Latest / CareTalk: Healthcare. Unfiltered. / Improving Patient Support with Conversational AI

Transcript

Improving Patient Support with Conversational AI · CareTalk: Healthcare. Unfiltered.

• 0:00
  Today's guest is a lifelong lover and solver of puzzles and problems.
• 0:04
  He loves the complexities of data and real world impact, but can the AI chat bots from his company, Hyro, solve patient frustrations in interacting with the healthcare system?
• 0:14
  That's a tall order.
• 0:20
  Welcome to CareTalk, America's home for incisive debate about healthcare, business, and policy.
• 0:25
  I'm David Williams, president of Health Business Group.
• 0:28
  And I'm John Driscoll, senior advisor at Walgreens.
• 0:31
  Well, today's guest is Israel Krush, co founder of HIRO, whose mission is to revolutionize conversational AI.
• 0:38
  Meanwhile, join the fast growing CareTalk community on LinkedIn, where you can dig deep into healthcare, business, and policy topics, access CareTalk content, and interact with the hosts and our guests.
• 0:48
  Or they're chatbots.
• 0:50
  And please be sure to leave us a rating on Apple or Spotify while you're at it.
• 0:54
  David, why are we talking about chatbots and how does it relate to healthcare?
• 0:58
  You know, I was intrigued when I was reading about Israel's company and he had, uh, he, he founded a company, we know he's a scholar because he founded a company with hypotheses, uh, which is something we would, uh, we would love to do.
• 1:09
  And it was really about chatbots and, and voice assistants being the future of human computer interaction and the traditional ways of doing it.
• 1:16
  We're no good.
• 1:17
  That was back in 2018.
• 1:18
  We're in 2024 now.
• 1:20
  So I want to know, like, how's it panned out?
• 1:22
  Yeah, absolutely.
• 1:23
  Uh, so you're right.
• 1:24
  I think that when we started the company, we had two hypotheses.
• 1:28
  One is that, uh, natural language interfaces, so chatbots and voice AI systems will be the dominant interfaces when it comes to human computer interaction.
• 1:36
  I think that today, you know, if the buzz around Chat GPT and large language models.
• 1:41
  It's it's a matter of time.
• 1:42
  Of course, we're going to talk with technology.
• 1:44
  Of course, we're going to type into technology and the technology computers, phones would be able to understand us in our own human language.
• 1:54
  So this hypothesis today, I think, is less of an hypothesis.
• 1:58
  This is a reality with regards to large enterprises, um, in the healthcare ecosystem, but generally speaking, adopting Um, chatbots and voice AI assistants and being able to deploy and maintain them.
• 2:11
  And that's the second hypothesis.
• 2:13
  We believe it's going to be very hard for them to do so.
• 2:16
  And the thing that you see that even post at the buzz, we're like, what, 18 months now, um, after, uh, call it like the media attention that we got with the release of JGPT, and you still see very few production ready chatbots.
• 2:32
  Patient facing or end user facing, facing chatbots and voice assistants that are available.
• 2:38
  So the hurdles to actually deploy and maintain those are still very hard and that's exactly what, uh, Uh, we try to simplify and, and help.
• 2:47
  So Israel, if you think about the consumer approach of, of whether it's claims or answering basic questions, chatbots are not only increasingly more pervasive, in some cases they're more sympathetic than the actual people on the phones, and yet there have been real hurdles and roadblocks.
• 3:10
  And, and sort of, uh, a high level of, of concern about engaging directly with patients around clinical issues.
• 3:17
  What's the big difference?
• 3:19
  Language is language.
• 3:20
  Knowledge is knowledge.
• 3:21
  What's, what's different?
• 3:22
  So, uh, yeah, absolutely.
• 3:23
  To your first comment, John, uh, you know, chatbots or voice AI assistants, they never get tired.
• 3:28
  They never get frustrated.
• 3:30
  They can speak with you 24 seven, whenever you call them and they'll have the same tone and they can be emphatetic these days also with some new tech.
• 3:40
  To your point, Watson, the question, even before getting into clinical use cases, let's just talk about administrative use cases within healthcare.
• 3:48
  I think one of the main things to consider here is the cost of mistake.
• 3:53
  So let's think about the chatbot that helps you buy, I don't know, clothes, t shirts.
• 3:59
  So if you request a blue t shirt and you get a purple one, you might be upset, but it's not the end of the world.
• 4:06
  The cost of mistake is not that big.
• 4:08
  So I'm going to be upset with my purple shirt, but I'm not going to be upset if they take out, like, my kidney instead of my liver, am I?
• 4:14
  C correct.
• 4:15
  No, absolutely correct.
• 4:16
  Maybe this is, uh, you know, um, uh, a very extreme example, but even scheduling an appointment, um, in a time slot that doesn't exist, And you getting into the clinic to find out that it's closed and the physician that you wanted isn't there, that's like much more frustrating and opens a lot of lawsuits against, uh, the healthcare facilities.
• 4:37
  So the cost of mistake is probably like a one big factor of why you don't see them a lot in production at these days.
• 4:45
  So, I mean, I, I get, you know, these voice bots, they always, I mean, they don't get tired, but also what happens is they, sometimes they try to be empathetic and they'll, and they'll say, you know, they'll say, say your whatever and I'll say it and they said, could you say it again or, you know, please say this and I was like, Oh, I'm sorry.
• 5:00
  I can't understand.
• 5:00
  It's like, shut up.
• 5:02
  And at the same time, I do find that the, you know, the chat GPT is particularly, it can be very empathetic.
• 5:08
  And in fact, The bar relative to like a healthcare administrator or a doctor isn't that high.
• 5:14
  So it can easily be more empathetic than them.
• 5:16
  Does it, does it end up playing a role?
• 5:18
  Can it play a role in, you know, personalizing care and actually improving engagement, these possibilities?
• 5:24
  David, maybe.
• 5:25
  The way to think about it in Israel, to frame it a little bit more broadly, the large language models, which kind of drive chat GPT, maybe you could kind of explain what they are and how they feed chat GPT to sort of contextualize David's question.
• 5:40
  Yeah, absolutely.
• 5:41
  So large language models, at least today, are very much statistical tools.
• 5:46
  So, um, they don't have deep reasoning as we as humans have.
• 5:50
  So when you ask me a question, I actually think about what I want to answer to you versus like a large language model will think about what's the next word in the highest probability.
• 5:59
  And all of a sudden, you know, because it was trained on so much data, we got like to this magical point in which the sentences makes a lot of sense.
• 6:08
  And then to your point, we can align them to be sympathetic and we can align them to not break certain rules.
• 6:14
  Um, And they think that while reading these answers or hearing these answers, it all makes sense.
• 6:23
  You know, there's a big hallucination issues with them today.
• 6:27
  We talked about custom mistake.
• 6:28
  So the AI will sound smart and sympathetic and will schedule an appointment for you, but with a physician that doesn't even exist.
• 6:37
  So what's, what's, what's the difference between David's hallucination and Chad GPT's hallucination?
• 6:44
  Like, how do you define it?
• 6:46
  It's a good question.
• 6:47
  I think that, um, I don't know, like regarding David's hallucinations, but, uh, to give you like a good analogy with human beings is like a six year old when you'll ask a six year old to do something.
• 6:58
  Um, some of it would be based on, on ground truth.
• 7:02
  And sometimes the six year old would be shy to say, you know what, I just don't know.
• 7:07
  So we'll make up something, and that's how you need to think about it, uh, with large language models.
• 7:13
  They don't know that they, they lie, actually, um, if you continue the conversation and say, you know what, this physician is actually, uh, not existence or like disappointment is not in their schedule, they will apologize and say, you know what, you were actually correct.
• 7:27
  Oh, that's very different than David.
• 7:28
  David never apologizes.
• 7:30
  So the hallucinations are a problem, and you talk about high rope controlling hallucinations.
• 7:35
  It's due to like, just like.
• 7:36
  Put the thing in a, in a headlock or what does it mean to control the hallucinations?
• 7:40
  Yeah, absolutely.
• 7:41
  So in healthcare, uh, we talked about the cost of mistake.
• 7:43
  And I think another thing that we didn't talk about is the fact that like, this is a regulated industry, very much highly regulated industry, which is part of why the cost of mistake is so big.
• 7:53
  So now, uh, I'd say the new buzzword, especially like in regulated industries that want to deploy AI is responsible AI or in, in our area, responsible conversational AI.
• 8:05
  And I think that.
• 8:06
  You know, I am allergic to buzzwords.
• 8:07
  So I like to say, like, what does it mean for me?
• 8:11
  And when I think about responsible conversational AI, I think about three main pillars, which are explainability, control and compliance.
• 8:19
  So explainability, why the AI replied the answer that it replied.
• 8:23
  And again, without getting too geeky, Uh, you know, large language models are large machine learning models, which are large black boxes, inputs, outputs.
• 8:33
  You don't really understand what's happening inside.
• 8:35
  So how can we make it, uh, more explainable to why did I recommend this physician?
• 8:41
  Why did I gave you, uh, this information about your headache and so on and so forth.
• 8:47
  And it's not like, uh, that's partially to deal with the hallucinations issues.
• 8:51
  And it's not like, uh, you can Eliminate it entirely, but you can definitely offer citations and paths to how the I deduct the answer control to your question is how do I balance between the generation, you know, Jenny I generative AI.
• 9:09
  So how do I balance between the I generating an answer to?
• 9:13
  You know what?
• 9:13
  This is a sensitive subject.
• 9:15
  I don't want you, dear AI, to generate an answer.
• 9:19
  I want you to give me the same exact response each and every time.
• 9:24
  Uh, the main, uh, the, the, the simplest example here is when you are in an emergency intent.
• 9:30
  So when we identify that you need, like, to get to the ER or call 911, We don't want the AI to offer any type of treatment or diagnosis besides telling you it looks like it's an emergency situation, call 911 or get to the ER.
• 9:48
  OpenAI have been somewhat allergic to actually limiting What's effectively a Ferro model, a model that is learning even on the questions that you're asking.
• 9:59
  Um, Anthropic has, a different competitor has worked to sort of organize and try to make more explainable the way the logical, the way the models are driving answers.
• 10:12
  It's just two different approaches to a similar problem.
• 10:16
  ChatGBT is more open ended and has obviously got more users.
• 10:21
  I think what I hear you saying in healthcare is rather than let the models try to answer every question in every place, you're actually putting stops, Uh, controls and, and effectively, um, guardrails that, no, you can't answer certain of these questions because the range of answers is, is, is, is, is too risky to risk the wrong answer.
• 10:46
  Is it, is, is that the right way to think about it?
• 10:48
  Absolutely.
• 10:49
  We actually.
• 10:50
  Absolutely.
• 10:50
  talk about guardrails or safeguards in terms of, you know, like today there are a lot of AI companies.
• 10:57
  Some of them are basically, let's call open AI or entropic or any other large language models.
• 11:03
  In healthcare, you just cannot do that.
• 11:06
  So the question is, what are your guardrails?
• 11:09
  For example, we use a knowledge graph, again, without getting too geeky.
• 11:13
  What we do is, uh, we, we can tap into the physician directory and scrape all of the physician's information and restructure it in a knowledge graph form.
• 11:22
  And then when I have questions about the physician, I know that the large language model isn't going to go to the world wide web and search for the data, but is going to search it Through the knowledge graph.
• 11:34
  Just to be clear, a knowledge graph is, would be a, a form of a, of a database.
• 11:39
  Only a little bit more elaborate with a, with a few more pathways.
• 11:41
  But it's, it's a, it's a, it's a data, it's a, a data structure that you vetted.
• 11:47
  And manage so that therefore it's safer.
• 11:50
  Is that the right way to think about it?
• 11:52
  Absolutely, and we reorganized it in a way that um, you can think so let's take a find the physician use case Um, so the main entity in this knowledge graph would be the physician And some of the attributes of this physician would be their specialty the insurance plans that they accept the locations They accept in and you can actually visualize all of this information.
• 12:13
  So when i'll say You I'm looking for a cardiologist to speak Spanish and accepts Aetna insurance in the Upper East Side.
• 12:19
  I'll see John as a physician that is a cardiologist that accepts Aetna and is in the Upper East Side, and whatever else I said.
• 12:28
  But that's how you guard the data.
• 12:30
  So you won't make up physicians just to satisfy an Aetna.
• 12:35
  There is a John Driscoll, who, I think it's a pediatrician on the Upper East Side, but I never could have gotten into medical school, unlike David. Yeah, I'm sure I, I could have gotten in, I don't know what I would've done once I got inside there though.
• 12:46
  Look, I have a, I have a brother does that now.
• 12:48
  Speaking of John and not John Driscoll, there was another John that you spoke with recently, John Brownstein, and I saw you had a, a webinar on, on responsible ai and, and John was actually been a guest on our show.
• 12:59
  And I don't know whether he's responsible or not, but he is a creative thinker.
• 13:02
  What, um, he's responsible.
• 13:05
  We love John.
• 13:06
  And please don't forget to listen to re listen to that podcast.
• 13:09
  That was a good one, John.
• 13:10
  He's always smiles through every, whatever we say to him.
• 13:12
  Kind of like you Israel, in a, in a sense, whatever we throw at him.
• 13:15
  What was that webinar?
• 13:16
  Like when he, any takeaways, uh, from that on, on the responsible, uh, AI side from John or others.
• 13:21
  Yeah, I think that john and Children and Boston Children's like are very thoughtful and very advanced with everything that they do with, uh, with large language models with the collaboration with open the eye.
• 13:33
  And I definitely don't think that you should take an example from john because, uh, They're very advanced in terms of, like, the resources that they can put in.
• 13:42
  Um, and I think that, uh, the nice thing about the conversations with him, and obviously that's not our first conversation, is, um, the depth in which they actually started experimenting with large language models for a variety of things.
• 13:58
  Um, now I said don't take an example from him because, um, it's, it's really, where we started, it still is.
• 14:08
  Hustle to deployment and maintain a good enough chatbots and voice assistance for various needs, especially in a patient facing.
• 14:17
  Setting.
• 14:18
  So, um, unless you have the time, the resources, both from like a capital perspective and from a technical perspective, uh, to actually get very much into the weeds, you probably want, um, some sort of a partner, uh, to, to help you like navigate that.
• 14:36
  And, uh, yeah, I think that Boston Children's is in a very unique place, uh, to be able to both find the partners, but also do a lot of experimentations by themselves.
• 14:47
  So we're fortunate enough to actually live down the street from Boston Children's, and my kids have gone there, and one of the things that, and we've done some work there, and one of the things that struck me is that AI, but also other sort of technology should be able to do, is enable anybody, wherever they are, even if they're not just down the street, but to be able to tap into that kind of expertise and to be able to project it further.
• 15:08
  You see people there from all over the world.
• 15:09
  There's only a few that can come.
• 15:10
  It's very expensive.
• 15:11
  You have to wait, et cetera.
• 15:13
  But how can we tap into all of that knowledge, not just randomly that's generated, but actually that they've generated and bring it, bring it forth.
• 15:20
  Hasn't happened so much to date.
• 15:21
  Does AI, that wasn't necessarily the conversation with John, but does AI enable that to a greater extent, or are we, or is that a different direction?
• 15:30
  Yeah, I think that, uh, SAI matures, it's going to be, um, you know, the most professional, uh, skilled workers, uh, knowledge workers that, that we've had.
• 15:43
  Um, that means that they can query various large databases and return with coherent answers and really create efficiencies.
• 15:54
  Um, and, and more access to care, access to knowledge, uh, generally speaking.
• 16:00
  So in healthcare, which is a very complicated area for a lot of us patients, you know, to grasp both from, you know, the clinical side, but also in terms of like how it works, you know, the payers versus the providers versus the farmer companies.
• 16:14
  It's, it, it is, um, still very problematic to navigate.
• 16:20
  But Israel, I, I do think that David's asking a slightly different question, which I've actually seen some evidence, David, that there is.
• 16:28
  I mean, there's a, uh, an MD who's a chief data officer in a hospital who's got a very rare form of cancer, and he's actually using the hallucinations, those in those, those sort of probabilistic jumps in the model to help test whether new forms of a new combinations of cancer treatments might help accelerate his healing.
• 16:51
  And he, because he's a PhD in data science, as well as an MD and understands that he can, he can, to Israel's point, control it.
• 16:59
  But I do think that in addition to a learning, learning faster, which I think was your question, we also may be able to tune some of these, what now is a hallucination and potentially control them as Israel's suggesting into insights.
• 17:14
  And maybe we'll try to get that doctor on our, on our podcast.
• 17:19
  But I, I think it's a really, um, it's a subtle question.
• 17:22
  And I think Israel, and you've got to speak to this, but I think we're still figuring out the models.
• 17:27
  But there's a lot of runway.
• 17:28
  No, absolutely.
• 17:30
  I think that, um, if we're looking at this from this perspective is, you know, um, when people thought about what are the first use cases that AI is going to solve, I can tell you that.
• 17:43
  The last use case that people thought that AI is going to solve is creativity.
• 17:48
  And all of a sudden, most of us on, on, you know, on our personal use, use ChatGPT for creative thinking, right?
• 17:56
  Like, it help us brainstorm, it help us come up with questions, it help us come up with workshops that we want to have with our leadership teams, and so on and so forth.
• 18:07
  So to your point, I definitely think that it's not necessarily like hallucinations, it's like even Mutation to genetic algorithms like, um, that's part of how we as human humans evolved.
• 18:19
  And it looks like there is a sub part of algorithms called genetic algorithms that uses this mechanism to actually come up with new novel solutions.
• 18:28
  So where can it go?
• 18:30
  I don't know yet, and I don't want to guess because it seems like our guesses in the past were very wrong, but it can definitely help us achieve I'd say like breakthroughs that we weren't even thinking about.
• 18:44
  So last question.
• 18:45
  We've been fairly positive here about AI as a good force.
• 18:49
  And on the other hand, we've seen that some nurses in particular have been raising the alarm bell over the use of AI and the introduction of AI.
• 18:57
  We actually did a A popular show on it.
• 18:59
  I think it was one of our highest rated episodes about that.
• 19:03
  What is your take Israel on why at least some nurses may be kind of raising the alarm and even taking labor action when they see AI being introduced?
• 19:13
  Yeah, so it's not only nurses.
• 19:14
  You heard about the nurses, but, uh, so, so we, we sell voice AI assistants, right?
• 19:19
  Why do you think the call center manager thinks about that or the call center agents?
• 19:23
  And they think that, uh, generally speaking, Every time that, um, a new technology wave, a big technology wave, like a platform shift, is happening, there's a question of, is it going to eliminate, uh, a lot of jobs?
• 19:36
  And, and what it's going to do to these populations.
• 19:42
  And, and I think that what we've discovered in the past is that while it's eliminating some jobs, Creates a lot of new ones, so I can end it in an optimistic manner.
• 19:52
  But because you you requested the pessimistic one, then I will also share that.
• 19:58
  Um, the pace in the change of pace, um, that we're experiencing with AI is unlike anything that we've seen in the past.
• 20:09
  Um, so the concerning factor is, um, are we as humans going to be fast enough to adapt to the change that AI brings, given like the technological breakthroughs that happening really every week?
• 20:23
  Um, and in the past we had enough time.
• 20:27
  Like with the, again, internet, cloud, mobile as, as a platform shifts, we were able to adapt to it because it took like several decades.
• 20:36
  Now it's going to take less than half a decade to see a lot of implications of AI.
• 20:41
  So that's the question, like, will the human, race be fast enough to adapt to this new pace of, uh, technological shift in the era of AI.
• 20:53
  Excellent question to end on.
• 20:55
  And I was going to ask another one about, uh, whether podcasts can be replaced with AI, but we'll save that for another, uh, episode.
• 21:00
  We're podcasters.
• 21:02
  Yeah, exactly.
• 21:03
  We can't, we can't do that.
• 21:04
  Well, that's it for yet another episode of CareTalk.
• 21:07
  We've been talking today with Israel Krushy, co founder of HIRO, revolutionizing conversational AI.
• 21:13
  I'm David Williams, president of Health Business Group.
• 21:15
  And I'm John Driscoll, senior advisor at Walgreens.
• 21:18
  If you liked what you heard or you didn't, we'd love you to subscribe on your favorite service.
• 21:22
  And thank you, Israel, for joining us.