We Know It When We See It: Defining Glaucoma

And thanks to you both very much.

That was a really, really great way of summarizing those, those proceedings that you worked so hard with.

I-- And Ben, what you just said too speaks to the question I think I was about to ask on behalf of all our listeners, like why now? Certainly, this is an age-ages-long problem of not being able to get a definition for glaucoma, so what has precipitated this? And all the people you mention, um, the NEI, Michelle Hribar, um, I, I, I gotta say, I've been wondering for a bit whether our good friend Mike Chiang, now heading the NEI, our clinical informaticist of ophthalmology himself, may be sensitive to this need for, uh, a real operational definition as we embark on further big data approaches.

But this isn't just for research necessarily, and that's, Sophia, to your point, the working groups included like, well, what are we making this definition for? I figured, uh, for our audience, I might give like a few, um, you know, not nec-necessarily use cases, but just like examples of why it's hard to come up with a definition, why it's been like the end of time we've tried.

Just, uh, you know, are we-- For example, is it hard to get a definition, and either of you guys are free to answer this or both, uh, do you think it's the heterogeneity of the disease that resists a, a formal definition or the imperfection of our tests, all the variants in visual fields, the variants of tonometry? Or just the fact that different use cases, you know, are we talking about research? Are we talking about big data? Are we talking about what your community ophthalmologist sees when a patient walks through their door and they have a conversation, "Do you-- do I or do I not have glaucoma?" What's, what's the hard part? It's like all of it, right? [laughs]

Yeah.

Yeah, I think you've, you've really hit on many of the aspects that make this a very difficult disease to classify, right? It exists, uh, along a continuous spectrum.

So where do you draw a line and you say that this is disease, but prior to this is just a glaucoma suspect? Uh, also an important takeaway from the workshop is how we internally represent the disease, right? Rarely do we look at somebody and just say, "That person definitely has disease or does not." We think of the presence of disease along a probabilistic spectrum, right? This person is, uh, likely to have the disease.

How likely? Sixty percent, seventy percent.

We have an internal representation of this, um, and it differs from person to person.

So to have, uh, multiple people get together and agree on this probabilistic representation of disease, uh, is an inherent challenge in addition to just the, uh, the diversity of the presentation, uh, of the condition itself.

So this was one major issue that we needed to, to tackle.

You just think about how you learn what is glaucoma.

Well, we have conceptual definitions, right? It is a progressive optic neuropathy with characteristic structural and functional changes.

Uh, that's fine, and perhaps that's even enough, uh, when you're teaching somebody, right, on a case-by-case basis.

But to, to be able to apply this in a reproducible fashion in a structured scientific environment is a completely different challenge.

challenge.I will say that the workshop, um, wanted to focus, uh, our work on glaucoma definitions as applying first to research, you know, kind of a research-focused approach.

But even within research, there's so many different kinds of research.

There's including patients for clinical trials, there's defining glaucoma for AI studies, there's epidemiology, there's defining glaucoma for screening studies.

And so all of these different use cases were something, um, you know, that the workshop really considered and discussed.

And, and even though, you know, we wanted to focus on research sort of as a first, um, field, there was definitely a sense that whatever we arrive at should be kind of aligned with how we practice glaucoma clinically, even if we didn't necessarily want to, you know, propagate a complicated computational framework into direct clinical practice, right? But it still has-- had to make sense, um, from a clinical standpoint.

So all of that-- those were, like, some of the considerations that we had to debate and discuss and balance.

And Sophia, I'm curious too, um, from a data science standpoint, like if-- especially for these kinds of research you're just describing, like how hard is it for a researcher to do the research they wanna do when the disease itself is just sort of, to some extent, our own expert judgment rather than explicit criteria [chuckles]?

It's very hard.

And I think part of the, uh, challenge that arises from, um, kind of the expert judgment is that that can really vary across different settings, different populations, who's grading, how many graders.