Using Artificial Intelligence for Diagnosing Dry Eye

During a first-of-its-kind session at the American Society of Cataract and Refractive Surgery annual meeting, speakers discussed ongoing efforts to develop and apply artificial intelligence (AI) to the diagnosis, treatment, follow-up, and prognosis of multiple ophthalmologic conditions.

In this exclusive MedPage Today video, Karl Stonecipher, MD, of Laser Defined Vision in Greensboro, North Carolina, describes the application of AI in the diagnosis of dry eye, using a software platform known as CSI Dry Eye.

Following is a transcript of his remarks:

I’m Dr. Karl Stonecipher. I’m a clinical professor of ophthalmology at the University of North Carolina. I’m a clinical professor of ophthalmology at Tulane University, and I’m the medical director for Laser Defined Vision physician’s protocol and physician’s protocol cosmetic. So that’s a lot of things that that we do, but the most important thing that I do right now in terms of this is the dry eye stuff. And primarily, years ago when I first started practice, dry eye was pretty boring and nobody really wanted to do anything and you really didn’t want those patients, and it was kind of hard to treat them because we didn’t have a lot of options.

And so, today, kind of like glaucoma, dry eye has become sexy because there’s a lot more opportunity. As a matter of fact, Miebo [perfluorohexyloctane ophthalmic solution] just got approved yesterday [May 18] with regards to a new drop in our dry eye regimen, mainly around Meibomian gland disease.

So as this plethora of new treatments is coming in, we looked at an opportunity to see how artificial intelligence can fit into this model. And I was invited to work with a group called CSI Dry Eye. Their website is www.csidryeye.com, and primarily this group is looking at artificial intelligence, deep learning, neural net, whatever you want to call it at this point — in terms of how we can better improve first the type model and the severity model. And then once we get to that, we’re working on the diagnostic model. And then last but not least, we’re going to ultimately get to the treatment model.

So as you guys know, AI is a very, very misused, I would say, word. And so I originally started working with neural nets with computer topography and we work with augmented analytic engines with regards to IOL formulas and calculations. So how is this different?

Well, what we’re doing is we’re using what are called support vector machines. And what we’re trying to do is improve our diagnostic skills. Basically we’ve got kind of nine buckets and those typical kind of things like aqueous and evaporative and mixed and everything you’re used to. But we also got co-conspirators and mimickers.

And so on the front end it’s really nice to not to have to sit with your patient for an hour having them answer a lot of questions. So we have a 50 patient questionnaire that goes through pretty much everything you can imagine. It talks about what their medications are, it talks about what different environments do. And so that is your first link that comes into the system already before you’ve even seen the patient. So it saves your technicians a huge amount of time.

So if that was the one thing it probably did, boy, that would be worth it right there. But beyond that, then you and I look at, I don’t know, what do we like for a subjective term? We like the OSDI [Ocular Surface Disease Index]. We like the speed. I mean some people like both and that’s what I do. But so we have this 50 patient questionnaire, we’ve got the OSDI and speed. And what I really like those two subjective components for is to teach me if I’m getting better, is the patient getting better, am I doing the right thing? Because if I went backwards, then I can say to the patient, hey, that treatment didn’t work, let’s try something different.

So we’ve got all this subjective information and then now what are we doing in terms of diagnostic? You do what you want to do and you don’t have to go buy expensive machinery to use this product. And primarily the product is an online adventure. I do consult for them, I do research for them, and they’re not paying me to do any of these talks or anything. So full disclosure there.

But it is a product that you can use whether you have a LipiView or whether you have meibography. You don’t have to have any of that. You can basically go through and put in what it is in your clinic that you use, whether that’s simply tear breakup time or subjective lissamine green staining.

But what’s most important about this software, I think, is, if the technician calls a one to one or a mild to mild, you and I are going to call it the same thing. So it doesn’t matter whether Cindy is actually doing the testing or I’m doing the testing, we both can go one Mississippi, two Mississippi, three Mississippi, or use a stopwatch on a tear breakup time. So we really need to do the same though, on, I don’t know, something like lid wiper epitheliopathy, or something like floppy eyelid syndrome.

So in the actual objective part of the testing, we actually have pictures. So what you call a mild is what I call a mild, what you call one is what I call one because we’re all looking at pictures and basing on that.

Now where the software is shining is on the severity model and the type model. We’re actually using support vector machines. And those of you that are not into AI, so to speak as I wasn’t before I started this project, basically these are a way to analyze the data in an artificial intelligence way. But we only have about, I don’t know, 25-26,000 assessments in the bucket. You know, we really don’t get to real neural net and deep learning to about 150,000, but we have a lot of physicians now contributing to the dataset and the more the merrier. And the more you put into the diagnostic side, the better it’s going to be in terms of outcomes and treatments.

So whatever you objectively see, put all of it in that you have in your toolbox. And if you have LipiView, put in the lipid layer height or the partial blinks or whatever you think may add to your armamentarium.

So lastly, where we’re kind of weak and to go there, our type model or severity model are in the 0.7 and above range for severity model and 0.9-0.94 and above range for the type model. And so those numbers are really good, but they’re only going to get better when we do what — put more data in. And then our next adventure is to come up with a diagnostic leading to a treatment.

And that’s what everybody wants. Now we have thousands and thousands of treatments in already, whether that’s tears, whether that’s telling the patient to blink, whether that’s actually doing low-light therapy, or whether that’s doing intense pulse light, or whether that’s doing radio frequency or some type of other interventional system. We put all that in and then that’s going to ultimately be able to be looked at from an artificial intelligence standpoint and see how we’re doing.

You know what I really think in kind of a summary and bottom line is my patients really like this, my staff really likes it, and it’s making me a better diagnostician, which is making me treat these patients better.

So let me give you, for example, because everybody always loves the for example. I see patients that are referred by doctors. So I have 495 doctors refer into me. So the majority of the patients that I see have already been through a slew of treatments or topicals or antibiotics or something else. And so when I get them, you know it’s kind of at the end of the road. So I’m kind of sitting here thinking, okay, what has this person not done?

So my technician isn’t spending a lot of time, I’m not spending a lot of time, because they’re kind of telling me beforehand what they’ve already done. So on the front end for me it saves me a ton of time. On the back end, I’m learning a lot because I didn’t know what I didn’t know. And the biggest thing what I didn’t know is how many doggone drugs caused dry eye.

So to summarize, it helps me be a better diagnostician. It helps my staff to get through a very difficult exam quicker. And the patient feels like hey, somebody’s listening to me now, somebody actually really cares about my problem. And yes, some of these dry patients can be a little bit more challenging, but the bottom line is the bottom line. And some of these patients can be very rewarding because as my good friend Marguerite McDonald will say, dry eye could be some of the best income in a practice if they haven’t already started.