As deep learning and artificial intelligence (AI) play an increasing role in daily life, it is worth looking at the unpredictability of glaucoma in this context. From diagnosis to treatment response to progression to surgical success, glaucoma is more perplexing than easily understood.

One of the strongest risk factors for glaucoma, disc hemorrhage, is associated with a 6x higher likelihood of developing glaucoma over time; however, only 13% of Ocular Hypertension Treatment Study (OHTS) participants with a disc hemorrhage developed glaucoma.¹ In fact, most models assessing glaucoma development or progression use data (eg, discrete data points from 7 visual field tests) not commonly available to most clinicians and still cannot predict the lion's share of progressors. Even models that seem more powerful, such as the OHTS/Early Manifest Glaucoma Trial model, the “predictors” of glaucoma that matter are really signs of existing glaucoma. We have to cheat to make our models powerful enough to be clinically useful.

The risk factors and strength of models for glaucoma development and progression are essentially the same. This leads me to believe that all the big data in the world isn’t going to replace eye examinations and testing for glaucoma detection and progression monitoring, because predicting these is still challenging. Consider this data point from the HORIZON study, where mild and moderate glaucoma patients were randomized to receive either a cataract surgery alone or a cataract surgery with a trabecular meshwork bypass stent/Schlemm canal scaffold: at 5 years, the stent group had a 2.4% cumulative risk of incisional glaucoma surgery vs 6.2% in the cataract surgery alone group (P=.027).² Interestingly, there was no way to predict from baseline data, for either group, which of the early glaucoma patients would go on to need incisional glaucoma surgery in such a short period of time.

As for treatment response, is that predictable? Many glaucoma studies have shown that the only consistent predictors of response to any therapy (medical or laser) are a higher baseline IOP and possibly corneal properties such as thickness and hysteresis. We still don’t have any meaningful ways to pick a successful therapy for our patients based on any type of risk factor or demographic data. Treatment side effects also remain largely unpredictable.

In the era of AI, I have 2 thoughts about a natural language processing program trying to understand glaucoma. The first is, “Good luck.” The second is, “We could really use the help.” Perhaps glaucoma’s complexity is too much for the human brain, but a computer model that dispenses with logic, fairness, or reason will have more success. In any case, I’ll take all the help I can get. GP

Dr. Radcliffe is a clinical associate professor of ophthalmology at the New York Eye and Ear Infirmary of Mount Sinai and a surgeon at the New York Eye Surgery Center.

 

References

1. Budenz DL, Anderson DR, Feuer WJ, et al. Detection and prognostic significance of optic disc hemorrhages during the Ocular Hypertension Treatment Study. Ophthalmology. 2006;113(12):2137-2143. doi:10.1016/j.ophtha.2006.06.022
2. Ahmed IIK, De Francesco T, Rhee D, et al. Long-term outcomes from the HORIZON randomized trial for a Schlemm’s canal microstent in combination cataract and glaucoma surgery. Ophthalmology. 2022;129(7):742-751. doi:10.1016/j.ophtha.2022.02.021