Healthcare Solutions: Vision and Roadmap

[01:45]
Sean: Thank you very much for that, Don. So imagine yourself a healthcare CIO. You’ve got disconnected systems. You’re looking for that trusted data to use. You’ve got all these innovative technologies like generative AI that you want to be able to use, but you’re not quite sure how. It’s almost like you’re caught up in a tornado.

[02:52] Woah, where am I?

Witches: You’re at the gates of AI City. This is the home of the Great Wizard of AI.

Witches: This is where AI has solved all of healthcare’s problems.

[03:03]
Sean: Oh wow, that’s fantastic. I’m a CIO at a hospital, and I’ve got so many clunky workflows, disconnected systems. I know AI can help. I just don’t know how.

Witches: Well, you’ve come to the right place.

Sean: Can I come in?

Witches: No. You’re not ready.

[03:23]
Sean: What are you talking about? Of course, I’m ready. There's “Ready” signs all over this place. That’s got to count for something.

Witches: Well, first you’ve got to learn about the foundations of a good data and AI strategy. That’s what’s most important.

Witches: If you learn all that, and you beg us a little bit, then maybe you can come in.

Sean: Okay, I got it. I’m game. Where do I begin?

Witches: You’re going to start your learning in the forest, where you’ll hear all about the importance of trusted health data.

Sean: Okay. Trusted health data. Got it. I’m off.

[03:56]
Witches: Don’t you think it was a little mean to send our CIO on that journey?

Witches: No, we’re witches. We’re supposed to torture people, right?

Witches: Okay. So, if we’re witches, then should we cackle?

Witches: What an idea. You start.

Witches: [Cackle]

Witches: What was that? Okay, let me try. [Cackle]

Witches: We’ve got to practice. Let’s go practice.

Witches: Embarrassing.

Trusted Health Data and AI

[04:32]
Sean: Hi there.

Lion: Hi. I’m the lion of trusted health data.

Sean: Well, the witches sent me your way to learn about trusted health data, but I want to learn about AI. What does that have to do with AI anyway?

Lion: Well, a lot, actually. How can you trust your AI if you can’t trust the data underlying it?

Sean: What do you mean?

Lion: Let’s go back to basics. In healthcare, it’s important to know which patient or member is which, right?

Sean: Of course. Fundamental.

[04:56]
Lion: So let’s say I have a patient, Margaret Hamilton, coming in for an appointment. I’m going to search for her, and we’ll see if you can tell me which record is hers. Let's search Hamilton.

Uh oh, I’ve got a couple of different records here. Looks like they’re all in Wichita. They could be different people. Let’s try clicking into them. Maybe the clinical data is the same.

Okay, you’ve got some diagnoses. She's go an allergy, a couple of meds. Great, let’s go back to that list. Go to that second one over there. Okay, so this one has lab results, but none of the diagnoses, meds, or allergies that we had in the other record. So, it’s entirely possible these belong to totally different individuals.

All right, Sean, let’s go back to this list here. You tell me which record is our Margaret Hamilton.

Sean: Oh, I don’t know. I’m scared I’ll guess wrong.

Lion: That’s my point. If you trusted your data, you wouldn’t have to guess. You’d just know, and so would your downstream AI applications. I can’t tell you how many people have come to me and said that they’re bringing their data together when they’re not actually integrating it in any meaningful way. They’re just storing records side by side. If you’re going to have data from multiple sources, you need an enterprise master person index, or EMPI solution.

If we’re going to understand what’s happening here, let’s take a look at our InterSystems EMPI.

[06:44] So this is the worklist. So ech of these entries represents a pair of records that may or may not belong to the same individual and need to be matched by hand. And look Sean, here are all those Margaret Hamiltons we saw earlier. No wonder we had so many different search results.

Sean: Wow, that’s a lot of record pairs to match by hand.

Lion: Don’t worry. I’m not going to make you match them all by hand. With InterSystems EMPI, you can use our combination of probabilistic algorithms, machine learning, and for records that don’t have enough data in your system to actually make that determination, you can leverage a curated external reference database of identities, addresses, and things like that to make that decision. That’s called referential matching.

[07:22] Let’s apply some of those tools to our system here and see what happens. Back to that worklist. Let’s refresh it. All right, Sean. How’s that look?

Sean: Wow, that’s much better.

Lion: See. Let’s try searching for Margaret Hamilton again in our clinical viewer.

There she is. Let’s click in. Look at all that data there.

[08:00] So in case you missed it, let me explain what just happened. We had five different records for Margaret Hamilton – Marge, Margarita Hamilton. And using all of those tools in InterSystems EMPI, we were able to decide that those actually all belonged to the same individual, and we then brought all that data together in one unified record of care. That was done automatically. So now you don’t have to worry about long, scary worklists, and you can trust that all your data for Margaret Hamilton is right here when you need it.

[8:27] Sean, trusted data is the foundation. You can never use AI in healthcare if you can’t trust that your health data is complete and accurate. Imagine trying to tell AI to create a care plan for Margaret here, and you have a record that doesn’t have her allergies, or her current list of meds, or any lab results. Or even take a step back. Imagine just trying to predict if she’s going to show up to this appointment, and you don't even have a list of all of the other appointments she's scheduled, and whether she's shown up to those.

I actually can’t think of a single application of AI that doesn’t involve good, clean, trusted data. An InterSystems EMPI gives you that full picture that you need, so you can have confidence in the data you’re using and have the courage to take action.

Now, do you understand why the witches sent you here to me?

Sean: I totally get it now. Trusted health data is super foundational. But now that I have trusted health data, what can I do with it?

Lion: All kinds of things. I hear that scarecrow in the cornfield down there is doing some cool stuff with actionable analytics. Why don’t you go over there and see for yourself?

Sean: Okay, great. Thank you.

Lion: Thank you.

Actionable Analytics

[9:45]
Sean: Okay. Trusted health data, super foundational. Let’s see what we’ve got for analytics.

[09:50]
Sean: Hey there, Mr. Scarecrow. I just learned about trusted health data, and the lion said to come over to you to learn about actionable analytics.

Scarecrow: Welcome. You’re in the right cornfield. I have been working on something to utilize all this data in a meaningful way. We know we can trust our data, but we need to take the next step and make it actionable. I’d love to pick your brain on how it’s coming along.

[10:16]
Sean: Great.

Scarecrow: So, recently, I learned some alarming statistics around hypertension, and that reminded me to get my blood pressure checked. So I went to my doctor’s office, and while I was there I picked up some literature on managing hypertension.

Turns out, one in two adults in the US has hypertension, but only one in four of those is managing the condition effectively. At the same time, hypertension is the number one modifiable risk factor for things like heart disease, kidney failure, and stroke. It’s costly. It’s deadly. It’s very common. But more importantly, it’s fixable.

And this isn’t just a clinical issue – it’s a population health opportunity worth solving. Effectively managing hypertension can yield three times the ROI in avoided hospitalizations and complications.

So let's take a look at what I’ve come up with. So this dashboard pulls data from Health Insight, performing calculations and aggregations to derive meaningful insights. We now have information on medications, lab tests, visits, etc., in the form of an an easy-to-use, intuitive dashboard.

What do you think about this, Sean?

[11:24]
Sean: Well, I think it looks great, but what does it tell us?

Scarecrow: We can identify patterns, highlight gaps in care, and support targeted interventions that improve outcomes not just for individuals, but across entire communities. Let’s dive in, and we'll start by looking at the distribution of patients by hypertension status. So you can see that this data is across multiple facilities. This is fairly in line with what you'd expect. Nothing out of the ordinary. But let’s choose one particular facility that stood out, and not in a good way.

And now you can see that this facility is not doing so great. I wonder what could be the reason behind such a high proportion of Stage 2 hypertension patients here. I think we should investigate further and look at the population characteristics for these patients.

[12:10] So here we have the demographic data for all patients associated with Oscare. Let's take a look at their age distribution. You’ll see that they serve an older population, and that might explain why hypertension is more prevalent.

This area also has a higher-than-average poverty rate, which can impact lifestyle decisions around diet, exercise, and stress. All these factors can impact hypertension. If you take a look at the map over there, you can see that this is an urban area. They tend to have better healthcare access, but despite all of this, patient outcomes are not improving.

Now, I know I’ve shared a lot of information so far. Let’s try to make this actionable. Here’s a list of all patients associated with Oscare. I feel confident this is the right group to target through patient outreach programs. Wouldn’t you agree?

[13:24]
Sean: Yeah, makes sense. But all those tornadoes can’t be good for their hypertension.

Scarecrow: Good point. Having all this data and being able to leverage it is truly putting your data to use. It’s a way to give your health system a brain.

Sean: Well, this is great. I understand now how you can take action on data through dashboards. This is super helpful. But what else can I do with it?

Scarecrow: Well, I heard there was some cool AI stuff going on over at ISC University. Ask for the professor.

Sean: Perfect. Well, this is great. Thank you.

[14:12]
Witches: I wonder how Sean’s doing on his journey.

Witches: I’m sure he’s having a grand time. Who doesn’t love meandering through the forest learning about data and AI?

Witches: Should we check the crystal ball to see how he’s doing?

Witches: I don’t have a crystal ball. Did you bring the crystal ball?

Witches: I don’t have a crystal ball. Oh, but I did get this snow globe at the Orlando airport.

Witches: We’re poor man’s witches. All right, let’s give it a shake.

Witches: I think I see Sean in there. Or, is that a dolphin?

Witches: No, not Sean. I see his red shoes.

Witches: Okay. So it looks like he’s learned about trusted health data and actionable analytics. So I think that checks everything off for me. Should we let him into AI City now?

Witches: Still no. First of all, aren’t you supposed to be the Wicked Witch? What happened to that?

Witches: I forgot.

Witches: And my popcorn’s not empty. Let’s have him learn more. Let’s teach him how to put AI into an application and how to use AI to connect a community.

Witches: Okay. I think that's an excellent plan.

Witches: Hey, how are you getting home?

Witches: I just got a new broom, but you know how it is. One minute you’re soaring through the sky and the next minute, bam, there’s a barn right there.

Witches: A broom? That’s so 80s of you. I want a giant pink bubble. I think that’d be cool.

Witches: That’d be good... for you.

End To End Applications

[15:50]
Sean: Hi there. Are you the professor?

Professor: I am. I am the owl professor.

Sean: Fantastic. Well, I was told to come to you, and you could teach me about putting generative AI into practice.

Professor: No question. I'd love to show you. Let's take a look. So here we are in Margaret’s chart, as we’re all familiar with, and we have a great opportunity to review her chart and see all her data. But what if we could do more? More than just looking at it from a category-by-category perspective, and really jump into what AI has to offer.

So I'm going to move here from Margaret's standard viewing chart to the the AI Assistant. Now when I get into the AI Assistant, you can see right away I have some options of promtps that give me data that I can ask for. I can also ask questions and get targeted answers if that's what I needed. But let’s just start at the beginning. Let's start with medications. Now with medications, I get returned a list. With that list, it's about like for like of what I might see int he viewer today.

[16:55] I can see what medications she’s on, but I’m not getting a lot of value yet. It’s pretty simple. Let’s go a step further. When I go a step further, let’s just take a look at labs. Now in the viewer, we know that labs give us most recent first and just working on their way down.

But I want the AI assistant to do more. I ask for the AI assistant to group it by categories. And in the categories, I can see hematology, chemistry, the way clinically I think about them. That gets me further on this path. But it doesn’t go all the way. Let’s keep going.

So let’s pick another one. Let’s pick conditions. Now, conditions I happen to know clinically are a combination of problems and diagnosis. So now we’re mixing data together. I’m also asking the AI assistant to group them by categories, by body systems, respiratory, cardiovascular, and I could keep scrolling down. In addition, I’ve asked it to actually infuse more information by giving me SNOMED and ICD10 codes, and you can see them listed.

[18:08] Now, that’s great, but what if this isn’t giving me everything I want to know. We know Margaret has high blood pressure, right? She’s shown up, and we want to be able to act on that.

Well, let me see what her most recent blood pressure is. Let’s see if it can extract that. And you can see actually her most recent blood pressure is 150 over 90. So she is still high risk.

You know, I think I probably need to think about Margaret’s profile in a way that really assesses her full cardiac risk. Let me see if it can put that together for me.

One of the great things about a demo that’s live is you get to see all my typos, and you know the AI assistant can manage that.

And so here you can see they’ve started pulling together in the AI assistant diagnosis, medications, lab results, the things that are really relevant to cardiac profiles themselves.

[19:17] Now when I think about who might use this, you know, I don’t think this is just for me. Do you think others could, too?

Sean: Oh, absolutely. My docs would love this.

Professor: So when I think about that, let’s go ahead and create a new prompt to get that available to all of my users. So to do that, let me go in as an administrator. So now I’m going to move away from the end user application that we think about and move on to the system settings themselves.

I have a lot of controls here. You can see deployments, prompts, roles, data. There’s a lot of configuration opportunity here for me to define what my end users actually have access to.

Let’s go ahead and start with roles, because it’s so incredibly important when we think about the capabilities of AI that we know we can do it with the controls needed to be applicable to the right users at the right time.

[20:08] So here I have the opportunity to actually configure what roles see what data and what messaging they receive, and in fact what prompts they receive themselves, what they should have access to. And being built on top of HealthShare itself already has that consent rule and engine that allows for the control of data.

Now that’s the role configuration piece. If we dive further into data, what you can see is that I can also control, down to the data element level, what data is available for my users. This actually defines what I can feed into an LLM to return back answers.

Now, this is one piece of the larger conversation around how data is processed as we move from our core products into an LLM and back, and all of the controls and the creation of the appropriate response.

So, if I were to choose, for example, that I wanted to include prescription status, I would simply save it, and from then on that data element would be available.

[21:20] But what I’m here for is prompt configuration, to be able to create that cardiac risk assessment in a way my users can access. I want to know also how often they access it. So I need to make sure that I have a change log for every action taken into the system, to be able to review and visualize.

When I do this, I can see that I have all of the entries that I’ve done, all of the data and responses that were returned. In fact, I could actually drill down further and have a record of every single piece. If I want to do analysis on this, I simply would generate an export of this file and be able to do additional metrics.

[22:04] Now, let’s go ahead and create that prompt. When I go into prompts, you can see that there’s already preconfigured available options, some of which I clicked earlier in order to be able to show you what’s possible.

But let’s go ahead and create a net new one. In this particular case, what I said I wanted to do was create a cardiac risk assessment. And to do that, I’ll just simply type it in.

[22:40] And so to prevent you from having to watch me type and type and type, I’m going to copy and paste the rest.

But here, I can actually define it further, in addition to just simply naming it for what my users will see. I can also do date range lookbacks and configuration there. In fact, I also have the controls in this to define what data is fed in.

If, for speed purposes today, what I’ll show you is just default selecting them all. But if I wanted to only include subsets of this data, maybe problems, diagnoses, like I mentioned in conditions, medications, labs, and leave out others to account for better performance, token utilization, I have the ability to do so.

And of course, it’s also important to say who can access this. In this particular case, front office staff don’t need a cardiac risk assessment. I’m thinking about offering this to my PCPs, my primary care providers. And so that’s what I’ll do.

Now, how long do you think that took me, Sean.

Sean: Oh, not much time.

Professor: Exactly. With no coding, low code, right. We talk about low code. This is a no code experience. But clinical expertise. I’m able to create this prompt and make it available to my users.

Let’s see how well it did. Now, as I pull this up, you can see it immediately displays here. Let’s see how well it returns information. Because what I fed into that prompt was guidance around some of the common screening tools available for cardiac risk assessment as well as some guidance around what labs are pertinent, what medications are pertinent in order for it to give me a more robust response.

Now, this I think is incredibly helpful and gives me a broader response and answer than when I just simply free text it in, talking about some of the prompt engineering necessity to really get a good comprehensive answer.

This I think is really helpful for my primary care providers. But you know what. I bet it could be useful for other specialties. Do you think?

Sean: Oh, absolutely. My docs would love this.

[24:52]
Professor: And when they love it, they might actually want to configure it by specialty. And by specialty, right, you can create something that says this is for my pediatricians, this is for my geriatricians, this is for any number of people, and effectively create your own viewer.

I think that’s a really great opportunity. But when I looked at Margaret’s assessments, I have to say she has cardiac risk, and clinically I would think about actually doing further testing. In fact, I might consider a stress test. But stress tests, you know, Sean, I’ve got to tell you, prior auth is a challenge. And so I don’t know if you want to really go through all of that pain.

But before you consider it, what do you think of the prompt builder?

Sean: Oh, I think it’s incredible. Our clinicians would love that, prebuilt prompts and prompt builder. Are you kidding me. Unbelievable. We could use this today.

Profoessor: Excellent. I’m so glad to hear you say that. But, you know, I have to say, I need to wish you well on your journey for prior auth, because I really think that stress test is necessary. And I think Jim might have some ideas over there in the forest.

Sean: Wow. Fantastic. Well, thank you. This was great.

Professor: Thank you.

[26:02]
Sean: Hey there. Are you Jim?

Jim: Yes, I am.

Sean: Oh, great. Hey, nice tie.

Jim: Yeah, it’s Tin Man silver.

Sean: Tin Man silver. Very appropriate for the forest. So, Jim, I was told that you have a cool generative AI tool that’s going to save me a ton of time.

Jim: Indeed I do. The last few days we talked a lot about prior authorizations being a bottleneck in healthcare. Much like the Tin Man, prior authorizations have lacked heart, preventing payers and providers from truly connecting.

So with generative AI, InterSystems is helping to accelerate connectivity between payers and providers.

And the prior authorization process today starts like this. It’s form driven. Frequently it’s done on paper. A PDF is faxed or sent to a provider. This PDF is disconnected from the EHR. It requires the provider to do duplicate work when all this information is already within the EHR.

But using generative AI, we’re able to change that. We built an AI questionnaire generator where we could take a PDF or other form, scanned or HTML, ingest it, identify intent and structure, and create a FHIR compliant questionnaire that is machine readable, fully structured in a digitized format that can be exchanged between payers and providers, making their lives far easier.

[27:35] And if we actually look at the FHIR JSON that was created using generative AI, we can see how that aligns with the PDF form. Focusing in on one section, the chest pain or suspected angina checklist, we can see now that it’s been translated into codified values with descriptions.

So having the FHIR questionnaire in a format that uses structured data means that it’s able to be launched from the EHR using a SMART on FHIR app, querying all that information using FHIR, accelerating the prior authorization process. And not only that, we’re actually able to take this form within our questionnaire builder and verify it and edit it to actually make it more usable.

So for instance, we’re able to take one of the drop down items, NPI, change it to myocardial perfusion imaging, actually making it more user friendly.

But once it’s verified, we can exchange it using our electronic prior authorization solution. It then can be launched from the EHR as a SMART on FHIR app, pulling forward all of Margaret’s information from the EHR, her registration, coverage, any prior exams or preexisting conditions into the prior authorization form, preventing errors, reducing work for the providers, streamlining the submission process as well.

[29:03] So InterSystems electronic prior authorization is better connecting providers to payers. It’s making them compliant with regulation and making them ready for automation. So we’re making the prior authorization process faster, smarter, better connected, and far less painful for everyone.

Sean: Wow. Well, that is just incredible. I mean, you are connecting the payer and provider ecosystem with a headless workflow and using generative AI to actually generate a form and populate it.

I mean, my prior authorization folks are going to love this.

Jim: Yes, you could say it’s almost wizardry.

Sean: Very good. Well, thank you. That was great.

Jim: Thank you.

[29:53]
Sean: Well, I’m back. That was quite the insightful journey that you sent me on. I learned a lot.

Witches: What did you learn?

Sean: Oh my goodness, I learned so much. So it starts with making sure that you’ve got trusted data. And that starts with EMPI and referential matching, where you can actually resolve duplicate records and those records just fall off your worklist and the patient profile just grows.

And then there’s this thing called machine learning, where you can actually further reduce your worklist with that.

But most importantly, I learned that once you have trust in your data, you then have the courage to take action on that in downstream apps and analytics and AI. And with InterSystems EMPI, you can be confident in your data.

And then I learned about how you can take data and action it through dashboards. And you can imagine in healthcare we’ve got so much data. We’ve got multiple patients and members, multiple providers, and we can bring all that together through a dashboard pulling from health insights.

And that’s your brain, where you can actually analyze and find meaningful insights in the data.

But then I learned about generative AI. And I learned how you could take generative AI and prompt responses and give voice to your data. Unbelievable. And this is powered by simple to complex workflows with prebuilt prompts that deliver summaries of patients.

And then I learned about prompt builder. Prompt builder. Oh my goodness. You pick the data you want to ask questions of and then it returns a natural language response in the context of the patient. I mean, unbelievable.

And then lastly, and maybe not most importantly, but another one, I learned how we can connect the payer and provider ecosystems around electronic prior authorizations in an API based headless workflow.

I mean, incredible. And you can take that further using generative AI again to take a paper prior authorization, turn it into an electronic prior authorization in questionnaire format, and workflow it to complete between payers and providers until it is done. I mean, are you kidding me. Unbelievable.

I have learned so much on my journey. So I now humbly request entry into AI city to meet this wizard.

Witches: Fine.

[32:48]
Wizard: I am the almighty wizard of AI. You have entered the great green palace of AI city. Who goes there.

Sean: Whoa. Is that the wizard.

Witches: I think that’s just Dawn with a new fancy microphone. This wannabe wizard needs to be cut short.

Wizard: This is the great palace. Only folks that understand how the foundations of data and AI work together.

Witches: Hey Merlin, we see you.

[33:26]
Wizard: Hey Sean.

Sean: So you’re the wizard. I thought you might be taller.

Wizard: I am indeed. I’m five foot eight and three quarters. I think that’s fine.

Sean: Oh my goodness.

Wizard: Well, I am indeed the wizard. Yes.

Sean: Well, I’ve heard that you are super powerful and super intelligent.

Wizard: I am super intelligent. I almost passed the InterSystems hiring test.

Sean: Impressive indeed, Don.

Witches: Don, I think you got like a four.

Wizard: I got four and a half. But I am super intelligent and powerful only because of what’s behind the curtain. I leverage AI agents, basically. They help me run AI City. Basically, schedule the flying monkeys, maintain my hot air balloon, plan celebrations, that kind of thing.

[34:17]
Sean: Well, that’s great. But I’m just a CIO at a hospital. I want to use AI agents to kind of superpower my organization and then give my users and have them be aided by intelligence.

Wizard: You can do that actually in healthcare. Let me explain what agentic AI is just briefly, and then we’ll show you the impact it can have in healthcare.

So we’ve all had a good run with ChatGPT the last few years. And Gen AI has, for all the wonderful things it’s done, basically just focused on writing. You ask it a question, it writes a response. You give it an audio file, it might write a summary of what it’s hearing, that sort of thing. It’s been writing.

And the idea behind agentic AI is we basically add two verbs to what we’re asking LLMs to do. The first is to call tools. So an LLM could directly schedule an appointment, or book a bed, or send an email, or do something specific, call an API, so to speak, and get something done. So LLMs can start to do things, not just write.

And the third verb, in a sense, is they can plan. So they can plan out a whole multi step workflow. So given this situation or given this request, here are the four things that I need to do, or here are the three things. And then depending on what happens on the third thing, here’s what happens next.

So agentic AI is really taking LLMs from writing simply to these two other things, where they can be even more impactful. And it’s going to make a difference in healthcare as well.

So to give you an example of that, I’m going to bring to the stage my friend Dimitri Fain, and he’ll show us a demo of what it looks like in healthcare.

Sean: Great. Thank you.

[36:03]
Dimitri: Thanks, Don. Hi, Sean.

Sean: Hi there.

Dimitri: So Don’s given us a great example, a great explanation of agentic AI. So here in the EHR business, we’ve been experimenting with agentic AI combined with real time voice analysis to see if we can build a truly transformational user experience in the EHR.

So what I thought we’d do today is take you through a demo of a prototype system of what our EHR of the future could look like, and then I’ll come back and explain a little bit about what we’ve seen.

Sean: Great.

[36:40]
Dimitri: So our patient, Margaret Hamilton, has arrived at the emergency department. She’s unfortunately a bit sick and in respiratory distress. So Margaret has arrived in respiratory distress, quite ill in the emergency department. Luckily, we’ve got a very good doctor with us, our own chief medical officer, Dr. Jonathan Teich.

So I’ll bring Jonathan to the stage now. Jonathan is, in addition to being our CMO, also an emergency doctor in real life. So Jonathan, I’ve loaded up the EHR. I’ve brought Margaret’s record up. You’re welcome to take a look, and just tell me what you need me to do.

[37:18]
Jonathan: Sure. Well, thanks, Dimitri. You know, this patient looks pretty sick. I think that I need to go see her right away. But the problem is that I don’t know her. She just came in. I’ve not seen her before. I can review her record, but her record’s pretty full of data. You’ve seen all the data that’s accumulated there.

So I can review, but even in the best of EHRs, it’ll take me five or six minutes to go through all the different things I need to know.

But what I do have is an AI assistant and a really great prompt. And that prompt is going to be able to suss out what her chief complaint is and filter everything out so I get just the information that’s relevant to what I need for this patient. So in ten seconds instead of five or six minutes, I’ve got what I need.

[37:56] So look at this now. Let’s get clinical here. So her heart rate is up. Her respiratory rate is up. This doesn’t sound very good. She was seen in the hospital about six months ago for the same thing.

I’ve got her medications. Even a little bit of an extra assessment. This is pretty much all I needed to know out of that whole chart. And the AI did that for me. So let’s go see her. I think it’s time to examine her.

Dimitri: Sean, could you bring our patient over.

[38:25]
Margaret: Hi, Sean.

Sean: Hi there.

Jonathan: So hi. I’m sorry you’re not feeling well. Let’s talk a little bit, get a couple questions answered, do a little exam, and then we’ll start treating you right away.

So I see you’re having an asthma attack. When did that begin?

Margaret: It started about two days ago, and at first my inhaler helped, but now it’s definitely not enough.

Jonathan: I see. Have you had any fevers with this or any coughing?

Margaret: I haven’t really taken my temperature, but I’ve been coughing a lot.

Jonathan: Does any junk come up when you cough? Anything greenish or brownish?

Margaret: Yeah, it’s kind of rusty colored, I would say.

Jonathan: And having any chest pain with all of this?

Margaret: Only when I cough.

Jonathan: Okay. So according to your record, you’re on a couple of medications. You’re on Singulair and you’re on a salbutamol inhaler. Is that right?

Margaret: Yep, that sounds right.

Jonathan: All right. And you’re still smoking. Is that correct?

Margaret: Yes. Unfortunately, I do smoke about a pack a day, and I know it’s really bad for the asthma. I’m really trying to quit.

Jonathan: It’s going to make a big difference if you can quit once and for all.

I understand you took a flight from Boston to Florida recently. Did you have any leg pains or anything else on the flight?

Margaret: It was a long flight, but no leg pain. I honestly felt okay until I got here with the crazy humidity.

Jonathan: Absolutely. All right, let me just listen to your lungs for a little bit here. Okay. So I can hear wheezing really all over your lungs. So that’s typical for a bad asthma attack. And I also hear some reduced sounds here in the left lower lobe. So I’m concerned that you might have an infection.

So let’s do a couple things. I’m going to order a chest X-ray to see if you have an infection. That’ll come back in a little while. And also, let’s get you started on treatment. So we’re going to start with a Combivent nebulizer. And then we’ll order a salbutamol nebulizer, 2.5 milligrams. We’ll do that every twenty minutes for an hour, and then we’ll see how you’re doing.

And also, because you said you were having pain when you cough, let’s give you some ibuprofen. We’ll order some ibuprofen, six hundred milligrams.

Hopefully that should be able to help with several of your different symptoms. Does that work for you?

Margaret: Yes. Thank you, doctor.

Jonathan: All right, let’s go get your X-ray now.

[40:46]
Dimitri: So let me just jump in for a moment, Dr. Teich and Sean, and explain a little bit about what we’ve just seen. So this is truly agentic AI in action, in real time, operating as they’re speaking.

If you look on the right hand side of the screen, that’s not a screen we typically show to the end user, but because this is a technical audience, we thought we’d share the agentic dashboard to give you a sense of what’s going on behind the scenes.

You can see a number of highly specialized agents. So for example, there’s the line annotation agent, which is looking for clinical concepts in the text stream as it crosses through, and you can see they’ve highlighted it and color coded it.

There’s an actions agent that takes those clinical concepts, matches them in the database, and tees them up for Jonathan to then execute them with appropriate human control in the record.

And so this is agentic AI.

We then thought, what if we could take this one step further. What if we could humanize the agents in a way that makes the interaction even better.

So, Jonathan, this is a pretty complex case. Do you think you could use a little bit of help with it?

Jonathan: Yeah, she’s not getting better, so I think we need to get some help. Maybe one of our AIs can help with that as well.

Dimitri: Well, let me introduce you to our AI agent avatars. You can go ahead and ask them anything.

[42:07]
Jonathan: Hey, I see a medical expert. Hey, I’ve been treating this patient with nebulizers, but she’s not getting any better. Is there something else we can do?

AI Avatar: Let me check the patient’s file to see what else is needed.

You’re likely to need more than just salbutamol. Here are the ACEP asthma guidelines recommending magnesium sulfate and methylprednisolone. Do you approve?

Jonathan: Yeah, absolutely. Let’s go give her both of those. That should help.

[42:44]
AI Avatar: On screen, you can see the new results that came a minute ago showing pneumonia in the left lower lobe.

Given these findings, along with fever and low blood pressure, sepsis is likely. IDSA guidelines recommend starting vancomycin and cefepime. Do you approve?

Jonathan: Yes, let’s give her both of those. You know, I hadn’t been considering sepsis, but that’s a great idea. Let’s add both those antibiotics.

[43:18]
AI Avatar: Lastly, I see there’s a drug alert. Let me transfer you now to the pharmacy agent.

Pharmacy Agent: I noticed ibuprofen was ordered for her pain, but she has an allergy to it. I’ve recommended substituting it with paracetamol, five hundred milligrams instead. Do you approve?

Jonathan: Oh, yes, very much. Let’s do that. Sorry, I didn’t have a chance to check the allergies, but that’s much better.

Pharmacy Agent: All medication orders are complete. Do you need anything else?

[43:53]
Jonathan: Well, yeah. This patient needs to be admitted to medicine. Can you guys handle that stuff as well?

Admissions Agent: I see you’d like to admit her. Let me check availability and prepare the admission plan.

I’ve arranged everything, booked a bed in general medicine, created the admission request, and prepared notifications for the ward. Do you confirm?

Jonathan: Yes, absolutely. That’s amazing. This is like 10 minutes of work you’ve just taken from me. That’s wonderful.

[44:34]
Documentation Agent: Since you’re admitting her, I’ll create the documents you usually prefer for these cases. Please give me a moment.

The documentation is ready. Would you like to sign the visit now?

Jonathan: Sure. Those notes look good. I’ll have a chance to review them also. Thanks for doing all of that for me as well.

Documentation Agent: I’ve completed and signed off the visit documentation.

[45:02]
Dimitri: So, let me just jump in one more time. What you’re seeing here are AI generated video avatars that are connected to our agents. So the video avatars themselves are not the agents. They’re just humanoid representations, but behind the scenes they’re doing real time speech to text and text to speech connected to the AI agents behind the scenes.

You’ve seen specialized agents for both clinical administrative purposes. And I think a great example, Sean, is the admissions agent, who had to do sort of a complex multi step workflow. Find a bed that’s free, call an API to put the patient in the bed, actually notify the ward, send a message to the ward, and get everything ready.

That would have saved you an enormous amount of time.

Jonathan: It’s wild. It sort of knows the entire sequence of an emergency visit. It knows that we need to do the admission here, the tasks that are involved. Just having these different specialized agents are great.

[45:57]
Dimitri: Yeah. And look, you know, when we implement this for real, I’m not sure whether we’re going to use the video avatars. You can see you can switch them off at the top if you like. You might choose text only. We’d like to give the users some choice there.

But you and I both found them quite creepy when we started working with them, right? But I have to say, having worked with them for some time getting ready for today, they grow on you, don’t they.

Jonathan: Well, they do. Now, I don’t know if I’d stick with the video, but quite honestly, having someone speak to you is faster and more efficient, even than a great chat window.

It’s like you’d have in a real emergency room where your consultant might be over your shoulder saying, you know, don’t do this, don’t do that, do that instead. So, having that kind of normal conversation was actually pretty good.

[46:42]
Dimitri: Yeah, I think that, you know, multiple modes of communication tends to wake you up and drive your thought process. So, it’s just one more option and we think it’s quite fascinating.

So, Sean, what do you think?

Sean: Oh, I think it’s remarkable. And I tell you, I think a lot of our staff would really appreciate the video avatar. To your point, doctor, it would really humanize things. And some people just may be turned off by it and choose to use the regular agentic AI. Either way, it’s saving me time.

Dimitri: Yeah. Do you think your users in the hospital would like the time savings. Jonathan mentioned this saved him 10, 15 minutes.

Jonathan: At least.

Sean: Yes. No one wants time back. Yeah. Everyone wants time back. Come on.

[47:19]
Dimitri: Well, certainly here in AI city, we all want time back. Thank you very much. We have one great advance. One last avatar to show you before I leave. And it’s one that I really think actually may be the first one that’s used.

So, we’ve also generated a patient support avatar. So, Margaret here has been admitted to the hospital with pneumonia. She’s, you know, not feeling comfortable.

So Margaret, we’ve programmed this agent to understand the demographics in our system. Also has full access to your clinical record, everything that’s going on with you. You can speak to her naturally, however you like, and just ask her any questions you might have.

Margaret: Is she programmed to look like me. Is that part of the service.

Dimitri: I don’t really see that, but she does look a lot like my colleague Judy. It’s funny.

Dimitri: Anyway, go ahead and ask her anything you like.

[48:29]
Care Companion: You’re receiving nebulizer treatments every 20 minutes to help your breathing. This treatment uses a machine to turn your asthma medication into a mist, making it easier to inhale deeply into your lungs. This can quickly open your airways and relieve your symptoms.

Would you like me to explain your treatments further or do you have another question.

Margaret: What nebulizer am I taking?

[49:06]
Care Companion: You are currently receiving salbutamol through the nebulizer. Salbutamol is a medication that helps open your airways, making it easier to breathe during an asthma attack. It works by relaxing the muscles around your airways so air can flow more freely.

Would you like to know more about this treatment or do you have another question?

Margaret: Yes, actually. Does it have any side effects?

[49:38]
Care Companion: It’s normal to feel worried. I’m here to support you.

Salbutamol can sometimes cause mild side effects. The most common ones include shakiness or tremors in your hands, feeling your heartbeat faster or stronger, palpitations, headache, mild nervousness, or restlessness.

These effects usually go away soon after the treatment. If you notice anything unusual or feel worse, let your care team know right away.

Would you like to know more about your treatments or do you have another question?

Margaret: No, thank you. So, I’m all set.

[50:08]
Dimitri: Well, Sean, I hope you enjoyed our tour of the EHR of the future. Thank you very much.

Sean: Thank you. That was wonderful.

[50:29]
Wizard: All right. So, what did you think of this journey today.

Sean: Well, I tell you that was wonderful. I learned so much about how to build a comprehensive data and AI strategy and we could totally use it to transform our hospital at home.

Wizard: That’s great. Well, that was the idea. You know what you learned, what we tried to teach you is align with our roadmap. It all starts with a connected ecosystem on the bottom. You need to take your community, your data silos, and pull them all together, building on that trusted health data.

So, you can’t just collocate the data. You’ve got to bring it together, normalize it, match patients, that sort of thing. And we help you do that.

Then you put it to use. Analytics is a good use case, like we saw with hypertension. That’s a really good example where population health activities can make a positive difference in a community.

And then we saw the AI assistant, basically digging through a big chart, building your own viewing experience, really tailoring, in a low code or no code manner in this case, your own way to interact with your data.

And then finally, we gave you a little picture of the future. You know, what AI agents could be all about. Taking on bigger tasks, not just “write me this,” but “get this patient admitted,” or “do all the follow up for this appointment,” or “do all the documentation.”

That’s the idea behind AI agents – taking Gen AI up to that next level. So, hope that was all educational for you.

[51:57]
Sean: It was. And I tell you, I feel like I am ready to put this to use.

Wizard: You feel like you’re ready now, huh.

Sean: I do.

Wizard: Well, I have a little secret for you. Since your hospital had already been using InterSystems technologies, you were already ready.

We just wanted you to learn that for yourself. Plus, the good witch here thought it’d be a good idea to send you on an hour long journey.

Sean: Oh, geez. Thanks.

Witches: You’re welcome.

[52:24]
Sean: But I think I’m ready to get back to Kansas now and put this to use.

Wizard: Ah, Kansas. Very nice place. I’ve seen it in the movies.

Sean: Yes, indeed. Well, that’s where my hospital is, where I want to take this back and help them transform. So, how do we get out of here anyway.

Witches: Oh, we can take my broom. I’ve gotten so much better at flying it. I mean, I haven’t had an accident in hours.

Sean: I think I’ll pass on that one.

[52:49]
Witches: We can use my brand new pink bubble. I parked it over there.

Sean: Oh, fantastic. Let’s do that.

Wizard: Very good. Bye bye.

[53:05]
Wizard: The end. So, I hope you enjoyed that silly little journey through our vision and roadmap for healthcare solutions.