Healthcare Interoperability & Agentic AI Demonstrations

Three tools, each with a clear job. First, map it. We're building a data atlas, a single source of truth for every mapping definition in the system. Out of the box or custom, see what's mapped and what's not and how the data flows between any two standards. When you need to map add a mapping, you can add it right there to the data atlas.

Second, build it. The codegen engine reads those mapping definitions. You define what needs to happen right there, and the engine produces the right transformation code for the job.

Third, use it. The universal transformation API gives you a single simple interface to every transformation in the portfolio. You ask once and you get back what you're asking for. No more digging through code libraries. What used to take a specialist weeks or months now takes days or hours.

The API is the first tool in the toolkit that we're shipping. Let me show you what it looks like. This is a CCDA clinical document. More than 8.6 million of these move across US healthcare every day. As you can see, they're very dense and very complicated. It contains a lot of information, problems, allergies, medications, immunizations.

Getting one of them into fire today is manual work. You have to know the right transformations. You have to wire them together yourself. With the universal transformation API, it's just one request. I tell it what I want and I tell it what I'm sending. Clean, structured fire. Ready for an analytics pipeline, ready for a clinical application, whatever comes next.

The ability to preview out of the box transformations via the API is shipping in the next few weeks, and that's what you just saw. We're targeting early access for the data atlas and the codegen engine by the end of this year and we're building AI into the framework as well. AI recommended mappings, AI generated semantic definitions and smarter code generation. We're building it with your input and we're looking forward to making it better over time.

If you want to go deeper, you can catch the SDC session from yesterday. We started this conversation talking about trust in our data. Without trust in our transformations, you can't have trust in your data. The unified transformation framework will give you the ability to make that happen. Easier to find, easier to build, easier to use. Map it, build it, use it. That's the framework.

And now I'd like to hand it over to Usha and Kristen who'll show you what happens when you start to make use of that clean trusted data moving between providers and payers. Thank you.

Thank you, Adam. Now let's see how this trusted data can flow across stakeholders in the healthcare ecosystem. Stakeholders whose applications may speak very different languages and the benefits that this data flow brings.

Consider a scenario where a patient has just completed an appointment with their primary care physician. The physician wrote up a prescription and referred them to a cardiologist. Once that patient leaves the facility, the provider may never really know if the patient picked up the prescription or if they made an appointment with the cardiologist. Perhaps they switched providers or moved out of state.

Exactly. The provider really has limited visibility once a patient leaves their office. But there is one member of the healthcare ecosystem that does know that information, and that's your health plan.

So let's say that this patient comes out of their PCP appointment. They make that appointment with their cardiologist and followup and then they go on vacation a couple states away and end up in the emergency department. What follows them throughout each of these clinical encounters is their health plan. So it makes sense that a health plan is uniquely positioned to be a source of clinical data for their members.

But where does the clinical data come from? How does the health plan get it? One way to get that information is directly from providers in their network. For instance, a payer may have Epic payer platform, a solution from Epic that makes rich clinical data available to a payer from providers who have the Epic EHR. Using InterSystems payer connector. This data can be brought into other applications within a health plans ecosystem such as quality and performance. All of the clinical and administrative applications that perform the various functions of the health plan.

One of these integrations is the admit discharge transfer notifications. Payer connector can receive these ADT notifications from Epic payer platform and route them to an application such as care management.

But why does a health plan care about ADT notifications? Can't they just get that information from a claim? Yes, but that claim may reach the health plan months after the encounter and by then the window for timely follow-up care has closed. And it's not just ADT notifications. The payer can also receive consolidated CDA documents which give visibility into the recent encounter of the member with the physician and the ability to track actions.

So for instance the physician who did not know whether the cardiology appointment was being made the health plan will be able to track that and perhaps the health plan was also planning to call the member about the annual flu vaccination. Now there's the opportunity to combine those two reminders into a single outreach.

That makes sense. It would definitely be a lot easier to just have one phone call or one touch point to take care of everything for a single member. And imagine how that could scale for a health plan that has hundreds of thousands or even millions of members.

Well, to that point, one of our customers has created a clinical data repository, a single source of truth for all of their member data using HealthShare unified care record. Not only does this create a longitudinal health record for every member, it also enables analytics at the member population level. So a health plan might look across its diabetic member population and notice that a majority of the members have not had an HBA1C test in the past six months. Now the health plan could reach out to those members and request them to have that test completed.

Without the clinical data underlying these dashboards, the health plan cannot be proactive about improving member wellness. Member care is a huge part of this. But there's also an opportunity to use this data to improve the efficiency of internal processes.

Absolutely. So why don't we take the example of risk adjustment? This is a process by which health plans have to reattest to the diagnoses of their members on an annual basis and provide clinical documentation to support that. And this is true even for lifelong chronic conditions.

If you had an amputation 10 years ago, your health plan still has to reattest that you still have that amputation. So let's see what happens when we can actually introduce a clinical data feed into the risk adjustment system and potentially reduce a lot of that payer provider abrasion that is experienced when payers actually have to call up the provider and ask for this documentation every year. So here we have a health plan that has a list of members here and we can actually go in and look at all the different risk scores for these patients. We can organize them by the highest risk delta and we can look at a specific member as well.

So she has a demographic risk score based on age and gender and then she has some potential diagnoses that need to be verified and need clinical data to support them. So this one here is sepsis for example. And if we view the clinical documents that are available, we can see a sidebyside view of the raw CDA data that has to be used for auditing that has to be submitted. And then we can also see on the right the friendly formatted CDA format for the clinical reviewers to take a look at. And what's great about this is that you can pull data from all of your providers into a health plan, right? So this really improves the process efficiency.

But what about that scenario where the physician wasn't really aware about whether that cardiology follow-up ever happened?

Yeah, absolutely. So let's say that you get all of this information into your health plan. You're actually able to change it and send it to a provider. So this provider makes a request and they can get access to all of the clinical and claims data that the payer has. So if we go back to this example that we showed earlier, right? We can see that we've actually closed the loop here. So we have clinical data flowing into the health plan and then we're able to return that clinical data out of the health plan and that can go to providers, it can go to the member and can even go to a different health plan if a member chooses to switch to a new plan. So we've really closed the loop.

So this bidirectional data exchange is really a win-win-win for the provider, the health plan and the member. The provider can better coordinate care with insight into the care journey outside the walls of their office. The health plan can also better care for its members and also improve operational efficiency. And for the member, there is a better experience and improved outcomes.

Absolutely. But up until now, we've been talking about individual transactions between a single health plan and the member or the provider or even another health plan. But there's millions of these data exchanges that have to happen every day. So, let's get Patrick up here and see how this works at a national scale.

So, I'd like to talk for a few minutes about nationwide patient data exchange here in the United States. I'm talking about national networks and frameworks such as what Dr. Keane mentioned yesterday, TEFLA, but also Commonwealth, CMS aligned networks, and of course, Care Quality. Not to mention state and regional hies across the country. Americans have medical data stored across multiple providers for a myriad reasons. You could have moved, you could have changed employers or for many other reasons.

In my own personal case providers in Seattle, Miami, Chicago, and Boston all have my medical data. In fact, chronologically in that order. Go Seahawks. I'm not popular with the Patriot fans.

While the national networks and frameworks have and continue to evolve as the foundation for national patient data exchange, they are really the starting point. From here, a number of factors contribute to making data usable and effective for end users. Enter Health Gateway. InterSystems Health Gateway is our fully managed cloud service that works for you as an automated execution layer at scale to connect, query, retrieve, and deliver data in usable, actionable formats for users. So whether it be Fire R4, parse Fire R4 resources or several other formats, the data is made actionable and usable.

I'm going to show you now an example of how we do this. Who's seen The Pit on HBO Max? Great show. Highly recommend. So, I'm going to take you on a quick journey in a busy ER. So, Dr. Sarah Murphy is an ER doc. She starts her shift and as she peruses the list of triaged patients, she sees Dr. Tom Waverly who's brought in by ambulance with chest pain, complaining of chest pain. And Dr. Murphy looks through, sees that they've taken some vitals for Tom. Doesn't appear to have too much other information. She's going to check see if there's any patient history. And probably that's due to the fact that Tom might be from out of town, on vacation, or simply a net new patient.

Now, when she looks at Health Gateway, it's telling her there's data from five additional external sources available for Tom. So, she clicks on that button and voila, Health Gateway has gone out, retrieved available records for Tom Waverly from other providers. She can now see Tom's medical history from previous encounters, including documents, meds, allergies, lab work, etc.

While this is really powerful, it's also overwhelming. It presents a challenge for a busy clinician trying to find relevant information for the patient. There are so many documents, results, notes, and so on spanning years of care across multiple organizations. I wonder what we could do to improve upon this. I'd like to invite my colleague Ken Coneton onto the stage to see how we could streamline this further. Welcome Ken.

So, Ken with Health Gateway, we go out, we retrieve from various networks available records for a patient. We frequently pull in hundreds of documents for a single patient.

Okay. You might be interested to know that our record up till now for a single patient was over 4,000 documents.

That is an impressive number of documents. I think I can help you with this problem. So, in November of last year, we released HealthShare AI assistant. Using generative AI, it empowers clinicians by quickly surfacing answers and generating summaries tailored to their role. So, Patrick, you've done a huge amount of the heavy lifting here. You've gone out, gathered all this data together to create a picture of Tom, but in your pit example and in the emergency situation, you're just not going to get time to read 4,000 documents. You just want that key information brought immediately to you so you can look at it. Right, Ken?

You're absolutely right. So, what can we do to bring these capabilities together?

So, we've been working on embedding the AI assistant within Health Gateway. So, the gateway viewer now has a chart called AI assistant. And when you select it, it knows who you are and what your role is. And based on that role, it will kick off a default summary for you tailored to your role. So in this case, we're getting a summary for Tom's emergency attendance. The summary opens with all of the urgent information you might want to assess straight away. So in Tom's case, I can see he's got an alert for a difficult incubation. And I'm also being told that Tom has been monitored for aortic root dilation and there are abnormalities in his echo. So this is really something you're going to want to drill into.

Other important information like if there's a DNR in place, mental health, safeguarding, infectious diseases, any possible deterioration are automatically checked to give you an immediate sense of the situation. A range of other sections are also available for a rapid summary for the user. We can see the previous ED attendances. This is summarized from both structured and unstructured data. So it gives you timeline of Tom's attendance to the department.

We also have allergies, active problems and medications displayed in reverse chronological order. Flags for high-risk medications or if there's changes to medications or if we see a consistency in the patient's record against what's on in the structured data, we bring that to your attention. We have vital signs and trending any recent investigations for radiology laboratory with quick summaries to give you information and other sections such as surgical history, mental health and social determinance if there's any advanced care planning in place and immunizations. All presented to you for quick consumption.

No need to click through screens, no need to go to different places to find information and no need to context switch to try to gather a picture of Tom's situation. So Patrick, we saw a little earlier that Tom was being monitored for the aortic root dilation. So let's drill into that a little more. So we can ask the AI assistant to give us more information on this. Once I ask that question of the AI assistant, the first thing it will do, it will decide what data I need to answer that question from both structured and unstructured sources.

So in this case, I can see there was an echo done six years ago and a consultation note. The important information is highlighted to me from that document. And if I select the source, I can now see that important paragraph highlighted within the wider context of the note. So now I can get a full picture of Tom's situation. So Patrick, it looks like Tom has not been followed up for a long time for that aortic root dilation. So we really need to investigate that.

Ken, this is really powerful. I think that our Dr. Sarah Murphy in the busy ER would be very appreciative of being provided that information so quickly. So just to recap what we've just witnessed. So we've retrieved a patient's medical history from five other facilities. We've pulled that in and parsed it nicely in the viewer. We've then used AI to create a one-page summary for the patient. What's more is that the AI helped us find a really important piece of information buried in a note from years ago that turns out that may very well be clinically relevant for the care of this patient. Amazing. Really great work. Really great work, Ken.

So, what else are you working on? We are working on a ton of other really cool stuff all around unstructured data, context engineering, and a lot more. But another area we're putting a lot of work into as an organization is Agentic AI. So, I'd like to invite Dimmitri up on stage who can walk us through some of that great work they're doing.

Okay. Sorry about that, little bit of setup. So, for those of you who were here last year, you may recall that we presented a demo of a prototype of an EHR with Agentic AI built inside. And what I thought we would do this year is follow up with kind of the next iteration of where we believe Agentic AI is going, not just in EHRs actually, but in end-user applications, particularly in healthcare. So what you're gonna see is a prototype. It's not ready for production, but it really leverages a lot of the advances we've seen in technology and our own advances in our R&D labs between last year and this year and starts to address some of the themes we've heard about this week.

So, while you watch this demo, couple of things I'd like to call out. You're going to see assistants become co-workers. The difference being with an assistant, you ask a question, you get an answer, you ask it to do something and it does it. A co-worker is an entity to whom you can delegate tasks and in some cases multi-step and quite complex tasks. You'll see that the co-worker is actually supported by a team of agents and each of these agents has a specialized set of skills. Those skills, by the way, are defined in markdown files, which means that they can actually evolve much faster than traditional development cycles. So they can kind of learn and grow with you as things adjust and change.

We're looking at building systems that are proactive, anticipating the user's need. Personalization is super important to us and we'll see that quite a bit. And in fact, there's a little secret about the UI for most of this demo that I'll show you in a moment. And of course, everything is grounded by human oversight, governed by evidence, governed by data that sits in our system. So, I'm going to invite my colleague Danielle to the stage. She's going to help me with this. Danielle is a nurse in real life and she also works for InterSystems. So, Danielle, you're the charge nurse on our surgical ward today. So, maybe you can start off by explaining what a charge nurse does.

Yeah, absolutely. So, as the charge nurse, I'm tasked with having situational awareness to what's going on that unit. That means I need to know how many patients are on the unit. What's the acuity of those patients, meaning what level of care do they each require? I need to know if I have any discharges planned for the day and I need to be able to manage any requests that come in for new inpatient admissions.

Okay, great. So, I'm just your guide in this journey. I don't really have a role to play, but I'll run the system and support Danielle in this. So, we'll log in. So, I'm set up as a nurse, and you'll see right away the agents are going off and finding data and pulling together an overview.

So, it's a nice overview, but it would be helpful for me if I could get some more detailed information.

Okay. So, to do that, let me introduce you to your co-worker. So, you have an agent, and the agent is supported by a team of experts. Each of those have quite specific skills. Now, I mentioned we believe in personalization and you should have a bit of a relationship with your agent. So, I don't love the name "my agent." So, Clara Barton was a nurse during the Civil War, founder of the American Red Cross and has very strong ties to Washington DC where we stand today. So, in honor of her, let's call the agent Clara. Hey my agent, can you rename yourself to Clara, please? I love that choice. Clara Barton's legacy of compassionate care and pioneering nursing is exactly the spirit we need on this ward. Let me make that change now.

So, Clara, Danielle has arrived on the ward to start her shift. She's the charge nurse and she needs a full detailed ward overview to start her shift. Can you please produce that for us? Got it. Danielle's starting her shift and needs the full picture. Let me pull that for her now.

So, you're going to see this pattern repeating multiple times. The UI agent delivers the detailed ward overview for Danielle now. She'll have the full patient board, vitals and status snapshot in just a moment. There's a coordinating agent that's calling multiple sub-agents to pull data and do their specific tasks. And finally passing it to a UI agent who's producing the UI.

So, have a look at this, Danielle. Let us know if this is a little bit better. Yeah, this is much better. This gives me a detailed patient list. I can see I have several patients who are ready for discharge today, but more importantly, I see that I've got two patients that have some sort of unstable flag on them that requires immediate assessment. Would it be possible for me to see more information for what's going on with J Anderson?

But before we do that, let me just jump in quickly and just point out that below the list you've got sort of what's planned coming up for the shift, what's happened in the past. And just to point out that actually producing a list of patients from most sick to least sick is not an easy task and this is something that traditional EHRs struggle with.

So the fact that we're able to order by most critical, there's some AI insights there and I mentioned earlier the UI has a bit of a secret. So the UI agent is actually dynamically generating these screens based on markdown files and we've included a source tab up here so you can see what this looks like. So this is the actual output of the agents that gets fed to the UI agent who has a set of templates and knows what to do. And this is for example all patients sorted by acuity here. This is where we define those things. So again, I think you can see that this leads to really high degrees of personalization and much faster agility in terms of delivering things.

So you did ask for a deeper dive on James Anderson. So let's click on here. And again, you'll see the same pattern, the master agent coordinating the sub agents, getting the data, bringing it together, sending it back to the UI agent, but it's a different template because it's a different purpose. So here's your patient snapshot for James.

So this snapshot is excellent. It tells me exactly what I need to know. Oh, I've got a patient that's been admitted for diabetic ketoacidosis. It's using up to the minute vital signs and lab data to tell me that this patient is actively decompensating. Normally, I'd have to go trolling through the record and try and put the pieces of this puzzle together for myself. Now, I know exactly how to target my assessment when I walk into that patient's room.

Okay, so I think you better go see James and the other patients on the ward. Grab your phone. You can use the app to take notes while you're doing the ward rounds and we'll use our ambient processing to bring it back into the system. So meanwhile, while Danielle's off doing that, I noticed that Clara's got her little hand raised. She needs something. Hey Clara, why is your hand raised?

Let me check what's flagged right now. We have a new admission coming to the ward and need to free up a bed. Carol Taylor in bed S104A is our best candidate. She's been medically cleared for two days and her billing clearance just came through. Would you like me to start discharge planning for her?

That sounds like a good idea. But why don't you go ahead and do the discharge planning and when Danielle's back from her rounds, we'll ask her to approve on it.

So, this is a more complex documentation agent working on the discharge plan for Carol Taylor now. It's the same. It'll be ready for Danielle to review when she's back from her rounds.

Discharge plan. Carol Taylor, ready for review.

So, Danielle is back. Welcome back, Danielle. While you were away, we got notification that we have a new patient arriving from the emergency department. Clara has recommended that we discharge Carol Taylor to free up a bed. So, she's produced a draft plan for your approval. So, I'll bring this up. You can see up here there's a readiness section. So there's a series of clinical stability tests that have passed. Her billing just cleared. There's a proposal here to book a follow-up appointment and we could change that to a different category if we want. She doesn't appear to own a car, so we're proposing to book an Uber for her. And we've also drafted a discharge summary, which will go into the record, and a patient education letter. Looks like Clara's native language is Spanish. So this letter is actually in Spanish. How does that look?

Everything looks in order to me here. I say we go ahead and approve it.

Go ahead and approve it. Okay. So when we approve this again, same process, the agents pulling it together back to the UI agent and we're now at the completed. So we can see the appointment's been booked. Proposal to send an SMS reminder. Transport's been booked. Looks like it's coming in an hour. The discharge summary we saw before. The patient education letter. Anyone need an Uber? We don't actually need this.

Should I cancel? I should probably cancel the Uber. Huh? All right, we'll cancel the ride.

So, you've arrived at the end of your shift and you now need to hand over to the next shift.

Yes. So, it would be helpful for me if you could generate a report for me for all those things that I first talked about. I need to be able to pass over a report to the charge nurse that's coming on shift. I need to be able to tell that charge nurse who are the patients on the unit, what's the acuity of those patients, what level of care do they each require, how many discharges are still pending, and how many new admissions have we admitted during the shift.

Okay, great. Hey, Clara, can you produce a handover report for Danielle so she can hand over to the next shift on it? The team is reviewing the shift. Documentation agent will compose the final handover report for Danielle. Now.

Shift handover drafted ready to review.

Okay, so it looks like you can see we've freed up a bed. We've got an SBAR summary for your most unstable patient. The things that are time critical that are coming up in the next hour, recent changes, we can see that Carol Taylor's been discharged and so forth. Does that look okay?

This is excellent. It's everything that I need to be able to concisely give a handover report.

So having been a nurse, what do you think of all of this? Would this have helped you?

Yeah, I love this. Love to have been able to have something like this when I was doing bedside nursing. Be able to have a team of co-workers with me at every interaction, whether it's an assessment or delegating task. This would have been a real game changer.

Yeah, we're really excited and I'm excited not just about the advanced clinical intelligence, but also the ability to do the more mundane tasks, the booking, the Uber checking checklists and things like that. And this is all not about replacing nurses, but about giving nurses and doctors and everyone else more time to take care of the patients. So, thank you Danielle. Appreciate your clinical insight.

So, all that's left for me today is to introduce our next speaker. So, hey Clara, could you I've got a really important guy coming up to talk now. Could you help me introduce him?

I'd love to help, but introductions are outside my wheelhouse. I'm here for clinical ward support. That's all you, Dmitri. Good luck with the speaker.

So, I knew she was going to refuse and that's a really important point, right? Because this is not chat GPT who's going to try and please you and do everything. This is a controlled engineered system that's got a specific set of tasks and if you try and go outside the task, it's not going to like it. So with that, thank you for your attention.