The Future of AI in Healthcare: Opportunities and Challenges with Isaac Park

Hello and welcome to the latest edition of the Techstrong AI Leadership Insight series. I'm your host, Mike Bezu. Today we're with Isaac Parks', the CEO for Keebler Health, and we're talking about the use of AI agents in healthcare and just how ready is that industry for this next wave of technology?

Hey, Isaac, welcome to the show. Hey, Mike, thanks for having me. You cannot go anywhere these days without somebody telling you about their great new AI agent.

And there's one every day now, but it's not clear to me that we as, as a society, nevermind particular industries are ready for this, and especially in terms of the business process engineering that might be required. But you work in the healthcare field. What's your assessment of what's going on here?

Yeah, I think people are trying to figure it out and certainly they, they see the opportunity in many of this, in much the same way, rather than in other industries or verticals across the larger ecosphere. I would, people are trying to figure out how to string together sort of these complex workflows and apply some degree of, of mimicry on human judgment to sort of relieve the burden. Right.

On clinicians, on back office staff in the healthcare world, I think what they're running into, or the problems that are being run into is how complex or how circuitous and how blended. I would say that that workflows can be between creative tasks and structured rule set tasks. And so I, I think it's when it, when there's two sort of, uh, like paradigms try and have to mesh, that's when it gets really tricky with agent workflows.

Yeah. One of the challenges that I think I'm seeing folks struggle with is they're starting to realize that a lot of these AI agents are probabilistic in that sense that it's a best guess and a lot of the workflows are deterministic in the sense that they need to be done the same way every time, especially in healthcare. So what's your sense of, uh, yeah, the understanding people have of what an AI agent can do?

'cause I think the first time they encounter their hallucination, they start to lose faith altogether. Yeah. I I think that's certainly true, and we run into it all the time when you're talking about, you know, workloads that are deterministic are to go in, you know, more structural rules at things like, I don't know, billing or, or lab values, stuff like that.

Right. And I, and I think what's interesting is, uh, sort of the, the, the learning that people have to do around how generative AI or large language model powered or, or transformer powered technology and sort of therefore the sort of non-deterministic tooling nature of all this. That it's still just a tool, right?

It needs to be used for the appropriate use cases, right? And deterministic rule, set based workload, it's just not the appropriate tool. It's way faster to use a calculator for math than it is to ask chat, right?

To do a math problem. Um, so I, I think this is where I was kind of going with like this combination between when you have a really complex workflow that involves a ton of creativity, where a non-deterministic approach might be really useful or valuable and unlock a lot of things that have previously been unable to be done. But then the sequential or, or downstream steps or even multiple steps require injection of deterministic or rule based things when it gets married and mushy like that.

I think that's where people really struggle, right? With trying to say, okay, this is a one stop solution to just throw it a transformer or an agent and say, go, um, and healthcare has a lot of this. And what is your sense of people concern that AI is coming for their jobs?

Or are they looking at it more like, there's large swaps in my job that I just don't like doing and I'm hoping that AI will come along and take care of this for me? Yeah, that's, that's a hard question to answer mostly. 'cause I, I would say that paradigm where those, those, that's an open question for not just healthcare, but writ large, right?

I think knowledge workers in general are all kind of asking that question, what's gonna happen? Um, if I had to sort of guess or predict or anything like that, it certainly, there's gonna be job transformation change. Like I don't, I don't, I don't deny that that's not some something that will occur.

I think really the, the question is how right, and the only paradigms we really have to really pay attention to are past technology shifts, right? And much the same way that you've seen, I would say, I don't know, the internet or cloud-based feeding or prem on-prem versus on-prem software or, you know, even looking at things like mobile or social, like just recognizing that this all large technology shifts to some degree providing our access, providing an increased tooling, right? That might remove a current workflow that a human is currently supporting.

Um, in many ways, when you look at sort of this past paradigms, what's interesting is that, yeah, they, they nixed some jobs, but you know, if you looked forward maybe 3, 4, 5, 6, 7 years, it actually created many more, right? And so it is a past para that's happened before. I, I want venture to guess that it's gonna happen again, right?

And so the, the really what's gonna happen is like that the jobs will change or more job will create or how, and, and that, that that's the best way I could probably, I ideate a prediction there without, you know, coming to some other high degree judgment. What's your assessment of where are the senior execs on this curve? Because I'll talk to them and uh, I'll get reactions that go anywhere from believe it when I see it to, we are so AI agent happy that, you know, we're already counting the number of employees we're not gonna need anymore.

And they're kind of like thinking that everything that they do could be done by an AI agent. Yeah. Uh, I mean, I like most things, I'd say there's probably a blend of truth in the middle, right?

Or the truth is really a blend of the two ends of that spectrum. Um, I certainly do think that there's opportunity for business leaders who have cost centers that need to reduce, right? And, and maybe even their sort of operating budget requires that they have to do that.

And this sort of age centric revolution not opens the opportunity for them to do that. But I also think that, you know, even in like some of the industries that we're in, right? When we're talking about like, clinicians who are doing pre-visit planning, there're just, there's aren't enough of them.

Like there's, so there are so many more open jobs right now, right? Than than even humans to fill those seats or roles. Like we have also business leaders who are saying, oh my gosh, this is gonna allow my existing workforces right, to do way, way bigger, faster, stronger, and, and, and fill the demand that's there with the team that I have, right?

So I, I don't think it's gonna be a straight binary, like it's gonna be cut. So in, in many ways, I guess what I'm arguing is that it's probably contextual, right? To what is the exact set of problems that, that these business leaders are trying to solve with this tool.

The other thing we're all kinda looking at is the rise of robotics, which are driven by AI and healthcare. There might be some opportunities to build these robots and use them for, uh, home care or whatever it may be, but what's your sense of where are we in terms of robotics in healthcare? Ooh, great question.

One that I don't know if I am entirely qualified to really answer. I'm certainly much more of a software guy myself. Um, but I am watching it 'cause that's so curious.

I am like on the edge of my stupid, Ooh, I wonder what these robots can do. Um, the funny thing is that there's probably also much to the surprise, there's probably some leading edge things that I know remote pro like remote process monitoring or RPM, um, in our industry is very, very big deal, right? There's big in this large movement to, to treat patients, to treat the more acute advanced illness patients at home or in a, a setting that is not in a hospital, right?

And that requires a whole set of infrastructure. That means that those biomedical devices or, or, or proxies for biomedical devices or, or clinical care folks are in the setting with them, right? So there's already hosts of, I mean you may not wanna call them robots, but they're, you know, devices that are, that are doing a lot of clinical support for these patients that are all integrated and, and wrapped together with software.

It's probably not that far stretch to say that, oh yeah, that's kind of a robot, right? And if you have all those things sort of powered through a transformer based product that uses judgment and can signal humans to say, Hey, go do this, go do that. I, I think there's, there's probably a ton of efficiency gain to be done there.

Just question of power and when, What about the security aspect of all of this? Is that still kind of an afterthought as usual? Or are people thinking about this more aggressively and upfront?

No, there's deep concerns, right? And caution around security, um, for using generative AI products on, on clinical documents and, and, and patient health and rightly so. I think right, that's one of the premier considerations.

Anytime we, and even our business, anytime we integrate or or talk to a, a protection customer that is top of mind forefront, like how do we protect, you know, the, the sort of downstream artifacts of, of patient information. How do we make sure that nobody's getting hand on on those items or documents and, and all of this in light of being, being able to say, okay, can we use these tools to improve their outcomes? Can we help them be healthier?

Can we, given longevity, can we improve the quality of life? What use cases Have you seen involving AI in healthcare that are kinda are working that stand out to you as something that, you know, maybe you wish other people would copy or at least emulate? Yeah, that I think the, the easiest way to sort of think about this is, um, the, the landscape shift that we're seeing on these large language models is really the ability to access unstructured information to reorganize it or informalize before you can generate it.

And whereas I almost liken it to, you know, the large language model is really to unstructure text what the calculator was to like rule space engines are bad, right? The, the, the sort of speed at which you could analyze that kind of mathematical data with a calculator is now what we can do with unstructured text. So I would say anytime you have a, a workflow and in the clinical workflows, there's a lot of these scenarios where you have bulk corpuses of unstructured text and you're trying to draw out patterns or mimic patterns or established patterns or career processes around that unstructured test text, that's where we can really fly, right?

So things like navigating any HR or understanding, um, what's happening in a medical record, in our case looking for evidence for chronic disease over a host of data that is just unstructured preform text. Um, those are all things that are really viable. One of the, I think more, I would say its advanced or leading edge cases of using transformer technology and clinical workflows would be dictation software, right?

So having something listened to a, a clinician in office and visit and then summarizing it into an actual clinical note that is structured the right way, that captures the right nuance and appropriately says, okay, these are the idioms. I'm gonna translate that into this clinical, like language that makes sense for them, that that saves so much time for clinicians, um, that it, it's been pretty amazing to see. Conversely, have you seen anybody try to use AI for a use case that's just not gonna pan out and others should not follow suit?

Yeah. Uh, that's probably one of more like a judgment call. We're so early that like, uh, you know, you also have this sort of selection bias where prob people have probably tried a lot of things and have not have not worked and we've never seen it or showed up.

Um, I, I think there's, um, the, the closest thing I can get to you is, is this idea around like, uh, replacing the physician in their clinical judgment, right? In the same way that you were seeing, uh, I would say a parallel paradigm around, uh, software agents or software engineering agents trying to replace software engineers, um, on their own as just software agents. They're not, the efficacy just just isn't quite there yet, right?

And I think similarly with clinicians, I don't think a, an an aid, a clinical agent on its own doing diagnoses and prognoses and sort of, uh, preparing a care plan on its own is going to really outperform a clinician alone. Uh, what we've seen is that there are benchmarks that are passing around these structured data like, like passing medical, medical licensing apps, things like that. But the, again, that's all still just structured traditional structural rule sets to try to figure that out.

Um, the same sort of, I think, uh, productivity in improvements, um, in software engineering co-pilots is, is actually happening, right? And so I think when you're looking at sort of this, uh, the evidence out there for, um, you know, software engineering co-pilots are amplifying like you're really good engineers and then kind of taking your, your not so good engineers and making them worse with a copilot. It's actually kind of similar with clinical diagnostic tools that are AI based.

It's actually accelerating the speed and judgment of really great clinicians and it might actually be slowing them slowing down the junior ones. How will we manage all these AI agents? 'cause ultimately there's gonna be a bunch of them and they're all gonna be performing specific tasks and I need to orchestrate them into something that feels like I'm creating an end-to-end workflow.

Um, and of course they have to do that alongside the humans who are doing various things that the agents cannot do. Um, is there gonna be some sort of orchestration framework for all That? Well, it sounds like you're gonna go build one.

It's definitely that the opportunity in the future. I, I completely agree with you. I, I think orchestration across all these agencies is, is pretty critical.

I mean, I was just looking at like, uh, we're just starting to see some of these like MCP protocol, um, power products and, and clinical software use cases. I, I, I think you're gonna see an adoption curve have been a lot slower in healthcare technology than you would anywhere else. Um, I'm really curious to see sort of that secondary, tertiary level of like agent to agent and kind of workflows or technology enable our protocol enablement.

I you're just gonna need a level, a critical mass level of infrastructure on software access is across healthcare systems writ large in order for that kind of an ecosystem to really take off. I have my doubts as to how quickly that will happen. Um, but certainly would be excited and ecstatic to see all that kind of stuff shake down so that like most regulated industries, I'm, I'm really curious to see, you know, what are the sort of different in industry pressures or, or business pressures or regulation pressures that would force our entire healthcare industry to move in that direction.

Totally. Or in totality. I, I don't know.

And then that's something we're gonna be watching pretty closely. I feel like there's a lot of people experimenting with various things, and that's all great, but I can't help but wonder sometimes if, if, you know, they're basically mapping out a project that's probably cost about a million and a half dollars to execute, given the cost of GPUs and everything else to replace, you know, two people making $50,000 a year. So do we have a, a good sense of the map and the economics of AI and healthcare?

Uh, I don't think so. I mean, to be, to be honest, i I, this is something that I've been, it's on my list to go do some research on, but I'm trying to figure out sort of the power law dynamics around, you know, compute and, uh, tokens and what it actually costs, like our ecosystem and energy to make, to make those calculations versus, to your point, a human that has just been trained over time to do those exact same like judgements. I, I actually have no idea That's on my, that's literally on my like t do list of other things to explore.

I'm really curious to know, you know, how that, how the economics of it all shake down. 'cause uh, my suspicion I know is a lot of it is obfuscated around sort of these foundational model companies that have, you know, pour a lot of investment into these tools, but therefore, you know, it's not always clear, right, what the actual cost is, right. For, for some of these tasks that we're trying to automate.

Ryan, folks, will, you heard it here. Maybe we need to start with the ROI and work backwards. 'cause otherwise we're just playing around with a bunch of AI stuff for grins and we'll see what happens later.

It's really interesting. Yeah. Yeah.

Hey Isaac, thanks for being on the show. Yeah, thanks Mike. Having appreciate it.

ai video series. You can catch this in other episodes on our website. We invite you to check them all out till then, we'll see you next time.