This transcription is provided by artificial intelligence. We believe in technology but understand that even the smartest robots can sometimes get speech recognition wrong.

The 229 Podcast: Building AI-Powered EHR Tools and Educating Staff with Dr. Stephon Proctor

Bill Russell: [00:00:00] Today on the 2 29 podcast.

Stephon : it really got the juices flowing for a lot of people to see just how useful embedding something like an LLM into your EHR could be in terms of an opportunity.

Bill Russell: My name is Bill Russell. I'm a former health system, CIO, and creator of this Week Health, where our mission is to transform healthcare one connection at a time. Welcome to the 2 29 Podcast where we continue the conversations happening at our events with the leaders who are shaping healthcare.

Let's jump into today's conversation.

All right. It's the 2 29 podcast and today we are joined by Stephon Proctor PhD with children's Hospital of Philadelphia, associate Chief Health Informatics Officer for Platform Innovation. Stephon, welcome to the show.

Stephon : Hi, bill. I'm honored to be here.

Bill Russell: I'm looking forward to the conversation. We actually got connected. Because I read the LinkedIn post that you [00:01:00] did following, essentially bringing all these disparate data sources together, putting a chat interface on top of it, a generative AI interface on top of it, and essentially allowing.

The clinicians, at least in a test environment to be able to interrogate the medical record and get a generative AI response, which I thought was really fascinating and worth discussing. I'd love for you to talk about that. And we are gonna talk about that a little bit. But first what is a full stack clinical informaticist, which is what your profile says.

Stephon : So for those who don't know, an informaticist. At heart is really good at bringing the clinical and the technical together. We're really like the bridge builders between that and there's different levels to which we have expertise in things like the technology end. And for me, because I really enjoy technology, I'm really self-taught in everything.

I describe myself as really understanding how everything works from the clinical end getting it into something like your EHR, [00:02:00] how to redesign and understand how that works, but then even going much deeper into. How are all that health data ending up in a data warehouse and how to use either programming or SQL to really understand that?

And so in many ways, I feel like I have a very end-to-end view of what happens in the office all the way down. You know, to the database level and everything in between, which makes it really helpful. So if we are designing a new clinical interventions or new tools, I have an understanding of where all the pieces need to connect to make it work.

Bill Russell: How the data resides, understanding the workflow, understanding how the clinician is going to utilize the system, gives you a distinct. Viewpoint of how the data's flowing through the system, and then how to get the data back out of the system.

Exactly. And that brings us to the conversation today. And there's a lot of discussion right now around how generative AI is gonna be used in healthcare. And I love when I see people who [00:03:00] are experimenting, pushing the envelope. And I'd love for you to explain a little bit of what you and the team were doing over there at CHOP with regard to generative ai.

Stephon : So we're doing a lot of things. On one hand, we are implementing a lot of the generative AI products that our EHR vendor Epic is developing. And on the other hand, we've started to do some prototyping and experimenting with creating our own apps. The LinkedIn posts that you're talking to.

Was about an app that I developed called Chipper. So to kind of take a step back like most organizations that were giving their clinicians access to generative AI tools, we had our own instance of open ai. We called it chop, GPT. It was a web-based way for you to ask clinical questions in a compliant environment.

And me and a lot of other clinicians felt like. The friction caused from going to your EHR to this webpage was, you know, great. You know, I think as you know, if you have a user go from one [00:04:00] thing to another, you're really gonna lose them in between. We really want everything to be integrated and seamless.

That's really good design. And so for a while, instead of copying and pasting things in between. Chop GPT and Epic. You know, I had that, you know, you remember like those infomercials, like there's gotta be a better way. And that's exactly kind of where it started. And so I started to poke around and I had known for a while that with other vendors and tools that we have that you can embed tools or even into Epic.

So then that got me thinking, well, could we embed chop GPT into Epic? And if so, what would that look like? And then I started to expand that further, that if I could get that in just as a webpage, but that's not gonna be good enough, right? Clinically, I want it to be aware that when I ask it a question, it knows about the patient that I'm talking about, not that I have to kind of provide that information.

So then I started to explore what technologies [00:05:00] allow. This application to talk to Epic. And then we use Smart On Fire, which Epic has that built out for a lot of vendors to be able to use for their products. And so again, going back to that full stack informaticist, I really just used a lot of the skills I had from web development to SQL queries to you know, and then using AI tools on top of that to help me code that.

And so I think that was kind of another. Aspect of technology that it's come so far that someone like myself who, you know, I'm a dabbler, I like to dabble in everything and AI really like ups the level to which people who really kind of have a general understanding of technology can take it to the next level.

And I'm really hoping that for informatics, they can do the same thing. So essentially I built this web app that allows us to query and ask questions about the patient. Everything from. their medicine, their lab values, [00:06:00] their demographics, but then even use generative AI to start generating things, letters, asking questions.

And then we thought, well, could we go beyond that? Could we integrate it with other sources? Right? Could we have a query? Things like an external database PubMed to help us look for medication and so on. And really what was amazing from that post is just to see the huge reaction that we were getting from a lot of people who o e, really wanted something like that.

And two, it really got the juices flowing for a lot of people to see just how useful embedding something like an LLM into your EHR could be in terms of an opportunity.

Bill Russell: Well, I would assume that a lot of people are. For Epic or Cerner to roll out something like this? what kind of timeframe do you think that's

to happen. And by the time that happens, what kind of things do you think you'll be able to do at Chop? Yeah with regard to a tool [00:07:00] like this.

Stephon : the recent UGM Epic actually announced that they were gonna go live with a similar type of interface where I think they were gonna call it live Insights or conversational search.

And essentially it was a similar type of product where it would be embedded in Epic. And you have, would have access to all the data that it currently has now. And so they have put it out at some time in 2026. And those are always subject to change. I've seen some demos of them. I've spoken with some of their developers.

It is pretty impressive. I think it is something that I think many of us were surprised it already wasn't something. But clearly they had al also been thinking about that as well. I'm also not surprised that if other vendors, you know, especially the ambient scribe vendors would start to do the same thing as well.

'cause really, ultimately what I did wasn't groundbreaking in terms of technology. I really connected the dots that were already there.

Bill Russell: I want to talk about the technical architecture a little bit. Yeah. Because I mean, you talked about Smart on [00:08:00] fire, you talked about standard SQL queries, essentially.

You know, the question is, how are you gonna provide the context for the LLM to make sure that it can find the information it needs, and then how are you gonna provide the guardrails? So talk a little bit more. Yeah, I mean, is it, are you using like a rag backend? . yeah.

Stephon : Yeah. So when the application is loaded. You know, smart On Fire is providing information about the patient, like who they are and everything like that. And then every time you query it, essentially what the LLM is doing is looking at what the user typed and then using a series of tools to try to figure out, well, this person is asking about vitals.

Let me do my vitals tool and then fire a query and then come back with what that person's looking for. So that's really how it's getting that information. So it's always a series of instructions. And then the LLM. Grabbing that information directly from the patient's chart.

Bill Russell: Yeah, so that's interesting.

That's similar to some of the tools that are out there the make and the innate ends and the whatnot. Essentially you [00:09:00] have a, I mean a brain in the middle of it, but essentially has a series of tools that are sitting there and based on the question you ask, it says, oh, I know what tool you want me to use, I'm gonna go, I'm gonna hit this tool.

This tool does essentially a basic or, I mean, not a basic, but it does a standard query of some kind looking at the data set and returns to the information. How do you provide the guardrails to make sure it doesn't. Get outside the lines. I guess

Stephon : even when I spoke with Epic, they also had the same concerns as well.

And that is something I think we're still, US and other organizations are still trying to figure out is what are those guardrails? what are the. Things that you're trying to prevent and then how do you respond to that? One of the biggest areas that we still are exploring is how do you prevent some of this data from going out, right?

So if you connect it with a tool such as PubMed or UpToDate or up whatever third party, if you send a query, it is possible that the LLM might send PHI out [00:10:00] to that other source. And then the, you know, you've essentially, you know, leaked that PHI and so you have to think about whether you have some other technical layers that are identifying PHI before it really goes out the door, so to speak, so that you don't create you know, a breach.

Because like with many of us, you are required to have something like a, BA, a business associated agreement. To make sure that you know the information that you're sharing with patients, that vendor is gonna protect that information. So we haven't solved for that yet. I also spoke with epic as well and just to try to see what they're thinking about as well.

And they have the same concern. Their version wouldn't be doing anything like calling any tools that would go externally. So as long as you don't have your tool. Going, sending information outside of it, you've already kind of created that layer there. The other then is, well, how do you provide protections on what the LLM can provide?[00:11:00]

And that's general, that's not specific to this, right? You or I could use ChatGPT and you know, put in information and probably get medical advice. That's not helpful or inaccurate. So that's a general consideration that. Everyone has to think with about, with LLMs, but really at the most basic layer, you wanna be thinking about privacy of our patient information.

Bill Russell: I was gonna talk about adoption and workflow integration, but at this point it sounds like you have a series of champions that are utilizing this tool or really testing the tool at this point, what, how many people are utilizing it, how many people are looking at it, and just getting a feel for how it works.

Stephon : So we haven't put it into production yet. this was really a prototype and as we sat back and evaluated, especially when we saw what epic had announced at UGM, we really had to think about what's the amount of effort in terms of time, costs, and resources to further to develop this. If they're going to eventually do something anyways, so [00:12:00] right now we've sort of put it on a pause as we start to kind of get a layout of what you know, epic and other vendors are gonna do and see whether a more general use case is what we want.

Or do we wanna develop the app for a very specific use case. This is often a debate you see within, do you you know, let's the vendor take care of something that's general, which Epic tends to do, or do you focus on a very particular use case that you know for the time being they're never gonna focus on Right.

It's a certain subpopulation or a certain workflow that is very specific and only affects a small percentage of users.

Bill Russell: from a generative AI standpoint, so you have the web interface that people can copy and paste and it's HIPAA compliant and protected. You know, how much usage does that get at a health system

Stephon : We are seeing growing usage of that. I think some of it is hand in hand with first give them access to the tools, but I think the bigger thing is really promoting the literacy of that. I think a lot of users either are [00:13:00] unfamiliar with how to use it, and that is, you know, one of those things that helps drive adoption.

So at Children's Hospital of Philadelphia, we have what we call a school of ai. And so we are actually doing in-person and virtual workshops to teach our employees what is generative ai, and then how to use it for clinical and nonclinical use cases.

Bill Russell: what does that curriculum look like and what kind of people are coming to that course?

Stephon : Yeah, so we actually just had a hackathon a couple of weeks ago, and it was broken up into three groups. We had operational we had researchers and we had clinicians into that. So we're really trying to be very broad because as you know, generative AI can really focus on a lot of different workflows, so we're keeping it very open.

I think today we've got over a thousand users who have. Took part in our various curriculums and they are everything from, you know, what is generative ai, you know, what are LLMs to, how do you use it to help with clinical summaries or how do you use it to help with drafting [00:14:00] documents and all of that.

Bill Russell: what do you focus on right now from a clinical informatics standpoint?

Stephon : Yeah, so implementing or discovering and implementing and evaluating epic's generative AI tools. So we have about nine in various stages of being enterprise being deployed across the enterprise to being in small production pilots.

We were one of the first organizations to go live with the in baskets, AI drafts or art. And then from then on we started to expand into the note summarization for inpatient, outpatient. We're also working on ambient scribes and finding a lot of success with that. So that's really where a lot of my focus has been on more recently is just looking at each of these use cases and evaluating them and seeing where it works.

Bill Russell: Yeah. What's it like going to UGM and then coming back from UGMI? I mean, you're talking about nine they rolled out what, 150 each year for the last two years?

Stephon : I just actually asked them for another presentation. They have 131 on [00:15:00] their roadmap that's including like what's available and what's in their future.

So it is a great deal. As you can tell, they are just looking for every opportunity to put it into a variety of workflows which is really exciting. At the same time, it's coming really fast, and so in some ways you have to kind of decide how much can you keep up with that pace.

Bill Russell: is the demand more a pull from the organization or are is the is the organization, does it have a governance group that's looking at it saying, you know, these five are the next tools that we wanna see implemented?

Stephon : So from the top down, we have a lot of support. You know, even within our strategic plan AI and automation is a pillar of that. So it's great that we have that support there. And then in terms of governance, we do have an AI governance group. And so all of our AI features need to be presented there when, where we're getting feedback on how to construct the pilots, how we're gonna monitor it.

Looking at safety any legal issues and everything like that. A lot of [00:16:00] the scoping for these actually comes from me and my team looking for what are those next things that Epic has available for us. Sometimes Epic will reach out to us and say, Hey, we have a new feature. We want you to be an early adopter, and we will partner with them really closely on.

Piloting that. Other times we have to wait for that feature to be maybe piloted with another organization before we can step in line and say, we wanna look at that feature as well.

Bill Russell: I I do wanna come back to adoption, you know, is it, Sort of assumed if you go to work at Children's Hospital of Philadelphia that we're an innovative organization and we're gonna be rolling out new things.

Or do you, or, I mean, how do you get gain adoption? How do you make sure that you don't overload the clinicians with, I don't know, with too many things?

Stephon : Yeah. I think we started to socialize. This innovative mindset a couple of years ago when CHOP invested a lot of time and money into what they call an epic refuel, and that's essentially where you overhaul your Epic [00:17:00] system to bring it as close to what Epic calls their foundation model, really the way that they intended it to be built, because we knew that setting that foundation would allow us to add layers of innovation on top of that.

So once we started to move through that and maintain that, it was really inevitable that. AI would be another layer of features that we are looking at. On top of that you bring up a really good question about this overload though, right? If people feel like there's AI in too many different areas, you know, , is that gonna lead almost negatively in, in some way, right?

Like you're trying to help them be more efficient, but now you're giving them too many tools. And that's something we definitely have to be cognizant of every time we roll something out. I do hope, and I have seen in some ways that there's a little bit more of a coalescing of these tools that you might have separate tools, but eventually be they become a similar tool or it becomes a similar ecosystem, so you don't really feel like you're having a separate tool for note summarization and then something for your in [00:18:00] basket, but that when you're in the in basket, it's summarizing notes while you're in the in basket.

Bill Russell: The promise of technology is that eventually it's gonna fade into the background and be invisible. And we're essentially going to get to a point where we're just talking to a computer and yeah, there's new features and things happening, but we don't even know that there's a new tool that's been introduced.

I'm curious when you guys, so you, I mean, everybody I know at CHOP is just fun to talk to. You guys are on the cutting edge of a lot of things. I was wondering, when you guys sit around and talk about what the future. Might look like for a clinician at Children's Hospital of Philadelphia, say three years from now.

I mean, is there how do you describe that? What do you think it might look like?

Stephon : It's going to involve. A lot less friction is my hope. We are actually in the stages of building a new tower. It's gonna be 28 floors inpatient tower, and from the ground up, they're already thinking about how can we build this room in terms of you know, the layout, the lighting, the technology so that it is a lot more [00:19:00] digital.

Right? Do we use, do we have microphones that are available? Do we have cameras that are available? In a sense that, to which the room is in a way. Digitally enabled for new technologies such as ambient or maybe in the future when you start to have AI that starts to. Not only hear which is where we're at now, but see right.

Could you imagine that? You know, clinicians, and nurses are walking into a room and there's a lot less of interacting with the computer and interacting with the patient and having the technology take care of a lot of the documentation or teeing up of the orders that you're already doing in the system. So I hope is what you start to see is.

Not just, I have an easier way of communicating with the computer, but that the computer is actually doing a lot more agentic things. And Epic sort of highlighted this as well, is that the first stage was getting the AI to write things for you or draft things, but how do we get it to start taking certain [00:20:00] actions for you, like teeing up orders or reminding you that something needs to be done and taking you to that activity.

So I think that's what we're gonna start to see is. A lot less friction of you feeling like you have to manage the EHR and move between screens and all of that, but that a lot of that is done with you and then you're, you know, approving that and spending a lot more time in you know, direct patient care.

Bill Russell: Stephon, I wanna thank you for your time and thank you for getting out there on the edge with this stuff. I appreciate it. Yeah, love the experimentation, love seeing where you guys are taking things and others. I think that community epic the people outta Stanford and the stuff that you're doing is moving things forward.

It's gonna be fun to watch in the next couple of years, certainly. Thank you. Thank you.

Thanks for listening to the 2 29 podcast. The best conversations don't end when the event does. They continue here with our community of healthcare leaders. Join us by subscribing at this week [00:21:00] health.com/subscribe.

If you have a conversation, that's too good not to share. Reach out. Also, check out our events on the 2 29 project.com website. Share this episode with a peer. It's how we grow our network, increase our collective knowledge and transform healthcare together. Thanks for listening. That's all for now.