Show notes
Scot Morris, OD, and Rehan Ahmed, MD, on where artificial intelligence and eye care meet.
- AI as medical detective: A study from Boston Children’s Hospital and OpenAI in which a deep-research model generated hypotheses that helped solve 18 previously undiagnosed rare disease cases.
- Tiny recursive networks (TRNs): Compact, hardware-embedded specialist AI that uses less energy and could offer fast, chair-side diagnostic help.
- Adaptive test sequencing: How AI and connected at-home devices could replace the fixed menu of tests with testing tailored, in real time, to each patient.
Transcript
Welcome to Real Talk on AI in Eye Care with your host, Dr. Scot Morris. And I'm co-host, Dr. Rehan Ahmed. In each episode, we discuss current news, add in a little AI education, and debate innovative AI tools and topics that will change the industry. This week's News of the Day takes us just outside of ophthalmology proper, but stay with me, Scott, because the implications of this lens squarely in our world. This is kind of like an episode of Unsolved Mysteries, a little bit. I love that show growing up, and this is kind of like the AI version of that meets healthcare. So just last month, researchers at Boston Children's Hospital for Orphan Disease Research working with OpenAI, but they published a paper in the New England Journal of Medicine.
And the model, they developed the model, and they turned it loose, basically, on these series of unsolved histories, like 376 previously unsolved rare disease cases. These cases that, you know, diagnosticians, doctors, experts have been looking at for years, they've already been through genomic sequences, expert review panels, you know, all the diagnostic tests that you can imagine. And what the model did was pretty impressive. It came away with 18 new diagnoses. So, 18 new diagnoses. 18. It sounds, you know, at first you might say, oh, just 18, but 18, these are not new cases. So, like, a few deals just thinking about. OpenAI's O3 Deep Research Model, they weren't asked to diagnose anyone, it was asked to generate hypotheses.
Kind of like what clinicians do, to synthesize the patient's phenotype, you know, they were given all the paperwork, the clinical notes, genetic variants, inheritance patterns, and they looked through all the literature that was out there. And then what's great about this, it's kind of like the AI plus the human, because then the human experts say ranked what the model came up with, and then the physicians made ultimately a final call. So, AI is to propose a clinician that disposes, right? So, like I said, 18 out of 376, 5%, but this is a group of patients that have been through their ringer, right? They've seen all the experts. There aren't fresh referrals. These are really in truly unsolved cases. It's like, you know, like, just like that show.
And here's the thing, it didn't cost anything. This was basically at marginal cost, and they got new diagnosis. So, think about the stories underneath here. There's one patient, her name's Kyra, she lived with an unexplained muscle weakness since she was nine years old, for 20 years. And then the model connected her case to a frameshit variant of a gene, of a muscle gene, a myofribular myopathy, and the team confirmed it clinically, and they returned the result to her family. So, 20 years resolved by reanalysis that no one else really had the bandwidth to do. So, the directors, he hadn't really put a great quote, he put it plainly. He said, researchers can't possibly keep 8,000 different diseases in their heads, neither can they.
Yeah, so think about all the retinal diseases we see, because inherited retinal diseases, that's just like hundreds of genes, thousands of variants. And, you know, now we're in a world of gene therapy, you know, and there's a lot of stuff which started with external, and it's no longer academic. There could be a world, Scott, where we may be able to manufacture a treatment for a single individual patient with AI. I think we're going to see that in our lifetime. I really do. I really believe we're going to see that in our lifetime. I hope so. So just a few caveats, as always, retrospective study, it's totally unblinded. And the researchers themselves have said this is not ready for clinical deployment yet.
But here's the architecture is proving that hypothesis generation with human expertise. This could be a template, right? And we're talking about, we always say, you know, it's not AI that's going to replace you, it's AI, the physicians who don't use AI. So this is a great example of AI plus clinical judgment coming back with diagnoses for patients who desperately need them. Isn't that kind of crazy? We talked about this exactly two weeks ago on our podcast of what AI plus HI would look like. And then this gets published. Yeah, really, really fantastic. I'm really looking forward to other studies that sort of show this in eye care because I think it's definitely going to happen.
It exists. I know it's sort of happening with my colleagues where we're using this to help us with diagnosing the unusual cases or the zebras. We're just helping you with routine stuff. So this is a great example. Scott, what do you have for us in the knowledge bank? Well, I'm going to take that little one step further because I think one of the challenges we always talk about is how big these large language models are and how do we access them and all that type of stuff. And so I'm going to take the opposite. We're going to go to what is a tiny recursive network or a TRN. And these tiny recursive networks has kind of been hitting the news a lot lately because they take a lot less energy, a lot less power, and they're a lot more focused.
It's like having a sub specialist. Instead of having Claude or Gemini that looks at all the internet, these are going to be much more, they're going to be a specialist. So another thing about this, so we're going to take kind of what you were just talking about. If you've ever tried to diagnose like a complex multi-layered pathology, like I was thinking about it earlier today, like early stage glaucoma, you know, only to find that our diagnostic software now misses a lot of stuff. I mean, what's happening between retinal nerve fiber layer thinning and collagen architecture and vascular changes at the laminar carboxa? We have no idea. Well, these tiny recursive networks are a little different. And I want you to think we're going to approach these in a couple of different fashions.
So, you know, unlike these are massive power-hungry large-angle models we're doing right now, these TRN reuse the same compact set of computational layers over and over in a recursive loop. Think about having layers and layers and layers and layers of information about just eye care or just glaucoma or just nerve fiber layer or just retinal, you know, pick a retinal layer. It essentially studies this data multiple, multiple, multiple times, refining its own conclusions every step it does. So now there's these little TRNs we'll be able to sit there and look at, let's say a glaucoma and go, well, let me look at it from a nerve fibers perspective.
Let me look at it from an architecture perspective. Let me look at it from a vascular perspective and start putting those pictures together to see, can we look this as a more complex package? Now, that also changes what we're doing a little bit. And, you know, so imagine the ability for us to have, you know, right now the challenge is we got to go, we got to either get on our phone or get on a laptop in front of a patient, put that information in. And, you know, that takes time and it doesn't always the most professional thing. Well, imagine having this super lightweight, ultra-fast diagnostic assistant that's put directly into our phone. And instead of waiting for cloud-based processing or software or internet speed or whatever, we get instant high precision pattern recognition right at our chair side, specific for the device and or diagnosis we're looking at.
So I think when we start thinking about using this technology in devices like an office tablet or a soot lamp camera or a retinal camera, an OCT, instead of using that information to go to a large language model, these TRNs are going to be something that we see embedded in hardware. So a different form of AI than large language models, without a doubt, it's something that may make our processing much more effective in terms of small devices. So tiny recursive networks, I know it's a pretty complex subject. I hopefully gave you a good 60-second two-minute observation or viewpoint of it. But keep your ears out for these tiny recursive networks that I think we're going to hear and see about them a lot more, not just in terms of integrated into the hardware, but also maybe in the future integrated into some of our differential diagnosis.
Which kind of brings me right on to what I want to talk tonight about fireside, is kind of a novel concept. I was having this discussion with one of my student doctors the other day. He asked the question, he said, well, Scott, he goes, you know, one of the things I'm always frustrated with is we do the same test in the same order on every patient that walks in the door. He goes, why don't we have something that tells us what test is going to get us to the best answer, the fastest? I said, well, funny you should do this, funny you should ask that, because that's called adaptive test sequencing. And that's a relatively new concept in that maybe in the near future, and I don't think it's maybe, I think it's for sure going to happen, when we have some AI model that does triage, we've talked about that before, Rehan, whether it be in office or out of office.
Now, all of a sudden, maybe when the patient walks in the door, the patient gets the diagnostic testing area or center, maybe it's not in the office, as we've had discussions before. Maybe now the first test you do is not visual acuity, or not pupil check, or not a sit lamp, but maybe the first thing you do is no CT, or an ERG. Because the AI says, yeah, based on the differential, the fastest way to split our answers, to split our choices or to further differentiate this is to do test six on our normal scale, you know, in terms of our numbers. So I think adaptive test sequence, I wanted to talk to you a little bit about that today about, you know, what you see, Rehan, as opportunities with that, and then where your natural gut says, yeah, I'm not sure if I like that.
Yeah, I think this is a great topic, because for most of the history of medicine, certainly eye care, testing is like a fixed menu, right? It's like the restaurant, it's the same thing, every single time, same visual field strategy, same imaging, same follow-up interval, you know, whether it fits them or not. So I love the idea that you just talked about how adaptive testing flips that. It's basically the right test for the right person, in the right way, decided in, like, real time. In real time. What we already know about the patient, and AI is making that possible. I think it's a great idea, and it's great for, we sort of talked about, you know, these large companies that have scale that are seeing a ton of patients, large groups, or even the individual doc who is very busy to have it.
A sequencing that is sort of informed by it. Now you ask them where I see issues. I mean, I think the one concern would be second order, right? It makes sense here when clinicians oversight, but I can imagine a future where it's used primarily in the initial parts of the exam and screening. And maybe it says, like, a certain test actually doesn't make sense. Maybe it sees, it takes up, you know, there's a fundus photograph, and it sees a totally normal optic nerve, and it decides that, you know, the evidence to pursue, no matter what, you know, the visual field, supposedly says that this patient does not have glaucoma or whatever, and certain testing sequences maybe are not done.
So I think that's kind of interesting. Let me play devil's advocate. Let me play devil's advocate. So think about glaucoma, right? As we all get in a habit, and I don't know what your pattern is, you know, for severe glaucoma, I'm probably doing an OCT in a central visual field every three or four months. For moderate glaucoma, maybe every four to six months. For my glaucoma, maybe less than that. But where's the science in that, right? I mean, I look at this and go, will adaptive test sequencing, will that decrease unnecessary testing? Or maybe it's going to tell us that we need to test more often. More often, exactly, especially now based on the individual, especially now as these testing screening at home with smart devices, the cost of testing may go down to near zero.
I mean, you may not really need a clinician to review every single test, but have an AI model with the assistance of obviously a clinician summarizing, you know, there could be a way where testing is actually a lot easier now, even though you're doing it maybe an order of magnitude more frequently. We've seen this now with OCT, where patients have an OCT at home, and they may be doing OCT every few days. Same with visual fields, these VR headsets. We may go to monitor things at a very fine level with much more testing. With much greater frequency. With much greater frequency, with actually overall less work. And you think about that, and I know we're kind of going a little off subject here, but I get kind of excited about this is these at-home OCTs and visual fields.
You know, it'd be interesting to as we start getting more and more data about this is going, well, hey, over the last two weeks, my blood pressure has been up 10 or 15 points, and or I'm not sleeping as well, and or my VO2 max is down, or my CO2 is up, and guess what? My visual field pattern has changed. Things that we have no idea about right now, no idea. Just some crazy thoughts. I think we're going to see your glucose levels correlated with your OCT to see if you have macular worsening your macular edema. I think having these connected devices will overall give us much more insight. And of course, this is what the health industry or your ORA rings and your what is the other one called a WOOP. Yeah, that's why they're having these billion dollars.
So it makes it makes a ton of sense. I'm really excited. I'm really excited about this. And I think adaptive testing fits just so well into the ecosystem of connected devices, because what are you going to use to help the year after testing? Well, it's these connected devices that will feed into that to help you determine what tests to do and when to do it and who to do it on. And how often to do it. Yeah, so it's a great topic. Great topic, Scott. Another great week. We look forward to speaking with everyone again next week. You can find all of our information at AI in Eye Care dot com or feel free to reach out to us at Ray on REHN at AI in Eye Care dot AI or Scott SCOT at AI in Eye Care dot AI.
Thank you guys for listening. And we will talk to you next time. You've been listening to Real Talk on AI in Eye Care your weekly podcast to keep you informed about AI technologies revolutionizing eye care.