This episode of Real Talk from AI and I Care is brought to you by Barty Software. 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. We're going to cover a couple of note worthy news items that have come out this week and then Rehan's going to cover a little bit about what is a large language model. And then we're going to get into kind of talking a little bit about scribes for those of you who may already be using AI scribes or those who are thinking about it.
Maybe hear a little bit about what Rehan and I are going to talk about before you make any of those decisions. Well, so the interesting news of the day, there's two kind of interesting things that come out this week. First of all, in fact, it was just launched two days ago. Microsoft is kind of unveiling its Microsoft Dragon Co-Pilot, which is its first AI assistant for clinical workflow that kind of takes natural language voice dictation capabilities and different ambient listening capabilities and takes those and puts them together for healthcare.
Now, it's an interesting thing that you have a big fish like Microsoft get behind this. But interestingly enough, as we're going to talk a little about what the advantage of disadvantage of scribes and we'll kind of see where that lands between the two of us. And then the next interesting thing of the day is Google just got FDA approval. In fact, it got approval today, March 11.
For its new smartwatch as an FDA device to FDA device to track heart monitor and be able to feed that information directly into cardiologist office to build it track what. How heart rates work and pace and all those types of things. So we're starting to see some of the wearables that we talked about in the last week's podcast. We're starting to see some of the wearables be now something that are FDA approved devices for the medical management of different conditions, maybe not an eye care yet, but interestingly enough in cardiology, it's there.
So that was a couple of interesting news pieces of the week. This episode of Real Talk is brought to you through the gender sponsorship of jobs and publishing. Publishers of leading sources of information in the eye care space. Ray, tell us a little bit last week reference large language models.
Tell us about what one of those are what what a large language model is. Yeah. So I really love these segments because I think it brings us all up to speed on these terms are just sort of thrown around Scott. And we don't always know, including me, we don't always know what we're kind of talking about at a little deeper level.
I'm sure most of you have used chat GPT if you haven't please use it's free. It's it's really amazing, but there are other services like Gemini from Google or a service called Claude, but and there's several several more and Grock from X and of course deep seek, which I wrote about in the in the journal. So LLM stands for a large language model. It's a model.
It's a type of AI that's trained on massive amounts of text when I when I mean massive. I mean basically the whole internet. We're talking books, articles, YouTube videos, other websites, research papers. And I just have to put a plug in it.
There is of course some lawsuits, especially from like newspapers like the New York Times saying this was done without permission, but just putting that aside, talking about research papers from basically every journal ever. But here's the sort of the key thing. It doesn't actually understand the way humans do instead it learns patterns of language. It's like the example I like that is it's like auto complete on your phone, but super charged on steroids to the maximum to the empty degrees.
So when you type a text on your phone that predicts next word, LLM sort of do the same thing just on a much bigger scale. They take the input. They predict the next likely response, not just next word, but within the context of the paragraph within the context of what you're talking within the context of the last question. And it uses its training data and it generates response.
That sounds human or Frank it sounds however you want the output to be. So the question is how do they think and the answer is of course they don't it feels like it's thinking, but it's not reasoning like a person is not forming a opinions or original ideas. It's just really good at predicting at what should come next in the sentence. Imagine you're playing a game and you have to guess next word in the sentence by say some like the sky is dot dot dot.
You probably guess what would you say Scott the sky is. Blue course you're an LLM right because you've seen millions of patterns just like this. That's what they all I'm doing just entire paragraphs essays conversations. This is why I could sound human but can make mistakes or even and this is where I have problems.
It confidently gives false information. This something called hallucination. I'm sure many of you have heard this problem because again, it's just predicting the next response. So where do LLM shine, where do I use them a lot and I'm sure you I would love to hear how you use them.
It's fantastic at summarizing complex information. It writes emails reports and content super quickly. In fact, I was just I was reading a sort of a competitive landscape article I wrote just three years ago. And it was like, it was actually on top of it or an opsy.
And I thought man, I chatted with you that it's on the same thing I did that I just spent like weeks on. So it's really good at that. It's very good to answer questions and explain topics in a conversational way. I often say, hey, explain this to me like I'm like a five year old and it will explain to you like you're a five year old.
It's really great. And I use it a lot for brainstorming creative writing, publishing emails. I basically use it every day. So what are things to know?
It doesn't actually know things. It recognizes patterns that we went over that. It just it could be biased because it's human generated data and it has human generated data has biases. So it will just we'll just copy those biases and you have to be careful about that.
It will make things up like I mentioned and it doesn't understand motions or context and sort of the way human does. But overall, where is all this going? It is definitely getting better. I started using chat EP at the earlier iterations and now it's at a model called for or maybe the next one's coming out.
Scott, I think you mentioned that last week and they're getting really, really good. So I encourage everyone to use them. I use it often within healthcare in terms of I'd use one today to write an op note at least to help me write an op note. It was fantastic and I think it's just getting started.
So that's your quick couple minute background on LLM's. Feel free to research on your own, probably within LLM. It will give you a kind of information about on its own. So until next time and the next topic, I encourage you guys all to work on it yourself.
The best way of learning is really by doing in this in this area. Scott, which one do you like? Yeah, I kind of mentioned this a little bit last week. If you'd asked me a year ago, I would say, okay, I dabble and chat you be a little bit.
Now, I mean, I use Claude for some things and I use Gemini. I use Gemini a lot in place of a search engine. I use chat to be tea. I'd say I use a bunch of different ones and I'm starting to get the feeling like I use one type for one thing and something different for another thing.
But I use them all day. I mean, whether I'm helping write an article or proof an article or doing research or a virtual ad board or I just I'm finding so many different uses to give me just great ideas instantly. I'm sure I could sit down and 30 minutes and figure it out or I could do it in 30 seconds and have some things I never really thought about like we're going to talk about AI scribes in here just a second. And you know, I was really helpful to go, oh, I hadn't really thought about that and then it gives me some context.
I use them as kind of a brainstorming help rayon more than anything. So yeah, I literally look at it now and go, I don't know how I'd get you the day without it. This episode is brought to you by Barty, the AI powered all in one EHR built for I care with AI scribe voice over internet phones websites payment processing in our CM all bundled into one system. Barty is redefining modern I care spend less time clicking and more time with patients at Barty dot com.
Well, let's talk a little bit about kind of talking AI you know, we talked about some of other podcasts rayon and I have these conversations I have a much like you ran I have these conversations on every day with some really interesting people. About what's really truth and what's real and I think sometimes we want to talk about all the raw raw good stuff. Today, maybe we're not going to do so much raw raw. Yeah, today is going to be about kind of the realities of AI scribes and what is ambient AI and is it what we want to use.
I don't know. What are your general thoughts on that one? Yeah, this is where I think there's been so much hype and promise of of scribing because if you talk to any doctor on where they feel like they have a lot of pain points. It's documentation and it's chart documentation and you think well, what this is where an AI scribe would make a lot of sense and the reason we say AI because we've been using scribes for you know decades I had a scribe at a prior practice.
It was fantastic. The scribe is basically a real person oftentimes, you know, maybe you do someone in their 20s coming in and they listen to the conversation and they basically write down what you're doing and they write down the charting and they're really good at it and the. The scribe is a real person. I mean, it's fantastic.
I have we have four scribes in our office and I couldn't practice without them because I really want to sit there and spend two minutes, three minutes, four minutes five minutes, concentrating on the patient and not worrying about turning my back and writing something down. I just can't even tell you the last time I turn my back to a patient wrote notes on a chart or in an electronic file. I just never do it. But you know, I think there's goods and bads and scribes right and you said, you know, my scribes, my best scribes, the ones have been with me for 10 years and they pretty much know what I'm going to say before I say it most of the time and they've documented I get done OK, you got all that and they're like yeah we're done.
And so there's no more of that crazy documentation that we did before and now all we're doing is taking that human piece and maybe moving it to some of the I stuff but I think there's problems too, you know, is our people can do this but sometimes there's the data that they take in the data to listen to is incomplete or it's inaccurate because. You know, AI is still learning how to how to have a conversation and some of the nuances in speech, I know that when I speak the first word out of my mouth, the first syllable is super fast and AI sometimes has a hard time doing that so I've had to learn how to slow down certain words that I say, otherwise the AI can't pick it up. And I think there's medical jargon like there's traditional jargon that we might say, oh well everybody knows what an I well is right but AI looks at it and goes, I don't know what I well means right in fact I just did it and it's running a transcript of our conversation and I said I well and it didn't know what it was it missed it you know so so perfect example right I think that there's some of the data is incomplete and accurate and then there's fluency issues I mean I was on a conference call. Early this morning with a group from Israel and the scribe that I was using did an amazing job transcribing it despite some of the accents and fluency issues that we are going on but on other times that you know I've conversations last week with I have a clinic in in Mexico and I was speaking to my to one of the staff people there and the AI that was doing it had a horrible job because her accent is not traditional Mexican Spanish and it was having a really hard time with the fluency issues.
It's a really hard time with the fluency of getting it right so I think there's problems still with the technical issues I don't think those problems are going to last forever though Rayhan I think that these are it's getting better all the time. Yeah this is I think a lot of things in in AI I think directionally we all know where it's going it's heading in the right direction but to be honest at least in in ophthalmology and I imagine the same in optometry I don't know of a single status friends I don't know of any ophthalmologist and I don't know what I'm doing. I'm just going to talk about the ophthalmologist and I did a little survey just to talk about I don't know anyone who's actually using an ambient scribe AI ambient scribe and frankly I don't know if any system and there's several out there's a system called like Suqui AI bridge, on meddics a lot of different ones maybe for for sort of other medical specialties but not have integrated with our EMRs as far as I know to my knowledge I would agree with you nobody's there yet. And I think that there's some good reasons for that I mean I think that you know we didn't really talk about data privacy you know and people go what happens if there's a data breach well you could say the same thing what happens if there's a data breach in your EHR and there's a lot of things that I don't know.
And there's been enough of those over the last few years all that information's out there well sometimes conversations carry much more content than what we might type into it because we I think when we listen and have a conversation with the patient many times will take the short points to put into the EHR but the wider more detailed description might be in our brain. But we don't write it down and I think that's good and bad right I think the ambient AI may at times pick up things that we kind of forget to write down which should be really helpful and at times it may pick up things that we don't want to write down kind of double edged sword on that one. The other thing is I think a lot of the EMR systems make it really hard in fact I don't think I know they make it really hard or their customers to create apps or integrate with their systems I'm probably think of one of the big ones in particular I've heard a lot of small. And I know which one you're thinking about the other big one that I didn't go yeah they don't they don't want things to integrate.
And so I think that's been a hurdle for some companies but we're getting we're on the way I'm just excited about when we're finally going to get there but that's definitely an area of hype where we have not delivered on the promise and so I'm looking forward to hearing from companies that are working on this and how the larger community can help them get there. I agree and you know I think the last little part about that is I think there's still some hesitation among providers probably as much as there are companies about well if it does record something wrong and especially as we start leaving I were just talking about this before we started recording is you know there are certain apps out there right now that I used to record lots of my pod or lots of my discussions with other people so I can take better notes than me sitting there right and not paying attention to the conversation. But in the world of medicine if it does transcribe something incorrectly who's responsible is it the company is it the AI scribe is it the doctor who didn't prove check it and when we move to a model where some of that context of what's in conversations will be used to train generative AI later on a couple years down the road well if it was wrong and now you're training the generative AI on wrong data who's responsible there's a lot of ethical questions. You know that are built into that and the bias that it will record of doctors making decisions that are either incorrect or biased that's that double it's sort again but maybe there's going to be more transparency to that too and maybe that's going to change the way cares delivered I think I think there's lots of positives potentially about the future of scribes but there's some ethical and privacy issues that I think we we have a little ways to go to work out yet.
Definitely agree I think great conversation on scribes a lot more to discuss in the future. You've been listening to real talk an AI and I care you're weekly podcast to keep you informed about AI technologies revolutionizing I care. We would like to thank birdie software for supporting this episode of real talk from AI and I care.