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. Hey, everybody, it's Scott. And this is Rehan. Welcome to another episode of our weekly podcast.
So today, you know, we always do our knowledge bite where we kind of do, let's call it our weekly download of some kind of technology, fact or description that we feel is kind of important to understand. And today we're going to talk about data mining. Now, there's not something like that needs hard hats or pickaxes or anything like that. But it's what actually happens inside the silicone chips of every computer that exists.
And you know, it's simple as data mining is kind of like, oh, it's a process of discovering hidden patterns or anomalies within these massive data sets of information. Think of it like a good way to think it's like a big sift, right? You pour in all of this raw, disorganized data points. And then we're going to talk about this a little later.
We get there a fireside talk, but you know, it's kind of like you can pour in all your grocery or seats or weather patterns or all your social media clicks and you use these very sophisticated algorithms that are pre-programmed, at least now the pre-row, maybe someday they won't be pre-programmed. Algorithms is shake away all of the stuff that let's consider it noise until we just find the little gold dust that we're looking for that'll actually go through the sift. So let's think about how data mining or let's call it digging actually happens. Well, usually follows a couple key steps.
First step, there's classification. So where does the system, I should say, where the system organizes data into groups, it kind of starts saying, okay, are you a receipt or are you as we used before a social media click. Then there's clustering and that's where the computer groups finds groups that share similarities you might not even notice, it might say, well, hey, let's look at buying patterns as it relates to weather patterns, because I live in up to the mountains on days that it snows grocery stores are pretty empty, right? That'd be a cluster going, hey, when there's a certain weather condition, other things might or might not happen or traffic accidents when there's bad weather traffic accidents go way up.
So that's clustering and then it the third step is called rule learning, which means it kind of think of well learned, we're called if then statements, it's kind of like that, where a supermarket might say, hey, people who buy snow shovels on Friday are also much more likely to buy salt for their driveway. Because there's no storm coming. Now, this isn't, you know, this kind of stuff isn't just for scientists is everywhere, I mean, it's how it's how you're streaming service works right I was just watching Netflix or not Netflix, we're watching prime or watching the pit, you know, in the streaming service knows what we're going to watch and now it's bringing up all these 90s, you know, ER and everything else saying, okay, we know you're very interested in medical drama. That's the classification and it's going to look at clustering, what else was no a while in, well, he was also in the librarians and so that comes up and it's okay, maybe there's there, right?
So in healthcare, we use it to kind of predict patient risks by analyzing thousands of medical records, you know, it says let's plug everybody in and gosh, well everybody that takes a stat and drug is typically within six months also on some type of metformin type of medication. Well, that's an association where you classify two different groups and you clustering go, wow, there's a whole bunch of people that cross references, you know, so it or you can't go in finance right it's something that says hey, there's fraudulent credit card transactions that are happening because, you know, all of a sudden you bought a TV in California, yet you live in Colorado and you just made another transaction at a grocery store. Some of it's not working there right so data mining kind of thing is the server, but it's really the engine behind what we call today's information agent it turns all of this raw data that we pour in into usable knowledge and that's what data mining is and so hopefully that you guys learned a little something today about data mining because you're going to hear us talk about that a lot in the next few weeks as we get start diving a little deeper into neural networks and some of the data that we're going to talk about. The data that's out there in the real world so hope everybody learned a little bit about that, but ran today I kind of wanted for fireside and and you know I try to always I think you do to we try to always be really positive, but I want to talk about what happens when AI fails.
You know, you can fail in lots of different ways, but I think the biggest reason it fails isn't a technology problem, it's a knowledge problem and it's a knowledge problem, I think that I listen to people go I try to say I and it didn't work and I'm going to go back to a Bill Gates comment for many, many years ago, he said technology, it exposes strengths and weaknesses in workflow if your workflow is not good, it's going to expose your weaknesses and if your workflow is great, it's going to make you stronger and I really think that AI is the next evolution of that, Brian is that I think we're going to start seeing because there's all these new technologies AI embedded technology where you see some of them fail. I'm going to see a lot of fail, but I don't think it's AI's fault, I think it's because we don't understand what's really happening in our systems, which we might call workflow. Yeah, I think that's exactly right Scott. So what we're talking about in our fireside chat really, it could be summed up in a phrase that it's not the algorithm, the failure mode, often in any type of technology implementation.
It's usually not the technology, but the workflow in which it's being placed right and I think the best example of this comes from the diabetic right right not the space and and obviously artificial intelligence applications are in. So Google Health, they developed an algorithm, this is back in 2020, which of course in the world of AI's ages ago, but, but DR was the first application of AI in. In healthcare and specifically with an ophthalmology, it was sort of years ahead of everything else, so we have some good data on this, they developed a really impressive algorithm showing well over 90% on sensitivity specificity for DR diagnoses. So they took this to they partnered with the health system in Thailand, they took it to Thailand.
Oh yeah, the anticipated the hope was that it worked great in the lab, it should work great in the field. And of course it didn't what happened while there all sorts of systems issues right. One, if the lighting in the clinic isn't great, you're not going to get great images. Two, if and this is maybe a little bit of technology, if the if the algorithm is trained on a bunch of people from Wisconsin and Minnesota, it's not going to work the same way that it does in Thailand.
Three, if it depends on say internet connectivity, well, if the internet goes down, what's going to happen then right for if there's a if it flags something. That needs to be followed up well, did you did you design the next steps in that workflow, well, what happens if there's a positive it wasn't clear that that was done right and so this is this is a failure mode really of technology. Sorry, it's a failure mode not of technology or the AI, I mean, I was still doing very well, but it's the application of the systems based version of it. So that in my mind, and we talk about this all the time, you guys that that's more important than than the actual technology itself, because now we were here, you know, six, seven years later.
The AI's are all really great at diabetic right and opathy, but we're still seeing, you know, there is not has there is the the up tick and the CPT code of nine triple two nine, which is the AI, AI code is not being done as much as we all thought it would be. Right, and I think this goes this is not just a Thailand problem. This is a problem, I think across, you know, across the world of trying to figure out where the algorithm is going to sit within the within the workflow. And I think the next big issue it's what I tied to DR, but it's going to be a culot mix right and everyone, you know, and their mother is talking about a culot mix and you know, we are both super excited about this.
I mean, the fact that we're talking about predicting cardiovascular events, biological age Alzheimer's and other neurodegenerative issues, kidney issues. But there are so many embedded failure modes Scott and so many virtual problems with that so much. And you know, it's, it's a people are working on this and we're thinking about it, but like let's just name a couple right. They're the reimbursement void, I think here for a kilometer is huge, right.
We're talking about not a clear way of how this is going to be paid for. Really expensive technology with potentially a less than $20 up reimbursement fee. I mean, I don't know how many people are going to jump into this world and the cost of sushi with for 20 bucks. So I think that's going to be a challenge.
I think that's a hurdle. And there's another risk too. There's something like the orphan referral risk where, you know, if this is being deployed and, you know, in retail locations and other. Other score, quote, non non primary health care situations.
There could be flags where, you know, it be there's no established pathway. Suppose you're, you know, you're somewhere in a retail location and a flag something for cardiovascular risk. Is there a proper referral pathway that's already been charted out maybe maybe not. And what happens really who's responsible for that.
Right. And this is a duty to act trap. This is where it generates a ton of information. Now whoever did this has a duty to communicate and document.
If it says you got high risk for having a cardiovascular event in the next, you know, next year. Well, it's a problem if the patient's not informed and if the doctor's it means only as good as someone using that information. Right. And then it also happens another failure mode is like, well, what happens when a neuro a neurologist, a neurosurgeon, a cardiologist gets this information.
Like, well, you know, okay. And AI said this like what in a ball in a ball. Yeah, and today are they had they, they agree with that. Do they trust that.
And so there's a whole issue, his whole issue there. And there's a ton of other failed potential failure modes. I think what we are hoping to surface in this, and this section is that. I think it's addressing and knowing the potential failure modes.
Like once you know that they're there, you can do something about it. Right. The problems were so invested in so interested in technology. And I'm guilty of this too that you kind of don't think about the quote unquote boring aspects of workflow until it really just sinks the entire project.
Well, I just kind of think you know what and you and I keep saying this, but for everybody in the audience, I want you to remember this phrase. Tech is just a tool. But maybe workflow is really where we need to focus. And so I think over the next couple of podcasts that we do, we're going to spend a little more time talking about some of the systems.
And where AI will be successful and where it might fail. Not because of AI and not because of the tech, but because the workflows broke. And you know, part of our job, I think around you know, you know, you talk about the part of our job is to educate the community. And I don't want to see it fail.
And people go, oh, it was the tech when it was really. So, right? So for the audience, I think you're going to hear us talk a little bit more about workflow and different areas of the practices, different areas of clinical care. As we kind of start breaking down like, hey, here's the I that's out there.
And here's what you can potentially fix. But here's what you need to fix before you do the implementation. And first step in fixing any of those workflow or technology issues is checking out our website AI and I care.com and subscribing to our weekly and multi X, we send out notifications and emails once or twice a week. We try not to spam you.
And we're also active on LinkedIn. You could reach Scott at Scott. That's with one T as C O T at AI and I care. AI.
And I'm always reachable at Rayhan R E H a N at AI and I care. I'm not AI pleasure. I always got looking forward to talking to you again next week. Sounds good.
Everybody have a great day.