How Health Data Interoperability Can Improve Patient Care

[J.B. WOGAN]

I’m J.B. Wogan from Mathematica and welcome back to On the Evidence.

This episode focuses on data interoperability in the health care space, which is part of a broader conversation happening at Mathematica, and at agencies and organizations in the public and private sectors, about how technological advances we’re seeing could make greater volumes of data available faster, at lower cost, and maybe even in higher quality formats, which would mean making data more useful. In health care specifically, this digital transformation has implications for improving the care patients receive and keeping people healthier, which ties back to Mathematica’s mission of improving public well-being.

Over the past decade or so, we’ve seen the widespread adoption of electronic health records across the United States, but all too often those data can’t be shared automatically, and even if they can be shared automatically, they may appear in formats that aren’t useful, or that need to be manually changed in order to become useful. A recent webinar hosted by Mathematica’s Health Data Innovation Lab examined strategies and tools for achieving greater data interoperability.

The conversation was moderated by my Mathematica colleague, Steve Linthicum, who is an expert on Health IT and data exchanges. The participants were Dr. Abel Kho, Steven Gruner, and Nicholai Mitchko.

Dr. Kho is the director of the Institute for Artificial Intelligence in Medicine at Northwestern University.

Steven is the president and chief executive officer at HealthWare Systems, a company that specializes in software that automates health care processes.

And Nicholai manages solution partner engineering at InterSystems, a health care data technology company.

If you’re new the podcast, please take a moment to subscribe. We’re on YouTube, Apple Podcasts, Spotify, and elsewhere. And if you like this episode, consider leaving us a rating and review. It helps others find our show.

With that, we’ll move to the webinar. We start with Steve Linthicum posing his first question to the panel.

[STEVE LINTHICUM]

All right, so to kind of start things out, I'd like to discuss some of the primary challenges that health care organizations face as they work to extract value from their interoperability efforts. I'd like to start that discussion with what are some of the significant barriers to effectively utilizing data once systems are able to exchange data. And I'll start with Dr. Kho. Dr. Kho, in your work on the Capricorn project, what barriers did you deal with in being able to use the data once it's been aggregated?

[ABEL KHO]  

Yeah, thanks, Steve. It's great that people are interested in interoperability. This has been a passion of my career all along. And part of that has led to this Capricorn project, which is the Chicago Area Patient-Centered Outcomes Research Network. And we're coming up on 10 years. And what that is, it's a network of, you know, nine institutions, which have come together under a shared governance and a steering committee to interoperate data, primarily for research purposes, all sharing a common data model and some common technical infrastructure.

But one thing that we've learned over the years, actually many things, is that interoperability is just the first step. And data quality is a never-ending challenge. And so once the data moves over or is brought together or in some cases computed in parallel, you still have to make sure that the quality of the data is high. You also have to understand whether or not the data is representative for the problem or population that you're trying to study or trying to get insights about. And then as soon as data exchanges, it loses its freshness. And so how do you keep that data up to date and fresh? Privacy concerns, always a concern. And it's something that we've been very passionate about, making sure that institutional privacy is preserved, as well as the patient privacy.

And institutional privacy, I would say, is something that people oftentimes overlook. I mean, there's regulations around the sort of patients with HIPAA, but institutions have their own concerns. And that leads us to governance. And so governance is never-ending. Leadership changes, institutional priorities change. And so how are you ensuring that once data is exchanged, that you're still governing it around the initial problems and going back and ensuring that any future purposes or uses of it are also equally well governed? So if achieving interoperability is achieving a highway and road system, then using the data is like driving the car on it. You always keep your hand on the wheel. You're constantly monitoring what's going on after those roads are built. So it's a never-ending process.

[STEVE LINTHICUM]  

Great. Thank you, Dr. Kho. You bring up some very important issues like data quality, potentially inconsistent data standards that affect the data once it's been exchanged and how you can leverage that data. Nicholai, what have you experienced in these areas with the clients you've served?

[NICHOLAI MITCHKO]  

Yeah, this is a great question. To continue the analogy, Dr. Kho, that you gave about interoperability being a highway system, you can think of data quality, data silos, and inconsistent data standards similar to the issues we have when we're driving. Data quality is like hitting a pothole. A data silo is like a closed exit. And inconsistent data standards would be like changing from driving on the right side of the road to the left side of the road. So these issues severely impact our ability to use the data -- doesn't impact the ability to get the data all the time. And some examples I've seen pretty frequently in the field are actually really simple and really common, but they can really hinder, they can create those data silos and data quality issues. So one of the examples I see over and over and over again, and I'm sure Steve would second my opinion here, is that I see diagnosis in the encounters in an EMR.

And so if I don't have a diagnosis in the right field, and I try to send my data and be interoperable, what that means is the receiving system probably won't file it correctly. And that is creating a data silo. Another common thing is just code set differences and customizations to data standards like HL7Z segments and FHIR extensions without using a guide. And all of this really leads to what integration tries to prevent, which is barriers in care and missing data. So, to kind of conclude, it is very much like a highway. It's like a pothole, a closed exit, or changing the side of the road. And just ensuring you have data standards and interoperability is step one, and then getting to your destination is step two.

[STEVE LINTHICUM]  

Thank you very much, Nicholai. And as we kind of just continue our discussion here about challenges, Steve, I recall we worked on a project together a few years ago that ran into issues using data with pairs once we aggregated it from several dozen primary care clinics. I'm sure you remember this project. Can you talk about that and how that hurdle was overcome?

[STEVE GRUNER]  

Sure. Yeah, it was an interesting project in that, in many respects, it resembled a local HIE for primary care physicians. And the whole goal was to exchange information so we could roll up quality measures and report them at the aggregate level while also maintaining the ability to do so at the individual organization level. And so, in addition to dealing with seven different EMRs with all sorts of different types of deployments and configurations that they chose, there were five different pairs, two different labs, Quest and LabCorp that we were ingesting data from, and a regional HIE. And so all of this was brought together into a clinical data warehouse.

And ultimately, the goal was that by sharing that information, we would eliminate a lot of the calls, a lot of questions that would come from the pairs that the local primary care physician or their staff would have to respond to. And so that controlled access was a big win. And then we went through the process of validating that through the individual pairs. It became obvious that to really do this effectively, we had to go through the whole data aggregator validation process, which we did. And at the end result, we were able to exchange all this information. But the big lessons we learned along the way was that, you know, in spite of creating or facilitating the ability to do this, there was still a lot of hesitancy around -- you know, Dr. Kho had mentioned HIPAA and the hesitancy at the different levels of that level of sharing information.

Very inconsistent from one site to the next and their willingness to do so. I would even say that we learned that there was some lack of trust in terms of the quality of information from one site to the other, and then the willingness of that site that's receiving information to actually bring it in without first going through their own validation. So the ability for us to share that and reduce some of the redundant care wasn't as great if they weren't willing to trust the information that was being supplied to them from that initial primary care physician. But, yeah, it was a great project in terms of really exposed to a lot of different moving parts and then tying that all together and creating that clinical data warehouse.

[STEVE LINTHICUM]  

Yeah, that was a tremendous effort, you know, and what we saw when we did, you know, connect those seven, I think it was seven different EMRs, 50 plus, you know, different clinics, and then wanted to exchange that data with those payers to support those HEDIS measures, and they wanted to validate every single feed that we had. So you can just imagine how untenable that was to us from a manpower perspective. So really going through that NCQA process to be able to create that whole validation process for the system level as opposed to just the individual feeds was a tremendous win. So thinking about some of these challenges, you know, I'd like to kind of put this out there to all three of you guys. You know, what kind of strategies can organizations employ to help to overcome the challenges with using the data once it's exchanged? Can you guys provide some thoughts about that?

[STEVE GRUNER]  

I could jump in here. So the project we just discussed, one of the more common methods for exchanging information to kind of bring that information together in the clinical data warehouse was the CCDA. And so that format, you know, was able to be produced by all the different EMRs and EHRs, but there's wide variation in terms of the quality of that information as we brought it forth from each location. Some of that had to do with the configuration itself and how they chose to deploy the EMR. But in a lot of cases, it was the behavior of the physicians, and I think the example was given earlier about the diagnosis code.

Well, having information in discrete fields is a lot easier to contend with than if some of that information is buried in the notes and comments section. And so we found that a lot of the quality-based measures and information that was actionable were in those notes section. And so as a result, you know, we had to do a lot of transformation to get that information into a more structured form. And then once doing so, then that became easier to go through that data validation process because they're looking for that information in a more consistent way across each of the supplied data sources.

[NICHOLAI MITCHKO]  

I would agree with Steve that really ensuring that you have a target of what you're going to do with the data, starting from the clinical standpoint, and really focusing on a, you know, clinical first, use case first, sort of methodology and ideology, and then tackling the interoperability and why you need it. And that will inform, Steve, like you mentioned, the individual fields in the NCQA that you need to pull out from unstructured data. And I think this certainly follows up to, you know, some of the next discussion points, which, you know, will include AI and unstructured data and how do we even handle that.

[ABEL KHO]  

Yeah, I think that's really important. I mean, you know, a lot of times we sort of -- you can't just build like sort of generalized data sets. I think having a problem you're trying to solve in mind is really important. The other thing I would just add is that it's really important to also engage the people who are involved in the process. Like, I think Steve had mentioned, there's primary care providers involved. I'm in a primary care provider. If the problem you're trying to solve is somewhere along those processes, what we found that's very helpful is to actually observe and see and maybe talk to some of those people who are sort of involved in the data chain, like the creation of it.

And from there, you can oftentimes identify like, hey, you know, what we're getting downstream, it's a reflection of this other strange process upstream. If you don't know that, you'll make sort of strange conclusions about that. So data quality is more than just conformance to the standards. I mean, if you don't pay attention, people can put all sorts of monkey business up front, it'll still conform. It'll look like a compliant HL7 message or a FHIR format, but it's just not valid data. So engage the people that are involved in that data creation too.

[STEVE LINTHICUM]  

That's great. And actually, that's a great segue into our next topic area. Once we've identified the challenges, let's shift our focus now to the technologies that can help address these issues. I'd like to start with Dr. Kho, in your work leading the Institute for Artificial Intelligence and Medicine, what role do technologies like AI, machine learning, and data analytics play in making interoperable data more actionable?

[ABEL KHO]  

Yeah, I mean, so, you know, I'm obviously I'm bullish on the potential for AI. You know, where it intersects with interoperability, what we've seen is we're really excited about the potential for AI to reduce the extraordinarily laborious process of making data interoperable. So converting it into conformant data sets, standards-based -- you know, anyone who's done this work realizes how much effort that takes. I mean, you'll have teams of people at each institution writing ETLs and trying to transform things into these common formats. We've been excited. We've been piloting some efforts with industry, startups, with our students. And what we're seeing is that a lot of these AI tools can really effectively transform raw data, electronic health record data, into a common data model probably better -- actually, it is better -- than what people can do because it never tires and doesn't make mistakes like that.

The other thing is that, you know, we also see potential because increasingly AI tools are being used at the point of data creation. So for example, note-taking assistance like our DAX co-pilot at Northwestern, that's generating notes. It's a machine generating the notes on behalf of the clinician who's just basically speaking sort of with patients and it's capturing all that. That's another point of like consistency that we can create in terms of data quality. The students that we're working with right now have also run some interesting pilots where they can take a bunch of data, electronic health record data, both structured, unstructured data, and then create sort of these multidimensional representations of that data, like how large language models work, cluster that together, and it actually does a really good job of creating clinical concepts that a clinician can recognize.

And it does so with higher sensitivity than if we were just to be bound by only data that is captured in diagnosis codes or structured data. So I think there's a tremendous potential for improved quality of the standardization, increased inclusivity, and upfront, upstream data standardization. And then finally, on the data analytics side, as I mentioned, sometimes you don't -- I think Steve alluded to this, institutions don't always trust each other. And so, are there ways for us to analyze data without moving the data at all? Because each time you move data, you oftentimes create another replicate set. And that's just not good practice. I mean, from a privacy standpoint, governance standpoint, that's not great. You don't want all these data sets out there.

So I do think that what we've been piloting in Chicago successfully is ways of analyzing data across sites without ever moving the data. But again, data quality is important for that process to work, something we've been funded by NSF, initially now by DARPA, to implement real-world secure multi-party computing. We've got operational across five sites. It seems very promising. But again, you still have to have that underlying data quality because each person can't see under the hood for the data of the other institution. So you have to make sure that the conformance and the quality is high. But distributed analytics looks like a potentially nice future as well.

[STEVE LINTHICUM]  

Great. So I'd like to kind of jump off of that and ask a question to Nicholai. How can health care organizations integrate these technologies into their existing systems? That's a big issue, right? People don't want to just -- you can't just rip and replace everything that you have. You've got to be able to integrate these in. So how do we do that to enhance data utilization?

[NICHOLAI MITCHKO]  

This is a really good question. It's probably the million or billion-dollar question, I think, of the day. And so, I like to think and InterSystems likes to think about AI technology sort of integrated in three ways. And Dr. Kho mentioned the first two, which is add AI at the point of care. And so, I call this adding AI at the point of transmission. So the creation of data using an AI, maybe it's an ambient device that's listening to a patient and a doctor interact at the point of care, maybe it's any other number of use cases that AI can use, can be employed today. That's integrating right when the data gets created. And doing that, you set yourself really apart from all the other data quality issues, really data entry or data creation issues by using AI to enforce a standard.

The second thing that I think is really important is making sure that that point of care integration is done with a human in the loop design. And what that means is that you want to keep your clinicians in their workflow and have a note pre-populated. But the clinician still needs to review it and the clinician still needs to sign off on any data to ensure that it's accurate, and that new data and insights are relevant to what just happened.

And then the final thing is I would call it a deep clinical integration where you can use AI is to enhance data as it comes across the wire. And this is the second point Dr. Kho mentioned is as data moves from system to system, there is a lack of trust. And it's really just because those data formats can be different. You know, I can get a CCD from system A, I can get FHIR bundle from system B, and I can map them to the same format, but things end up in different places. So using AI to ensure that the data quality is correct at transmission as well, and to clean up data to create a proper warehouse, a proper operational data store. Those three things I would say are the first three places a health system should think about adding AI to their sort of interoperability toolbox.

[STEVE LINTHICUM]  

Great, thanks Nicholai. And, you know, there's certainly more than AI technologies out there that can really help us to leverage data once we've exchanged it. Steve Gruner, I know your company is deep in workflow and process automation technologies. Can you describe why these technologies are important in extracting value from interoperability?

[STEVE GRUNER]  

Sure. I think we've alluded to some of this already in some of the examples that were given. But, you know, when we talk about, you know, at the point of transmission and being able to apply, you know, enforcement of some standards at the point we're creating this information. But it also has to be in place where I sometimes call it point of origin, because, you know, we may not be able to control that first piece, right? At least not yet. And so if we're going to not impose change on the provider community and still accept some of this unstructured data, well, where we receive it, it becomes really important. Is it coming in through a fax server? Is it coming in through a data fee? Is it coming in through, you know, traditional snail mail? All of those are still happening at a very high clip in health care. I think the last percentage I saw was 70 to 80% of all information coming into the health care system is still unstructured. So our goal is always to impose structure on it. We've kind of talked about this in some of the previous comments.

And, you know, making sure we're getting it right the first time, because if we recognize the information as we receive it, and then we classify it, we index it, we extract the actionable data, we start to build a repository of information that is truly managed content. A lot of people think of content management systems as, well, as just an electronic filing box for my scan documents. The reality is that if we know what type of information is in each and every one of those objects that we store, now workflow automation and automation of any kind becomes much more viable. A lot of workflows in health care are deficiency-based, or they're triggered off of the absence or presence of certain information. And if we don't know what we have, then we don't know what we can automate.

And so if you impose that level of data governance on all information coming into the system, now you can impose or automate workflows that are driven off of the receipt of that information. I sometimes call these triggers, right? So this triggers the start of a workflow. And so one great example of this, and then I'll hand it back, is that correspondence automation. Information coming into the organization where somebody is asking us to do something. It could be a refund request. It could be a payer asking for additional documentation. Well, today, somebody is going to read that, interpret it, decide how to respond to it. Maybe they'll kick off an electronic workflow to then take it from that point forward. But all of that can be automated. You know, you can interpret that request. You can extract the actual information. You can even kick off the automation if we've done the proper work on the front end. I have a fully indexed repository of information.

If they've asked me for an itemized bill or they asked me for a medical chart, they've asked me for a doctor's order, I can draw those objects down instantly, hand it to that human in the loop as a fully populated, ready-to-go response. And then if you still want that person to validate it, approve it before it's released, you have that option. And then over time, you might choose that some of those are -- we don't ever really change and we're constantly just kind of rubber-stamping and out it goes. You might decide to allow those to flow out automatically. So a lot of automation can kind of keep going up the ladder the more we do on the front end in capturing and indexing and storing this information.

[STEVE LINTHICUM]  

Great. Thank you, Steve. I'd like to kind of wrap up this kind of question session and open it up to all you guys. Are there any other emerging technologies or techniques that show promise in further improving the use of interoperable data?

[NICHOLAI MITCHKO]  Yeah, I would jump in here and say that, you know, when we think about interoperability as the two steps, step one would be connecting the data, step two would be using the data. There still is a lot of room to improve on the connection side. And one of the emerging technologies I see, and we're certainly experimenting with, and I know that others are, is AI-assisted data mapping.

So when you don't have a trust between two systems, and you have schema one and schema two that should be the same, but they're not, just using AI as an assistive technology to be able to map that data correctly at the transmission level and the master level will help you accomplish that step one of interoperability, connecting and collecting all the data. And then the actual use in improving the interoperable data, you can get to faster as opposed to spending your many, many man hours just connecting to feeds.

[ABEL KHO]  

Yeah, Nick, that's exactly what we've been really excited about as well. And we've seen some good examples. It works amazingly well. I mean, it cuts the time for, and more importantly, the consistency by which people map tremendously. There's still a human in the loop, I think, to your point, Nick, but it takes out a whole swath of the simple mappings and conversion to, like you said, a scheme or a common data model. It takes care of all that. It's oftentimes you can convert from one data model to another very quickly. It's been really exciting. One caution is that we sometimes in health care can over specify a concept. Like, so, for example, say a person had a diagnosis code, they have diabetes.

Well, it turns out if you take all the information around that person, it might have been just ordered for a lab test or what have you. They don't have diabetes. And if you add all their notes in, they used to have it, they don't have it now. So I think that one of the things that we see the value of large language models using unstructured data is sometimes you can include more data and get in many ways more accurate, but not over-specified clinical concepts. You can say this person probably has diabetes, and sometimes that's better than saying for certain this person has diabetes in weird ways. So I think there's a lot of potential. And the AI, interestingly enough, has some of that looseness in how it defines things, which we kind of like.

[STEVE LINTHICUM]  

Great. Thank you. Thank you, Dr. Kho. So I'd like to shift gears kind of one last time here. And I know we've sprinkled in some real-world examples in some of our discussions so far, but I'd really like to kind of reinforce some of that because that's how we learn, right? We learn by doing. So I'd like to just kind of give the panel here a chance to share some examples of health care organizations that have successfully navigated these challenges of data utilization post interoperability.

[STEVE GRUNER]  

I can jump in here at this point. So we recently were asked from a client to deploy a prior authorization automation solution, and they believed that the time that they had sufficient information to pass that information to us, and we would simply issue the prior authorization request, grab the result, and then push that back into the EMR. During the discovery phase, what we identified was that a lot of the information required at that point in time hadn't yet made its way into the EMR. So what we ultimately wound up doing was putting a document automation solution at the point where these orders were coming in.

And so we were then extracting that actionable information that was needed to satisfy that prior authorization request, but it also had a secondary benefit, which was we could impose those standards at the same time because today, you know, or prior to the solution going in, if that order came in, it was incomplete. It was a sign missing some key data points. Somebody would have to read that, review it, and then reach out to the physician and try to resolve the issue.

Well, now that can automatically be returned to sender because it doesn't meet the required standards for us to accept that order. And now that then becomes part of kicking off the pre-admission workflow and the scheduling workflow, but we're not bothering our staff, asking them to work an order that ultimately we can't do anything with because it's incomplete. But it's a case where you're using different technologies to achieve your ultimate goal. And this particular case, without having to impose change on the referring physician, which was something that the hospital didn't want to do at this point in time.

[ABEL KHO]  

You know, maybe I'll jump in quick. And so, Steve, you used to work at State of Illinois. So this may come as an example, but I've had the good fortune to help and get involved with public health at the city, state, and federal level with CDC and some of the government agencies. And I've been really encouraged and excited to see there's a tremendous amount of work and thinking around interoperability at the city, state public health, and at the CDC.

And oftentimes this conversation would be right at home there. You've got experts, really great people, like sort of, I would say, heroes of public health who are working on this exact problem. And post-pandemic, people have realized the importance of interoperability. And there's some resources towards there. And so they're talking about using TEFCA as a guidepost for use of implementation of standards. They're talking about interoperability of systems and how do we simplify systems. They're making recommendations around how do you change the financial incentives and policies to enable these things to happen.

So there's actually a tremendous amount of exciting things happening in the public and population health sort of domains. But yeah, so surprisingly, maybe it shouldn't be surprising, city, state, federal agencies are actually doing a really good job with this idea.

[STEVE LINTHICUM]  

I'm glad. Let me just jump in here real quick. I was with the Illinois Health Information Exchange about 10 years ago. Unfortunately, we weren't completely seeing that big picture you're talking about, right? We were pretty focused on the state public health office and being able to exchange clinical data between hospitals across the state and for those hospitals to send syndromic surveillance data and immunization data to the state public health department. Certainly, the thinking has been much bigger now and broader that being able to share that data across all those different agency levels, et cetera, that's just tremendous how far that's come. Sorry, Nicholai, for jumping in there. I know you had a response here as well.

[NICHOLAI MITCHKO]  

I think kind of the experience we've seen is this, Dr. Kho, like you mentioned, is the ability to create, especially with COVID, a registry, right? So with interoperability, you connect all these data sources, you have it in one nice place. What can you do with it? Well, we've certainly seen the federal government say, we're going to try cancer moonshot. So how do you figure out how to cure cancer? Well, you first need to figure out who has cancer. So that's a registry. And one of the use cases we've seen is with foundations and the federal government, and recently the Cystic Fibrosis Foundation created a next generation registry. And we help them use what's called computative phenotyping. It sounds complex, but it just means applying an algorithm to define individuals with and without a disease state.

And typically what that means is, look at their unstructured data, look at their notes to figure out, does this person have cystic fibrosis or CF? And giving someone who has these large amounts of data the ability to interact with their own data in a registry, or be able to review a, let's say, 7,600 pages of clinical records for one patient, you can do that with AI as well assisting. So I would just say that the CF is probably one of the premier examples of actually doing something with that type two interoperability.

[STEVE LINTHICUM]  

Great. Thank you, Nicholai. And to kind of wrap up this section, I would like to put this question out to all three of you guys. What lessons can other organizations learn from these examples and key takeaways for health care leaders looking to optimize their data exchange? What's kind of, you know, in your mind, each of your minds, what's kind of the one key thing to really keep in mind? Steve, would you like to jump in there first?

[STEVE GRUNER]  

Sure. Yeah, I would say, you know, first off, understanding kind of the source of truth for the different data sets and data sources that you're dealing with. Not everything is going to be in one consolidated repository. So being able to tap into that distributed network of data sources to build a response, let's say, or to aggregate information in such a way that now I can respond to the query that I was posed.

So you know, when we look at different automation opportunities, we're usually walking into an ecosystem. There are multiple solutions, multiple data sources, and the source of truth is not always the same, depending what information we need. So it's applying the right technologies to go and get the information when it's needed and provide it where it's needed so that we can automate those workflow solutions accordingly. And I would also add that, you know, a lot of people look at that process and think it's daunting that I'm going to have to take all this unstructured data and I'm going to have to map it all out and do all this different work. And that may be true 5, 10 years ago, but the technology has really evolved with AI to the point where it's not as monumental a task as you might think.

There's a lot of assisted tools that go into that process that make it much easier to recognize that next document to interpret it before I even have to get involved. So I would encourage everybody as you're looking at information flow coming into the organizations, you know, there's a lot of technology available that will help you impose structure on that information and make it actionable so that it assists all subsequent processes and creates efficiencies for you and every down-path operation after that.

[NICHOLAI MITCHKO]  

I would say that's very similar with our experience and kind of, I think what you're describing, Steve, is that total integration is a journey, right? It isn't a project that you say, all right, I start and then end. To think organizationally, you have to think about this integration as a journey, especially as you think in maybe six months, 10 months, there will always be new data. There will always be new codes. There will always be new fields that you'll have to handle. And there's always formats that will be emerging for that data.

So if you set up your technology and organization so that you're best equipped to handle those new changes, that is what I describe as the journey and having the right software is part of it, having the right organization is part of it. And the mindset that clinical first, and then research second, make sure your doctors have what they need, and then make sure that the research side also has what they need to ensure they're both doing their job correctly.

[ABEL KHO]  

Yeah, and I emphasize what Nick said there, the last bits, like, you really have to prioritize understanding, you know, the values of the people and institutions involved. You have to prioritize getting a deep understanding of the problem you're trying to solve with data. And that sometimes means getting your hands dirty and talking to the people that are most affected. And then it is a journey. So you know, I'm from Chicago, so, you know, Ferris Bueller, I think, said, you know, life moves pretty fast. If you don't stop and look around, you're going to miss it. Well, with interoperable data being the case now, I mean, data is moving really fast. And so you do have to, like, stop, look around, so you don't miss any of the problems and opportunities that data presents.

[STEVE LINTHICUM]  

Awesome. All right, great. Thank you, gentlemen, for, you know, working through those questions with me, providing those kinds of insights. We would like to kind of move into the Q&A portion today. Looks like we have a question from -- oh, it just moved. There we go. Question from Aparna Keshaviah. The question is, given HIPAA constraints and other challenges to data sharing that were mentioned, what are your thoughts on leveraging distributed data networks in lieu of trying to unify multiple data sources to obtain insights from these multiple data sources?

[ABEL KHO]  

Yeah, so Aparna, I appreciate the question. That is exactly what our network is. So we are big, big fans of distributed data networks or federated data networks where everybody holds on to their own data and you only bring together data when you need to. And so, as I mentioned earlier, you know, we've been piloting and demonstrating that you can effectively analyze data across a federated network without moving data at all, while the compute is all done in the encrypted space so nothing is revealed. You don't reveal anything about the patients, the compute, the data, or the institutions. That's really appealing. It's not easy to pull off because you have to -- you know, there's some computational cost to that. But that seems really an effective method. And then you still have to have governance, though, because there's going to be instances where you may not bring all the data together, but you might bring in some subset or metadata that allows you to come to an answer.

And so I think, you know, we'll see more and more of these federations happening. I think there's a real interest from some of the federal funders we've seen. But again, they still have to have high data quality. That means everybody has to maintain their data with high enough fidelity and quality and freshness. Because, you know, in the federated network, you can't necessarily look behind the hood of the other person's car. So I think it's a little tricky there.

[STEVE LINTHICUM]  

Great. Thank you, Dr. Kho. Looks like we've got a couple of questions that are somewhat similar from the transportation sector. So let me put this out there and see what you guys think. From Jana Lynot and Al Benedict, these are very similar questions, so we'll just kind of work through them. Question is, my sector, transportation, is exploring how to interoperate medical and human services transportation providers in the community to create greater transportation capacity to ensure patients get to their appointments and other destinations. Do any of the speakers know of health care sector agencies, companies, individuals exploring this from your sector?

[NICHOLAI MITCHKO]  

Yeah, I can jump in here. This is actually a really common problem. So no show rates, which is the percentage of time that a patient doesn't show up to their appointments, cost health systems, lots of revenue, and it's actually, they can be as high as 40 or 50% in some locations. A few ways that you could tackle this is work directly with the health system. There's a messaging type called a scheduling message, an SIU if you're technical. Where, if you can tap into those feeds, you would be able to be notified of when a schedule -- obviously, some of the patient data would be obscured for your use case. And be notified of where and when to pick the patient up, where they should be dropped off and how they might handle this.

The health system is very motivated to provide this integration because they are reducing the no show rate by doing this. And that's why we see, and at least in the major cities, particularly in Boston, where I'm located, most of the health systems offer you a free Lyft or Uber to your appointment. So certainly, I'd say this is common and, you know, Steve and Dr. Kho would welcome your insights as well.

[ABEL KHO]  

Yeah, I mean, certainly our health system thinks and cares about this a lot. You know, every major city, it's tough to get around, frankly, you know, and you see that. I do know that, like, in Boston, Uber and Lyft, I think we've had some conversation with them in the past, and they see themselves as logistics companies. And they actually have, I think, offered services specifically around getting patients to their appointments. The health systems themselves oftentimes have some insight into this. They can run reports and identify where patients are coming from, where their population is drawn from. And in follow-up surveys, you can also identify oftentimes when that was an issue.

In fact, now I think we have a social determinants of health wheel that lists, asks the question, did you have difficulty getting to your appointment, or how did you get to your appointment? Oddly enough, one of the questions, one of the answers isn't, I drove myself to my appointment. So that's odd, but it has other ones there. So this is a huge issue. I think it's an important problem that I think commercial companies are trying to solve, and the health systems are certainly incentivized to solve.

[STEVE GRUNER]  

Yeah, and if I could add something to that, it's also not just for a clinical purpose, right? We also have the situation where a patient might be scheduled to come in to meet with a financial counselor, or that patient may need to attend a Medicaid hearing or a disability hearing. And those are all things that ultimately benefit the facility in terms of having a reimbursement source for care that's provided. But that won't necessarily show up in some of the scheduling systems, but it needs to, because that also is a challenge. And most organizations that work with third parties, or in some cases do it themselves, will arrange for that transportation to ensure that patient gets to that appointment. Again, not that it's clinical in nature, but it's still an important appointment for them to keep because it helps the hospital get paid for the services that they're providing.

[STEVE LINTHICUM]  

Yeah, and certainly, you know, this topic kind of gets us into an area that we should probably have a whole other webinar on around social determinants of health, right?

[STEVE GRUNER]  

Yeah.

[STEVE LINTHICUM]  

You know, there's many more things other than an appointment to get to that really drive, you know, a person's well-being. So transportation is a major, major part of those social determinants of health. I'll stop there because we can go on forever about that. We're just about to the end of this one. So let's see, another question here from Anne O'Malley. A question for Dr. Kho about the need for the data to be of high quality. Could you walk us through an example of how current AI looks to the primary care physician on the clinical end? For example, if you are seeing a patient in the office who has various problems, conditions, you're going to enter data into the EHR. How would present-day AI enable you to focus on the patient rather than on your computer screen?

[ABEL KHO]  

Yeah, so thanks, Anne. Great question. I've been fortunate that I've been able to be an early sort of user of the automated note-writing device. In our case, we're using the DAX Copilot product from Microsoft. And I've watched it go through iterations. It actually is incredibly high quality. I would say it reduces my cognitive burden so much so that -- patients consent to if I can use it or not. And if they don't, it makes it much harder. I mean, I do save significant time.

The quality of the notes actually is super high now. So when it started, they had some issues, but now it's actually, frankly, better than notes I would write myself most of the time. I still have to review the note and check it and all that kind of stuff. But that has led us to come to the realization that those low-level tasks, like transcription and note-writing, are likely to be replaced in the future, if the budgets are right, with automated methods. And again, that creates an upstream consistency point, or at least in process, for how these things are generated. Unstructured data now, too. So the note, unstructured data, is going to be generated by a computer.

That's a really interesting space we're going to get into then. And one thing we've also realized is that the value of these notes is not that, again, it specifies specifically that a person has this or this or this, but it can interpolate between the spoken human language and come up with sort of a general concept that the person probably has something or has this or this. I see that as, oddly enough, you can have high-quality, clinical, useful data without being over-specified. That may be hard to explain that. But it's high-quality from a clinician viewpoint and not asserting something that may not be there.

[STEVE LINTHICUM]  

Great. Thank you, Dr. Kho. I'm going to take one more question here. Let's see. I got one from Ngan MacDonald. Do you see adoption of interoperability improving with the increase of value-based models of care? Steve Gruner, this might be near and dear to your heart.

[STEVE GRUNER]  

Yeah, I think that certainly the interoperability was a key part of that project. And ultimately, we were able to address some of the inconsistencies in terms of how the information is being captured at these different sources. However, the technology that Dr. Kho is talking about would have eliminated a lot of those problems because at the point that that information is being entered, you could impose that data governance. You could impose those standards not just on structured information, but everything that they're entering into notes now can be made actionable or in some cases even mapped to those structured data points.

So we definitely see where had this technology been available seven years ago when we were doing that project, we would have been able to take greater advantage of it. But it continuously evolves. We talked about data as a journey. Technology is the same way, right? We constantly see improvements there, and we have to stay on top of that and leverage those when we can because it eliminates a lot of the manual things that we have to do around data governance and manual processing of information.

[STEVE LINTHICUM]  

Awesome. All right. So as we're wrapping up this discussion, I'd like to thank our panelists for sharing their insights. It's clear that while achieving interoperability is a significant milestone, the journey to fully realizing the potential of health care data is ongoing. The challenges are substantial, but with the right strategies and technologies, health care organizations can transform data into a powerful asset. Thank you to our audience for your attention and participation today, and I just wish everybody a great rest of your day. Thank you very much.

[ABEL KHO]  

Thanks, everybody.

[NICHOLAI MITCHKO]  

Great. Thanks so much.

[J.B. WOGAN]

Thanks for listening to this episode of On the Evidence, the Mathematica podcast. In the show notes, I’ll include links where you can learn more about Mathematica’s Health Data Innovation Lab and watch previous webinars about artificial intelligence and data governance. If you liked this episode, please consider leaving us a rating and review wherever you listen to podcasts. To catch future episodes of the show, subscribe at mathematica.org/ontheevidence.