on the total sample would be needed to show effect; second, how many were required by including only those in the subgroup; and third, the number needed based on top quartile risk.

Results exceeded the researchers’ expectations. For a 3-year clinical trial demonstrating a 30% effect size, 1656 patients would be needed to achieve 95% confidence interval (CI 1638 to 1674); but using subgroup individuals only, the number would be reduced to 903 (884 to 922) with the same CI. Sample size was further condensed to 636 (625 to646) participants who were selected by the AI model.2

“So, as a proof of concept, I think we demonstrate that an AI model can recognize; based on baseline visual field data, patients more likely to progress,” Dr Chen explains. “And for a clinical trial, you want to target these to minimize your resources and costs in order to best evaluate your neuroprotective agent or drug under study. So, I think this is very promising, because it potentially reduces the burden and resources required to conduct such a trial.”

AI Could Streamline Trials 

In glaucoma care, most clinical trials currently underway are directed toward devices and pharmaceuticals that can manage IOP. But a higher than average IOP is not necessarily indicative of the disease. “I think it really gets down to what glaucoma is as a disease: the definition of glaucoma being an optic neuropathy,” Dr Chen says. IOP is a risk factor. “And at the end of the day, it’s simply that, a risk factor.”

The FDA’s criteria for approving neuroprotective agents is visual field assessment — VF loss representing worsening functional vision — the agency defines VF loss as 5 VF locations or more displaying significant changes of at least 7 dB.  “They want to make sure what we’re evaluating translates to meaningful outcomes for the patient,” he explains.

No neuroprotective agents are yet approved by the FDA, Dr Chen reports. One agent that initially demonstrated promise is memantine, an N-methyl-D-aspartate (NMDA) receptor, or glutamate receptor antagonist that has been used for dementia in Alzheimer’s disease and other illnesses. In phase 3 clinical trials, 2,298 participants from 128 ocular clinics were enrolled, with data reported in Ophthalmology.10

Using standard automated perimetry, VF progression endpoints showed significant loss in 5 or more locations after an 8-week period. However, no differences existed between case and control sets — after 4 years, and a cost of approximately $100 million, according to Dr Chen, adding that phase 3 studies published in 2018 “took a lot of effort and a lot of resources, which left the industry with a bit of a sour taste and a little reservation to take on (neuroprotective) projects.”

Dr Quigley says that, in fact, no studies on neuroprotection have been submitted to the FDA. “One large past study was not successful according to what the authors wrote up and so there was no reason to ever submit it. The drug didn’t work.”

Machine Learning as an Internal Control Group

Dr Chen’s research uses an approach that corresponds to current FDA criteria for neuroprotective drug approval based on visual function. Another new machine learning model developed by Mark Christopher, PhD, and colleagues uses structural features to estimate decreased population sizes — but in this case for early-stage clinical trials.3

The novel deep neural network uses inputs of retinal thickness scans to forecast how a region treated with an agent would have worsened without the therapy, according to this cross-sectional investigation published in Ophthalmology Glaucoma.3 The dataset incorporated 1096 eyes of 550 participants in the Diagnostic Innovations in Glaucoma Study and African Descent and Glaucoma Evaluation Study. Spectral domain OCT was used to scan thicknesses at a 4.1 mm diameter around the ONH and 30°x25° centered on the macula, in 4 distinct hemiretinal regions: superior ONH RNFL semicircle, inferior ONH RNFL semicircle, superior macula GCIPL map, and inferior macula GCIPL map.

Using 1 hemiretinal macula GCIPL and ONH RNFL paired scan set of the case eye and all images of the fellow eye, the network was trained to predict the remaining hemiretina of a patient’s case eye. Specifically, regions representing untreated areas are referenced to estimate how a treated region would have progressed if not given the proposed agent — thus accounting for inter-patient variability, and functioning as an internal control. 

Investigators calculated that in a conventional trial, 4109 participants would be needed to show a 25% drug effect in a 2-year trial using GCIPL thickness, as well as 7085 for RNFL depth. With the deep learning regression, though, these numbers would be reduced to 287 GCIPL and 623 RNFL predictions, representing a reduction more than 11-fold.

This model does not use longitudinal data imaged at various time points, and depends on agents that act on a localized region of the retina. Due to the structurally-oriented endpoints, these predictions are less suited to phase 3 clinical trials, “however, their use as exploratory endpoints in earlier stage trials could help increase the pace of research into neuroprotectives,” the analysis explains.

Potential Neuroprotective Drugs

A limited number of neuroprotective agents are in the study pipeline. “There are a couple here and there that may have mechanisms of action beyond IOP lowering, but that does not tend to be the focus of the pharmaceutical companies at this time,” according to Dr Chen.

An AI model can recognize; based on baseline visual field data, patients more likely to progress. And for a clinical trial, you want to target these to minimize your resources and costs in order to best evaluate your neuroprotective agent or drug under study. So, I think this is very promising, because it potentially reduces the burden and resources required to conduct such a trial.
Andrew Chen, MD

Brimonidine, an alpha-2 adrenergic agonist, may have an advantage in this regard — an IOP-lowering medication, it also exhibits certain cell-protective aspects. An interventional, randomized clinical trial at Wills Eye Hospital, in Philadelphia, used OCT-angiography (OCT-A) to evaluate peripapillary blood flow in 35 patients with primary open angle glaucoma (POAG), normal tension glaucoma (NTG), or ocular hypertension (NCT03323164). The pilot study included 2 arms: brimonidine tartrate 0.2%, and timolol maleate 0.5%, with the trial period during 2017 and 2018.

An earlier nonrandomized, prospective trial published in the American Journal of Ophthalmology tested brimonidine 0.15% to evaluate whether it improved retinal vascular autoregulation and short-term visual function. Participants (N =46) with NTG underwent blood flow assessment of a temporal retinal artery in seated and reclined postures. The 23 who showed vascular dysregulation upon position change were treated with the test drug. “After the 8-week course with brimonidine, 14 of the 17 patients who completed the study showed a return of posture-induced retinal blood flow changes to levels consistent with normal retinal vascular autoregulation (P <.0001),” according to the paper. There was, however, no change in 8-week VF outcomes for those receiving brimonidine.11

To date, the neuroprotective effect of brimonidine in humans has not been definitively proven and the FDA has not approved it for clinical use.

Options researchers are exploring outside of brimonidine include several types of neurotrophins including the NT-501 ECT implant. This sustained release device elutes a soluble ciliary neurotrophic factor (CNTF) receptor in an effort to improve the visual field index as well as potentially affect contrast sensitivity and change the ganglion cell layer thickness. A phase 2 trial into this intraocular implant is ongoing ( Identifier: NCT02862938).

Researchers are also investigating the neuroprotection and neuroenhancement potential of recombinant human nerve growth factor (rhNGF). Although phase 1b trials show no statistically significant short-term neuroenhancement, the investigators say the nerve growth factor had “strong effects” in preclinical models, warranting analysis for efficacy in a neuroprotection trial.12

Another compound being investigated as a possible neuroprotective agent in glaucoma is citicoline, which investigators say “has demonstrated activity in a range of central neurodegenerative diseases, and experimental evidence suggests it performs a neuromodulator and neuroprotective role on neuronal cells, including [retinal ganglion cells], associated with improvement in visual function, extension of the visual field and central benefits for the patient.”13

In a 2020 study, researchers suggest that citicoline “may prevent the deterioration of membranes in neurons and inhibit the apoptosis that occurs in neurodegenerative processes.”13 Additionally, a 2022 study of inner retinal cell function using pattern electroretinography (PERG) testing shows that daily oral doses of a fixed combination of citicoline 500 mg and homotaurine 50 mg can improve functioning in the inferior and superior quadrants independent of IOP reduction. The study authors say this indicates a potential option for neuromodulation in patients with glaucoma that could prevent progression.14

Several individualized gene and stem cell therapies are also under review.15

AI Is Inexpensive to Implement

Recently, with advances in trend-based analysis and the AI model, studies may be becoming more cost-efficient. “That’s now an attitude that’s coming around and changing due to things like Dr Chen’s paper and the paper that I was associated with,” Dr Quigley explained.” You don’t need thousands of patients for many years and hundreds of millions of dollars if you design it properly.”

The model at UW was built with open source code and requires no specific platform or application on which to run, simply a modern computer; no code-specific technology limitations.

“But it does require access to visual field data that has to be exported from the Zeiss Humphrey Field Analyzer,” Dr Chen clarified. “The visual field report the end user normally sees is not sufficient.” He added this model could be adapted to work with data from other VF analyzers.

Dr Aaron Lee, an AI study coauthor added that the analysis was performed on an open-sourced visual field dataset anyone can download and use. The link for the UWHVF dataset is posted on Translational Vision Science & Technology.16

The code by Dr Christopher’s group will also be shared if a request complies with data protection rubric at the affiliated institution, University of California, San Diego, the study notes.3

Advances in AI May Add Integration

“Our selection methodology was based on the patient’s visual field and basic patient characteristics,” Dr Chen said. “And for clinical trials, you presumably have a little bit more information.” Other variables to factor into the model could include comorbid diseases, family history, and pachymetry. Potentially, these additional elements would reduce the sample size needed even more.

Integrating structural characteristics may also increase the model’s preciseness — such as data reflecting changes in the lamina cribrosa (LC), now of increasing interest. A new study published in Ophthalmology Glaucoma reports that when IOP is reduced, the LC moved either slightly inside or outside the eye — more consistent changes were apparent in the structure’s radius, radial strain (Err)=-0.19±0.33% (P =.004); and thickness, anterior to posterior strain (Ezz) = 0.94±1.2% (P =.0002) when IOP decreased.17

Whatever the selection model in a neuroprotection trial, Dr Quigley strongly advises that researchers never lose sight of IOP lowering for all participants, even though reduced IOP makes the neuroprotective therapy harder to prove. “Because we know for sure that pressure lowering is beneficial to the patient, so we won’t be ever comparing neuroprotection to pressure lowering,” he said. “We’re going to be comparing neuroprotection against no neuroprotection in a group of patients, all of them having pressures lowered.”

“Promising Avenues” for Patients, and Society

Since 2009, when the FDA’s event-based criteria was published, trend-based analysis of progression between groups has also been shown as a promising option to reduce sample size in clinical trials. “There is abundant evidence that trend-based analysis of visual field data is going to be much better in a variety of ways than event-based,” Dr Quigley said.

In addition to AI allowing for less costly clinical trials, it may someday also be a common diagnostic tool. Machine learning will never replace the human provider, Dr Chen believes, although it stands to help clinicians catch subtler signs faster. Since patient history, visual fields and OCT images are routinely gathered, the data is there for integration into new AI models. “Now, how that can be done is going to be an area of research and debate,” he added. “The Holy Grails of AI are: 1. Helping to screen patients; 2. Helping to identify those who have been diagnosed with glaucoma that have worsening glaucoma.”

Algorithms are progressing toward remarkable accuracy, and helping clinicians with swifter diagnoses. Combining tests is an option, “but the most convenient would be if a single test can be extrapolated using AI to predict the other,” according to an article published by the American Academy of Ophthalmology. An AI model built into OCT software could also predict visual field loss and lead to earlier treatment, according to that report.18

Dr Chen and his colleagues are planning upcoming work on the next steps for machine learning; testing the model with other database test sets, and/or integrating other patient values into the code. Another method might be to combine the AI with traditional selection methods, all “very promising avenues.”

Notably, AI just may be able to get neuroprotective medicines into the hands of clinicians sooner. “If AI can deliver that promise, then yes, potentially the pipeline to develop and procure; to prove, test and approve new drugs, would be shorter,” Dr Chen said. “And be less costly overall for society.”


  1. Martin KR, Mansouri K, Weinreb RN, et al. Use of machine learning on contact lens sensor-derived parameters for the diagnosis of primary open-angle glaucomaAm J Ophthalmol. 2018;194:46–53. doi:10.1016/j.ajo.2018.07.005
  2. Chen A, Montesano G, Lu R, et al. Visual field endpoints for neuroprotective trials: a case for AI driven patient enrichment. Am J Ophthalmol. Published online June 28, 2022. doi:10. 1016/j.ajo.2022.07.013
  3. Christopher M, Hoseini P, Walker E, et al. A deep learning approach to improve retinal structural predictions and aid glaucoma neuroprotective clinical trial design. Ophthalmol Glaucoma. Published online August 26, 2022. doi:10.1016/j.ogla.2022.08.014
  4. Otori Y, Kusaka S, Kawasaki A, et al. Protective effect of nilvadipine against glutamate neurotoxicity in purified retinal ganglion cells. Brain Research. 2003;961:213-219. doi:10.1016/S0006-8993(02)03951-3
  5. Osborne A, Wang AXZ, Tassoni A, et al. Design of a novel gene therapy construct to achieve sustained brain-derived neurotrophic factor signaling in neurons. Human Gene Therapy. 2018;29(7):828-41. doi:10.1089/hum.2017.069
  6. Osborne A, Khatib TZ, Songra L, et al. Neuroprotection of retinal ganglion cells by a novel gene therapy construct that achieves sustained enhancement of brain-derived neurotrophic factor/tropomyosin — related kinase receptor-B signaling. Cell Death & Disease. 2018;9(10). doi:10.1038/s41419-018-1041-8
  7. Pease ME, Zack DJ, Berlinicke C, et al. Effect of CNTF on retinal ganglion cell survival in experimental glaucoma. Invest Ophthalmol Vis Sci. 2009;50(5):2194-200. doi:10.1167/iovs.08-3013
  8. Aref A, Bhatt A, Conner I, Asanad S. Neuroprotection in glaucoma. Updated April 22, 2020. Accessed August 12, 2022.
  9. Quigley HA. Glaucoma neuroprotection trials are practical using visual field outcomes. Ophthalmol Glaucoma. 2019;2(2):69-71. doi:10.1016/j.ogla.2019.01.009
  10. Weinreb RN, Liebmann JM, Cioffi GA, et al. Oral memantine for the treatment of glaucoma: design and results of 2 randomized, placebo-controlled, phase 3 studies. Ophthalmol. 2018;125(12):1874–1885. doi:10.1016/j.ophtha.2018.06.017
  11. Feke GT, Bex PJ, Taylor CP, et al. Effect of brimonidine on retinal vascular autoregulation and short-term visual function in normal tension glaucoma. Am J Ophthalmol. 2014;158(1):105-112. doi:10.1016/j.ajo.2014.03.015
  12. Beykin G, Stell L, Halim M, et al. Phase 1b randomized controlled study of short course topical recombinant human nerve growth factor (rhngf) for neuroenhancement in glaucoma: safety, tolerability, and efficacy measure outcomes. Am J Ophthalmol. Published online November 12, 2021. doi:10.1016/j.ajo.2021.11.002
  13. Gandolfi S, Marchini G, Caporossi A, Scuderi G, Tomasso L, Brunoro A. Cytidine 5′-diphosphocholine (citicoline): evidence for a neuroprotective role in glaucoma. Nutrients. 2020;12(3):793. doi:10.3390/nu12030793
  14. Rossi GCM, Rolle T, De Silvestri A, et al. Multicenter, prospective, randomized, single blind, cross-over study on the effect of a fixed combination of citicoline 500 mg plus homotaurine 50 mg on pattern electroretinogram (PERG) in patients with open angle glaucoma on well controlled intraocular pressure. Front Med (Lausanne). 2022;9(4):882335. doi:10.3389/fmed.2022.882335
  15. Khatib TZ, Martin KR. Neuroprotection in glaucoma: towards clinical trials and precision medicine. Curr Eye Res. 2020;45(3):327-338. doi:10.1080/02713683.2019.1663385.
  16. Montesano G, Chen A, Lu R, et al. UWHVF: a real-world, open source dataset of perimetry tests from the Humphrey Field Analyzer at the University of WashingtonTrans Vis Sci Tech.2022;11(1):2. doi:10.1167/tvst.11.1.1.
  17. Czerpak CA, Kashaf MS, Zimmerman BK, Quigley HA, Nguyen TD. The strain response to intraocular pressure decrease in the lamina cribrosa of glaucoma patientsOphthalmol Glaucoma. 2022;S2589-4196(22)00120-X. doi:10.1016/j.ogla.2022.07.005
  18. Kuriakose A, Moore DB, Akkara JD. Artificial intelligence in glaucoma. Updated August 11, 2022. Accessed August 12, 2022.