Subscribe to The Podcast by KevinMD. Watch on YouTube. Catch up on old episodes!

We delve into the intersection of large language models (LLMs) and health care with David Lareau, a health care executive. Since the debut of ChatGPT, the health care sector has been abuzz with anticipation and intrigue regarding the potential applications of LLMs in medicine. However, amidst the excitement lies a critical discussion about the inherent limitations and risks associated with these powerful tools. Join us as we explore the challenges of LLMs hallucinating inaccurate information, the perpetuation of biases in medical data, and the importance of integrating clinical relevancy engines to refine LLM outputs.

David Lareau is a health care executive.

He discusses the KevinMD article, “The promise and pitfalls of large language models in clinical applications.”

Our presenting sponsor is Nuance, a Microsoft company.

Together, Microsoft and Nuance are leveraging their rich digital technology and advanced AI capabilities to tackle some of health care’s biggest challenges. AI-driven technology promises to revolutionize patient and provider experiences with clinical documentation that writes itself.

The Nuance Dragon Ambient eXperience, or DAX for short, is a voice-enabled solution that automatically captures patient encounters securely and accurately at the point of care. DAX Copilot combines proven conversational and ambient AI with the most advanced generative AI in a mobile application that integrates directly with your existing workflows.

Physicians who use DAX have reported a 50 percent decrease in documentation time and a 70 percent reduction in feelings of burnout, and 85 percent of patients say their physician is more personable and conversational.

Discover AI-powered clinical documentation that writes itself. Visit https://nuance.com/daxinaction to see a 12-minute DAX Copilot demo.

VISIT SPONSOR → https://nuance.com/daxinaction

SUBSCRIBE TO THE PODCAST → https://kevinmd.com/podcast

RECOMMENDED BY KEVINMD → https://kevinmd.com/recommended

GET CME FOR THIS EPISODE → https://earnc.me/dSskf5

Powered by CMEfy. 

Transcript

Kevin Pho: Hi, and welcome to the show. Subscribe at KevinMD.com/podcast, and get CME for this episode by clicking on the CME link in the show notes. Today we welcome David Lareau. He’s a health care executive, and today’s KevinMD article is “The promise and pitfalls of large language models in clinical applications.” David, welcome to the show.

David Lareau: Thank you, Kevin, it’s great to be here.

Kevin Pho: So let’s start by briefly sharing your story and journey.

David Lareau: OK. I was in Baltimore, I had a billing company, we did RCM stuff mainly for ambulatory practices. One of my largest clients was the private internal medicine practice for many of the docs who headed up the Johns Hopkins Department of Medicine.

They were pleased with what I had done for billing. They came to me and said, we’re looking for what at the time was called an EMR, we want an EMR. And I said to them, I don’t know anything about that. And they said, well, we trust you, why don’t you see what’s out there and maybe you can at least introduce us to those people?

So we set up, for the third Wednesday of every month I would bring in an EMR vendor to my office, about four or five of the docs would come, they would look at it. And after about eight to 10 different vendors came through they said, you know, Dave, these don’t work the way we work, they don’t think the way we think, all they do is, they’re all check boxes and stuff, we don’t work that way. Thank you for your time, but, nice dinners, let’s forget about it.

So I went back to them, I said, what are you looking for? I told one of my salespeople, and maybe three or four months later he was at a conference, he comes back, he said, I met these guys from this little company, Medicomp Systems, that I’m still with. And when you describe to me what the docs were looking for, which is something that was based on what’s the patient’s problem, show me all the information I need about that problem, let me treat the patient and move on to the next one, and these guys showed it to me.

So I came down to Chantilly, drove down from Baltimore, looked at what Peter Goltra, the founder, was doing. And on the way back, this is in the mid 90s, remember those big car phones that were like bricks? I called my wife, I’ll never forget the conversation. I said, you know, Mary, I found what I want to do with the rest of my life, but it’s not what I’m doing now, sorry, it’s this.

So what I had seen was an approach to turn clinical information into data and use it to inform the provider at the point of care. I took it up to Hopkins, I showed it to these guys, they said, boy, this has real potential, needs to be built out a lot, but I think these guys are on the right track. So I convinced Peter to let me come work at Medicomp, and I’ve been here ever since.

Kevin Pho: So health IT doesn’t have the best reputation among the clinicians that I talk to on this podcast, and pretty much every clinician that I talk to. What is one misperception from the health IT side that you would like to clear up?

David Lareau: Health IT has been focused for many years on the gathering of enough information to justify a procedure and produce a bill for it. It’s almost as if the management of the patient and the ongoing management don’t matter, just what did you do today and what justification do I have to provide to bill for it.

And those are very powerful financial transactions. I mean, it’s hundreds of billions of dollars flowing through our system based on that premise. So the systems naturally have followed the money, and the money was for billing, for transactions, fee for service processing, putting information in buckets, organizing it to get paid. And the structure of these systems is still based on that premise. It’s opening up now.

But physicians don’t go into a room to create a bill, but it’s been the focus of systems to make it easier for them to do that and justify it. So that’s one of the reasons that clinicians are so frustrated. It serves the financial needs of the enterprise but not necessarily the clinical knowledge needs of the providers.

Kevin Pho: So today we’re going to talk about AI and large language models, and obviously it’s created a seismic shift within the health IT industry. Today’s KevinMD article is “The promise and pitfalls of large language models in clinical applications.” So tell us what your article is about.

David Lareau: My article is how generative AI, particularly large language models, make it easier to do certain things and more difficult to do others.

It’s now become easier with these ambient systems to produce a narrative note, documentation for a narrative note that goes into the system as free text, and then possibly you can use NLP to get some data out of it. But it’s really made the generation of text notes and the burden of documentation of text notes much easier. So now there’s more textual information going into these systems.

And if you think of a chief medical officer of a hospital or a large medical group walking down the hall, the thing I always hear from docs is, these systems are terrible, can’t you give me something that takes away some of the burden? You hear a lot about clinician burnout.

And generative AI with large language models solves one of those problems, and one of those problems is getting information, documentation, into the record. But that documentation is still computable physician or clinician oriented data, yet at best the NLP that processes that text creates these disparate sets of codes, LOINC for labs, RxNorm for meds, SNOMED or ICD, CPT, et cetera. Those still are not a computable platform. It’s just a bunch of codes built on that whole premise that what we’re actually doing is trying to capture fragments of transactions, not organize the clinical care of the patient and each of their conditions.

So generative AI and large language models make it easier to get text into a record, but they don’t make it easier to care for the patient yet.

Kevin Pho: So as a physician, why do I need to care about that? I feel that these ambient AI systems, it makes one job in my life easier, it makes my notes faster, more polished, more or less accurate after checking. Why, as a physician, should we care about what you’re saying?

David Lareau: With the switch to value based care, you’re going to be working, particularly if you’re in primary care or you accept a risk-based contract, you’re going to be held responsible financially and clinically, of course, but financially, for how well you manage the chronic conditions of those patients, particularly Medicare Advantage or other value based programs. That’s going to affect your star ratings, that’s going to affect everything.

And you still have to be able to treat a patient, say for chronic renal failure, and quickly find the key indicators in that record for chronic renal failure. And you’re not going to do that by flipping through pages and pages and pages of text. You’re going to need a usable clinical interface to filter that information diagnostically, because you’re now being held responsible not just for seeing the patient and billing for that encounter, you’re being responsible for how effectively you manage the chronic conditions of the patient. That’s what Medicare Advantage and these other value-based programs are all about.

So producing the documentation is just the first step in caring for the patient. The other is managing all of that information in the record that helps you say, is this condition under control, is it getting better, is it getting worse, what’s changing, et cetera, what do I have to do? And does my documentation meet all these ridiculous requirements for quality measures, for the management, evaluation, assessment, and treatment for Medicare Advantage, for the hierarchical condition categories? All of that is an additional burden that just generating text notes does still not solve.

Kevin Pho: So how do we get past that gap? Because generative AI certainly makes a note more polished. And in terms of what we currently do, how physicians produce the correct documentation to fulfill the requirements for value based care, what more needs to be done?

David Lareau: Well, at this point one of the common approaches is to have to do post encounter review, have somebody review it, run it through another, possibly a generative AI process, to say, hey, this note did not have a plan in it for this neuropathy, you’re required, the documentation requirements, you have to have a plan.

So in some cases the reviewer calls the doc and says, hey, you forgot to do this, can you contact the patient or do this? The reviewer cannot update the record, the provider has to do it.

So the thing that we’re trying to do is put the information in front of the provider at the point of care, which could be on the phone, could be what we’re doing right now, or could be in person, so that they can complete everything at that point in time, not have to revisit it, not have to contact the patient, and not get dinged at the end of the year because they didn’t meet the documentation requirements.

But to do that, the tool that you give the clinicians has to be able to diagnostically filter that information, process it, and present, oh wait a minute, this patient is over 65, nobody’s done a fall risk assessment on them, here, might as well do it now. You can catch it later, you can do it later, it’s better to do it while the patient’s right in front.

And I think the issue is the systems aren’t geared to do that. They’re now geared to create the note, but there are lots of other things that need to kind of close the circle on the care of the patient.

Kevin Pho: So generative AI tools really is a small first step in terms of creating what’s needed to document all that we need for value based care.

David Lareau: Well, you and I might think it’s a small first step, but for the docs it’s a big first step, because it removes one of the primary problems they have with this, which is, oh my God, I’m spending all this time on documentation and I’m still not getting anything for it. But now at least they’re not spending as much time on the documentation.

What’s going to be different is, with the adoption of the Fast Healthcare Interoperability Resources and the widespread use of that, and the pushing of that into the industry and the opening up of these systems, it will be possible to take that text, turn it into data, and make it actually usable without the burden of having to create the text by taking hours to do it.

So it’s a small first step, but it’s a big improvement in the quality of life for these docs. The next step is turning that into computable, usable, normalized, standardized information to manage a patient’s conditions. And that’s what we’ve been working on for the last 40 years here, in building those tools to do that.

Kevin Pho: So tell us where we are today in terms of the specific challenges that large language models have when it comes to ambient documentation. You talk a little about that in your article, about hallucinations, about issues with compliance. Tell us of the challenges that the technology currently faces.

David Lareau: OK. One, hallucinations, and what people are now rebranding as misinformation. That’s sort of a feature of large language models, not a bug, in that they’ll look for information until they find it. Tools are getting better and better at filtering that stuff out and identifying it.

But while the large language models are good at creating text, the requirements for reporting things like quality measures and managing those conditions are very specific. There’s very specific data elements. For example, in our engine there are over 150,000 data points in our engine that feed our algorithms for, were these quality measures met or not. Those data points are much too specific to get by relying strictly on generative AI and large language models themselves.

Getting the text in is OK, but you have to turn that and have to report very specific data points. We’ve been playing with the large language models for a little over a year to see, can we use those to get to these specific data points? And we found that we can’t. We still have to use our 10 million plus mappings between phrases and specific codes that need to be reported.

Kevin Pho: So the approach you’re taking is that you’re using some type of custom large language models to look for the data points that would satisfy requirements, and we’re not quite there yet. Is that what you’re saying?

David Lareau: We’re not quite there yet with the large language model. So we’ve built over 40 years a curated expert system that, for probably about 10,000 diagnoses, says here are the relevant concepts for that, and here’s those concepts mapped to the requirements for coding and reporting, I mentioned earlier LOINC, RxNorm, et cetera.

Let’s take the large language models, let’s run them through natural language processing, identifying and normalizing on our concept library, so that we can filter it diagnostically, process it for quality measures, process it for the requirements for value based care, and present it in a normalized view to the provider at the point of care, to say, OK, that’s great, we’ve got all this information, now these three additional things need to be done and you’re finished.

So it’s an exciting time to be doing this kind of work.

Kevin Pho: Given the speed of the evolution of AI and large language models, do you see us getting there in the foreseeable future?

David Lareau: I don’t see AI and large language models doing everything you need to do to meet all these requirements, to filter this stuff diagnostically, et cetera. Everybody says two to three years, I think it’s going to be more like five to 10, because these things are evolving, they’re evolving faster than the large language models can adapt to.

Kevin Pho: So give us your crystal ball prediction. What do you see in the next year or so in terms of the evolution of health IT and large language models?

David Lareau: I think it’s going to be assumed over the next year, 12 months from now the assumption will be, documentation will not be as much of a burden, producing documentation will not be as much of a burden as it is now.

With the adoption of FHIR, with qualified health information networks, the opening up of the systems, the forced opening up of the incumbent systems to share data, there will be more and more information, not necessarily data but information, available to providers, who will now need even better tools to filter it, present it, and process it.

So we’re moving from, we only need documentation to justify a bill, to, we now need to turn that documentation into data to track how well we’re managing all of these conditions for the patient and meet their reporting stuff.

You will start to see more and more tools become available to do that over the next 12 months. But I still think we’re a little bit away from closing the loop from documentation to data to better care. But that’s where we’re headed.

Kevin Pho: We’re talking to David Lareau. He’s a health care executive, and today’s KevinMD article is “The promise and pitfalls of large language models in clinical applications.” David, let’s end with some of your take-home messages to the KevinMD audience.

David Lareau: OK. Change is coming faster and faster, get ready for it, embrace it. But for providers, always ask, how does this help me care for my patients better? Because if it helps me care for my patients better, the patients will do better.

And we still have to improve patient health one patient at a time, and that is a system with clinicians and a single patient. What tools are you giving me that help me take care of the patient and remove the burdens from what I do? And AI and large language models and ambient stuff is the first step, but there’s many steps to come. But embrace it, get ready for it, it will help. Help is coming.

Kevin Pho: David, thank you so much for sharing your insight and perspective, and thanks again for coming on the show.

David Lareau: OK, thank you, Kevin. It was a pleasure to be here.