Insight ON HCA Healthcare's Bet on Junior Talent — Why Inexperience Is an AI Advantage

Audio transcript:

HCA Healthcare's Bet on Junior Talent — Why Inexperience Is an AI Advantage

Sara Liao-Troth (00:01):

My dreams of my eventual side hustle of becoming an influencer is not really going to work because you require so much equipment that I have no concept about.

Jillian Viner (00:11):

I don't know about that. We walked by ... Yeah.

Sara (00:15):

I don't really want to talk about it. I just want to record traveling and things like that.

Speaker 3 (00:20):

You don't want to talk about AI?

Sara (00:24):

You need a little bit of separation of a church's state a little bit. It's

Speaker 3 (00:28):

True.

Jillian (00:32):

You're all good? Okay. Before we get started, I'm going to have you look at this camera and say your name and title as you would like it to appear in a lower third.

Sara (00:42):

Sarah Leal Troth, AVP of Data Science at HCA Healthcare.

Jillian (00:46):

Beautiful. And Jeremy, if you can hit my timer, thank you so much. All right. Sarah, thank you for being here. It's lovely to have you on the podcast.

Sara (00:59):

Thank you for having me.

Jillian (01:01):

Just to ground the conversation today, tell me a little bit about HCA and your role.

Sara (01:06):

So HCA is a very large healthcare system. We have about 190 hospitals, thousands of care centers across the US and the UK. And we provide care to a large number of patients annually. My role within HCA is to guide the data science strategy that provides the intelligence that's at the center of a lot of our AI-powered technology.

Jillian (01:36):

I understand that HCAC is about probably more than 44 million patient encounters a year, which is an obscene number, but also that's a lot of data. And for a long time, a portion of that data was kind of untapped. What changed?

Sara (01:55):

So what's really interesting is I would say about 50% of that data of all of those 44 million patient encounters is embedded into what's called free text or doctor's notes or orders. We have, with the advent of the electronic health record, a lot of structured data that's captured. And that is what really drives how we provide care today. Anybody in healthcare would be very familiar with that. But what's really interesting is that there's quite a bit of it that's sitting in notes that the nurse is typing because they're sharing it with the next nurse that's coming on.

(02:36):

It's freeform. When you think about when you go visit a doctor, a doctor at the end of that visit will dictate that person's summary of what happened to you and what are the things that they recommend. Those doctor's notes are actual real pieces of assets that's attached to the patient record, and that's just free text. It's free form. So you have a lot of information that's sitting there that is previously untapped because we had no way to process that in a meaningful way. But what's changed now is with the rise of generative AI, we've done a lot of experimentation. And what generative AI is really good at doing is it can extract, it can summarize, it could really pull data together in a way that used to be really infeasible, incredibly manual, and absolutely not able to be done at scale.

Jillian (03:32):

So what has that unlocked for you? How does that change either the patient experience or even what that means as a data scientist?

Sara (03:41):

Well, we're hoping that we can use it to improve the patient experience because I think in order to really deliver on a good patient experience, to do it safely, we need to provide a lot of intelligence to the care workers, to the clinicians that are taking care of the patient. A lot of times you need to do that in real time. You need to weave together a lot of different data sources, and you need to present it in a way that's very easy to digest for someone who has multiple patients, multiple concerns on the floor, and they don't have time to read screens and screens and screens of text. So we are hoping to be able to take a lot of that data that's generated and to process that, process that for many different reasons, turn that into information that gets fed into data science models, and those models kick out intelligence that then help the caregiver know, "I really need to focus on patient A right now and then patient B, not the other way around because there's something that needs to happen right now." Or to say, "You know what?

(04:52):

Patient A, B, and C are really important because we need to do this for them and we can then not have to worry so much about patient D, E and F because we've got them covered in this other way." That kind of aggregated level of intelligence is very difficult to do when you are a single nurse on the floor and you just see information about one patient or the various ones that you have. We're really looking to have that data and intelligence drive this very systematic view of the enterprise of the whole hospital as a living organism at that point in time.

Jillian (05:30):

I have to admit, when I think about hospitals and that workflow you described, the data scientist part of that equation never really comes to mind. What has this meant for your team? How are you thinking about how you need to staff or augment what your team is doing?

Sara (05:47):

Well, absolutely. Data science is a skillset that we need in healthcare because if you think about it, healthcare, in order to provide care, a hospital to provide care for the thousands of patients that pass through stores, you need to coordinate across thousands of clinicians. They all should know the same thing about the patient, but the patient is in the hospital for a very specific reason. So the information is complicated. It's ever-changing. You have new information coming in all of the time. That information is kind of in little bits and pieces. You go in and you have a chief complaint. You then have labs. You have this lab coming in, that lab coming in, that lab coming in. Somebody needs to read and understand that lab. Then they need to make a decision about what to do for you. Then they need to tell somebody to go do that.

(06:39):

Then somebody needs to do that. Then we need to then do additional tests to see is it working? This is coordinated across thousands of people, and we're trying to do it in a way that is specific and customized for every single patient. In order to have that happen, you need intelligence. So data science comes in because we can pull data together. We can summarize and aggregate. We can predict. We can really say, "This is what you should do next." That has an impact on the kind of data scientists that we are trying to build our team with. You need to have a really strong skillset in terms of the math, the quantitative aspects, statistics, a lot of the traditional machine learning. But now with the ability to unlock 50% of our data in all of these unstructured notes, you really need to have the skills to be able to use generative AI in a way that is very natural in your overall toolkit.

Jillian (07:40):

I want to press you on something that's a little bit sensitive because we've seen different points of view and headlines about certain roles potentially being eliminated by AI or levels of roles being eliminated. What is your viewpoint on the type of experience or the role that you would look for to hire because of AI? Is there a positive spin to be found here?

Sara (08:05):

Absolutely. And I would say that my perspective on it at this point in time is a little bit counter to what might be discussed in the overall media. Right now where you see a lot of headlines about AI is going to lay off thousands of software developers because AI can do it all. And I have no doubt about that. We are playing around with AI to see how that will make our software development more efficient. And I am fairly positive that Next is going to come, the same thing will come toward data science. How much of data science can be automated by AI? And I think that's a very valid question. Here's the thing, I'm excited for that to happen. I know it will happen, but I also need to have a strong team who are going to be at the forefront of not only making it happen, but figuring out how to take that and to use that tooling to be able to execute data science safely, accurately, and productively.

(09:06):

So you still need experienced data scientists to be able to adopt and figure out how do you make this technology really work in a way that really is working, right? Not just the hype, not just what looks impressive, but really what works under real constraints. But on the other hand, you can take advantage of having junior level talent who may not have had enough experience to get stuck in a way of thinking

(09:39):

And have them come in and be available to break the technology, to really push it to the boundaries of what can work and what cannot work. And so in that way, I've really started rethinking how I'm kind of shifting the balance of my team to be able to, in a lot of ways, have our cake and eat it too, but also adapt to the times. You still need people who have had the experience, who know what the math is supposed to do. You also need people to come through and to break things and not know how to do things the old way so that you can merge those two together and then hopefully come up with a really strong team that can very much adapt and flow with the times. We don't know how much AI is really going to change. It will change things, but I think the key thing is how it's going to change is what we're going to try to adapt for.

Jillian (10:39):

It sounds like what you're arguing for is you've got a group over here with years and years of experience who really understand intricacies, and that group of people almost have their brains hard coded into what's possible and what's not.

Sara (10:52):

Yes.

Jillian (10:52):

So you're leaning on this newer talent who has no idea the boundaries to push and try things. And then by bringing those two teams together, that's where you're expediting innovation.

Sara (11:04):

Absolutely. Absolutely. And in that way, I think the idea is that you get this kind of melding of perspectives, and hopefully that will happen organically as the team starts working together in that way.

Jillian (11:19):

I love that perspective. I love how fresh that is, and that gives some hope to people that are entering the workforce. Walk me through what that actually looks like when you have a new problem on your team. You hand it over to a more junior teammate. Do they get hung up on, how do I do this? Or do they just start putting things into ChatGPT or Gemini and they just figure it out? Have you seen any really interesting innovations come out of this experiment so far?

Sara (11:49):

Yes, I have. I have a great example for this actually. So we use this kind of junior level talent with the senior level guided exploration. In one particular example, we were looking to understand with spine fusion surgeries, which happened quite often in the US, the data that is kicked out of that information really tells you whether or not you have a single fusion of a spine or you have multiple spine segments fused. And why that's important is because if it's multiple, in a lot of ways, I kind of really do want to know, did we fuse five segments or did we fuse 10 segments? And the reason why that's important is because you need to find the denominator in order to be able to benchmark or use a lot of the data to compare apples to apples that has an impact on how we benchmark surgeon performance, how we benchmark the amount of osteobiologics that are used.

(12:52):

It impacts vendor strategy. It impacts so many different things in the enterprise. That information is embedded in free text in the postoperative notes. It's very confusing, a lot of nuance there. And we were able to use interns to actually, in a very targeted way, figure out, can we extract meaningfully and consistently how many levels of the spine were fused in order to create a data set that we then pair up with cost data, outcomes data for numerous people to use at the enterprise, whether we're doing research, whether we're trying to manage supply chain, anything like that. And it worked really, really well because we were able to have a very targeted specific problem for the insurance to go off and solve. They are overseen by geno data scientists who are, of course, working on their own projects and needing to deliver. The minute they find out something new, they kind of come back and get a gut check.

(13:55):

Is this the right direction? Is it not? Should you spend the next week on this? Should you spend the next month on it? Should you spend the next day on it? And it is through that process that you hone and iterate a capability that works. And we were very successful in doing that. This is actually something that was very accurate. And to me, as a leader of data science, what it also tells me is that this effort is fruitful because not only did we solve for this particular use case, and then I have data and intelligence to give to all of my business partners, but we also solved for what I would consider to be a core data science platform question. We have all of the data out there. It's in its most raw form. If you're willing to pay for the compute, you can go back and you can create custom datasets with enough history for model training specific to answer many types of data science problems.

(14:58):

And you can do that to specify these datasets for much of what you need without being stuck with saying, "Oh, well, we need to start collecting the data. Let's wait five years before we have enough to work with. " That's completely changed

Jillian (15:14):

Now. I love this model. You basically gave a very clear problem statement that seemed like an impossible task and you put an intern on it. So you got a solution for a very low price. You didn't have to put this to your highest level engineer.

Sara (15:31):

No, I mean, because oftentimes, and this is very common in enterprise, my highest level engineer, my highest level data scientists, they're also out there. They're solving the hard problems for a very important business value-driven problem right now. They don't necessarily have the time to play, if you will, in the nicest way at the point. So I'm trying to take some bets right now that I can put some interns on things that we want to see if we can get something out of. And then if it bears some fruit, then we figure out very quickly how we bring the team in to turn that into kind of a data science level capability that then everybody contributes to and then builds out. And that way you're minimizing risk at many different levels.

Jillian (16:20):

This is so fascinating. And it's such a great example of a company having that innovative mindset. It's so hard right now because everything is changing so fast. The ground that we walk on feels so unstable because every single day there's a new model release, there's new capabilities. How do you plan your technology strategies or just your programs or your work on with all of that uncertainty?

Sara (16:48):

I think the key takeaway here is that you can plan to some degree, but I might argue that you shouldn't be planning. I think you need to just get started. You need to get started with experimenting. You need to get started with understanding what new technology can do for you. Does it work for you specifically with your context? In healthcare for us, patient safety is very important. So there's no room for hallucinations. There's no room for the model to do something probabilistic. And so that dictates what you have to build around the solution to make it work. That's not necessarily relevant to every single person out there using generative AI, but it is relevant for us. And when the ground keeps moving on us because you have new things coming along, my biggest advice is you can never catch up to that. You can never catch up to that change.

(17:44):

So we are investing in the capability internally to be able to adapt to that change. So it is with these small targeted teams. It is with being able to get moving quickly so that you understand very specifically what it does for you and what it cannot do for you. It is trying to be opportunistic in a very strategic way and then to start building because if you just get paralyzed by waiting for things to settle, I think that that's the wrong answer right now.

Jillian (18:18):

I have to ask because you're in healthcare highly regulated. We're talking about patient data. There's very few things that are as sensitive as that. How are you using AI in such meaningful ways with all of those regulatory protections? How are you doing this responsibly? What frameworks have you put in place?

Sara (18:36):

I think from a data security standpoint, anything that is relevant to patient privacy, anything like that is just as relevant internal to HCA as it is external. So a lot of consideration in terms of patient data going outside of HCA, we don't really want to have any concerns with patient privacy. Internally, even for data scientists who are working with our internal data, we go through processes to make sure that you're stripping it of the PHI, you are obfuscating it, you are anonymizing the data sets so that you're looking at trends in aggregate. You're not really talking about a specific patient. And so I think that's really important. From a broader responsible AI standpoint, I think that really goes into making sure that whatever it is you're building is going to consistently do exactly what you're asking it to do, no more, and that it will keep doing it in a way that we are managing the inevitable data drift, model drift, output drift.

Jillian (19:46):

If there's one thing that you want a leader to leave this conversation with full belief in, what is it?

Sara (19:54):

I would say, especially from a data science perspective, we need to embrace the new world. We need to be ready to take advantage of it, and that means that you need to just get started.

Jillian (20:12):

Simple as that, just dig in. Sarah, thank you so much for your insights today. It was really lovely to have you.

Sara (20:19):

Thank you so much for having me.

Jillian (20:22):

All good. Perfect. I don't know if you can, but if you're comfortable, would you give us a soundbite about the relationship working with Insight?

Sara (20:31):

Sure. Let me think about that. I can say a number of different things. What would be ... Maybe give me some parameters, not parameters, like what is useful to you guys?

Speaker 3 (20:51):

Talk about being a partner, like the expertise of people, understanding your business and what you would hopefully expect from just a partner who comes in and works with you to build the solution tool.

Jillian (21:04):

Yeah. Anything that makes us different than other partners you might've worked with in the past. Yeah.

Speaker 3 (21:09):

Not just coming in and being transactive, but really what we call.

Sara (21:14):

Okay.

Speaker 3 (21:15):

And also all this will go through your team earlier. Does that sound like you're competing like, "Ah, we're not sure about that development chance."

Sara (21:23):

Okay. Think about this. Okay. Okay. Whenever you're ready. Oh, okay. I've had the pleasure of working with Insight for a number of years now, and there are two things that stand out to me in terms of the relationship that we've had. First is that Insight is one of the few vendors that I've worked with in my career, and I've worked with many, where I do very much feel that they do understand the problems that we're working with very deeply and have a level of technical expertise that they bring to the table to have the conversations about. So it's not starting over from you. It's not superficial in terms of understanding what we're trying to solve and how we're trying to do it. And I very much appreciate that because sometimes it's more fruitful to work with insight from a thought partner perspective than just a vendor perspective.

(23:09):

That's the first thing. And the second thing is really that I very much feel that the technical talent coming from inside is top-notch, and that is very difficult to consistently get in the marketplace.

Jillian (23:27):

That's great.

Sara (23:27):

That's very ... Okay. The idea stands, but I feel like my delivery wasn't very smooth. But anyway, yes.

Jillian (23:34):

That's great. One last thing, if you wouldn't mind signing our release form, and like Travis said, we will make sure that you and your team get