The session explores the exciting possibilities of how, when used appropriately, AI tools in the electronic health record (EHR) space have the ability to assist clinicians in numerous ways to improve clinical documentation.
I'm Thomaszelski. I am one of the planning members. I'm a faculty member in the Department of Medicine and very excited to, uh, introduce, um, our speaker today, uh, Doctor Jeff Sattler. Before I do that though, I want to remind you that please submit any questions you have at the end via poll everywhere. There will also be an opportunity if you want to ask questions verbally, there's microphones up there, so you can go ahead and do that. Um, the poll Everywhere app will come up, so just scan the QR code now and then you can ask questions and we'll get to that at the end. So with that, um, I would, uh, I'm very, uh, pleased to introduce Doctor Jeff Sattler, who's an internal medicine physician, uh, hospitalist and system medical informatics physician for Saint Luke's Hospital in Kansas City, now part of BJC West in our, uh, BJC family. He is originally from Saint Louis. Uh, he attended, uh, Saint Louis University High School just down the road. And then Saint Louis College of Pharmacy right here on our campus, uh, before attending medical school at Kansas City University and internal medicine training at the University of Kansas. His interests are in improving efficiency, including the incorporation of AI and electronic health records, improving the use of clinical doc uh decision support, and clinical documentation improvement. This session is going to explore, uh, how, when used appropriately, AI tools have the ability to assist clinicians in numerous ways to improve clinical documentation. So with that, please welcome Doctor Sattler to the stage. Good afternoon everybody and thank you, Thomas for such a great introduction. It's wonderful to be back in my native hometown and where I entered healthcare, not too far, just down the street, literally, uh, at the Saint Louis College of Pharmacy a few moons ago. So, uh, just a brief show of hands, who likes this topic? Besides me. OK, great. Who is here for credit, maybe doesn't care, no one's gonna raise their hand. OK. And then who thinks maybe AI could be a big problem and they have a lot of reservations? We got a few, OK. Including my cousin in the upper deck. So great. All right. Uh, you know, when, when you come to your native town, that the cousins do come out, including the second cousins. So this is important. Uh, anyhow, so I am thrilled to be here talking about AI tools in the HR, our experience in the West, and path forward. So just a couple of quick reminders, we are in a different epic instance than you guys are in the East. So we have some different versions and we have some different things that I'll point out as I go along. So. Um, I really do think that technology and AI can assist with both safety and quality when used appropriately, OK? I think that that is one of the ways that we go forward, especially in the quality and the outcomes aspects, but also in the safety aspects. I do not have any relevant financial interest to disclose, uh, so easy slide there. Some objectives. We're gonna go through some of the main types of AI just to kind of review that. Some of you will know them well, some of you may not know them so well. We'll go through how AI tools can be used in Epic to assist clinicians, not, I'll sprinkle in some applications for folks who aren't physicians or APPs, uh, just because I know we have nurses in the room and others in the room who are, uh, on the healthcare team. And then look at future use cases for how AI can be used. All right, quick introduction. I kind of mentioned how this ties into uh safety and quality some, but also ties into a lot of what we do. You know, in my opinion, when we went to electronic health records, we digitized, but we didn't really optimize very much. And it's taken a long time, 20+ years to try to get to a point where we might be able to start optimizing. The challenge is, is that we're optimizing on foundation of data that may not be very good, reliable, or maybe even consistent. We'll go, uh, and so we'll go through these things here as well, uh, just sections I'm gonna touch on, so I just put a quick overview slide and, uh, experience with early development, um, and there is a, quite a bit that we did at Saint Luke's on that aspect, and then the use cases impact future directions and then end with questions. All right, so I don't think anybody's gonna disagree with most of what's on the slide, but certainly things continue to rapidly evolve. So in 1900 in Brooklyn, it's horse and buggy. Weren't any automobiles. 1920, virtually not a horse and buggy, but a lot of automobiles. So 20 years. And if you see pictures of that, uh, which I didn't put in here, it's quite striking. And so it's been said that over the next 10 years there could be 100 years of tech advancement. I can't even figure that out. No, I, I don't know that the human mind can figure that out. I should say, not me. I mean, you know, I'm a techie guy, but that doesn't make me necessarily more powerful than others. But, so AI and tech are expanding faster than any time in human history. Uh, the goal is that together we're better versus either one alone. And that, uh, we want to improve patient experience and clinician satisfaction, looking at tedious areas, bottlenecks, things that we know are pains, have been pains sometimes for decades. Given AI assist and decrease burden, governance is very important, so I don't wanna overlook that, and this is just a quick slide on governance, but this is some of the guardrails that we can use to help ourselves out, and I think it was earlier, might have been Doctor Dunigan mentioned that, you know, one. You know, maybe large mistake with AI and we slow everything down because of the one example, instead of trying to think about how we can adjust some things, guardrails, mitigate risks, and continue to move forward for the greater good of what it can do maybe across hundreds, thousands of patients, etc. So, this is important. The National Academy of Medicine actually has a very good guide out. Uh, American Medical Association does as well. Uh, there's some here, uh, WashU that were involved in the National Academy of Medicine, including Doctor Tom Maddox. So, for what it's worth, that's a good reference. OK. So types and use. So we have generative AI which most of us are familiar with in terms of summarization. What I tend to think of as composing drafts, uh, reinforcing the importance to edit. Uh, and we'll go over that in, uh, in detail. Trend recognition, ambient voice, and also natural language processing. So essentially taking our old, uh, optical character recognition and natural language processing and letting AI start working with that. Machine learning and deep learning is a very common example. It has been around quite some time. And this has to do with predictive models, deeper trend analysis, and risk scoring as well as health analytics. Agentic AI is your agents or your bots. So there's lots of different ways that these are being explored, um, some early use, uh, especially in communications, that's the easiest use case, but also the closed gaps, provide suggestions or actions, they can answer calls. Uh, I met a, uh, OB-GYN who runs, uh, a series of OB-GYN clinics in Connecticut where they have a bot that answers the phone and people call asking for the bot by name. Uh, it's not like Alyssa, it was some other kind of name that was very comforting, and they call and they want to speak to Alyssa. What are you talking about? That's not, anyway, that being said, so there are some, some ways where folks don't even realize that AI is potentially answering the phone instead of the front office staff, etc. Uh, we can go via text, so, you know, there's a lot of texts back and forth, uh, refilling prescriptions, you know, making appointments, waitlists, you know, an appointment came up, etc. So, computer vision is very interesting to me. It's been, you know, some of it has been around a long time, obviously with EKG, radiology. Uh, cardiology, echocardiograph, you know, echocardiograms, etc. And where you can actually interpret not only still images but video. So imagine a scenario where you could have a, a true virtual sitter that has an AI assist that the AI is trained on that patient room and they know, let me rephrase, they being the AI, excuse me, pardon me, a little faux pas there. Um, and you know, where the AI actually knows patient is likely to get up, they just turn this way this many degrees, you know, they're this close to the edge of the bed, they're doing this, they're doing that. Um, also things more advanced potentially in terms of, uh, nursing being able to voice command back through that system that they emptied a Foley catheter and there was 800 mLs of urine output, and it just goes in the electronic medical record. No flow sheet, no going back to the HR, just done. Or in the instance that The dressing change didn't get documented. That day I nudges the nurse and goes, hey, you went in there and changed that dressing, but I don't see it marked in the chart yet. So, those are just a few things in terms of that. Robotic process automation, again, robotic-assisted surgery, other automation, claims processing, prior authorization, anywhere where we don't need humans to do the tedious process that is being done. Also operations, you know, lots of uh spreadsheeting in that. Uh, Thomas was sharing with me about a spreadsheet example last night as a matter of fact. And then multimodal where we actually have multiple areas and that is something new where we're using You know, graphical image interpretation plus generative things and, you know, those sorts of things. So, all of these are coming around quickly. There's probably more, but those were the main ones I wanted to touch on. OK, big long list. Don't worry, I'm not gonna read it to you. But this is what we have deployed right now in the west. And so we have uh in basket messages, uh, we have actionable follow-ups to radiology which has to do with nodules, lung nodules, etc. uh, AI tech assistant, ambient voice. And then the rest of these on the right side of the screen as well. One thing I do want to point out is the extracted SIG details. So what that is, is when we get medications from an outside source, not natively from the health system, whether it's the east or the west, and BJC it doesn't matter, is it comes in free text through Care Everywhere or through another source. This actually puts that into structured text so you don't have to click all of the buttons. How many people love clicking all those buttons? No one ever. No one. Um, you know, some of my partners still like to count mouse clicks. You know, they don't when they're on Amazon, but they do when they're in the electronic medical record, OK? Uh, there are also localized predictive models, so very, very huge area here in terms of sepsis, likelihood of discharge, patient class, other things where EPIC is coming in and saying, we've tested this out on however many health systems, but we're actually gonna take that model and tune it to your specific health system patients. And then if it's outside of parameters, it can be retuned. And so, um, we've done some very good work with sepsis modeling on that and also some other predictive models. Others, I throw this down there. This is Cosmos median length of stay. We'll talk about that briefly at the end. Uh, but actually using AI on a huge deidentified patient database, instead of just looking for trends in the database, actually using AI to extrapolate some things. OK. I didn't know how well this was gonna show up and I apologize. I know it's after lunch and we've got lots of numbers here, but you can see just a year ago, we we're talking maybe 20,000. Maybe. And now 10 times that, 200,000. Now, there is this big pink blob here and what does that mean? That means that we're looking at a post-visit risk adjustment looking at HEC and non-HCC diagnosis for waiting and for other things to make sure we're not under overbilling uh outpatient claims. So we've started that. So why is this really taking off here? Part of that has to do with we do have access in the West at Saint Luke's at some additional AI resources to have essentially uh a better price per token, uh, a way of uh turning on more things and being more cost effective. So, uh, that is something we just got in January. So a lot of this is taking off as we speak. So these are a lot of the things I had on the previous slide. And then over here you'll see there are thousands of examples of draft hospital course, ambient, there I think there's tens of thousands on the ambient line, uh, extracted sig details, one of the things we talked about, I think is almost 30,000. So, um, and that's, this is January. I took out the pink blob for this specific slide over here. So lots, and this is generative AI. So lots of generative AI going on. OK. Early development. Um, I'll try not to bore you with this, but this has been, uh, quite a bit of my non-patient care professional life for the last several years. So with inpatient examples, since I'm a hospitalist clinically, I'm gonna stick to this, and that is Epic approached Saint Luke's in April 2024, so essentially two years ago about draft hospital course and inpatient insights. And at that point, uh, we looked at what Epic was thinking about development-wise. They shared. Uh, their initial very early beta version, almost like an alpha version of software, to be honest, um, large language models back then were not as, uh, good nor complex, uh, as they are now, and we began some pilot, uh, testing in a support environment, so our sub environment in June 2024. There were lots of iterations, lots of changes. Notice it took The better part of almost 9 months, 8 months to get to production. Part of that was because of the durations needed on the software and part of it was because there were changes in large language models and we keep having to revalidate everything using chat GPT here, but you're going from chat GPT, you know, 3, 3.5, 4. You know, there's lots of uh changes there in terms of what the LLM can do. We're also doing prompt editing and working with Epic peer organizations, etc. So, finally, we get folks live in February 2025 with both tools and we're gonna go over those tools here in a minute. And just looking, it was validated, it's useful and accurate. We're talking 75 to 80+% of the time uh in both tools uh by different folks, hospital, cardiology, physicians, APPs, etc. And we had a, a rollout to production about a year ago. Whoop, let me go back, sorry about that. About a year ago here, um, and then eventually took it to our hospital medicine division in palliative care. So this is 240 users here. And then now rolling out to everybody here uh as of March 11th, which is 1100. So We do not include, at least for now, although coming soon will be some of the house staff, uh, some of the residents and fellows, uh, eventually medical students, although that is still being, uh, weighed out there. So just for the sake of completeness, I'll include that as well. Those folks are important. Early development, we did learn that working with vendor, uh, and, uh, you know, in terms of Epic as well as the peer organizations really was a very synergistic combination. We were able to get through a lot of this iterations and promptings, so on and so forth. The speed of development, I, I kind of touched on already was very rapid and allowed us to eventually uh go live a little, you know, faster in terms of finally getting things in early 2025 rolling. Uh, one thing I will also mention is the big delay between June last year and March this year is largely related to, um, token cost and contracts and those kinds of things. Looking at, hey, how can we justify the cost of doing this? You know, can we look at how much more efficient things get? Can we look at does this save time? How much time are we talking? How do we put a number on that? So that slowed things down even though we had the prompt, we had the user interface, we had everything else, we still had to wait longer than we would have liked, but to our financial folks who keep, uh, you know, good stewardship of all of, uh, those kinds of things, we had to wait. So, eventually we got to where we would uh like to be coming up here next week. Another thing I'll mention is that sometimes we have learned that default settings can lead to very long summaries. I apologize. I did that twice here already this afternoon. We'll try not to do it again. And so this is especially true in draft hospital course and I'll show you that here. This is a huge point and I probably should have picked a different color, but reviewing and editing the AI outputs, Transitioning the user from editor in chief from author. This is not something to where uh another professional wrote it and you can trust it and kind of summarize it real quick and move on. This is uh AI looking at in the case of hospital course notes. We all know that notes can be ambiguous, they can contradict, they can be outdated, they can be inaccurate. How on earth does the LLM handle that? How does it work through that? Took some time, uh, took a lot of feedback, which is another important thing down there below. We had lots of feedback to look through, uh, useful and accurate were kind of our gold standards. You know, this is an initial draft, it's useful and accurate. Is the summarization initially useful and accurate. There may be a few things that you would like format-wise or longer, shorter, etc. but those are the things that we focused on. And the last thing is, is different patient care environments and different roles require adjusting the tools. Hospital medicine use case is different than primary care. It's different than nursing. It's different than orthopedic surgery. It's different than name your role, respiratory therapy, you know, those kinds of things. So, quickly, we figured out even within this that, well, cardiology likes to see a little bit more of this. Hospital medicine, you know, likes to see a little bit more of that, so on and so forth. All right. Let's go through these specific use cases and I hope that these screenshots come through pretty well. OK. So this is one of the inpatient insights. Anybody in the room seen a patient story, inpatient insights from Epic? Thank you, Doctor Williams. Thank you. All right. OK, I got a couple. And my eyes are deceiving me. There may be somebody else out there that I may have missed. OK. So this is an earlier version of what this looks like. So this is kind of the Reader's Digest of the patient. This is why they came in and. Some history things about them in the past, what's going on here during admission. What most recently has happened with them here. It does give you also some new updates here. Hey, since you generated this, there's some new stuff you might wanna know. It also tells you this is 6 hours old and out of date. So, um, you can generate this down here. So this can actually live in multiple different places. So it can live in a patient list, hover, see this. Generate it from there, move on. You can also open up during, in, in the chart. So uh there are footnotes that you can uh hover to see some details and you can expand to see the references. You can click on it and it takes you straight to where it got the information. Notice with this particular piece, we're looking at notes, results, orders, everything that's in there. OK. Not just notes. So we'll get to the importance of that here, uh, in just a minute cause I will contrast it with draft hospital course. So, there are also a couple of other examples that I did not screenshot uh in total like this one because it's just the most comprehensive, and that is recent notes and recent events. So those look at the, it's filterable but we have 48 hours in terms of what we look at and recent events essentially will tell you vital sign changes, lab changes, needed appearing pain medication, antibiotics were changed, hemoglobin dropped. Uh, recent notes goes by specialty and will tell you, cardiology said this, nephrology said that, hospital medicine said this, nursing is working on this. So on and so forth, and these are configurable. 01 more thing. Cause, you know, I'm trying to make you guys seasick after lunch here. And that is there's a feedback button. So if you notice, you know something, just tell us. Tell us what's going on. That feedback button feeds through, it gets exported and we can see everything that was sent up, the prompt, what was sent back down and what the user is saying. So we can control that whole, uh, if you will, that whole cycle or loop. OK. So, what did we learn with inpatient insights? The bolded text can be lengthy. And sometimes people get a little glassy-eyed. I know the older I get, sometimes the easier it is for me to get a little glassy-eyed if I see a bunch of texts, but that's just me. And there's no graphical trends, at least initially. It can be useful in the inpatient side for pre-rounding or catching up on developments, and this is really where uh we have actually allowed our um care progression or uh case management nurses to have access to some of these summaries so they can quickly see what's going on with the patient if they don't know them. And you can share across generations. So if I generate it, Thomas can see it, and then nurses's doing care progression can see it and we would all see what date and time it was done and if it's outdated and can regenerate it and have the icon in our list and build a way to, you know, where we want it to be in terms of reports, you know, those kinds of things. This is very easy to, uh, manipulate that way or overview summaries, etc. So, you'll see here at the bottom. Uh, that we have some graphs here. So this is new. Uh, this is available in, I wanna say it was available in May, but they did an update in, uh, November 2025 hyperspace to where now in the patient summary, it tells you some of the events, but then it also points out what the blood pressure graph is doing with the graph. With, with the graph here as well as with text. So Epic is probably, you know, they're gonna rapidly expand on that to feed into things. If a chest X-ray is mentioned, you can almost envision a little icon here where you could click on and see the chest X-ray, you know, those kinds of things. You can work on these off-patient lists as well. This is note summary. So this is what, you know, medicine's saying, surgery is saying, nursing is saying about the patient just from the list, not even within the chart, just hovering and figuring that part out. Switching to draft hospital course. It's a very early version of draft Hospital course. There's lots of options up here. I mean, look at all these different options. I mean, how on earth would you know as a user which ones to pick? Well, let's go with the default and see what happens. But this is where we started. This is an early version. That I wanted to show on purpose just for the iterative purpose, uh the iterative explanation of how much detail went into developing some of these things and how long it took to get to where we're at before we can just start applying things, uh, left and right, which is kind of where we are now. Uh, the, uh, hospital course note shows up at the bottom. Uh, we figured out, you know, in terms of length, we figured out in terms of time, note types, and really the only button that Epic leaves is this output right now. We have maintained our length button that I'll show in the next slide. And so here, I know this is big and long, just bear with me. We do use the navigator-based hospital course note. I'm not sure that uh you all use that very much in the east. Help me here, Nancy or Thomas, do you guys use the navigator-based hospital course note? I don't think so. OK. So we do. So this at the top is what human has put in in terms of hospital course note in the navigator. And then at the bottom is actually the AI output. And you'll notice there's only two toggles here, length, summary type. You can copy it. And then take this piece here. And either bake it into what you've already been working on up here or copied into a different note. You can do different things with it. Of course, copying also means editing because we don't just want to copy and paste and be loose with this. But that being said, that all of your references are down here. I did expand that for the screenshot. And um this was the workflow up until uh we had an update to where you can bake it directly in the note. OK. So, source here is limited to notes and it can be longer if an unspecified length is chosen. So what I mean by that is This right here. We found that less than 10 is good for most instances, but the unspecified is a little too long for our tastes. But Epic likes that because they'd rather have it be longer than shorter, and you can always edit out what you don't want. So That being said, um, Does, you know, we talked about this editing. Epic has improved the prompt and there is a general prompt that we are using that the Epic organizations are using that was part of going through all of the early development was coming up with a prompt that works because who, who would know how to write a prompt for hospital course? Yeah, exactly, right, so software developers help and, uh, you know, just going through and validating that and working with them is much easier than trying to do it ourselves, never mind, you know, us with MBJC and watch you trying to figure it out and then somebody else trying to figure it out and everybody having disparate outcomes. Epic was very, uh, in tune with we wanted it to be useful and accurate, predictable. We want it to work. The workflow can be clunky in this right sidebar. So you'll notice a lot of what I showed with hospital course. I just showed the right sidebar. I did not show a patient list. I did not show another area where you could generate it in the chart. So I think that that is important to note. And then in October, there's a smart link where you can directly put it in the note. However, you can't toggle the length. So we're still a little miffed about that, pardon the expression. Sometimes you don't always get what you want, but you might get what you need, maybe. All right. Maybe. This is a study by Doctor Small and his colleagues at NYU Langone. This is actually looking at draft Epics draft hospital course. So this is in 2025 JAMA, and, uh, essentially the key points here are that over here in this section. They compared quality of uh AI assisted or large language model assisted hospital course notes for discharge summary purposes, the narrative summary versus human-authored. And they did some quality scoring and uh it's a great article. um, and they found that the LLM generations were actually better in quality than the human. So Now, you can argue about how many, how many they chose, OK? And 100 emissions, you know, you can argue about some other things, fine, right? There it is. And there's an earlier study that I did not put up here that was by a Dutch organization uh in Lancet that was actually they were one of the very early folks in uh in terms of international usage outside of the US and they found that using the draft hospital course saved their physicians. Over 5 minutes. Per discharge summary. 5 minutes per and some people would say, 05 minutes per discharge summary, you know, that's 5 minutes. What are you talking about? Think about how many discharges, think about how much time that is for healthcare professionals, for APPs, physicians, house staff, everybody else. It's a lot of time. Because I can share with you that if I save 80 full-time partners of mine 3 minutes a day, I have half my salary a year. So, uh, and I'm, I, you know, I'm, I'm reasonable, but I'm not cheap. All right. Fair market value, definitely. OK. Just wanna make sure we're above board here. All right. Some additional tools that we're using Ambient, uh, we have DACs in the west. I know you guys have a bridge in the east. Uh, we are using in the ED inpatient ambulatory environment. Uh, we do use drafts, uh, for MBasket, AI tech Assistant, outpatient insights, and I wanna touch on these a little bit as well. OK, so ambient. So we chose Nuance or DAX, Microsoft-based product, and uh we have been live on that pilot-wise for probably, it was a good year before we had enough uh and then we had to negotiate that. In terms of wide rollout, it was this fall. Um. It was really uh August through November, just here recently, but we did have a good pilot group in the ambulatory space that helped us out in terms of decreasing pajama time, decreasing note creation time. Patients noticed and physicians noticed and providers noticed that they're actually looking at the patients of. Yeah, I've all been, you know, I've been guilty at times of doing that. Um, faster close of encounters which impacts revenue, quicker you can bill. The patients, we really did not have a lot of resistance with consent. I know that is somewhat of a touchy topic, uh, something we talked about last night, something that legal folks are certainly, um, you know, sometimes you get different answers from different folks depending on who you ask and depending on which system or. Sometimes even, uh, which side of the same system you may be in. Uh, no, that doesn't happen with us. Uh, that being said, uh, clinicians do notice improvement in their documentation. I've had surgeons say, I actually have an HPI in my note. Like, it's actually there, instead of like patient presents with, you know, Inguinal pain, period. It, so, uh, you can use this in mono mode. So this is something I think is important that folks underestimate. What do I mean by that? I mean literally taking out my phone before I see the patient and pre-charging and just talking to the ambient experiences I'm pre-charting. And it will get a one sentence or two sentence like why I'm seeing the patient. I can do results review and it will bake that into the results section. Things that I would normally do, I'm just using my voice to track that and then I can go in and see the patient. And um again, have the AI turn that into text and capture that and also do my exam. Assessment and plan, much more complicated. Uh, we were told that the ambient would be coming uh in November of 2025. Crickets, exactly. Not here. And for those of us who use diagnosis aware note templates. Definitely use them. That's all we, that's what, I mean, I know it's crazy sounding sometimes, but we do a lot of problem-oriented charting and hospital medicine in Saint Luke's. Uh, the assessment and plain areas, I don't want structured diagnosis aware with free text. We want it one or the other, we wanna be interacting with the problem list, those kinds of things. Hopefully, fingers crossed, ambient improvements will help. And on inpatients, initially, we thought this would be best with histories and physicals and consults. Uh, but I have partners that are actually using it on every patient. Uh, these are, uh, partners that feel some pressure and, and I do this too. It's a, maybe it's just Amount of time in practice that we need to go see a bunch of patients and before we start doing notes. You know, we need to go see not just the 4 or 5 they're on this unit that we have, but we need to go see 15 or, you know, what have you. And it helps them keep track of things in terms of small details in the history, physical exam things, helps them make sure that, you know, the note is uh ready for them when they get back to batch their notes and do them. Um, so there's a debate in our group about batching versus not batching, but we won't get into that because that'd just be too boring. All right. I'm gonna skip in the AI text assistant. And so this is really word processing brought to Epic. Anybody seen this one before? OK, good. So this is the original text. Over here You see, we got, you know, 4 different sections of data brought into a paragraph. So essentially doing what any, you know, Gen AI summarization tool would do whether it was co-pilot or you know chat GPT or whatever. Take this and give me a summary. And then you can easily say, hey this looks great. You can adjust it and say, hey, make it longer, make it shorter, do this, you know, do that and revert. So Epic is starting to bake some of these things in to the EHR. Which honestly is really just an operating system these days for healthcare. We only had about 1000 uses, which is low. We didn't announce it when it went live because we're worried about how many tokens we were using, and we were like worried about the token budget. I was like, whoa, wait, wait, wait, we're gonna turn this on and not tell anybody and they're just gonna kind of find it. We're like, yeah, I'm like, really? There's like this little tiny icon that I didn't even show that I should have shown that is like the magical AI icon. It's like this, seriously. That you hit to do this. Otherwise, you wouldn't know it was there. You right click and you can do the same thing, but folks don't know this is there. Well, uh, the cat's out of the bag, so to speak, now, so now my partners know and the uses have gone up. Uh, but it's very helpful when you have, you know, lots of, uh, text data. Notice this is all text. This is not anything to do with, uh, lab results, fish bone results, images, or anything else, so. All right. Outpatient insights. I included this after some discussion last night. So, uh, one of my partners, Doctor Scott Russell, who is an internist in primary care, worked on this with Epic. And so let's start in September 2024. Uh, don't worry, I won't bore you with as many details as I may have with hospital course and inpatient insights, but it was initially, uh, focused on user-specific search or questions and then quickly evolved into essentially what's happened since the last time I saw the patient. And now is moving into problem-focused options as well. Time filtering can be important, especially with limits on how much information we can send up to the large language model. Remember, literally, it's like every character counts, and eventually you run into, well, if it's been 8 months since I've seen the patient, there can be a lot of information there that may return an error or a timeout instead of returning what we want it to return. So, That being said, Um, this is actually a couple of examples, so. Here we go. It should be pretty readable. So this is just what happened since last time I've seen the patient. The date is old. This is a, this is a very old example of kind of how things started. You'll notice that we have our note here, visit, signed visit, summarized notes here. I can focus on, you know, kidney disease, hypertension, cardiovascular disease, whatever. Um, and it will cue that in and then come up with the summary. Does the references just like all the other tools that we've talked about. You'll notice that there is a lead feedback here with thumbs up, thumbs down. So, uh, just in a slightly different spot. This is how Epic is remodeling it and offering a Problem-based approach. So what I like about this is, is you can actually see, pardon me for moving over here towards the Computer, but you can actually see here exactly what happened with the AKI from that last visit just by Expanding and hovering. You can also see what happened with the leg edema, and nothing had happened with the COPD since last time. I had to cut it off to make the screenshot, but this is gonna be available I believe in February 2026 release. It just came out last month and it is something you can turn on and turn off based on how people feel about it. All right. So what kind of impact has AI had? I'm gonna start with clinical impact. So some assistance for clinicians with tedious items, chart review, documentation, drafts, that's kind of what we covered. There is a give to get. So, what I mean by that is the user has to shift their mindset. They have to be willing to use the tool. They have to be willing to be an editor, which they should be already. I mean, has everybody seen something copied and pasted from one note to the next? Is there a copy and paste policy in our health system? Does it say don't do that ever? Is it enforced? OK, sorry guys, I had to go there, uh, but I mean that's what this is. And so the thing of it is though, this is not dragon taking my speech and turning it into text. This is not somebody else who's seen the patient who's a professional and clinician and trained and I'm copying that and carrying it forward. Sorry, carry forward, not copy paste. Make sure I'm correct about things. Um, and, but you have to review things. You have to do it. So, the other thing is we have to avoid increasing burden, OK? Just because we can do something doesn't mean it's the right thing to try to solve bottlenecks, to try to improve things, to try to do what's best for the patient or best for us. Uh, we don't want to create more noise, you know, signal to noise ratio is very much there. And uh we already have too many distractions. Uh you know, I mentioned it's taken 20 years to try to revamp what we digitized but really didn't optimize after all this time in electronic records. And um we can't create more problems. I've talked about the editor in chief already. Uh, I do want to also mention that AI can surface some things that we could easily miss. We talked about ambiguity. Notes can be out of date or inaccurate, especially with draft hospital, of course, or things that rely on notes. Very important that you understand what the input is, at least be aware of it, generally speaking. Many different areas are being explored by Epic and other vendors. OK. One little quick, I'm gonna flip around on patient impact, and that is The stuff's in patient portal too. You can now ask. As the patient can ask, when was my last MRI? When do I need to follow up with my cardiologist? I don't remember. Was it 6 months? Was it a year? Uh, they can ask questions and get answers by pulling things from inside the chart using the portal. Self-triage, um, using AI chat. Recently had to file an auto claim. Unfortunately, one of the kids hit a deer, but they're fine with the truck. And did it all through an app. And it all was chatbot. It was all just. Easy. Did it on Monday night at 9 o'clock, you know, it was done. Then they actually have to make a phone call. It was nice. Anyway, um, we're talking about doing self-triage. You can actually use existing triage protocols that you already have and have the chatbot help triage those patients. They need an appointment today, tomorrow, what do they do? Um, I did not include it, but there's a study in Annals of Internal Medicine, uh, within the last year, I believe it was 2025, that looked at urgent care. And uh AI versus uh physician triage for urgent appointments. And The AI was, was good. Matter of fact, it was a little bit better in some areas than the physicians. So However, when things change while you're doing the chat, The clinicians do better. You can also do pre-visit tasks, right? Does anybody really like receiving a, a bundle of paper in the mail you used to get before you see a specialist? You know, fill this out before you come, and then you get there and you gotta fill out more forms and all that. You can obviously do all that electronically, but not only that, you can get estimates for how much you might have to pay. You can pay your copay, you can do all those other things. Some of these we've experienced as patients ourselves, but some of them are coming. So I just wanted to make sure to point out those things too. All right. Where are we going and what kind of challenges do we have? Well, there's no shortage of areas. Another laundry list. I'll try not to be too terrible in terms of reading this to you. Ambient orders queuing, I'm gonna show you, but there's some sticking points. Uh, problem-oriented charting within diagnosis and interacting with assessment and plan. That's the ambient. That's been slower than usual. Uh, CDI inline nudges, uh, actually going to show, and, uh, we're gonna turn on some of those next week on the Saint Luke's side. Uh, result drafts within the in basket. So not message drafts, but actually drafting what the labs mean back to the patient. Creating patient instruction from notes, also translating those to different languages is a huge area. Um, so not only can you take it from the English note, get it to patient instructions written at whatever grade level you might need, but then you could potentially translate that to whatever language they may need. And I think initially Epic's gonna have 12 or 14 languages. Enough to, you know, it's a pretty good start. Uh, appeals on the back end for medical necessity. So, um, I don't think I include any examples, but we are starting that, uh, as well, uh, upcoming Wednesday. A lot of times, you know, that surgery wasn't medically necessary. Sorry, ma'am, you weren't bleeding enough to have that hysterectomy. OK. Um, and then backend things. I did not cover a lot of rev cycle here, but there are many, many rev cycle cases. Nursing is huge. Being investigated Simply put, obviously many more nurses than physicians and APPs, right? So, trying to offload the nurses, trying to think through how we can actually use Ambient to go through a flow sheet, to go through an admission database, to go through all these things instead of having to click, click, click, be tied to the computer. Context with problem list would mean also kind of like I had shown before with that outpatient insight, but actually just on the problem list that this is what's happened recently with this patient's creatinine and their CKD. This is what's happened with their heart failure. So if they go in the emergency department, well then, That's what's recently going on with that problem. Now, I don't even have to look and think, well, what's their baseline? What's their EF? what's their this, what's their that? Think if it even included whether or not they filled their medications. Different story. Of course, you got to take them too. Guilty as charged. All right. Uh, ambient voice, uh, for specialists and procedures. I talked about this as well. Um, lots of different use cases are coming up for orthopedics, for cardiology, for structural cardiology, for, you know, different specialties who need things, pediatrics, obviously. Uh, other things there, um, many different use cases. However, they're not all the same. You can't just take hospital medicine and apply it to pediatric, you know, folks that are pediatric that are in the hospital. You can't do the inverse either. Uh, so we're just trying to, you know, again, expand that and Epic does a really good job at doing that with different roles. We've seen this before, right? The brain for the nurse, rover for the nurse, but it's haiku on my phone. You know, very similar, but. You know, 2 different platforms depending on your role. Uh, medical necessity gathering for the uh. AI, excuse me, for prior auths. So a huge thing. They failed conservative therapy. They've had joint injections. They've tried physical therapy. Now they need a knee replacement. And instead of hunting all that down, the AI picks it up and then puts it in to get a prior auth from the insurance and there's less work to get that prior auth done. And we already talked about this at the bottom, but continuing to focus on those predictive models that have been around and localizing them and updating them and improving them. And one of the other things I didn't put on here is dashboards. So there's dashboards, uh, slicer dicer I did have on the previous slide, but dashboards and being able to create dashboards by simple language. Being able to create slicer dicer models by simple language and just, you know, essentially putting in what you want. And not having to worry about what data model you're using and all the pain and, you know, those kinds of things. Does anybody have pain with slicer dicer in here besides me? OK. All right. There's pain. There's even pain in the upper row. All right, upper deck. All right. It's not quite Big Mac land, that would be over here. Uh, but anyway, all right. That being said, This is ambient voice. And queuing orders. So you can see the patient's actually here with ankle pain, have an ankle sprain, and a contusion. And we're gonna get him a brace, an ACE wrap, crutches, and meloxicam. And those are queued up based on the conversation with the patient. Based on what we're telling the patient. This is very powerful, especially obviously for uh physicians and APPs. However, it's been a little tricky because, you know, is this in a preference list? What dose do they want? There's multiple different doses of meloxicam. Is it a 7.5 or a 15? Might depend on their age, might depend on their cracking. Who knows? You know, how, how do we trying to queue these things up correctly? How to pick the right DME order, though, you know, so this is actually Epic's native version. So, for those of you who may not be aware, Epic is creating ambient within Epic using Microsoft's backbone, using Nuance's backbone that is separate from DAX, that is separate from a bridge. So And they're doing some of the same things that uh Dax and Abridge are doing, and they're trying to bake it all in Epic vis a vis the operating model system and the Epic tried and true. If we give you 2/3 of the functionality, but it's in Epic and it's integrated and you don't have to interface it. People will buy it. People add it on or we'll just uh edit in whatever price you might be paying to start with until you, you know, for a pilot or something. OK. CDI inline nudges. So this isn't a diagnosis aware nodes. This gets into some technical things about provider templates, but briefly, on the left, we have an AI based nudge. And so this is Essentially, the AI reading the chart using natural language processing, noting that this nutritionist says, hey, this patient has moderate malnutrition. You see right here, hey, malnutrition. Now, what you don't see down here is moderate malnutrition. I know, not yet, but you see malnutrition. When you click that, you get a diagnosis calculator that you can easily pick moderate or severe. You pick moderate and you move on, if you agree. Then you can say, no, and this fits in right below where you're working with your other diagnosis and your other assessment and plan. So it's a way to try to nudge folks, hey, we don't have malnutrition on the chart. That can be a big weighted risk in terms of morbidity and mortality, and we do address it in the hospital. We do things for it to try to treat it, etc. This is an example of rule-based nudge that we had created custom that we had in our note template in a different format that disappeared when you signed the note. To tell folks, hey, look at whether or not this patient could have acute kidney injury or acute kidney failure, which is the same diagnosis in terms of uh lingo speak with uh DRGs and uh ICD 10 and all that kind of fun stuff. But what we've done here is created it into this nudge process, so it's nudging folks when they're working on their assessment and plan. It doesn't sit above it where they can ignore it and not do anything with it. Hopefully they'll do something with it here. This feeds into the back end where folks would actually be able to track it from a CDI perspective in terms of the CDI, uh, billing and coding folks. So not only do you not get the query, but we know whether or not the nudge was effective and actually had an impact. Which right now we're having trouble with because our nudges are mostly disappearing and we're not sure if folks are taking action on it or not and it gets, you know, more complicated that way. This is what this looks like outside of diagnosis war notes. I know this is super small and I apologize. I should have tried to just blow this part up here. So it's just a free text note. Just imagine free text over here, but we still have a nudge, we still have lab values, we still have other things here, and you can still copy it and put it in your free text note. You can still disagree and do other things with. So this actually works underneath this little section down here. So Epic's actually been pretty smart to do it either way. So for those um who have not yet taken up diagnosis where notes, of which I know there are, you know, quite a few for good reasons, I understand. I'm, I'm not holding it against you. All right. NLP we kind of already talked about, but to hit on a few things briefly, using AI to scan and read the notes. The text within the chart. And then we could look at and say, hey, your echocardiogram. shows a problem and you need to see a specialist. We can take free text on echocardiogram report, which may be uh potentially not in any kind of structured data with measurements, with EF with severity of valvular disease, and plug it in and Make it structured, and then potentially set up a referral to our cardiology folks. I do wanna thank uh Doctor Jeff Cislow for reminding me of this kind of a use case last night at dinner. So, uh, we can nudge the clinician. Hey, you, you mentioned in your note, you were gonna order that MRI, but it's not ordered yet, doc. You're gonna order it? Cause you know how long it could take to get done once you order it, especially if it's a weekend or a holiday. Uh, and we talked about trackable data from the unstructured free text. Free text has been the initial, you know, partial. Large problem with and limiting steps of, of what we can do with our electronic health records. If we can get that in structured trackable data, whole different ballgame. Looking at the appropriateness of current care, whether it's safety or quality. Looking at, you know, best practices, looking at avoidance of, you know, prescribing, uh, benzodiazepines and opiates together or maybe opiates alone, or prescribing Suboxone, uh, in terms of, uh, opiate um treatment. One last thing, and that's up at Cosmos, and then we'll get to questions. And that is Epic Cosmos is something that uh we do have access to in the West, but it's a de-identified base of 300 million patients, uh, database, excuse me. And it has median length for hospitalization. It keeps track of all the events of all these patients. And Epic is starting to run uh advanced AI uh using powerful uh GPU and other computing power on this huge data set. You could take somebody like myself potentially, once this is fully developed. And look at what's the likelihood in the next 5 years that I have a cardiovascular event, cerebrovascular event. What, why did that happen, happen because of these things. Look at rare diseases. Look at treatment outcomes in similar patient groups in terms of, well, this patient responded to this antipertensive, but isn't gonna respond to that because they're similar to these other folks that have had this outcome. You can also look into clinical trials and some other research options. OK. We'll continue to have waves of AI continue to expand. I use waves because sometimes it moves faster, sometimes it moves slower, and we've seen that. There's a rapid development of use cases as we've talked about. And then I think awareness of the shortcomings and bottlenecks, as well as the uh governance areas can really help us inform where we want to go. And that shift in user uh perception that we talked about. With that, I'll get the questions. I didn't quite leave 10 minutes. I apologize. OK. Is there a concern for AI usage and early career medical professionals versus developing critical skills without AI assistance? Thinking of the cognitive atrophy termed in the recent MIT study. So, great question. This is something that uh I thought about quite a bit. Um, so a couple of points to that. One is, is that, uh, folks are already Using AI in their personal lives. They're already looking things up, uh, medical, uh, educational folks, allied health, you know, those kinds of things. And the amount of knowledge that we need today is huge compared to when it was even when I trained. And so we think of it probably more as an assist, but it's something we certainly have to look out for. The goal here is not to lose your human critical thinking skills. The goal here is actually to not have as many tedious tasks. We still have to be critical thinkers. We still have to think through differential diagnosis. We still have to go through those routes. We do not want to replace that, that piece of thing, those, uh, parts of things, excuse me. Do you think increased disparities in healthcare already is a gap between well-funded and under-resourced patients' access to healthcare? Um, I think it could actually help access, much in the way that telemedicine helps access. And so if we can get AI with enough bandwidth and other things out into areas where folks either don't have the resources, maybe they can be geographically, uh, you know, impaired or even, you know, even obviously, it can, multiple different levels of what you're despair. is whether it's geographic or whether it's financial or whether it's, you know, what have you. But I think that with the right technology, we can actually improve access to care. Although we have to be cautious of the haves and the have-nots in terms of even health systems, um, being able to afford the AI and some of these other things. So it opens up a lot of good questions. Are there reductions in note writing time experience for equally among HM docs? Do efficient providers seem to have less benefit? I think we're probably talking about draft hospital course there. And um yes, if folks are really efficient and they keep that hospital course note updated daily, uh, they're not gonna have as much at the end, but you can use the AI to help you intervally keep it up too, and that decreases your time each day. So, yes, it's, it's still there. It's probably not quite as much if folks are really on top of their patients and keeping up interval summary notes when they go to discharge them. Um, Oh, I lost track of a question here. Environmental impact to this level of AI usage, boy, it's a great question that I don't know yet the answer to in terms of things. Um, I think for now, we're trying to go in a, in a, in a, in a reasonable pace. You know, I mentioned some limitations. Uh, there are some bigger issues there that certainly we need to think about. Um. Large gray area, safeguards. OK. Safe safeguards, OK. Under the 21st Century Cures Act, EHR and hospital administration, they are excluded from FDA regulation. With this large gray area, what safeguards have been put in place by Saint Luke's and BJC for data privacy, health information, security, etc. So we are only using the AI right now within EPIC. Uh, Epic's, uh, cloud-based infrastructure, and our, um, cloud-based infrastructure are, uh, fully secure. And so that's how we handle that part of things. Uh, hopefully that answers that question. Can a, a reasonable ROI case for IA tools be made that doesn't involve providers seeing significantly more patients? Yes, yes. And the answer is decreasing burnout and turnover. So, if a physician, nurse, nurse practitioner leaves, it's 3 to 5 times their salary to replace them. That's a lot of money. I mean, for most physicians, that's over a million dollars a year. And then it takes 3 months to replace them at least and then how long for somebody to get up to speed who might not be familiar with the area and all that. It's a huge thing. So Doctor Michelle Thomas, who's actually in the BJC, uh, you know, area here in the east, did a wonderful thing about, uh, ambulatory sprints and how it decreases turnovers, uh, turnover in terms of that and using that as the ROI, um, case studies de-identified. I don't know about that for, so when using Epic Cosmos, can you use any of the identified data to create case studies? I do not know the answer to that question. I will have to follow up on that one. When using ambient notes, oh, sorry. Do physicians feel more engaged with the patient? Yes. Both feel better. I actually have a human connection instead of staring at a screen and pardon me trying to read the questions as they're juggling up and down the screen. Uh, and you can see right now I'm having trouble interacting with you guys. Uh, and we're just answering questions. We're not even talking about, are you short of breath today? Uh, do you have any chest pain? Have you lost any weight lately? How do you feel about the Cardinals this year? Does that make you depressed? OK, let's not go there. What do you think, Steve? Could be a problem. Yeah, could be a problem. Yeah. All right. Um, how has the use of AI affected the investigation of malpractice claims? Does an attorney see the use of AI as a tool? So this is something we've talked about with our legal regulatory folks. So essentially, Uh, the author's responsible for the content of the note. Doesn't matter what's in it, doesn't matter if it started in the land and was edited, whether it was dictated, transcribed, whether it was dragon, you're responsible. So, that's the risk. You know, you have to realize that and that's why we're trying to, you hear my editor in chief, you hear my, you know, browbeating some folks there. When might the East have similar technology? Good question. There's some early testing going on with some of the insights in draft Hospital course. Um, Doctor Jeff Cslow would be able to answer that question better than me, and I know he wasn't able to, uh, attend, uh, this particular session, this talk. Um, how does this work with cognitive offloading? So great question. Uh, you know, we all have patterns and epoch that we go through for looking in a chart, and we don't want to change because we think our way is still the better way. We don't wanna make the shift AI in that. But every time you go down that pathway and click through that chart, every click, every look, every scroll, it wears on your brain and you do wear out. So there are some studies on that. I just don't have one, quick to quote. So, and we're gonna be out of time, guys. So I'm gonna have to leave it at that, I believe, right, Thomas? Yep, great, uh, great timing and thank you for a great talk.
