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, hello everybody, this is Dr. Scott Morris. This is Real Talk.
Today we have a special guest, Dr. Steve McNamara. He's at the University of Colorado. Steve, welcome to the show and I'm going to definitely have him do his own introduction here in a few minutes, but we're super excited to hear your thoughts about how we bring bench to bedside.
But before we get started, we always do this knowledge bite for all of the audience in terms of learning one new thing. And today our concept is going to be what is prompting. I was having a conversation with my colleagues the other day and they said, I just don't even know how, when it comes to gender, AI, to do prompting. And I don't even know what prompting is.
And they said, well, it's just the way you talk to AI. It's the way you get AI to do what you want. And you can think of it like Chachi, Biti or Gemini, not really as a search engine. I think we need to keep quick thinking of as a search engine, but this really advanced, very literally mind, literally minded assistant.
And your prompt is basically how you're going to coach your assistant to give you the output that you want. You're going to give them instructions, you're going to ask them questions, you're going to input text. And this is how you coach the AI to give you the information you want. And we've talked about this before, you know, that there's all these very powerful, AI's that are large language models or LLMs if we talked about before.
And these are just predictors and they've been trained on these massive amounts of data, text from the internet, book and code. All the stuff we talked about last week with Rayhound and I were talking about the Achilles heels of AI. Well, what are the advantage of it? Well, think about it this way.
When you give it a prompt, you say, write me a short story about the detective loses their keys, we little creative today. The AI doesn't know at the detective story instead its processes whatever prompt you give it. And then says, what's the next word, next sentence, next paragraph, next chapter that it comes in, it builds those things out. And so there's a couple of things we think about when we do prompting, when you get on Gemini or GPT and you say, okay, this is what I want.
You have to give very, very clear instructions. I always think it's like teaching a five year old what you want them to do and say, just be very clear. Don't say, well, talk about AI. Say, write me a 150 word summary on the ethics of AI or how to prompt.
Second, context, give it necessary background. You might say, you know, I am a college professor specializing in ancient history or I am an ophthalmologist thinking about surgical intervention innovation. This tells the AI what persona or role or how they think, how the AI thinks about the process. Number three is constraint.
So set limits say, you know, give them, I want this to be formal or I want this to be in, I want this to be casual. It must be a bulleted list. It must be under three minutes. It must be under a thousand words.
Give it, give it limitations. Say, you know, it's like when you go to give it to hockey, say, you've got three minutes. And they say, well, I prepared for an hour. Well, that was my fault for not telling you to be three minutes.
And then you give an example. And sometimes you say, well, you know, show it kind of some pattern you want to follow. Use the encyclopedia to prep encyclopedia Britannica to give me a description or put it as a journal club as an outlet. You know, in short, those are the four bathing instruction context constraint and an example.
I think many of the GPT type of generative a eyes are not needing so much of an example anymore. But the first three are very important. I think in short, a prompting is kind of that new language of communication with machines. It's how we teach a machine.
An AI to give us the output we want. And the more specific and clear and contextual instruction is the better you can unlock the full potential of AI. The biggest thing I always tell people when it comes to prompting is try. Get out there, do it.
If you use something like Gemini, we can show where you can click on their tools and show. Show me train of thoughts. You can learn how it's processing that helps you be better at prompting the more you do it, the better you get at it. So hopefully you guys all learned a little bit about prompting and what that means in terms of a knowledge bite.
So as I said in the introduction, I am super happy to have one of my new friends, Dr. Steve McNamara. And interested enough, we grew up. We went out after we talked for a while that we grew up not very far from each other.
And we actually went to competitive schools in Illinois. Steve is at University of Colorado and Steve, I'll let you do a little introduction about yourself and what you're working on before we kind of get into our discussion of the day. Yeah, sure. First of all, Scott, thanks so much for having me on the podcast.
I think that I've been looking forward to this for a little while just because it's it's nice to kind of be able to talk about. Some of these topics in kind of like an informal setting and to kind of elaborate on them a little bit more. But yeah, I'll give a quick background. As you mentioned, I'm originally from Illinois grew up there, spent most of my 30 years there.
I went to Augustana College for undergrad and then Illinois College of Optometry for Optometry School. And then I did a little bit of private practice, a little bit of corporate practice in Optometry. When I moved out to Colorado, I did practice as well for a little bit, but then I joined the University first in clinical research. And then our PI of our lab, J.S.R.Y.
called Pathy Kramer, moved here from Harvard and kind of set up her lab. And I was fortunate enough to join. I kind of taught myself during all the downtime in the pandemic, some coding and web development and data science. And I really kind of fell in love with it.
So it's it's really just kind of serendipitous that I was able to kind of merge all of these passions and kind of get to where I am today. Well, awesome, Steve. I am so looking forward to we for the audience. We've had a couple different conversations and we go all over the place.
I think Steve brings this unique combination of years of clinical care and now he's in research facility can kind of those things together. So we want to talk to a little bit today about what it takes to bring AI and many innovative type of technologies from bench and I either research part of the world, the development part to tools that you me and the entire audience can use. So we call it bench to bedside and so I'm super looking forward to it. So tell me a little bit about how that process works, Steve and maybe we can go into a little bit of what are some of the challenges.
And then what what are some of the advantages that we can look forward to as a profession as an industry. Yeah, sure. I think that you know, there's a lot of there's a lot of talk about kind of the hope and the hype of of AI. But I think that sometimes what what we don't want to talk about is kind of the reality check of it, you know, how do we get this thing into real practice because, you know, I think on the research side of things and in the academic side of things.
We're developing these models, we have the data, we get papers out and we kind of highlight the hope and the hype. So we're going to we're going to detect disease earlier, you know, we're going to have better access to care. It's going to be more consistent, more efficient. You know, I think that there's there's some realities though that once the publications over, okay, you know, like the confetti is cleared, we got to clean up the gym.
How do we actually kind of, you know, put put this into practice so that it's actually impacting our patients. Yeah, I always kind of we joke about this. We the hype of AI, we agree and then you know, I always think about when we hear these great ideas and the answers, when is it going to come available and the answers always someday. That's about as far as we yet.
So how do we shorten the someday to the two day. Yeah, I think one of the things that that I think I maybe bring to our team, Scott, that you kind of highlighted is that, you know, someone that's been kind of in the clinical practice realm where, you know, you walk in and you know you got a 25 patient day and you're you're kind of looking at your book and you're like, okay, how can I, how can I make this better? How can I, how can I make it more. More streamlined for myself, I think that that's one of the things that I can kind of bring to our data scientists when we're kind of talking about our models, you know, or how is this going to be implemented.
That's one of the things that I always kind of give them when we're kind of first talking about these sort of research concepts. You know, they say, hey, I got this great idea for a model. I got this excellent data set, you know, it's diverse. It's rich.
I want to do this and I want to use this methodology. You know, I let and so I say, okay, well, how are we going to use it. And that's some of the perspective, I think that you need when you're kind of starting these things off is that it's all well and good if you're able to kind of develop something and it's able to get, you know, a point nine AUC point nine. I've got five AUC point nine, five AUC, but I think that you have to have those kind of follow up questions is how is it going to be used.
What populations is this going to be able to serve. You know, how are we going to be able to implement this, you know, from an ML ops standpoint. You know, I think that's kind of the unsexy thing that people really don't want to think about, but like, how are we going to have this on a platform where it's going to be able to reach people. I think that brings up and it's always that theory of no different than any other type of strategic planning and business Steve is that you know, you got to start backwards and go what the problem I'm trying to solve and does the problem actually need to be solved.
And if it does need to be solved. And if this tool really solve it, you know, can we build it where it solves it is this tool create more problems. So I think, you know, like, like I said, any type of strategy got to work backwards. And so, you know, how do you guys when you're in a research lab, how do you look at this, how do you, how do you go through and make those strategic decisions about what projects really worth working on and how do you get some of that feedback.
One of the one of the biggest strengths that we have here in our department of ophthalmology is that our clinicians are pretty invested in partnering with us and kind of saying, hey, these are my needs. Just because you know, we have a lot of sub specialties. So like UVitis where you know, it's it's like they're saying, this is the this is the imaging that I'm getting. It would be really great if I could use it in this way or hey, I'm getting, you know, an optos wide field.
Both a, you know, a color and also an auto fluorescence. And you know what I'm having to do. I'm having to open up the last three exams in different browser windows and then putting them, mosaic, and them all kind of on a screen and then kind of looking back and forth really quick, you know, is there any way that we can streamline and you're like, oh, yeah, well, let's try to see if we can register those images. So you would be biased in that process at all, right?
Yeah, we're trying to do all in 60 seconds. Exactly. And you know, okay, then times that by however many patients that you have. And then yeah, is there any changes were good in the first place?
Exactly. So stuff on the mirror. That's exactly. Thanks.
Look real life. I mean, there's lots of challenges to you. I mean, there's a lot of challenges and how do you, how do you develop the tool you just said, you know, that's machine vision for the audience, what Steve's really talking about is, how do you build a machine vision that makes our job easier. Well, obviously you can see the advantages of how that would make life easier if you truly have good, non biased, valid imagery, how it can save us a ton of time.
And maybe prevent us from overlooking stuff that's just too microscopic for it or to minute, minute for us to be able to see with our just our naked human really pathetic eyes. Yeah, and then I mean, you know, it then it also comes down to intervention timing and is it worth it cost benefit analyses to is like, well, I don't it from the looks of this, it does look like actually this lesion has gotten a little bit bigger XYZ. So I think that yeah, it's, it's definitely something that we know that the need is there. And then I think at least working in a large institution, maybe this is where it could be better for, you know, some of the private practices or maybe corporate practices, because we have, we have a lot of red tape that we kind of have to work with security across hospital.
But then we also have the university side. So it's kind of setting up meetings after meetings after meetings to make sure that things are secure that we can implement them within the firewalls that were able to kind of access the data that we need or were able to access the internet if we needed as well. So there's there's a lot of kind of. I love that we're being, you know, I always think about I was a book I read a long time ago.
It's one of those things on the entire book. Can't remember the title of book, but I remember this phrase meetings are a cul de sac that great ideas go down and are never seen again. You know, so exactly. Well, that brings us to really good questions.
I mean, I want to get into the what makes data good part of this, but I also want to talk to Steve about. So layout for me and for the audience like let's use that example where you were you were UVI specialist who I actually refer to a lot, but I have these complicated UVI decisions since you guys are kind of in my backyard or I'm in your backyard, depending on how long does it take from the time that your UVI especially comes. Hey, Steve, you know, it'd be really cool if to the time you can give them a working product. Let's see, there are no hiccups, no obstacles, no challenges, right?
Oh, man. Yeah, right. That's not real. Yeah, that would be great.
Just for the audience. The thing is that, yeah. Just for the audience to understand like, what's the step by step and maybe big picture first, like how long would that process take? Yeah, I think that I think one of one of the rules that I've always kind of heard, especially in data science and that I've kind of learned in in my own in my own master's courses is that data science is is 80% data extraction and cleaning.
So, you know, that's that's kind of half the battle right there is, are you able to extract your data? Are you able to clean that data efficiently so that you're able to kind of utilize something I know everyone's probably heard the garbage in garbage out kind of thing. So if you're if you're not able to work with data that's at least some degree of of of of cleaned before you start working with it, then you're you're really not going to going to have a viable product at the end or something that even, you know, works half of the time. So going back to your kind of original question, I think that the data extraction and the cleaning process is definitely something that we're working on kind of streamlining that, but I think that that's that's an issue and that's anywhere from, you know, if you're starting from scratch anywhere from like maybe maybe one to two months to kind of be able to identify those things.
I think one of the other things to that makes it a little bit more challenging is that sometimes we have at least an ophthalmology and when we're a big institution with, you know, epic utilizing an epic EHR is that we have these two different systems. So we have the EHR that's separate from the pack system that kind of houses all over imaging. So if we kind of want to marry those two things or at least try to get like, okay, does the date in this you have an API interface you need to create just so they talk. Exactly.
You and I talked about in some of our previous discussions, you know, I mean, there has to be a better way. I am not the world's biggest span of, I think, EHRs or digital filing cabinets. And I know that many of the EHR companies will use that phrase and, you know, not be happy with me, but the reality is we have to develop like how do we improve our data sets so we don't have to clean as much as we clean number one. And number two is how do we merge some of these technologies together my word of the month, right?
I know we're talking last week is connectedness. I just think there's no really great connectedness of the data and how do we get the data right. I mean, if you're spending one to two months and I'm keeping tabs this Steve, so when we're all done, we're going to give the audience like this is how long really takes right. But so if it takes one to two months to clean the data, it's like, can't we develop a system that has better data in the first place we don't have to spend so much time in the world of AI.
You would think that we should build up snap our fingers and the data is clean. Yeah, that's just not the way it works. And, you know, I mean, give me some 30,000, I know we're getting off on a beer here, but yeah, sure, you know what? If you could design a system that you wouldn't have to spend all the time cleaning data, what would that look like?
Yeah, I think one of the things that I've talked with some of our clinicians on and you know, one of them's like our medical director of imaging kind of thing is that. You're never going to be able to get 100% you know, great quality photos, no artifact. There's just going to be inherent issues, you know, with patient cooperation or with kind of media capacities that that's. There's dust in the air that day exactly exactly.
You know, and also with Colorado dry eye that's always going to that's always going to affect image quality. But just putting, you know, all of those things aside, one of the things that I've kind of talked about or floated is, you know, it would be great if on the front end. So at the point of capture, you kind of have your your ultimate photographers kind of thinking about that quality aspect or kind of your standardizing the process in terms of having imaging protocols in place. Because the other thing too is you want to have a reasonable, a reasonably consistent field of view as well, because you want to kind of standardize inputs as much as you can for some of the models.
So I think that either being able to label like what did I capture. So it what what's the structures that are in here or kind of what field of view did I have. So being able to for us to have that in the database as part of the metadata of the imaging. So that we know what the imaging is, what kind of capture was it, what are the settings that way.
And then it comes back to maybe really quickly, I don't want to go off on the tangent, but to kind of answer some of the questions is like we're very bad at data and I guess imaging harmonization, like we we need to really take a page out of like radiology's book. Because they've they've I think done a lot of the hard work of the die come harmonization and kind of making sure that across devices that things are being saved in a standardized or a unified way. Huge issue with that. It's a huge issue.
Yeah, it's a huge problem. And also, I think radiology's done a decent job at kind of marrying the imaging with the the EHR because like that's their whole that's their whole stick to and and kind of like any annotation process saying hey, the the the disease of interest or the disease feature of interest is in, you know, raster 25. That would be huge, I think for us to so my industry folks who are listening to this, I think what Steve say and I'm agreeing with is we need have some form of validity check this not just giving you this is it's this is the it's validity score, which nobody ever pays attention to and nobody writes down. That has to be if it's below a certain level, it's just as a not good photo, I can you don't give the ability to save it because just not good it's not helping anybody get to the image right to get to the to that standpoint.
Alright, so let's go back to our original question, so you have data extraction and cleaning, you said that takes one to two months and that's the big puzzle, maybe we can solve that in the near future, and I know there's a couple of companies. You know that are really working on that top con with their harmony, harmony software is really trying to improve some of the validity of what we do. So you get a you get a data set that's let's call it and I say it for air quotes, if since you guys can't see me or just listen to us air quotes of valid non bias data and that's. A huge problem in the first place, but let's say you get a set that's clean, yeah, then what happens to you.
So I think at that point, if we have a valid data set, I'd say usually before the data set, we're probably meeting with the clinicians, but let's just say we we we had to know how we knew what they wanted, we have a quick meeting with the clinicians, we say hey, we have this data set it's all cleaned. And then we were able to come up with, I guess the question that needs answering or the problem that needs solving that's that's usually pretty quick. Well, I say that's usually pretty quick, but when you're trying to get very busy clinicians and very busy. Eight of them in the same time, exactly when they're all going to conferences and traveling and clinic schedules, let's just say two weeks there.
And so we get a meeting on the calendar, we're able to get a good direction forward from maybe that 30 minutes to one hour meeting. We we're kind of clear on on what needs to be done. We can have maybe like a debrief with the data scientists and say, okay, what's our what's our methodology, what's our pipeline looking like. And then so maybe another 30 minutes, we decide on that.
And then I think we're at the easy part, right, we have our data, we have the question, we know we're solving. And then we were able to train and test and do all of that fun stuff. And how long does that train test revise repeat? That's the way we always teach people it's train test revise repeat.
How long does that process take? I think luckily for us, the I think sometimes it comes down to how much compute power you have in your infrastructure. Luckily for us, we have a good cluster here. We have a good cluster of GPUs that we can utilize and that kind of cuts down on on our training time and everything.
So that's, I would say, if we need to iterate a couple of times, let's just say that that's two to three weeks. So ideally only like one or two sprint cycles. And then we have, you know, a good model. So conservatively, let's just say a month.
And then we have a good model for whatever task that we have. And then now we're kind of that where we started or where we started talking about earlier, we have a model. Let's say we've we've published on it. And you know, it's it's it's out there in the world for everyone to read about.
And now, OK, now what? And this is kind of where we have been getting bogged down. I've had a lot of good conversations with. With one of my colleagues and right now, specials here, Dr.
Naurantj and my heart. And we've had a lot of talks about this. So there's this kind of like. And there's a fork in the road here and you're saying, OK, is this like a clinical decision support system?
Or is this something that needs to be FDA regulated? Is this helping me in some way. To do something automatic. Where we're going to need FDA clearance.
And or is this something that just supports my decision making? It's like adding to my decision making as a doctor because. I think you have two very different timelines here. And I think the clinical decision support system is it requires just more internal.
I guess internal figuring out of deployment and everything. But then I mean, if you're going to FDA route, then OK, you got to add on a clinical trial. You got to you got to figure out. And then the localities of getting all done is crazy.
And I think that brings up a great point for the audience. You know, I think that. Steve, you laid that out really nicely in terms of you get to a point you're like, OK, I have a product. But is this going to be an augmentative tool?
Or is this going to be a replacement tool? Right. And I think that that's one of the fears. And I think that most health care providers optometry, ophthalmology included are like, is it going to replace us in the reality is it's much harder to build a replacement?
Because once again, we'll go back to that is the data 100% perfect. And the answer is always going to be no way at least in our current format. So do we have it more? Are we going to go down the path of having as an augmentative, you know, tool to it.
Augments our clinical decision making gives us more information. And I think that we're seeing more and more of our industry go down that path. Just because we understand there's realities of validity and bias and, you know, we got into interoperability and those types of things yet. So cool.
So let's say you're going down like here comes the 30. Here comes the well, we're probably the million dollar question. Okay, so you have this great tool. You're like, Hey, this thing really works.
Then you got to get providers to use it and believe in it and trust in it and retrain it. And into you know, iterate to get it to be even better. How long is that like what are some of the challenges of that process and how long does that take? Yeah, I mean, I think the first one is that you need to get all of the relevant security clearances and I think that that's that's the one of the biggest issues is that in a large institution like ours.
And there's a lot of oversight. There's a lot of kind of disconnected offices that kind of need to be brought in. And I think that the other thing too is like we're. We're connected to this problem again.
Exactly. Yeah. And I think part of the problem too is that we're still kind of building out the. The infrastructure, you know, you're kind of almost trying to build the plane while it's taken off.
So it's definitely. Yeah, well, I guess at least from like a security standpoint, from like a deployment standpoint, then it's like, is this going to integrate with the PAC system? And is it going to be just kind of its own program, then doing into right software for that program? Where's that software going to be hosted?
Who's going to be maintaining that there's a lot of questions I think that kind of come up it comes up and then, you know, again, that's that's kind of going back to that. And I think that's the MLA ops kind of thing. You know, the the unsung hero that kind of gets gets things from the idea phase to the actual implementation phase. So I mean, I guess.
You're you're asking me how long that that takes and my answer to that is, I don't know. We're still kind of at that phase is like how long does it take to get something implemented. So I mean, ideally, if if all the infrastructure is in the right place. If if you're able to to kind of have everything set up and you're able to get all the right people in the same room, I would say maybe six months maybe that's a little optimistic, but I think maybe six months is probably.
I would say realistic without being, you know, too over optimistic, but. If you have hurdles that we've been kind of talking about, it can drag out, you know, it can drag drag out. I mean, there's you were like saying everything in the perfect environment, but I think the point about the Steve what I was trying to draw out for our audiences. I think right now and you and I talked about this in our pre meeting is that there's this great delay or I am going to bury my head in the sand and wait and see what happens.
Yeah, and I think that's really dangerous for industry because. And this is the reason I was so wanted to do this podcast with you is because. For all the audience who's listening, Steve just laid out in a almost perfect world like with no real hurdles of big challenges and money's no option, money's no issue, you know, everybody's on the same page going the same direction, you know, it's that right seat on the right, you know, right seat on the right bus going in the right direction with the right driver. If all that happens, you just laid out it's a year and so I think for the audience what I really want to bring to light is that there are lots of people like Steve's team, they're the team Steve's working on who are working on these things and you know there there's projects like this going on over where and you know people these great ideas and know that hey it's going to take a year of investment, but then when they happen things are going to move very quickly.
So I think right now our industry is so much in the weight and see meanwhile everybody's working really hard in the background like your team and everybody else to bring these products that are going to make. Help us make our jobs as clinicians so much easier and so much better and I think that we're going to come to a place where all of a sudden there's a whole bunch of these really cool tools and technologies out there in a very short period of time. It's going to be very difficult if you're in the weight and see group to adapt and adopt and so Steve and we try to keep these in somewhat of a or people can listen to them. I think so if you were to look at it's a let's break down some of those challenges so we talked a little bit about data quality and what bias and validity look like.
But I also think there's a data quantity issue right is how much data does it take to really have enough data you know is it a thousand images is it 10,000 images is a hundred thousand images. You know and then you run into where do you get all that data from and how far back is it and you know I mean if you're talking 10,000 images are you still using the same piece of technology you started with this three years ago or if you upgrade the technology now you're not talking apples to apples. I mean how do you guys get around some of those challenges. Yeah I think you know just just going back to the data quality issue you know I was I was talking with someone that you know they do AI but but in the finance world and they were just they were kind of saying well okay you work in kind of the AI in the medical space and so I'm just really surprised that you guys haven't made as many.
I guess advances right as many advances as we have and I'm like well okay let's when you have data like yours where it's purely you know numbers and there's bank transactions and everything's really clean. Then it makes it easier to kind of work with and to kind of do those things and to find patterns and you answer biological creatures that are not clean data centers. Exactly and and I think one of the other things if I'm if I'm being you know real about about us we're not always the best record keepers either and. As clinicians were not much of patients exactly exactly and so I think that that's that's one of the Achilles heels is that you know I think what what we require for a lot of the data inputs and especially as we kind of move in to.
You know reasoning and and and VLMs and stuff like that where there's there's going to be a text output when we're saying okay scatterdruzen atrophy that's not really great data for for giving context of like what's what's in the image where is this thing at what are we seeing and it's human interpretable but it's not really like machine interpretable for for these it's like Albert exactly so. Yeah exactly and and that's that's I think one of the things that outside of you know the imaging issues that we talked about is one of the Achilles heels of our our data quality is that it's not precise enough it's not rich enough it's not complete enough to kind of be able to do what we're I think what we have in our heads of what we want it to do. At the end of the day versus reality Steve perception reality one hundred percent and then I think you also bring up another good point and I think that part of it is to. Like we have this data fragmentation where our data is kind of locked away in all of these kind of separate places and it's it's tough.
We need to balance privacy and and kind of people's ability to access their own data and keep it private with the ability to kind of pull that data and I don't have the answer but I do think you know hippas a little it's a little outdated it was kind of it was kind of all drawn up at. In 1996 yeah 1996 I mean it's 30 years old we technologies changed a little bit I mean think about when hippas being produced I was graduating optometry school and we didn't we didn't have flip phones yet. Mm-hmm that's when it was built maybe we've come a little waste and sip a. I think so and I think you know it's not not bringing the issues with like you know government gridlock into into the conversation but you know even in a perfect world I think that that would need to be brought back to the table and kind of revise and and kind of updated to be relevant you know 30 years later since we've made such leaps and bounds and then.
Well and play devil's advocate on the steven I've been saying this for three years now and I've spent on various podcasts so I'll just repeat myself for the audience to probably get sick of listening to this is we have to rethink how and where data is stored. I believe that we need a reinvention of the electronic health record to be very different than what it is that we built in 2004 and 3 and 2. You know think about the EHRs are 20 years old and they're still for the most part the same structure infrastructure they were then until we can change that where we have greater. Greater what's a good word transparency and yet increase connect this connectedness and access until we can fix that it's going to continue to fight these same fights.
Yeah and you know not to be not to be incredibly cynical about it either but even where we do have these large kind of data pools at like say our institution at the University of Colorado where we have like a larger amount of patient volume that's coming in every day. I think that right now especially in academia the incentive is not to share that data the incentive is to keep it locked away for ourselves because. What if I if I make that data available to someone else in this kind of publisher parish environment i'm giving away my competitive advantage and I think. I'm going to be thinking about if you took every university who's got an ophthalmology clinic across the world.
How much data is that that's an insane amount of data that you know just that that now I can get in stuff that's not even completely clean you just have a numbers game right. Yeah and I think that we're we I agree with you we have this problem where everybody's like it's my data and we always talk about data is the new oil right and it is. But the reality is is that everybody's working in such a small data set that if we were able to share that data. And crazy thought everybody got credit crazy thought.
How would that change what we do right me how much faster would we move along the process and that once again that's my word of the month we're back to a connected this problem yeah and maybe a little bit of human ego. It's mine. Yeah yeah so I do this is a perfect segue because I do think the one one of the things that I wanted to bring up so dr Travis red who is at oh it's you he gave a grand rounds talk at. For us here at CU and one of the things that he brought up was this exact problem right and so we have this with one of our KPIs you know as as faculty at universities is what's your H index right and that's like how many times has your has your paper been cited by other people but he proposed like an S index so what's your S index what's your data sharing index.
Yeah I totally agree with you Steve that's that's there's a solution to this yes there was this everybody gets credit if you share your data and maybe your data is worth more if you share it then if you keep it to yourself. Exactly yeah and and so that way I think it's it's well it's about kind of like reframing maybe the conversation like okay well yes there's a little bit of human ego that kind of goes into my my H index is higher than your H index well. Can we incentivize the the kind of sharing and the the deep fragment or yeah the I guess like defragmentation of our data where we're kind of able to aggregate it and you know what at the end of the day you also get points for sharing your data instead of hoarding it away. It's a whole different version of evidence based medicine point system right now we get to feed everybody's ego and go instead of you being the first name on the paper your a name on the paper but you're the name of a paper that changes the whole world instead of what I would ever read.
Yep. So thank you so much for trying to I we can go up for another hour I still have like 20 questions I want to ask so maybe we should do this again in the near future because I literally have a list I think I have more. Questions now then when I started this conversation about where do we go from here so for the audience hope you guys got a little bit of an understanding of what it takes to bring augmentative tools or AI based tools from bench to bedside Steve thank you so much for your time. Thanks God appreciate it.
You've been listening to real talk an AI and I care your weekly podcast to keep you informed about AI technologies revolutionizing I care.