Welcome to Real Talk on AI and I care with your host, Dr. Scott Morris. And I'm co host, Dr. Rehan Ahmed.
In each episode, we discuss current news, adding a little AI education, and debate innovative AI tools and topics that will change the industry. Well, welcome everyone to another episode of Real Talk. My name is Dr. Rehan Ahmed.
Hey, this is Scott Morris. Rehan, it's been a relatively slow week in AI if that actually exists. No great news, no earth shattering news today. Just everybody's kind of get their heads down.
Heads down. Yeah, we're all, there's a big snow storm coming. So maybe we're all getting getting prepped for the freeze. That's for you guys in Colorado.
We got nothing, nothing. You guys are going to get more snow in Texas this week than we get in Colorado up in the mountains. That's crazy. You know, and we sort of a society sort of breaks down when and when temperature goes below freezing.
So we hope everyone has a safe few coming days coming up. And but I wanted to have you talk a little bit about AI models. That's something we've talked about a lot. We sort of referenced the term the phrase, but I don't think we've actually spent time really dissecting what that term means and what is an AI model.
It's definitely not a fashion runway model. That's different than that. But Scott, why don't you illuminate us on what an AI model really is? Well, I'll do my best in describing this one because it is something I've been thinking a lot about lately.
It seems to be keep coming up everywhere, right? Whether it's GPT or any of our podcasts or facial recognition stuff. It just seems like. So what actually is it?
You know, is it a physical model kind of like you said runway model? Is it just a big list of rules? Well, let's try to strip that away a little bit and see if we can't figure out what that is. So at its core, you know, an AI model isn't a robot with a brain.
It is just a mathematical program. And so think of like this, an A model is an algorithm that's been trained on some massive data set. We've talked about that before about pattern recognition and all this kind of stuff. But it doesn't know what let's we always use the analogy of a dog, right?
It doesn't know what a dog is the way you do. Instead, it's looked at millions of photos and learn that this arrangement of pixels with a tail and with ears and four legs and short is usually a dog. And so it essentially is this giant, big, complex game of let me match all the patterns. I don't know we've talked about that before.
So what's a model? Well, a model is saying we're going to build a mathematical environment, if you will. And that mathematical environment is what we're going to train our staff on. And it's, you know, instead of real life, it's we have a little bit more guardrails that we can create, whether that be good or bad.
And it usually follows kind of a three step model. So when you hear or three step process. So when you hear about modeling, know that in the background, the first step is we're building this model based on some data. Now that data may be thousands of piece of information.
It may be millions, it may be hundreds of millions. And that depends on the. Basically the volume or the scope of that particular model. And then you do the training where the algorithm and there's different levels of training.
I know we've done this podcast. We've done this knowledge by before in different forms of training, but as a general rule, the algorithm looks at all these and says, I'm going to make a guess. So let's use that dog example says, well, I think this is a sheep. And you know, the process goes, no, that's not a sheep.
It uses something called back propagation. And for any of you who listen to Ron is last time we had Ron on real talk, she did a great example or great explanation of back propagation. But if it's wrong, there's back propagation that says, hey, we need to nudge the weights because guess what a dog doesn't look quite so fluffy and it's not quite so round. It looks more like this.
And so it says adjust the math. So instead of, you know, hey, it's got four legs and it's fluffy. Maybe fluffy is not as important as the four legs with tailpiece. And so it adjust the math and then they do it again and they keep doing it and keep doing it and keep doing it over and over and over again until the model goes, yep, I can identify dog.
99 100% of the time. And then that's kind of the finish line. You know, once the math gets good enough and the weights are all correct. We then say, okay, now let's give this model new data that it's never seen before.
And we feed it a whole bunch of pictures. Never saw and says, okay, knowing what you know, tell me what this is. And does it get it right? And if it gets it right, great, the model's working pretty good.
We'll give it some more data as long as it keeps getting right. If it gets wrong, we go up. Guess what? Our model is not correct.
We need to go back and retrain it. Now, kind of the magic piece of this rayon and the audiences. It isn't that the computer's really thinking. You know, when we talk about artificial intelligence, we almost think like it's thinking.
And the answer, it's not really thinking. It's learning from experience. Now, that's a form of intelligence. And rather than following the rigid script, it's saying, let me just keep trying and keep trying and keep trying and keep trying and keep trying and make this unpredictable data.
More predictable by changing the weights until we get it right. You know, and anytime you see any part of AI, whether it be voice assistance or, you know, you're on Netflix looking for your next movie recommendation, whatever the case may be. No, there was a model somewhere under the hood that got trained time and time and millions of times over. And doing the math to predict what you want next.
And when you choose something, this is, it says, well, you've, you know, you've been watching. I don't know my wife and I watch and lay in man right now. It's like, hey, you like all the Taylor Sheridan stuff. And it predicts the Taylor Sheridan.
I go, nah, that's not what I want. I want to watch something fight with Robert Janero. The model goes, oh, man, we messed that one up. Let's retrain it based on whatever Scott likes that day.
And it keeps going and keeps going and keeps going. So this is a little bit about models and everything we do in AI is about a model. And I think in the future, and one of these I want to do in the near future is, well, what happens when the model isn't starting with real data. But instead it's being fed synthetic data.
So maybe that's what we'll do for knowledge by next time. No, it sounds, it sounds great. I think this is a great transition to go from sort of theory to application. And so what the fireside chat, what we want to talk a little bit about now is sort of how this actually pans out in the real world.
Meaning how do you use a model like an AI model and then impact and improve ultimately what we deliver to patients in the form of I care and health care. So what we're really talking about is the hybrid model of care and how AI is going to impact all of those. So when I think about hybrid care Scott, AI is going to impact before the patient comes in in terms of. Oh, pre screening pre identification of disease.
Of course, at the appointment with clinical decision support and then back end. So administrative support insurance authorizations, billing support to help to practice. So it's kind of like before the appointment at the appointment and then after the appointment. So sort of a simple framework that you can think of how AI will impact this sort of hybrid care and we already have hybrid.
That's a great way to think about workflow. And that's a great way to think about workflow. Yeah, and we've had a hybrid care before AI and we use technology, but it's just going to do it on steroids because you're going to have intelligence systems that can screen patients beforehand and then help you identify and decide what treatment you want to do for this patient. And then figure out how you're going to build them and do insurance authorizations and verifications, etc.
So there's. You know, I heard this great term. It's kind of explaining a little bit what you're talking about. I heard this term the other day and I heard it and I had to like replay the podcast because I was like, that's a crazy word.
What is that? It's digital. PHY GI T.A.L. When we start combining the physical world with the digital world, now we have hybrid care.
And I agree with you. I love you. That's a great analogy. How is it going to affect us before?
Maybe that should be our outline for this. How's it going to affect us before the visit? What's going to happen during the visit? What's going to happen after the visit?
So talk to me a little bit. Think. Tell me what you're thinking about what happens before the visit as we start thinking about a hybrid model. So I think we are starting to see these systems already being deployed, whether it's through a kiosk based system, like I check.
Or the one that's gotten a lot of attention recently that I'm really excited to have Alex on from I bought. I'm having a sort of a kiosk intelligent system that can identify patients with. You know, either refractive error or some type of ophthalmology. And then they get fed into the system into the patient into the clinic.
And you already know the key thing here is that you already know a ton about them. Right. This is not just a new patient calling. You already have some diagnostic information that's already part of the.
A part of the charting, you know, and it's probably an intelligent chart. And it's probably slotted with all the right diagnostics that should be done for that patient specific order at that time. And I know Scott, you spent a lot of time thinking about this. Oh my gosh.
Lots of times looking at. So you know, I think it's not like we've been doing models in terms of how do we do. Previsit differential and pre visit, preemptive. You know, test sequencing about what's going to be the most efficient way.
To get the right answer instead of us all doing the exact same test every way the same, you know, the same way every single time in the same sequence. That's not the most efficient way it's not even close to the most efficient way. But I agree. for the before visit, the pre-visit.
I think it's a great model for the next three to five years, right? But I kind of look and go, maybe even a little further down the line, is what happens when we can start taking data from other visits, your PCP visit, your neurology visit, whatever, and start integrating that with all of that information for the visit today. It's our goal. And we're going to look at this more holistically.
So I know I'm sure we're going to get into a calomix here in a minute, but it's almost pre-visit a calomix, meaning we're going to look at the whole body from the perspective of we have all that data. And then we'll start doing testing, which is what we think of traditional a calomix. But I think the hybrid model there is much of the pre-visit stuff is going to be done not in the office, but way before you ever get to the office. Oh, absolutely.
And I think this touches upon another fire side topic that we've talked about is data sharing. And this is a key element of the hybrid model. We're going to have to come up with systems that can actually talk to each other and thinking about fire and various other protocols of health care systems actually talk to one another. I know this is a bugaboo that we've talked about many times, but this sort of breaks down.
If that doesn't ever happen in a meaningful way that your PCPs and your hospitals health care system can't talk to the primary care doctor, can't talk to specialists, health care systems. And we are not reaching the full potential. And so I know we are both big advocates of data sharing and information sharing between various providers. And then the patient themselves having access to the practice.
Transparent access. Yeah, exactly right. All right, let's go to part two then, Ron. So you're talking about, okay, so what is the, when we think about the hybrid model, what is the, and I'm going to say in air quotes since we're on audio, the in office hybrid model look like.
You know, we've talked about this in terms of the office of the future. And I think there's a lot of connection between that where you have diagnostics sort of set up in a way. And it's done intelligently at the right time. And then you see the provider not behind the slut lamp, which is, you know, what we've been doing it for more than 100 years.
But in an office with a fireplace and a hearth or something. And it's just, and it's just them talking to you like a real human being. And so, you know, the hybrid, even that term kind of ballad of the hybrid model of care, it's like, as if we're taking two things, then we're sort of making some weird Frankenstein version. This is really just like, we're called like the better model of care.
Like it's just a better model where you have really smart devices, really smart AI and a very empathetic and experienced clinician. Man, that is, that is a great combination. It's not a hybrid. It's just, it's just better.
It's just augmented, truly augmented intelligence. You know, I wonder too, Rayon, and you know, for everybody in the audience, I'd love to get your feedback on this. I mean, you know how to get a hold of this at this point. Is that I look at that and go, maybe the office, the high, and I agree with them, that's your hybrid's the right word.
I just don't know of a better word yet. But, you know, will the patient actually come to an office? You know, we talked a little bit about this a couple weeks ago, is that, you know, will you have a diagnostic testing center that does all the testing, especially when we get autonomous sit lamps that are, you know, ubiquitous. I mean, that's three, four, five years down the road.
But, you know, we get autonomous sit lamps. We don't actually need to be in the same room to do the testing. We don't even need to be in the same country to do the testing. Will it become much like you said, where now the model of the clinic visit is virtual, and it's really about, are you, like you said, an empathetic provider that has all of that data at their fingertips, not just from this visit diagnostically, but what happened with their wearables?
You know, I mean, it takes telehealth and remote monitoring and digital communications and ties it all up and says, it doesn't have to happen in an office setting anymore if you don't need to have a procedure. Yeah. And so there may be some sort of centralized diagnostic suite set up. But even there, you wouldn't can imagine.
I just did something online where a robot, you know, a computer voice was like an AI was just talking to me for 20 minutes and I was responding back with questions and it was a type of sort of a continued education exam. And I thought, wow, this is gonna be the future because it was, its responses were dependent on what I said to it. And so in the same way, the intake will be done by some type of avatar or a computer listening device or an audio based LLM. So this is clearly the future.
And you know, one wonders where is it gonna, where's the human in the loop? And I think it's gonna be at all the critical junctions of that visit. All the place where we gotta be human to human for empathy and interaction in those things, I agree. All right, backside.
Once the patient leaves the quote unquote office, whether it be virtual or in real, you know, physical space, what does the backside look like? Well, from the, there's, before I go into the billing stuff, there's definitely a patient follow up tools that can be used through AI based interventions, AI using your drops, follow up after surgery, parioperative patients, compliance, follow up with reviews. You know, this is something that's really important to a lot of us, making sure that if they had a great experience, they sort of tell everyone about that. And that could be automated through AI and that's sort of a tangentially a hybrid model of care.
But really, I think what most of us are thinking about is getting the billing and coding and verifications all done properly. You know, this is the pain for automatically. And automatically. So we're not thinking about that.
And it gets done right every single time, based on what we did and what we said. And yeah, and the, and ideally, the chart would be updated on its own. It would be an ambient scribe that listens to the conversation that looked at all the charting. And we just sort of verify it, include our clinical decision making and it's done.
You know, and there's less burden some EMR aspects to our care. Yeah. And I will say, you know, and not to do a pitch for the project I've been working on, that exact piece for the last almost four months. And it's a possibility, Rayhan.
It is definitively a possibility. Our single biggest challenge once again, is the data share piece, right? Is there are certain entities out there that don't want to share that data. And thus it makes building that, the prospect of building that very difficult.
Because if you don't have transparent, you know, on demand access to that data, it's really hard to fix the problem. And I kind of wonder sometimes if maybe certain entities in our industry don't want to fix the problem. Well, that's a whole nother political discussion. That's a whole nother political issue.
But I think for us, the key take home message that we're trying to talk about is sharing is caring. And especially when it comes to medical charting and reporting and that's true because data not shared is data underutilized and it just sort of goes into a silo and no one ever takes advantage of it. So I think that's a key point, a great way to sort of end this fireside chat. And so we encourage everyone to share data, to subscribe to our podcast and to definitely give us a like.
You've been listening to Real Talk on AI and Icare. Your weekly podcast to keep you informed about AI technologies revolutionizing Icare.