Show notes
Scot Morris, OD, and Rehan Ahmed, MD, on the shift from piecemeal, “bolted-on” AI tools to fully integrated, AI-native practices.
The hosts break down a newly deployed AI-native eye clinic in China: its multi-agent system, optimization challenges, user-interface design and the “silent trial paradox” in clinical validation.
The conversation then turns to data retrieval and anticipatory computing: how predictive intelligence, heads-up displays and ambient systems could replace search queries and screen-toggling, leaving clinicians more time with the 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. All right, Scott. For this week, for the news of the week, we have something hot off the press. One that I think our leaders and the leaders in AI are going to find super interesting and we'll have a link to it at our website. Because for me, this is one of the more interesting papers that have come out. It's about a clinic in China that started an AI-native clinic, not AI bolted on, which is a solution that often people are doing, which are sort of piecemeal, but a truly AI-native clinic.
So I think it's really interesting to dive a little bit deeper in it. So let's start with a problem in China. They actually have a name for their health problems. It's called the three long, one short problems. So long registration, long waits, and long administrative procedures. So it just takes a long time for these things to happen. And then after that, it's kind of insult to injury. It's a super short visit with the consulting position. But it's actually not that different from here. I don't think that's that different from what we do. I love the long, short, short, because it kind of is like long intake, short testing, or even long testing.
No, it's a long wait. It's a long for everything. And then finally, you're in front of the doctor. 60 seconds. Yeah. And yeah, I've heard this point back. I have a cousin who saw a doctor I highly recommended, and she complained. She's like, look, I waited all this time, and it was literally three minutes in and out. I feel I understand the pain. And I guess the point the authors are making, something we've talked about for a while, essentially a better classifier, be it like a better diabetic classifier, does nothing for this three long, one short problem. You're going to have everything, you know, the read perfect, all the AI algorithms perfect, and if the patient still waits two hours to get it, or there's no referral pathway, if there's some issue that's wrong, the consulting doctors don't have time to actually evaluate the issues.
And so the question is, does it actually solve patient problems? It may be a piecemeal solution for one specific part of the workflow, but it doesn't actually give you any total efficiency gain. So what do they do in check? So they actually, they built what's called an AI native clinic. It's called AI tech. It stands for the clinic that was in near Beijing, and it isn't just one algorithm. It's actually a team of agents running across the whole visit, actually before the visit. So they have an agent that takes history before the patient arrives, one that triages and orders testing, one that pulls images together, one supports the physician decision, and one that handles follow-up.
And they sort of went through this in a design schema in the paper. And I think actually the most interesting agent is an agent called Agent X. And Agent X is the agent of agents, and it actually orchestrates between the different agents. So it's very interesting. It's the conductor, basically, right? It's the conductor. And it's the conductor, exactly right. I think that's a great term for it. And so the idea is that by the time the patient sits down with the alchemist, all the prep work is done, and it's not a bolted-on solution. It is integral to the entire clinic. And what I really actually like about this paper, it's not just like, rah, rah.
They talk about some of the things that didn't go so well, right? So once the model hit the real world, so they built this all, they actually found, they used what's called a foundation model. We've talked about it before in our inner meeting with Pierce. His team in NHS, they built the RET found, which is a foundational retinal image classifier. It was not that good on their clinic's own images. And so what they had to do, they actually had to use their own images to basically make that model a little bit better, right? So they optimized, and I think that's a term we talked about at the real beginning of starting these talks, they optimized that RET found image set from images of their own clinic to make the algorithm better.
So they had a percentage of success, maybe like 80%, and then up to 94%, the area under the curve, basically. The second issue they talked about, which wasn't maybe a great thing to talk about in general, but they talked about is that initially the usage wasn't great. The usage by the doctors? The usage by the doctors. The doctors sort of using some of the clinical decisions and support. And the reason was, it was just too many clicks, right? Actually, when they started the system, they started in December of last year, and it actually was under 4%. So under 4% was a very, very large iClinic. Under 4% of doctors were using in April. And not because the models got worse, but because there was way too many clicks, and so they optimized the image.
The user interface was the issue, yeah. The user interface. And they got it down to like maybe two to three clicks, and the work and the usage skyrocketed, right? So this just goes into saying, you know, this is not a build and sort of forget it. You have to be involved. There's a lot of design issues to think about, and so they had to really make the workflow streamline for the doctors. But the last thing I think that was interesting, they talked about a term I actually haven't heard before, the silent trial paradox. And what the silent trial paradox is that most of our validation studies in AI, they actually hide the AI's output from the clinician.
So in these trials, the clinician is basically not involved. They can't change anything. The diabetic eye classifier tells you whether you have referable diabetes or not. And that's great for telling whether the model is accurate, but it can't actually tell you whether the model is useful and do we care about it. So a tool that's designed to change decisions really can't be evaluated by making sure it changes nothing. Right. And so what I love about this paper overall is that, again, AI sort of bolted on as a silo. It's just like another screen to check. But having a clinic, and this just admittedly, this is the first stage is that having a clinic that's sort of thinking about this in a more holistic fashion and testing it and then measuring your results and being honest about that.
So for me, this is kind of the first step and direction of a truly AI native clinic. Super interesting to read about the good things and some of the works and how they're figuring it out. So they talk about the next steps, and the next steps are thinking about how does this affect the overall three long, one short problem. And they didn't get into that because it hasn't been enough time, but that's what they're aiming to tackle. So yeah, super interesting stuff. I'd love to get your thoughts on it. Yeah, I think I want to do the knowledge bite, and I think that might lead into what you were kind of talking about. But maybe for our later conversation, we're going to talk about the AI native clinic because I think it's a mental redesign that's coming.
And I think we should talk about that. So let's put that hold for just a minute. So what do you have for the knowledge bite? Yeah, so I've been thinking a lot. I was at a meeting this weekend. It's called the boss meeting. B-O-S-S is put on by Eugene Shatsman, Jamie Rosen. Yeah, and I mean, they've been on, you know, Eugene's been on our podcast before. And it was a really interesting meeting, but what kept coming up was, you know, I had numerous people would ask, well, where do you think search is going? And there were marketing people in there with a way more marketing background than me. And it was very interesting listening to their answer, and they said across the board, search is dead.
And I was like, huh, that's kind of interesting. And so I said, well, what do you mean by that? And they said, they come up with this term that I had never really heard of before, but it was more of anticipatory intelligence. And I thought, well, that's really interesting. And it was more of we have the ability or we're soon going to have the ability to not do search, but to basically do ambient on-demand predictive information. And so this is kind of going to feed into, I think, a little bit our talk later about the Native II clinic. But I want to kind of cover this because I think it's fascinating. I had to do a little digging on this one to kind of find out, like, what does the future of search really look like?
And is search really dead? So I want you to think about how you interact with clinical technology or really any technology now. I mean, you know, for example, a patient walks in, what do you do? You click and you type and you filter and you dig. And if you don't know the answer, you go look it up on Google, or maybe you're looking up on Gemini, or maybe you're looking up in chat GPT. But we live still, even today with the LLMs, we still live in the era of search where every answer requires us to actively go type something in to do a search. It requires our cognitive knowledge. And candidly, it requires something more precious than all that, which is time to be able to do this.
Time to develop the right question, time to send out the search, time to read the answer, and time to process the answer. And the reality is that I think the other term that somebody used was anticipatory computing, which is really this whole idea of predictive knowledge delivery. It's not something just for medicine or just for healthcare. It's really going to be for everything. You know, Rayon, when we first did our very first meeting where you and I actually kind of got to share the stage the first time, I brought up this term called APHIS, right? Adaptive Personal Health Information Systems. And I think maybe I had the A wrong, and I think maybe I had the P wrong.
I think it's Ambient Predictive Health Information Systems is kind of where we're going, not just in medicine, but overall. And the logic here is kind of really simple. Instead of us querying a machine, and instead what's going to happen, the machine's going to understand the context in real time of what we're really looking for. And in terms of clinic, it's going to surface some of that data from the patient record or from prior exams or from evidence-based medicine and then alert us about things we need to think about, do, find, purchase, even before we think to ask for it, right? And I find that way, and I think we're kind of already getting there, and I think last week we talked a little bit like, how do you talk to AI?
And I think as you start learning how to talk to AI, it starts learning how to talk back and starts predicting things you're going to ask for or things that you should be asking that haven't thought of yet. And I, you know, much like you, Ray, and I live on AI all day long, and the more I use, the more I'm finding out. It's thinking two, three steps ahead based on some of the other things I've done before. And so it's flipping this information. The whole big picture, it's flipping information retrieval kind of on its head. It's not data waiting to be found. It's data being delivered not on cue but before being asked for. And that's pretty cool.
I mean, when we think about how that's going to change things is that we're not going to be the data detective. We're not going to be spent toggling between screens and digging through, you know, previous consults with retina or corny or whatever from two years ago. Now, what was that? What was it that they had? What did they decide to do? This is something that friction, right? It's still very difficult. It's very time intense of what we do now. I think that when we get to this anticipatory computing, and maybe it's already begun, that this cognitive fatigue that we have and the burnout is going to go away because it's going to be delivered to us many times before we even think about it.
And, you know, I mean, and I think right now we're in the early stage. We've talked to maybe good or bad about scribes, you know, scribes or listeners. But what happens when they don't become listeners but they become processes and talkers, right? And they start saying, hey, you know, have you thought about this? And maybe it's going to be something obvious where they talk out loud. I think kind of relating to one of our prior conversations. I think we're going towards heads-up display and I think it's going to present a screen or do something where it's, you know, what's that called? Bone, bone, bone propagation. Yeah, yeah, thank you. Bone conduction.
Where it's processing and we as providers are going, oh, we're getting the answers so we can have a face-to-face crazy thought conversation with patients that is information being presented to us before we ask. Now, it's going to require some stuff, you know, there's some predictive inference engine engines that have to happen. And that's kind of what you were talking about in that Chinese study. Maybe some of these inference engines really weren't that great yet, or maybe humans didn't trust them yet, but that's like everything else that just takes time. And we've talked about how we're going to go from a role change, and you and I were doing a presentation a couple of months ago, that our roles are changing from seeker or searcher to editor, validator, and translator.
And I think that's a really big piece of this is, and I think this is a good lead-in to our fireside is, you know, I believe, Rehan, that the future of healthcare data isn't search box. It is an interface that knows what we need before we know we need it. So we can focus on providing better care to the human being and sitting in front of us. And I think that's going to be the foundation for the AI native practice that I, you know, I can't wait for you and I to talk a little more about. Yeah, this is a super, super interesting things you're bringing up. It's sort of mind, we had a innovator talk with Nitin, who started, you know, who started Maximize.
Maximize, yeah. And he's deep into this stuff and talking very long about the heads-up display. So for our listeners, please check out that. And if you're interested in this topic, he is at the forefront. Obviously, he was into AI in the 90s, you know, way back before any of us were even thinking about this stuff. Well, the thing about the technology that we saw we were in when we were in Milan, I mean, heads up with bone conduction and all of that stuff. That's not fantasy. The early stages are here right now. Yeah. Yeah. And I really like the idea of surfacing information before we even know to look for it. And it kind of reminds me of all the AI dashboards that are coming up.
If you look at any solutions, there's just so many dashboards. It's really interesting. But there may be a little bit dashboard fatigue. And I would just love for the AI to tell me what's coming down the pipe. And I'd love to see some examples of this as it develops. But talking about the AI Native Clinic, this is something I know you are not only thinking about. Oh, I'm designing it right now. I have blueprints sitting in front of me, Rayon. Yeah. And so if you're a clinician, let's get concrete. If you're a clinician in the United States, and suppose you're in private practice, you have some control over equipment you buy, and you're thinking about how to either cut down costs or improve efficiencies or have a better customer slash patient experience, what are the things you're thinking about?
I mean, it's hard to say, I'm going to start an AI Native Clinic and buy all new equipment and do all this sort of stuff. So is there a middle ground? Or do you feel like, no, you just have to go all in and say, I'm doing an AI Native Clinic? Well, I kind of have the optimist answer and the realist answer to that. I'm going to give you my optimist answer first, which is, I think there hasn't been anybody who's really developed the next generation model yet. That being said, I hope to see one in Colorado here in the next two, three years. But I think that nobody's developed the model yet. So it's all at this point in time, just us fantasizing about what it could be.
The Chinese paper, I didn't read that. I have to say, Rand, I'm super, I wrote down a note here for the audience. I got to get that for Rand because I got to read this because I think it could probably save me a lot of frustration and errors. I kind of developed the next generation of practices. I think, but until we have a model of what new and different look like, it's all in our heads right now. So how do we transition from all of the friction and frustration and broken model we have now? I think we all know what those problems are. That's first step. I think that none of us really have a good concept of what the future looks like. We talk about this model.
We talk about not search, but anticipatory computing. I don't think we know where the model is that we're going to yet. So developing the between A and C, A being now, C being later. How we develop the B practice, I think we don't always know the answer to that either, but I think that we've come to talk about we have to move that direction. So you're talking about, we've talked about all these different kinds of AI and our different podcasts and different conversations here and some of the things we've done is like, so what do we do in terms of intake AI? So the triage bots, and we talked about that a couple of weeks ago, what are we doing in terms of coverage and eligibility engines and what does AI look like for that?
In fact, one of the other big news pieces of the week was, and if you saw that the federal government now is like, they have a probe out, the Senate's created this probe to say, why is the insurance companies denying everybody with AI when the AI is processing and not even looking at all the data? So that's probably another discussion for news here in the next couple of weeks. But what is coverage and eligibility inference engines look like? What does diagnostic technology look like? You and I both have seen in the last couple of months, I mean, some technology that's just, I think it completely revolutionizes a practice. I was talking with one of my staff today, she says, well, what pieces of technology do we need in 10 years?
And I said, about one fifth of what we have now in a footprint one tenth of what we have now. A lot of it's going to be heads up, a lot of it's going to be at home, a lot of it's going to be in remote stations instead of having 32 slit lamps in a big practice, you're going to have four for procedural stuff because autonomous little lamps are going to be placed. I think that the biggest piece that ran to answer your question, what does the native practice look like? The answer is, I don't think we know, but I think we have to all start taking small steps towards that based on, and we've talked before, what's your biggest issue in your practice?
Go find an AI product or innovative product that's going to solve that issue. And slowly, you'll solve that issue, and some other new issue will pop its head up, for sure, there always is. There's always a new bottleneck. And then we're going to grab, okay, what technology is going to fix that? And we're going to slowly all move towards the AI native practice. But I think what you, I was hoping you and I want to talk about today is, what might that native practice look like? And I have some ideas. I mean, I think history does not happen in an exam room. It doesn't even happen in the four walls of a building. I think that's something that's going to have to be driven by AI.
I think that diagnostic testing, a large part of that won't happen in an office. I think it'll happen in either week. A couple months ago, I think we did the future of remote diagnostic testing in one of our real talks. And for those of you who didn't listen, I think that was a great podcast to get some information about what this means. But yeah, I don't think testing is going to happen in office. I'm not sure that all of education is going to happen in the office. I think a lot of it's going to happen, handheld, heads up display, all that kind of stuff. The one thing that I don't think is going to disappear, the location might change, but the process won't change, is the human interaction between a provider and a patient to say, let me help you, not data gather, as we've talked about, but let me help interpret what all this means and help you develop a game plan to do that.
I think the AI Native practice, I don't want to say it's wall-less because I don't think that's true, but what's on the walls and what's on the ground is very different. Yeah. So sometimes we talk about, and I mentioned this in the news of the day, it's not a bolted-on solution, it's a totally AI Native clinic, but in practical terms, and Scott, you mentioned this, no one, very few people have the capability to start a new, total AI-based practice. I think that's an interesting thought experiment. It's very speculative, and it's hard to even think about what that will be. That's expensive. And it's expensive and today in 2026. So for our listeners, I would say exactly as you said, I think for now, bolted-on solutions will give you a taste and may provide some efficiency gains and some cost reductions.
I guess the point of the article and what we're saying is that the transformation, the revolutionary type of changes that you would see would probably will not be delivered by these bolted-on solutions because the true efficiencies are when AI is embedded in everything. In everything. But that's kind of speculative. It's sort of fantasy land, and that's not the land that we live in. We live in a real clinic with real patients, with real techs. I loved what you said about, what do you call it, tech X or the commander in chief or the field general. I mean, but think about we're developing chatbots. We're developing eligibility engines. We're developing patient education bots.
We're developing scribes. What we don't have is the field general that controls. What did we say? We had a different name for it. But we don't have the conductor. That's what it was. We don't have the conductor yet. That piece, that brain, that algorithm set is still missing. But I think a lot of the other agents are starting to show up. Yeah. And so I'm looking forward to that future. I think we're going to need like image management software and our whole redesigned EMR and maybe not even an EMR. I mean, the way we're thinking about it, maybe I totally reimagined EMR. And I really like the idea of dislocating different parts of the quote unquote exam.
And what does that even mean? And the history can be asynchronous and not at the point. And so this is very much getting to fantasy land and sort of like not clear. But it's definitely not going to be the way it is right now. That's 100%. For sure. For sure. But I think we're getting glimpsed. And what I love about this paper is that we're getting glimpses. And I hope we're going to get a lot more papers about people doing it in different ways and get hearing about different experiences. So if you're listening and you are thinking about running an AI clinic or have an AI Native clinic or just trying out different solutions, please let us know and tell us what you think.
Scott, I think this has been a great, another great episode talking about AI Native clinics as well as the Knowledge Bite. As always, all our information is at ainicare.com. You could reach Scott S-C-O-T at ainicare.ai and myself, Rehan R-E-H-A-N at ainicare.ai. We look forward to talking to you again next week.