← Practical AI in Healthcare
Practical AI in Healthcare
S1, E13 - Part 2 of 2: Dr. Yin Ho's discussion of her new book, Rushing Headlong: Health IT’s Legacy and the Road to Responsible AI
Loading summary
Transcript47 lines
- [00:04]
A
Welcome to Practical AI in Healthcare, the podcast that cuts through the noise to spotlight real world solutions delivering real world value. From patient care to clinical research, from life sciences to patient engagement, we focus on what's truly moving the needle in healthcare. No hype, no theory, just practical insights where AI is making a true impact. Welcome aboard and let's get to it.
- [00:28]
B
Foreign.
- [00:34]
A
Hello and welcome to another edition of Practical AI and Healthcare. My name is Dr. Stephen Lapkoff and I'm here as I am every week with my partner in crime, Dr. Leon Rosenblit. Leon, how's it going? Hey, Steve.
- [00:44]
C
Doing great. Looking forward to an exciting conversation today.
- [00:48]
A
So today we are going to have part two of our conversation with Dr. Ng Ho. And if you'll recall from our last episode, Ng and I, our old colleagues from our time back at Pfizer many years ago and Inga has recently written a book. Her book is called Rushing Healthcare or Health It's Legacy and the Road to Responsible AI. And Leon, you want to take us through a little bit of a recap of the last episode and then we'll get started?
- [01:16]
C
Yeah, it was a really exciting discussion and started by talking about the history of healthcare it, which is long, tortured and somewhat dramatic. And we managed to raise Steve's blood pressure to medically unacceptable levels because, you know, he was, he was living through the PTSD episode of his career. Then, then we pivoted a little bit and talked about data and real world data and real world evidence, which is another very important application of data in healthcare and research. And I think what we're ready to do is today is sort of turn towards the last but not least section of the book and talk about AI. So, Ing, I'm going to start by just asking your general thoughts and reflections. What do you think about AI as a category? I've got, you know, scare quotes here that I'm making with my fingers in general and LLMs in particular. What is it that they're doing that other IT systems in healthcare and life sciences aren't doing well?
- [02:22]
B
These are great questions and they're hard questions to answer since this is a space that's just rapidly moving. What I can say about AI right now is it's running at such a rapid speed. We've put a lot of money behind large language models that are used in so many different ways. And most of the applications and tools have been built, have been built on these commercially available LLMs, of which it seems that every day there's another one. And the challenge with that is that there is a question around sort of quality control, around which should be the number one thing that most people should be worried about. And whether or not you want to apply an LLM to a commercially available, off the shelf LLM to something as sensitive and as personal as healthcare. And I think that this is really what the crux of some of the questions that are being played out. One of the things that it's always been very challenging is that right now it seems like the Marketplace is only LLMs, right? That unless you have the ability to create your own small language model or maybe create some version of a rag which may or may not be that helpful, it's still LLM based, you're kind of stuck. Yes, you have a choice of LLMs, you have a choice of foundational models, but the problem is you don't have any control over how they're developed and you have no control over what they're taught and what data is being used to refine those models. And hallucinations do happen. And since you don't have any control over the inputs and you don't really have control over the outputs, it does make it a little scary to say, am I willing to trust tools that for the most part will work reasonably well if it's something that's very straightforward and a code or an administration task or maybe even writing a note or something like that. But it suddenly becomes a little more concerning when we start thinking about the fact that we're going to allow something to become more part of the infrastructure within our healthcare system. And this is kind of where I really would like people to kind of pause and think about, like, not only about the large language models, but why isn't there a greater emphasis on creating what we call small language models, specialized, more proprietary models, of which there are controls around the inputs and on the training and on the data that's included in there. And that it's okay if it's not like, you know, really enormous size, because it's not about the size that matters as much as it's about the accuracy and about sort of the refinement that ends up occurring. And you can refine with smaller size models than you can with larger size models. So I do think that there's a fundamental question there that isn't really being addressed. Um, but that's a little bit my concern. But again, like I said, every day there's another model, every day there's another application tool, multiple application tools again, all being built upon the same set of LLMs. And so pretty soon we're going to end up with. I'm not really sure what we're going to end up with, but I, I am not surprised, nor am I upset about the fact that healthcare is being a little more cautious.
- [05:29]
C
Yeah, no, it's a really important focus that you're bringing to this on trust and validation. Right. So you have, you're introducing technologies that creating outputs of various kinds and we need to feel confident that they match the constraints of the problem, which includes safety. Right. Human health and privacy. Are there distinctions, I mean, I'm going to draw on something that you've highlighted in the book between different kinds of applications. Like you make a distinction between decision support versus decision control. How do you think about that? Where is our AI models safer in one context than the other?
- [06:05]
B
Well, do you know really what's funny about decision support versus decision control is it's a very human decision that sets things up in that way. It's like how much do you allow a decision to be made by something other than a person? Right. And that's really a. And it's based on that that determines whether you move from a position of supporting decisions like providing tools or whether you are actually trying to push toward decision control in which you allow more of the decision making or more of the recommendations to be gener generated not by a person. And I think that this is actually. But this is a human decision. This is a. Humans create systems. Humans are the ones who operate systems, create the rules around systems, the policies around systems. And I think that there's something that we have to also recognize about us as people is that we also are victim to automation complacency. We have, you know, we get tired, we will read a lot of things. And over time what can sometimes happen is it's easy to just hit accept. It's just easy to sort of allow whatever's coming through. Like that looks pretty, pretty good, looks pretty close. We're going to go ahead and accept it. And so it's because of these sorts of situations that really it's about safeguarding. How much do you want there to be an active decision that's being made, an active thought process that's being made by a human, by a professional, by somebody who's exercising judgment. That's where you keep things in the realm of decision support versus sort of saying, you know what, it's going to spit you out seven recommendations. You just kind of choose one of them and you know, kind of go from there and then you're now slipping a little bit more toward a decision control because maybe seven recommendations becomes two recommendations after a certain point in time. Maybe what happens is that if enough times it goes through, it suddenly just becomes one more step that just doesn't necessarily involve a human. Again, I mean, this is a little bit of kind of the concern around agentic AI, which is like how many steps will you daisy chain where you won't have a human in the mix is a little bit of down that slope around decision control. However, again, this is again a decision that can be made by people like we in the system can make these decisions about how far we allow a tool to be used and how much we use a tool's recommendations to be followed. And so I do think that that's a very human decision. The distinction between decision support and decision control tools are only tools.
- [08:29]
C
So wonderful. I'm going to hand it over to Steve because I know he's itching to ask a question.
- [08:36]
A
So, you know, one of the things that I find challenging about these models, and I think you alluded to it, but I'm going to try to sharpen the conversation on this one a little bit, is we don't necessarily know and you know, how these systems are being trained, especially at the general level. We don't know what information is being used as the fodder for helping these systems make the decisions. And, and the challenge here, I think is that if you, if you're taking information from the emr, just like secondary data usage is done and you're using it now to train, but it was never built to train. So you're actually taking what could be in theory mediocre notes to create the training set. It's going to train on mediocre information. And therefore the, the recommendations that it could make might not be the sharpest that they might be. And that leads to like, you know, you mentioned small versus large use case, sorry, training models or. And John Glasser, who we had on here a few weeks ago, alluded to this as well. And I'm wondering if you could comment on and go a little deeper on this one because I think this issue is really at the crux of the credibility, the explainability and ultimately people trusting that these models will provide decision support that's actually credible.
- [10:02]
B
Yeah, I'm happy to. I mean, there's a number of things in there. I mean, we'll start with the whole small versus large, right? I mean really it has a lot to do with what, how much the size of your model, the amount of input that you use it, how much do you train in it? If you use a large language model, it has been, it's been trained on the Internet, it's been trained on a bunch of different sort of inputs that there has never been a vetting of the source or there vetting of a quality of the data itself. And so in some sense you kind of hope that maybe the ocean of data is enough to kind of take out some of the worst issues around it. But generally speaking, you just don't have control over what is being assumed as being high enough quality to support the model. But if you were talking about something a little bit smaller, where you're talking about a small language model, where it's not necessarily running in the cloud, but running more on edge applications and things like that, in some interesting way, physically you actually have model that's within your control. And because of that you can also control what it is learning from. The interesting thing about medical data is the way it's captured, right? So if you think about clinical data, there's variability, as you pointed out, there's mediocre notes, there are great notes, there are lots of different ways of how people can write notes. And some people write it very well and some people don't write so well. And the challenge with that is. But what interesting about each individual is each individual has kind of like a style. And so what's fascinating is that if you can control this parameters around a model, you can start to understand how any individual might be saying something. In the same way you can start to think about groups of people and how they may be saying something. And that gives you the ability to really think about how you refine a model and just sort of saying that certain people may use these types of syntax or these types of terminology and other people may be using this, but they may be meaning something very similar. And I think you have more of the controls around the inputs. We just don't fully get how the LLMs work because there's this constant tension between the sort of authoritative way of training a model versus the creative of a model, right. You know, in some sense these are generative, right? So then in some sense, if you're tuned to be more generative, it's going to come up with all sorts of random things, right? And it's to be expected. You wanted it for to be creative, but the challenge is in something like medicine, you may not want it to be so creative, you would actually want it to be a little bit more consistent. And so I think that these kinds of dialing up and Dialing down is easier to control within a model that's a size and also within the sort of governance and of the sort of the data provenance within a system. And I think that this is kind of where we should be going. But you know, but that does involve a level of investment. It involves a level investment of creating that model that things will be based on that is going to require energy and require also like training sets and, and also a lot of building, which is not necessarily something that's that affordable. But honestly I would say is something worth exploring.
- [13:14]
A
You know, it makes me wonder about. There's. There's a company that we had visit with us at the DCI Network recently, Open Evidence. I'm sure you're familiar with Open Evidence. Yes, they. Mine. Mine is maybe not the right word, but they provide a large language model that sits on top of basically most of the current medical literature. And when you pose it a query, I've used it myself, it's quite good. But when you pose it a query, it's drawing off of these papers and it's trying to find the references that answer the question at hand, which means that how that's been trained to do that. And they went through some very significant iterations with their tool. Initially they had problems with hallucinations as well. They've seemed to have ironed those out. But it really gets to the question of how do we use these in intelligent ways? How do we pick and choose the right models for the right tasks at the right time? It's kind of like the same question we used to have with medicines, right? Right drug, right time, right place.
- [14:17]
B
I mean, I think that's just the challenge, right? We don't know enough right at this point in time, but we're actually willing to accept an output. And I think we have to remember how does an LLM or how do any of these sort of generative AI even work? They work by guessing, right? It's about guessing the next word or the next sentence or the next sort of thing that you're about to ask. That's really what it's all about. So if you limit the universe of how many things you can ask and you limit sort of like the training models around it, you're going to be a little more accurate and that's just kind of natural, that's sort of heuristically going to work that way. But also just simply because you have limited what's possible the error parent choices in there. And I think this is a little bit of kind of the part that there isn't maybe as much discussion and thinking about, which is like, well, what are the right parameters? Right? What is the level, how much more interest in there? Is there a vetting process that happens pre, pre training data or is there? No, but as long as it's within the system, we'll count it within there. I mean, I think these are things that are not necessarily being fully discussed because again, most of the tools that are still being sold are still being based on, you know, foundational large language models that exist commercially.
- [15:30]
C
I just wanted to drill into your technical detail that maybe will be illuminating. But the one way, so one way to think about the problem of increasing validity within a domain is the way you've described and actually is consistent with Ray John Glasser described it, which is smaller. Models that are more domain specific are going to make up stuff less, which is, I think, completely right, consistent with experience. There's another way to think about it, which is, look, an LLM is just a layer and a stack. Its goal is to derive a latent semantic space that represents language, not a domain of expertise. And you need something else that represents a domain of expertise. So I can imagine in architecture we have LLMs for language translation and then you have structured representations of causal models that are specific to an area of medicine or an area of physics, or an area of real world evidence that then verify what has been said against some kind of a structured data representation, a much narrower domain. Is that consistent with the way you're thinking about it? I want to make sure that we're not only focusing on making models smaller, but we also ask, do we just need a different stack? Right. Is there?
- [16:44]
B
I mean, part of it is a stack, but part of it goes back to something Steve brought up earlier, which is sort of the quality of the data. I mean, one of the reasons, and in the last episode we talked a little bit about why there's such a workaround to develop real world evidence. Because the data that exists in clinical care, the way it's captured, isn't structured. And because it isn't structured and because it's inconsistent, it requires a lot more effort to determine whether what actually did happen to abstract actual meaning from it in order to then run research on it. It's a similar problem here. The quality of the data is that it was captured for one purpose and it wasn't even necessarily captured equally the same for that one purpose. And so to them, think about sort of like, should I just simply use this as a way to. Because it happens to be one system and use it to train? Probably not. You're still probably going to have to do something with that data. And kind of one of the other challenges, and I think this is something that we talked about in the last episode, is because so much of the data that was considered valuable was the data that could be coded into a billing structure. You naturally have more structured information that exists around billing concerns and a billing mentality that's not the same as the clinical observation that might be captured in an unstructured and inconsistent and very idiosyncratic way of which you're going to try and train a model from. You almost can't do that. You do require utilizing models or utilizing methods to abstract the meaning of this data and then take that output and train. So it's kind of like there's still one more step. One of the areas that I'm pretty excited about is that there has been this new movement toward abstracting meaning from the unstructured notes that exist inside of electronic health records. But it's done on a question by question basis because there really isn't any other way to do it. You can't go back and say, I'm going to restructure everything for every possible possibility of every question that could be asked. You just can't do that. So it's still driven question by question. But what's interesting is question by question is that today we have the ability to train a model on what we're looking for and start using human abstractors to train a model or a tool on how to at least organize the information in such a way that we can start looking for commonalities or look for like groupings of patients, or looking for groupings of symptoms or grouping of things and suddenly recognize that we're actually talking about the same. This kind of abstraction capability is today done both in a human way. That's the only way. You kind of make sure that what you're reading means something, but also the only scale part of it is that you're basically using humans running models that can run through lots and lots of fields or lots of pieces of information, and then trying to create what we call labels and match it up against the labels. That's really what's happening today that I think is really exciting because that's starting to say we care enough about the. The quality of figuring out what the quality of this data means in order to then train. The problem is that the only people asking for the abstraction are just solving for a specific question. They're not or a specific research question. So it's still one off. You could say that maybe one way to start to think about this and a very responsible way for a health system or anyone to start to do it, is could we start to think about the most common questions? Could we start to think about the most common diseases? Could we start to think about the research questions that almost everyone likes to ask or wants to ask? And could we start to think about what we would do to abstract this data so that it could actually answer that and use that as the input for a new model? It would basically mean we need to do something called abstract the meaning out of the data before we actually use feed it in as any kind of structured data set that can be used for training.
- [20:47]
A
So ing, I want to pivot a little bit. You know, in the first part of your book, you know, you, you, you raised so many of the things that kind of went sideways as we both lived through that world. And, you know, you've raised the question in the second part of the book, basically, how do we avoid Groundhog's Day here? What are the kinds of things we could do to avoid some of those challenges that we, you know, that you and I have both lived through, that we've watched at, and to Leon's point, makes my blood pressure boil. Makes my blood boil or my blood pressure go high. Mixing metaphors. But you get the idea. How do we avoid these pitfalls from your perspective? I've got some ideas, but your book was enunciated it really nicely.
- [21:29]
B
I mean, I think part of the challenge right now is if we just leave things the way things are going, then it'll naturally just continue forward. The fragmentation will continue. Because, to be honest, the easiest place to apply AI at the moment is around administration, and because administration is the burden that has been that has sort of caused us all these issues in the first place with healthcare and with the healthcare IT system. If AI can be applied in that way, whether that's basically dealing with insurance claims or whether that's dealing with sort of paperwork, things like that, the problem is it just accelerates it. Now, that's the worry. The worry is that something that already was pretty burdensome and very onerous is about to. Is getting accelerated. Right? And that's because AI, it's easier for AI to run through claims. It's easier for AI to run through things that are much more clever and very. A little bit more predictable and things like that. And that's the worry, is that the money will all go in that direction, revenue cycle management will be the major area. All of the money will go in that direction to drive AI use cases. And this is where almost all of sort of the tool makers and everybody is going to go running toward. And that is kind of what you're talking about, Steve, which is basically, that's your groundhog's day, that we're just going to do it again and we're just going to do it faster, right? Or maybe do it at a bigger volume. Where I think we stand a chance to sort of not have that happen again is two things. One, I think there is a better awareness, or at least raising the awareness of where clinical data and clinical research are connected. Like, where did it. What is it that you want from that data in order to actually do research? I think that that sort of link has not been well understood or that gap has not been well understood. That's part of the reason it keeps getting bigger. But I think that this is actually a place where, similarly, as I mentioned about sort of abstraction capabilities, if you can start to derive more meaning from what's going on in a. In a faster volume, a faster and larger volume, you may also be able to demonstrate much quicker sort of links between care that is observed in one place and research that is basically being conducted on that care data. I think that's one aspect of it. Okay? So that's just putting the energy and effort in that direction. And I think the second part of this is to also remember that the people who are creating these new tools have traditionally been engineering teams. But one of the greatest things right now with kind of with AI is that people who don't have a technical background also can start to go pretty far. Right. They can already start to build technology without necessarily relying on large, large teams of engineers. And this is when something else happens, you sort of shortcut the domain experience directly into the product. Right. And I do think that there's an argument to be made that if there's the willingness to back this, you could start to create companies, you could start to create offerings that are new. They're not based on the old. Right. They're not based on an infrastructure of the past. They are. Maybe they're generated to do research. Maybe they seem like a parallel system, but they're probably faster to spin up. They're probably a little more oriented toward a better sort of care research purpose. And you may have a chance of basically driving forward an awareness and an understanding that data can be used in multiple ways, can be collected in multiple ways, and that the current system that we have today doesn't actually help fuel our ability to progress from an innovation point of view. It only fuels our ability to run more transactions at a rapid pace. I think that the fact that you can create the alternative makes a huge difference. And I think that raising the awareness, I think people are starting to become more aware about the fact that there is this linkage is going to start driving it to be maybe, maybe, you know, maybe this time a little bit.
- [25:31]
A
And boy, do I hope you're right about that. It's, it's, you know, I remember in 2004, I met David Braylor for the first time at one of the conferences. We're in the basement of the Renaissance in Washington and we're on the escalator. I remember asking him, how are we going to be able to tell that we've made a dent in this whole space in, you know, 10 or 20 years? And he looked at me and he says, I'm not sure we're going to be able to measure that the way you think. And he went on, so, you know, so hopefully, Ying, you're right. Let me hand it back to Leon. Go ahead, Leon.
- [26:02]
C
No, thanks, Steve. And I really think you make a powerful point about how the relative ease of automating standardized business processes, and this is a theme that's come up in our conferences, in our podcast episodes. The mundane often is easier to automate. And the reasons, I think are the ones you mentioned, the standard inputs, well defined processes, standard outputs.
- [26:29]
A
Right.
- [26:30]
C
We know how automation works in those domains. And it's the difficult stuff, you know, creating meaning from data, you know, figuring out how to treat a patient, delivering the right course of action, that requires human judgment. And we don't know how to automate it. I think there's some excitement around it. I mean, you've talked a little bit about the successes, I think, in automating billing, and I think they are very visible and, you know, I don't want to underestimate them. Can you, since we are focusing this podcast is practical AI in healthcare, can you. What else comes to mind in terms of big impacts that you either see or anticipate as proximate that AI tools can make in healthcare and in life sciences and real world evidence.
- [27:11]
B
I mean, one of the areas I was really excited when it first came out was Ambient Scribe or basically AI Scribe. Okay. Which is really just a listening device. Right. It's a listening device that basically like records a conversation and then just transcribes it almost immediately. And sometimes Also just sort of converts it into a note. And that's the thing that has gotten most physicians and clinicians excited because they feel like, wow, you've just saved us a lot of time since, you know, two thirds of our day is spent documenting anyway. You know, you just knock that out of the game. But what's starting to creep up based on that is the fact that there's some challenges with it, which is that no recording device can capture every nuance that exists inside of a. Inside of an encounter. Right. Sometimes there are more than one person in the room. And so sometimes what's being captured is it's hard to ascribe, you know, who said what and what context they were using with it. Sometimes physicians ask questions because they're not trying to get the answer direct answers. Maybe they're watching for body language or they're watching for the way someone thinks about an answer, or maybe their hesitation, or maybe this is somebody who's starting to suffer cognitive decline. It's takes longer to answer a question, whatever it may be. There are all these subtleties in behavior, in the way humans are, that someone is observing, but they're using language as a way to get access to seeing, to eliciting these types of responses. Problem is, if you use a listening device, it doesn't capture all of that. Right. It will just. It will. It won't keep the pauses. It won't do any of that. It'll just sort of summarize it. And perhaps the one thing that's the most unhelpful about is sometimes it tries to be a little too clever and summarizes for you. That's a danger. Your point? Because maybe you don't want the summary. You have documented a series of observations. You didn't want someone to just sort of say, oh, but patient's mother is doing great and has been living in a nursing home and whatever. And therefore, you know, we had a pleasant conversation about that. I mean, you don't want that kind of a summary. What you want is perhaps all the questions and the answers. So what you'll find is that people who are very astute observers end up losing time. They go back and they, they correct the AI, they correct the out output because they're like, no, no, no, that's not what I meant. No, no, no, that's what I, I was looking for. No, no, no, whatever. And then they go back in, and now they're spending time worrying, what else did you miss? What else did I not miss? So I do think that there are like, it's natural to want to sort of like, attack a problem by simply saying, hey, I'll. I'll just make the burden go down faster because I'm going to give you an assistant. That's a large way AI ascribed is thought of. It's an assistant. Right. But we also have to remember what is it that the physician is really wanting to do? The physician wants accuracy captured. They want their thought process put down. They don't want to have the burden of actually having to coat that information later so that it can get billed. They don't want to have to spend the time making sure that they've checked off every box necessary in order to, like, satisfy some regulatory metrics. What they want to be able to do is capture the information in real time and think about it and then actually be able to kind of go forward with it. Right. They don't want anything sort of impeding that. But we've not developed ambient scribe with that mentality of what does the physician need for their workflow? Or what is the patient going to need in terms of information? We have simply looked at it as like, oh, you have a time problem because you'd have to have, like, these notes scribed and like, and it's taking up two thirds of your life. So what I'm going to do is put you, give you a device that will at least, like, take the burden off by at least writing it for you. And I think, again, this is sort of a misunderstanding around work. Okay. And this is where it's easy to want to put forward an AI sort of an automation sort of solution, because it really does. There is a real burden. It does. It does solve some of that burden. But what we seem to continue to forget over and over again in health, it. And so back to your point, Steve, about sort of repeat the story again and again, is that we seem to always forget that we don't pay attention to workflow. We don't think about what is it someone's trying to do. You know, we only think in terms of what they do. Okay? And we tend to replace with what they do. We take the action and we replace it with the tech. Maybe the tech's better, maybe the tech's not worse. Doesn't really matter. The point is we're just replacing activities because all we measure is activities. We have gotten so accustomed to thinking in those terms, we forgot that there's something that's a little bit more of an interaction that we're trying to support. And that I think is maybe A long winded way of sort of answering your question that yes, for every way that there's a new solution that does seem to be solving one problem, we have to remember that it may only be solving a surface problem. It's not really solving the underlying sort of workflow mismatch problem, you know, so
- [32:04]
C
one, you know, if one way to approach the problem that you're describing would be with an architecture that learns from physician input, right, Is to say, oh, hey, Dr. Smith keeps correcting this thing and I am going to learn from that. Unfortunately, if you think of LLMs as the unitary thing, but that layer just doesn't have an episodic memory. So again, I want to come back to this concept of we're just missing some cognitive functions, man. We have a system with static weights that represent the way reality is and we keep running it over the same problem. Whereas what we are looking for is a system that will learn and converge on the value delivery for a particular use case in a particular setting. If you want to think about this in a context of patient empowerment, another area that Steve and I are really interested in is how do we help the patient contribute to their own care. Are you seeing anything in that area where AI is actually has promise or is advancing things?
- [33:11]
B
Again, I think it comes down to who are the vendors and what are they trying to do. Right. You know, in some sense I have seen some of it, which is really great. But you know, what they're simply doing is they're doing something that you would thought we would have solved a long time ago, which is that people, there are companies out there using AI and using standards to just assemble the patient record for the patient. You know, that's actually, strangely enough, the hardest thing for a patient today is still the fact that they cannot get all of their data. In theory, they can get all of their data that is different from the reality on the road. Right. You know, in theory they have rights to their data, in theory, that there's nothing you can say that they can't have technically access to or that they can't and they can't correct that they have the rights to correct. All of the laws have been enshrined to allow them to have that. But that is so different from impractical. Because if you are a patient and you have some kind of disease or some kind of, I'm sorry, some kind of very complex condition that ends up with a lot of data that gets included with it with films and, you know, lots of tests and stuff like that, your Ability to go get that information is really difficult. First of all, you may stay in one system, so you have to go around and go to different systems. Second of all, most of the images and stuff like that are still getting burned on a DVD somewhere that you have to go down to some sub basement and wait around for it and then you get to hand carry it to the next place. Lucky, you know, right?
- [34:35]
A
Groundhog Day again, Ben. Groundhog Day again.
- [34:38]
B
Right. You know, you're putting it all together yourself. And so really right now you want to talk about what patient empowerment is right now patient empowerment is the fact that there are companies now who are trying to do this on behalf of patients. Right. They're just literally assembling their data together. That's just step one, right. And then step two is like, okay, can I put some context around it? We're not even at the context level. We're still at the, let's see if we can like shorten this time frame of you getting all your data in one place. So I do think that a lot of times when people talk about like AI should be able to empower patients. It's only going to empower them if we allow, if we create the ability for them to actually access their own data to then have an actual say in their own care. But when we can't even give them that right now, then it's almost like it's theoretical.
- [35:20]
A
Yeah.
- [35:20]
C
I love, love your realistic assessment of the real, you know, day to day constraints. Right. You know, if patients, if patient empowerment looks like we give them a nice banker's box to carry all their charts around, right. They get to, it's like, you know, we're, we're not ready to talk about automation and cognitive enhancement. I'm, you know, before I hand it back to Steve, you do make a point in a book I wanted to elevate about how technology could either entrench existing players or accelerate disruption. I'm wondering if in this, in, in this area of, you know, patient empowerment and generally AI healthcare application, do you, you know, how do you see, see this playing out? Do you think this is going to help the epics or do you think this is going to help the scrappy new startups or some other kind of party?
- [36:11]
B
Well, I am on the side of the scrappy side, scrappy startups, to be honest, because I honestly believe that they're at least trying to solve a problem.
- [36:18]
C
You don't support the Empire and Darth Vader.
- [36:20]
A
Come on, you got it.
- [36:21]
B
Don't. But the Empire unfortunately, is going to continue to exist because at the moment they have the money to be able to do multiple things. One, they can continue to pay. Whatever regulatory requirements get brought up, they can meet those requirements. It's actually not that expensive, not so hard for them. It's much harder on a smaller company to be able to some of that certification. Even though you can understand the thinking behind certification, it still creates, in practicality a financial burden, a burden that also small companies can't necessarily meet. Most small companies are sort of sometimes building what they would call point solutions. Very difficult when the consul is over here. So I have to say that in some sense the problem is at the moment is that the entrenchment is not going anywhere, right? It's still going to be there. The financial and the regulatory all favor that. But the part that gives me hope, honestly, is the fact that it is possible to build alternative systems. It's possible to build other ways of doing what really need to be done. And the question is, is, will there be enough money that will go in that direction? Will there be enough of a movement of people wanting to have that in order? Because they understand now that their own data has a direct relationship to research, or their own data has a direct relationship, their own empowerment or their own ability to basically start to gain the wisdom of the professionals is possible for them. That's the only way you start to drive perhaps a parallel universe. The nice thing is you can in theory rebuild that faster than you could before, but the entire financial payment incentive system is still going to stick with the current incumbents. And I think that's part of the problem. Most recently there has come out with the new ways of paying the physicians and paying the hospitals. And this is a new CMS rule that has just come out. And even in this rule where they're trying to talk about basically decreasing the sort of variability around low, low back pain and also around heart failure and trying to even things out in terms of utilization of care. At the end of the day, if you read it all the way down, the entire documentation burden is still going to come down to the physician. The entire risk thing is going to come down to physician. And what's going to happen, all of your large systems are just going to try and make it easier for them to do more documentation, for them to do more of this, and so they will get even more cemented into the there. So I do have to admit it's going to take a bit of a hard pivot, but maybe it's about alternative financing Maybe it's about an alternative system. Maybe it's just really about a group of physicians getting together and rage building the next ehr.
- [39:07]
C
Yours are powerful for you just called it rage building.
- [39:12]
A
Rage building. I love it.
- [39:15]
C
That's our next hackathon, man. Well then, Rick, rage building. That's it.
- [39:20]
A
So wow, I love that term in. We're going to start to wrap up in a sec. But before we do, I thought, you know, your book was so insightful and it brought back so many memories and you've got a point of view that I think I hope that our listeners grab a copy of your book. I think you mentioned it just came out on Amazon a day or two ago.
- [39:41]
B
Yes, it did, just yesterday.
- [39:43]
A
Just yesterday. That's what I was talking.
- [39:44]
B
Happy day.
- [39:45]
A
So that's a great timing. But let's turn our eyes to the future and let's assume that. Let's pray or assume or pray that we can make some pivots here and go in different directions. What do you think the biggest different direction we might see? Let's say we revisit you and we contact you in five years from now. What's the world look like from your perspective, from Ingo's perspective? What's changed to the positive and maybe what's not changed to the positive? What do you think?
- [40:13]
B
See, I mean the most positive I can think of is I think that the opacity between where people's own experiences and the research that comes late is never been well understood. I know that that sounds. It's something. It seems to be a theme I harp on a lot. But I honestly believe it's because we've never understood how one links to the other, that there hasn't been a concerted effort to make sure that that exists. What I see in the next five years or what I'd like to see, I should say maybe I can't say. I can't predict the future the way things right now. You could almost argue that we might end up in the same place that we were, you know, now or that we're still in. But what I'm really hoping is that there are actually alternatives. That there are the fact that people can start thinking about their data in a very different way than perhaps they have in the past. Right. It's not just something that's collected and used to sort of solve their problem. It also has other value ascribed to it and that there are other ways to be able to apply that value for it to be whether it's to participate in trials Whether it's to donate your data into other ways, whether it's about ensuring that there are cohorts that you can be part of. I think that the fluidity of this data economy only works if you people actually have access to their own information, if they actually understand the different contexts of their information, that it actually has meaning has been derived from it, and that it isn't just holed up into the incumbents. Right. I think this is a large part of it. I think there's a liberation aspect of it. But it has to go somewhere, right? Right now, at the moment, there is no place for it to go. When you talked about the bankruptcy box, I honestly, that's, I'm laughing. But at the same time, that's exactly what exists. Right. You know, even if people have access to their information, they're printing it out, right? And they're sticking it in a box. It's still in this way. It's still. There's something not helpful about that. But if we could build up, if there could be an industry of tools and ability to help people start to make sense of information, to store it, for crying out loud, to do things with it. I think there's something different. I don't think this is the old personal health records of the past and which Google holds it for you or Apple holds it for you or anything else like that. I don't think it's about that. I think at this point in time, we may be able to move in a direction where perhaps the ability to control and access and to be able to share might actually be possible. I think that's the world we got to move toward.
- [42:33]
A
Well, you know, from your mouth to God's ears are to at least the, the, the, the folks who are out there, the scrappy young folks who are out there. We've had a couple of those young CEOs on our show show in the past couple of months. We've had folks from avo, a guy named Yair Saperstein, and we had a guy named Derek. Oh, I'm blanking on Derek's last name. Oh, Leon, what's Derek's last name? You remember uphand.
- [42:58]
C
I know it's Glass Health, but I
- [42:59]
A
forgot it is Glass Health. I forget Derek's last name. But these young guys are trying very hard to create the future that you're mentioning. And I hope you're right. I hope that they are able to be successful. I hope they don't fall into the same, same, into the same potholes that our generation fell into. Ing, I want to just wrap up and just mention one more time. The book is called Rushing Health. It's Legacy and the Road to Responsible AI. And again, it came out on Amazon just yesterday. So congratulations for that. It's been a long, long effort. I know. Ing, do you have any final thoughts or final words today before we call it a day? Day?
- [43:39]
B
I mean, I, I think people should be having conversations now about AI. And I almost, honestly, I would tell every single physician and every clinician, get smart about this, because right now, the last time any health IT infrastructure got built out, any. The last time there was any kind of movement in a direction of creating any kind of an infrastructure or system, the physicians were left out, the clinicians were left out, patients. It's hard for them to get involved, except for as a public. But physicians, there really is no excuse. You know, fool me once, you know, shame on you, shame twice, shame on me. There's a little bit of that this time. Come into the game, have an opinion. You know, it doesn't have to be all consensus. The point is you want to aisle it up and you want to be able to say that, no, no, no, we're not going to sit here and just accept tech that comes to us. We're going to have a hand in it. And we're not. We all don't know the answers, but at least we're trying to drive in a direction that's supporting tech. Patience.
- [44:36]
A
Well, I, I couldn't agree with you more. And I think that the, you know, I think everybody's got to wise up and figure this out. The, the, the, the notion that we could repeat the past so readily is, Is, Is so frightening to me. That wouldn't be ra. It would enrage me, frankly, if that happened yet again while I was still in the game, so to speak week. Oh, I found his name. It was Derek. Paul was the CEO of Glass Health. I just felt very obliged to mention him by name. Ing, it's been a pleasure to have you on the show today and the last two episodes, frankly, you bring a fresh perspective on both the past and the future. Our hope is that your book is quite successful, and we just want to thank you for being on the show, giving us your perspective and validating some of the things we've been hearing. We've had a lot of guests on the show so far, and you're echoing a lot and you're in good company. When people echo John Glasser, I think that that's. He's one of the most educated folks in this space, and he has been for close to 40 years, or maybe more than 40 years. And you're echoing a lot of the things he was saying over the course of this conversation. I want to say thanks to our listeners, and we hope that you found this to be useful, a good use of your time for the last two episodes. And we're going to say thank you all. This has been Practical AI in Healthcare and for Leon Rosenblit and for Ing Ho. This is Steve Lapkoff signing off, and we'll see you again next time on our next episode. Thank you so much. Thank you for joining us this week on Practical AI and Help Healthcare. If you're ready to go beyond buzzwords and hype and explore how AI is truly transforming healthcare, stay tuned for more conversations that get us to what works. Until next time, stay practice.