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. Hello, everybody. Welcome to Real Talk. This is Scott and Rehan.
We are looking forward to have today with some hopefully some interesting stuff. I'm going to cover a little bit of news of the day. What are you thinking you were going to cover Rehan for the knowledge bite? So today we are going to be talking about voice agents like Siri or Alexa.
Awesome. And then for Real Talk, the kind of our fireside chat, we're going to cover a little bit about a really big subject called data. Like where does it come from? What do we do with it?
It is essentially the food source for AI. And sometimes we've got to wonder where our food comes from. So here's a couple of things for kind of news of the week, I guess you would say. And there's been some fascinating stuff coming out in terms of companies and countries and areas trying to regulate AI.
So California, you know, they're really kind of trying to regulate their AI giants, as you'd say, with some of the new bills they're trying to pass about where. How much that where does that data come from and how can it be used? And then the UN, and this is when you know when the UN starts getting involved that maybe we're on the right track with AI, because they're saying, hey, they come out with a statement yesterday. About how there needs to be development has to comply with international human rights.
So it was a little vague on descriptions and comments, but it's just a general statement saying, hey, I think we need to do this now. On the other hand, the GOP just sent out a bill and that would come out this morning. They're launching a bill to say, how do we protect the AI company from lawsuits related to transparency? And I think that's when we'll start getting the government good or bad interested in giving us some bylaws, some guardrails, you know, about what are in those situations, situations, situations like medical or financial scenarios.
What does that look like in terms of there? So I think some fascinating, not necessarily like one particular company, but a lot of information saying, hey, I think the regulation and what the risk are people are starting people, countries, companies, industries, governments are starting to have those conversations about what do we need to be worried about? And what are the guardrails? So a little bit of news of the day.
Brian, go ahead and kind of knock out our knowledge by let's talk a little bit about voice. Yeah. So this is a really interesting topic. It sort of hits on sort of a theme or like a thesis that I think we've sort of been talking about a lot on this podcast and I think in the journal as well.
And that theme is essentially that AI is going to not detract from our relationship with our patients, but it's going to enhance them. And one of the ways we talked a lot about visual models last time and LLM's and we've been sort of building this sort of increasingly set, I think of sophisticated knowledge. And today we're going to be talking about another element where the technology is enhancing that relationship. And this is voice agents, right?
And not like the old voice agents were like you call and it's like a terrible experience of talking to someone like you talk to a voice bot on air. And not to pick on a like 18 to your whatever, like I think I was on the phone with them for a while. It's like a terrible voice bot and they're like, I know this is fake and I'm just going on my phone anything zero zero zero zero zero. That's like all the songs possible, right?
Until I get to a real human. No, no, we're talking about modern voice agents. I mean, I think like theory or Alexa, but even but better. I mean, I think those are those can definitely be better and some of the new stuff coming out by suddenly startups like 11 labs or whoever like a bunch of other ones.
They're really good. And in terms of health care, I mean, we can think about the tons of possibilities. But how do they actually work? So it's there's there's like basically six steps in sort of taking this speech and turning it into action.
And so the first step is sort of like this kind of basic idea, but like the wake word detection. That's, you know, every time I go to bed at night, I don't actually set. I had, you know, I have my phone next. I'm like, hey, Siri said an alarm for like 6 am, right?
And that's there. That's what wake word detection is. There's a lightweight model that's running on the device and only up this word has heard. It can start activating that speech.
So imagine you have an EMR. You could be like, hey, EMR or. Hey, Alexa, Google, I mean, they'll have their own wake word detection. And so that's sort of how it gets.
How it starts to know to listen, right, has to have some way. The second step is the speech recognition, right? And this is where the job is to transcribe that speech sort of into text. And the really interesting thing here is that it needs to recognize all sorts of different things, right?
Like accents, right? So I'm in Texas, right? There's a text in accent, but you know, there's also a Boston accent. There's a Southern accent.
There's a Midwestern accent, right? And then it has to distinguish between like kind of like we have things we don't even think about like a human like you would not like my five year old. We would totally get like. Between homophones like site s i t e like like a surgical site or research site.
Versus site like what we use in vision s i g h t right has to figure that out. And then of course handle ambient noises, you know street traffic people talking the background, etc. Right? So think of an example, right?
Like you're with you. You're like, Hey, can you refill my use a hay mark? Can you fill refill my I drop prescription at CVS? Right?
So then it has to convert that into being able to know that it has to do some type of action. Right? So it's using natural language, understanding to see what you actually meant it to do. So it has to figure out your intent and then figure out, well, okay, what are medications like I draw?
Right? Where's location that CVS? You may say something like what's the dose for Timilol? Well, it has to get an intent.
Okay, you're wanting to retrieve drug information or you're telling the EMR to schedule fall up to two weeks. Well, though, there's an intent to create an appointment. There's a timeline of two weeks. It could it could get even really good, right?
You may say something like, oh, this patient's out of drops again. And then sort of into it from that that that's a refill request, right? And infer. And there's all sorts of interesting things where we have to we're we again, the power of a eyes it's doing things.
That humans just sort of do right? And there's something like Bert like embedded. I forget it. Birds are very long acronym, but it's basically like.
If there's a sentence that says she went to the bank to sit by the river, right? Or she went to the bank to deposit a check, right? You the bank, the word bank there is is not clear. It's it could be it could be a bank that you deposit.
Checking the bank by the river, right? And they're just different different meanings, right? So you have to know what the word before after bank is that word and figure out what you're trying to say. So there's a lot going on behind the scenes to figure that out.
The fourth step is doing a dialogue manager like going connecting this through a computer term called an API which stands for like application programming interface, but basically it's a function, right? Like it takes that refill request and then it goes to a backend system. It does some context analysis. You may say schedule appointment with you know, Dr.
Morris, etc. And so it turns that all after figuring that all out and then it turns it into some type of action and integration with EHR or a PIM like a practice management system or clinical decision support. And then and then this is what we talked about a couple episodes. It actually has to compose some natural language generation, right?
Where it can be, you know, telling you your next appointment is whatever July 12th, it could be instructional telling you how to use I drops. And then it could, you know, I think the stuff that's getting really good, it could be empathetic, right? Okay. There's the real chance.
Yeah. So if you listen to the really good voice agents now, they add hesitations casual phrases. Just what it is right now those qualifiers right because you don't want it to sound robotic, even though it's a robot. So, and I'll say stuff like let's take a closer look and just just be normal, right?
A human normally talks, not not a robot. And then the sixth step is of course going doing the subtle disfluencies, like I mentioned, to add realism, make it relatable and then turn it text, you know, the whole text to speech model. And there's a lot of great tools online that do this. One of the, you know, one of the ones I like is the called 11 labs, but there's so many available now that are really great voice agents.
And the, we've talked about this before, but the opportunities are endless here. They're like becoming not, they're not just assistants, they're collaborators. And it's that that's the topic we've talked about way more than just receptionist where they're taking appointments. They're going to be in there with in the exam room, really helping you again enhance that relationship with your patients, you're not staring at your screen.
They're not even staring at their like examining guys, you, you're able to look at them and talk to them. And there's a voice agent that's assisting you in the background to allow you to spend even more time with that patient and deliver more personalized service. So that's a knowledge bite for for tonight. Super excited about voice agent and trying to just bring this topic every so we're going from from beginning.
So the first topic we were talking just about what is AI and what are algorithms and how we're just marching forward to to this and more topics in the future. I think that's fascinating stuff, right? I mean, I think about voice agents to go and you were addressing it from like what happens in the office, but what happens when our consumers are patients are starting to say, hey, I'm almost short on. I'm short on my met form and you know, Siri, whatever in case maybe could you order some more, some more of that and the voice agent takes that turns it into API sense at the pharmacy.
I do think that voice agents are really going to change the way the world works. And I think that's going to be the impetus for a lot of change is going to happen inside the clinical wall inside the commols of the clinic as well. I agree with you. I think that that's a huge part of the agentic AI of the futures what we do with voice AI.
Yeah, absolutely. And this goes goes so much further than just being receptionist. I think that's I think that's a low hanging fruit and that's where a lot of these tools are being developed. But the opportunities as they say are truly endless here.
Fascinating stuff. Well, so right today I want to talk a little bit about data. You know, it's been one of those things that seems to keep coming up in conversations I have with kind of everybody going well, Scott, that's all great. I mean, AI is wonderful, but where are you getting the information from and it makes me kind of pause and think where really what are the challenges we have with getting data because I think you and I've said this before.
And what is the is the oil of the future, it is the currency in which AI learns grows gets better and it's the currency how we work too. But right now I you know I always mentioned and I've says numerous times is I think that we're we're not even scratching the surface of data. I think we deal with less than 1% of all the data that's out there in all of our decision making and there's so much more information. So I think about what some is asking me that this weekend they say, well, we're going to get the data I say well, obviously there's the HR's.
You know, and then there's all the medical images we take in and I care we're so heavily on the imaging side more so than probably just about any profession in terms of imaging, whether it be OCTs or, you know, visual fields or topography or a barometry or. You can even say perimetry to some degree and I think as we start getting into and to your segment imaging and looking at cornyas and lenses and into your chamber angles and aqueous flow and occipit symmetry we talked about a little bit last week. You know in terms of what is blood flow look like and what's metabolites and I the data is going to be amazing if we can access it then you have like genomic data I mean if we think about you know I'm not trying to do a pitch for the failing 23 and me or anything like that. The genetics part of that is really fascinating as we get into personalized medicine and you know I don't know about you right I think we're a little ways away from the genomic data but but it will happen.
And then you know I think it was two or three weeks ago we talked about the wearables right I mean you think about now everybody's got a watch a ring and soon we're going to have true wearables in the eye industry which is glasses right I mean smart smart smart smart lens wear smart eye wear that kind of stuff. You think about all the data that's going to come out of those about what's our sleep pattern like does that have an effect on our IOP pressure when you can start coordinating that kind of data that's pretty spectacular in terms of clinical care but then you can also look on the backside is you know I want to do a whole fireside chat on. Ed is based medicine and value based medicine I think we're building towards that with some of our content but you know you think about though it's not clinical the data that can be offered in terms of looking at what codes were build how often they build what treatment seems to work and what treatment might be most cost effective is it is it always that a fourth generation floor of Quinoa is better than poly trim I don't I don't know I mean according to all the drug companies it is but we really have data I mean I don't know if we do and maybe that's going to be administrative data or billing data that helps us figure that out I think there's going to be a joining of the operational data with the true clinical data but there's a lot of places the data comes from and right now we can access almost none of it. Right I think one of the really interesting and they're you're right there's a ton of data one of the really interesting things is I saw this on informed consent recently for a research study where they talked about your data may be used as part the identified of course in an AI and that was like it's just one little paragraph that they had towards the end I'll come.
People are definitely getting more in tune with the value of data and the value of their data and so this concept I don't think I have any answers here but data ownership right when you go into a clinic you're my clinic and you get a fundus photograph right and I don't and I've heard so many different viewpoints on this and I haven't really heard a satisfying answer but this whole concept of data ownership right the patient the clinic the company that the device I know we talked about this before but I just think that's such a fascinating. But who are you right now we as providers on the data that's right but I think that's going to change we've talked about like a pys systems and when there's patients own their own data now I'm a sudden if you use it to patients get compensated for sharing their data that trains the AI. Yeah I think society is going to be more sophisticated to think like hey if my data if my wearable data my my eye data my health data is going to be used to train your AI to then sell more products and make some money using collectively our data do I get a you know are you going to send me a check. Do I get well it's almost like you're getting more interesting right I mean put your money in the bank and they pay you you know two cents every month for the money out of the bank.
For the money out of the concept right and I think people are going to our understanding the fact that you know it's one of the fine lines that if the product is free that means you're the product right for something exactly right. So you know I think people are starting to recognize that and then maybe the case that you the service of providing something so valuable that you're that the customer patients willing to give away that data. And that's you know I mean where do we sacrifice privacy versus data versus information I think that's a whole you know ethical comment we talk about but you know I I think about all the data that's like in an EHR right I mean so and there's the problem is is all these different EHR's none of them talk. None of them are accessible by anybody outside you know I think I've mentioned this before like my the surgeon that I refer to rayon he uses the same EHR I use but our EHR's don't talk yeah how is that possible in these days and times right that just there's all everybody's got their own proprietary data formats and you know their own coding system and all this is really done and we've used this reference before is create these huge data silos that nobody can access the information just.
Basically sits there in rots because we're not using any of it and and you know we got to figure out how do we get access to that how do we try if if maybe that's not even what we do I mean. You know the challenge with the HR's and I think we talked about this in our privacy conversation is that you know the challenge with the HR's is each are can't share that data because it's not actually the patient didn't give them the right to do that nowhere in the nowhere if I collect an EHR if I collect put. Information into my EHR nowhere did I ask the patient if I could use their data. Nowhere right so the challenge EHR's have from a legal standpoint is it's not really data that they even if it's anonymized it's not really data legally that they have access to to go train an AI and so there's going to be you know on this one of those I think I hope this doesn't end up me being an expert testimony in a court case somewhere down the road but I.
I think that EHR's can't use that data it has to be something patients give access to to build to use it you know and even if they could do that. You know I don't know how you you know your abbreviations for certain things might be different than my abbreviations for certain things and if the abbreviations aren't all universal then they I have to have more training more data more processing to build a say hey you know H. TN is the same as hypertension right. I think that's it and then there's all that unstructured data and that's one of my concerns we we talked about this when we talked about.
Scribes right that's a lot of unstructured data when you just start recording our voices and that's where the really good stuff is but how do we train a eyes on what unstructured data is important and what is not important so this free flow text of clinical notes that's really going from talk to text and there's so much detail that I mean there's so much information but what of it do we extract that's actually meaningful I mean what is it quantifiable is it qualitative I mean then we get into a whole you know that's like we talked about I think it was couple of weeks ago about NLP I mean that's a whole nother layer of really complex interpretation so I think the challenge with the current system we think about EHRs is unfortunately I think the data is largely unusable you know I think it's going to take as much time to figure out how to use it is to actually use it. I you know I don't have the answer for that I think that getting data from EHRs is going to be really tough. Yeah and I think there's and even once we do right AI systems gobble up data like there's no tomorrow right and so it's already been talked about have we hit peak data like have we have we already exhausted the internet and so there's all this talk about now synthetic data where AI systems are creating date now imagine like a fund this fund is you know fund is image sets now. Couldn't they just create a whole bunch of other fun because you don't have us so so you know you're trying to build a fund this image set for Parkinson's or Alzheimer's I just don't have enough of these cases well you know how would it work for you to create synthetic data right to develop that database that you almost like the virtual twin that that right now that exactly right right and so.
Do we create a virtual trend that creates its own synthetic data right right is that data real is that data real and more you know more play is that data representative of the data you're trying to model it off of because because if you're just hand picking a few samples for it to model off to create new synthetic data if there's something off on those well then you're just. Bicing the model to get you know toward that toward that handful of. But it was there may not be a study that we create right now you know I mean I look at studies and go. You know and this is me being the editor of a couple different journals over the years and having to read all the time I'm like if you really read a study most of them are preset to be successful before you enroll the first patient so.
Is this just the next evolution of clinical studies we're just creating the perfect scenario by creating synthetic data. I don't know they were raising more questions and then answers but you bring it up a great a great a great fireside chat. Well and then I think one more thing before we close that though I think that you know I also kind of wonder about you know there's all that data and medical imaging and I think we've had lots of conversations and we've had lots of guests on our. Cast them talked about what some of the future medical imaging is going to be in creating the AI you know the challenges of that sounds great but these files are huge right and even like you look at.
Really high def OCTs not mentioned what you get in like X raise or MRIs or CTs i mean you're talking gigabytes. Per patient and storing that transmitting processing it you know that's pretty computationally intensive I mean that's that takes some serious. Computer energy to figure that out you know so that's our first challenge and then do we have something I think about this all the time I was going to my student doctor today we're talking about we have an optimap. You know I don't know if you knew this round but optimap every night at midnight or whatever time same it's set in your particular practice it uploads all of its data.
To London to deep minded London where it's being analyzed and study but what happens if I didn't write in hey this is a cradle in this. Then it's having to learn from something it doesn't actually have confirmation that is or what if I say hey this is a retinal melanoma but it's really a cradle in this then it takes that picture says oh well. You know you say it's a retinal melanoma and it's confused with the data so who's the expert that's annotating that you know there's another challenge and then you know if you look at you we've had so many talks about this is you know you can take a picture on one OCT and take literally move the chair take a picture the same exact picture on a different OCT of a different type and you get told to totally different things because. patient positioning matters how they acquire difference in scanners difference in depth that there's focus to read at all of a sudden you know all the if the AI models trained on image a from scanner a.
But now of a sudden you look at it on scanner B and the data is different does the a model be able to doesn't know how to read that. There's all kinds of issues with their data as much as I want to be really positive about AI I think there's some data hurl hurdles that are going to take some time to get over. Absolutely but they're introducing probably most fundamental topic to AI and and you know our journals is this the problem that the promise the peril of big data. So I think I think you hit it right on the right on the head there.
I don't know I don't have any answers right on I think that you know EHRs have bias and we could spend a whole time on talk about bias one of these coming weeks too but I mean there's just so many. Garbage in garbage out you know GIGO that people talk about you know I mean data is a huge issue we've got to figure that out I mean even the best AI system we create if the data is crap going in it's going to be crap going in. So you know fascinating discussion I I don't have the answer like you know I I think we're going to figure it out as we go but the one thing we do know is if AI is having a hard time analyzing the data how are we as humans analyzing the data or are we even thinking about it. I don't know.
Right but that's where the ads are going to be coming in to help us hopefully and that's I think that's the main goal. Well hopefully everybody you guys learned a little something tonight we always enjoy our Tuesday talks when we kind of we usually record these on Tuesdays even though they come out sometimes different days. If you want to hear any of our real talks as well as our innovator podcast series and as in addition to all of the articles that are being written about AI and how it's going to affect the I care industry please tune in to www dot AI and I care dot com that's a job some publication there's so much information out there that if you want to know and keep it updated. You've been listening to real talk an AI and I care you weekly podcast to keep you informed about AI technologies revolutionizing I care.