Originally Published:
ChatGPT has gotten a lot of attention in recent months. AI and Machine Learning have the potential to be used in a variety of healthcare applications, including medical documentation, virtual assistants, diagnosis and treatment, drug discovery, and patient engagement. Here experts discuss the promise and the potential risks of AI, machine learning, and other conversational language models in healthcare. Specifically, they will discuss how the tools can be:
Ronald Razmi, MD, Chief Executive Officer, Kinders
Elise Kohl-Grant, MBA, Healthcare Engagement Advisor, Amazon Web Services
Divya Pathak, MBA, MS, Vice President/Vice Chair - Artificial Intelligence/Machine Learning & Analytics, Mayo Clinic
Megan Antonelli, Chief Executive Officer, HealthIMPACT
Hype or Hysteria Promise and Realities of AI, Machine Learning and ChatGPT in Healthcare
Megan Antonelli: I am very excited to have this group of panelists together. We you know, I think everybody's heard a lot about uh, G B T lately, but of course, ai, machine learning, all of that, you know, has been around for a long time in healthcare. So I have Divya, who I had the pleasure of meeting in December as well as, Elise, and Ron, I think we are at also at the digital medicine conference in December.
So you were very busy in December. it. Yes. and it feels like it was only yesterday, but here we are. So, but I'll let you guys introduce yourselves and sort of talk a little bit about the work that you do and get us kicked ofAndf.
Divya Pathak: Okay? All right. My name is Dya Patak. It's to be here. Megan. I lead, I'm the vice chair for AI Innovation and Enablement.
In the Center for Digital Health at Mayo Clinic is primarily focused on looking at AI-based innovation in applying it in the clinical business and technical problems at Mayo Clinic.
Elise Kohl: You could have introduced yourself for a lot longer, so that that's okay. Um, Elise, co grant, executive advisor at aws and I was a former CIO of the behavioral health system in New York. A lot of people don't know what like an executive health advisor is. So I help our international customers get ready to onboard to the cloud and help them navigate through their compliance and regulatory concerns.
Wonderful.
Ronald Razmi: So I'm Ron Rassie. I am a venture capital investor. So on the dark side. And by way of background, I'm a cardiologist who years developing digital health solutions. I was previously c e o of ACU for six years and now our current venture firm, we invest solely on digital health and AI in healthcare solutions.
So we have a 300 million fund that we're investing out of right now and a second 700 million fund that's being stood up as we speak. And. Entirely focused on digital health and AI and healthcare. Great.
Megan Antonelli: Awesome. Well, it's a pleasure to have you all here today. Divya, let's start with you cuz you know, Mayo's big organization doing a lot in this space.
You know, I think John Halanka was Yeah. Keynote here at our last, as the 2019. Yeah, yeah, yeah. Tell us a little bit about, you know, as we, we talked, we called this hyper hysteria. Tell us, you know, what is the reality of it? What, what are you guys really working on?
Divya Pathak: Thank you for that question, Megan.
let's me speak about what we do right? Mayo is as you all know, it's, you know, it's, it's a number one organization when it comes to patient care. And the foray into digital health and AI really started with our ceo Dr. John Rco Feria, who was a gastroenterologist by.
Ronald Razmi: He's not a cardiologist.
It's
Megan Antonelli: a problem. .
Divya Pathak: That's true. So he's a clinician researcher and he's a very visionary leader who really wanted Mayo to be the national leader when it comes to digital health and ai, and really to have a strategy called bold forward in terms of cure, connect, and transform. And with his him taking the role of the CEO and.
COVID that just reinvented everything when it comes to healthcare. There's been big investments made in Mayo Clinic in several France. One of the most prominent ones that you hear in the press is the Mayo Clinic platform, which is a business entity as part of the non-profit organization, really looking at how do we discover and deliver AI-based innovation to external partner.
Working in the, in the healthcare ecosystem, the second big investment that Mayo Clinic made was in the Center for Digital Health. Mayo hired the first chief digital officer Rita. K. She's leading the digital transformation strategy at Mayo Clinic, and there's no digital health when we don't talk about ai.
Mm-hmm. , so AI organization that is, that I'm leading now was actually formed within the Center for Digital Health to be this office of. To support AI initiatives across different clinical departments at Mayo Clinic. So we talk about specialties. Mayo is trying to actually look at AI based solutions across all specialties.
Cardiology, radiology, to be more cardiology is in that space too. So there are, those are more prominent initiatives that Mayo has established. And my group is supporting all these initiative. In the entirety of AI lifecycle, not just in the development of algorithms for any clinical problems, but also in translation of ai.
So how do we really take AI into the clinical workflow? So there's a lot of research when it comes to that space. Building frameworks that's risk covers, it's explainable, it's trustworthy, so we can really eliminate the PLA box of ai. So that's been a focus in my. . And last but not the least is the education component.
I mean, we talk about AI adoption and we still have clinical workforce, which is still under, you know, not everyone is understanding the implications on in use of ai. So being able to educate healthcare professionals make them literate enough to use ai not just in building ai, but also in in clinical adoption, white clinical adoption Education is another focus. So broadly in speaking off different terms, there's, there's more internal engagements, there's more external partnerships, and Mayo wants to be a leader in the space. Thank, you've done a good job of being here,
Megan Antonelli: so, yeah. That's great. Elise, how about you? What's your, what's your thoughts on that?
Elise Kohl: So, I'm thinking about the title of this panel, hype versus Hysteria, and. Folks are familiar with something called the hype curve. I'm gonna like MBA 1 0 1 where it kind of shows the hype versus how much work we've actually done in terms of leveraging technologies that have built in, you know, AI into it.
I think that the top right would be the hype and less work. I think we're actually over. Not over. We're definitely in the hype, but I do think that now we're getting into the nitty gritty of work where providers and hospital systems are leveraging AI a little more. Now, we are not fully in the trenches yet because, you know, like chat G P T, there's a lot of hype around it, which I actually think there should be.
But I am seeing now hospitals actually implementing these systems and in specific use cases. It's not everywhere. Yeah. Even though it's every, even though we read it everywhere, but we are seeing forward pass that are being taken using AI and just to kind of like set the record with AWS where we fall into this.
Possibly a lot of the vendors that you invest in are leveraging AWS services to build those models. So I wouldn't say, you know, we're not a software platform and here we're selling it to the doctor. We're actually like the infrastructure services that help engineers. Yeah. And holders build that. Mm-hmm.
right. .
Megan Antonelli: Now, Ron, you just you're finishing up a book
Ronald Razmi: on uh, my, yeah, I, I wrote a textbook of AI and healthcare during the pandemic. Right. Which is coming out this year. Great. What's it called again? It's called Making the World a Healthier Place using Artificial Intelligence. Oh, that seems
Megan Antonelli: wonderful.
Seems timely, . So tell us about that and tell us about, you know, was there anything in the process. Of doing the research that you learned that you weren't expecting? What are the
Ronald Razmi: surprises from it? Absolutely. I mean, and, and we're looking at these companies every day to invest, to decide to invest. I mean, our deal flow is I mean, I can hire 20 people and we wouldn't have enough manpower for all the deal flow.
We're seeing companies coming to us every day saying, we're solving this problem. We're solving that problem. Life sciences, health systems remote patient management. And so, , there's a lot going on. What I would say is, in my opinion the use cases are gonna start with unsexy, low risk. Mm-hmm. , backend, operational.
Absolutely. Things where you can improve operations in a hospital. Mm-hmm. , you can improve patient experience or communications, however, , the heavy duty clinical use cases, drug discovery patient management, and so forth. There's a long way to go. Yeah. And there's a long way to go because the state of data is very chaotic.
Yeah. When you come and say, look, I have a, I had this last night somebody was pressing me to look at their company again, company that collects data. Using a wearable and helps manage heart failure patients and they use AI to process some of that data. That's a great example for this discussion.
Look, you can collect whatever data you want out in the wild way, high ox, pulse, oximetry, you name it. That data has to go back to a care team. Yeah, it has to make its way into a data database where somewhere it's in the vicinity of all the other data, the patient's echocardiogram, labs last visit, and so forth.
Mm-hmm. , somebody needs to review not just that data. If you, I mean, when I was a practicing cardiologist, if you came to me and you said, My weight went up 20 pounds in the last week, and my oxygen saturation is down and I'm very tired, you know, all consistent with a heart failure exacerbation. First thing I had to do was figure out who this person was.
Yeah, what is your name? Why are you talking to me? I had to open the record, read through the chart, see what I wrote the last time I saw. . So that whole process still needs to happen. Absolutely. You can connect whatever data you want. So, and a lot of that information is unstructured. So it's not like you can get the data, combine it with e h r data, which, which is all nice and structured, and algorithm figures out what to do and tell special what to do.
This requires a lot of manual work, extra work by the clinician, right. If I've noticed anything in the last day here, there's a lot of talk about burnout, staff shortages. So now you live good luck with that. Develop all the AI you want out there if it's not integrated. Yeah. So this is, this is where we are now.
Yeah. Yeah. We have unstructured data. We have fragmented data, and th these algorithms have not been taken into clinical. and their, their impact on patient health and patient outcome. Hasn't been proven. Yes. Right.
Elise Kohl: So, so, oh yeah, I wanted to jump on it and I like, let's hone in on some specific use cases.
The three, I'll go, the three that I'm most excited about, and I'll go from least scary to most scary is automating administrative activities. Right. Which you talked about exactly. Actually health equity, I think there's a way that we can increase access to services. through ai and probably the most scary to people are precision medicine.
And so let, let me just talk real quick about the automating activities and you mentioned unstructured data. Data. Healthcare data is so messy. It's not like financial data, it's very, it's qualitative data, right? There's a lot of different information out there. I know. I come from a behavioral health background.
The heart of what happens with the patient is in the case. And someone mentioned a good data point here on this stage that the US case notes are three to five times longer than the rest of the world. I had no idea. That's really interesting. I'm gonna look that up. But imagine if we could use AI to extract some of that information that we see in the case.
Note to Translate to ICD 10 codes and procedure codes and diag, you know, diagnosis codes so that way we don't have to put it on that provider. I know we had a quality improvement department where one person read everybody's case note to make sure that it met the standard and I. That's just crazy to me that you have to read every single case note.
So if there's ways we can automate that I, I think
Ronald Razmi: that's really insane and that's, that cannot be automated now. There is too much context. There's like NLP history, like a lot of, but N L P N L P right now in healthcare is terrible. Like we have, we have better NLP now, but it's get, it's, it's better, but it's not there because it's not there yet.
A lot of the clinical notes. are abbreviations, right? Like I, I never wrote a full sentence. I mean, c o you know, all of those app abbreviations and there's no standard mm-hmm. . So NLP can get you, is that your experience? 60, 70% there are
Divya Pathak: you fi do you Yeah, yeah. I can actually totally relate to, I can give some real world examples of using NLP and administrative and scheduling tasks at Mayo, right?
I agree with both of you. I think Still exists when it comes to personalized and precision medicine. Mm-hmm. , where hype doesn't exist is the use of machine learning and AI-based techniques when it comes to administrative scheduling, triaging, clinical information processing, of which 70%, actually more than 70% of clinical data is all unstructured.
Right. That's totally that, that goes by the definition. So, , we have been working on N L P algorithms not just using the off the shelf commercial ones. Mm-hmm. only because of that context you mentioned and the linguistic and the semantic understanding of the data that differs even from one specialty to another specialty.
So what we are trying to do is and the industry is also evolving so fast that you see all these LA large language models, which is really the basis for the chat, G P T uh, chat. Our group is, and predominantly even in the field of NLP healthcare research what is happening is they're building more frameworks and models that can be adapted, contextualized, depending on the use cases, the use case to read a clinical note for finding out the patient's condition so that they could triage.
I mean, they could schedule that patient to the right service line versus the use. Clinical use of NLP to look at clinical notes, to identify the labs, the medications the problems and diagnosis. They're very different use cases, so we know there is no one nlp that fits all. But being able to contextualize N L P based on the use cases, based on the physician or the care coordinators need, requires a framework that we're able to build that could actually use as a plugin and plug in of N N L P algorithms off the.
Custom state of the art, but adapting to that framework to support a variety of those administrative tasks, especially when it comes to unstructured data, has been a focus. Yeah. And that's something we've been working on.
Ronald Razmi: And it's improving, actually. It's improving. I think one of the things about chat, G P T is an N L P.
Yeah. Model underneath it is a, is a step forward. Mm-hmm. .
Divya Pathak: So it's actually better, it's generated ai. It's better. Yeah.
Ronald Razmi: Better. However, NLP so far has done better in other sectors with the. Then it's done in healthcare. But the foundational models, yeah, are the transforming models are better and now they need to be trained for specific, specific in orthopedics, in cardiology and so forth.
That's where we are now, is the, the foundational NLP models are much better than they were three years ago. Right, or five years ago. And so
Megan Antonelli: within your organization and you know, in what you guys have observed, like when you. to sort of tie in what we've been talking about all day. I mean, in terms of the clinician and involving the clinician.
When you're looking at these areas where you need it Yeah. And where you can use it, how involved are they in that process? Okay.
Divya Pathak: That's the best part of me being at Mayo Clinic is my day job is to work with physicians, right? To understand not just their problem space, but also use their subject matter expertise in building a, and I know that we spoke about it earlier.
You know, there's so much technology advancement when it comes to AI and machine learning cross industries, right? Not, not just specific to healthcare, but what's more important for healthcare is to have that clinical subject matter expertise involved in building the AI models. And that involvement is, call it as a prerequisite for AI to be adopted.
Unless there is practice guidelines. That's every organization, Mayo Clinic, Cleveland Clinic, New York Prison. , each organization has their own practice guidelines on top of all the, you know, national standards that might exist. For example, diabetes, the way of defining a diabetes patient in Mayo Clinic, h Onec, you know, A1C has got to be greater than seven.
But actually the National American Diabetes Association's criteria is greater than eight. So there are so many of those guidelines that's very specific to an organization. So having p. Partnership call it as close partnership with the technical teams are so important. Mm-hmm. , and that's something that's a model we are trying to create within Mayo, right?
So it's the medical and the technical expertise in place to build the digital health and ai right. . And are
Megan Antonelli: you guys seeing that, I mean, you know, not just with the health systems, but also even the tech providers that you work
Elise Kohl: with. So it depends, and I think you probably can speak from experience. Yeah.
With the vendors, I don't see as much of that, to be honest. Yeah, I mean, I think we saw it on stage a little bit where the perfect model is to have that subject matter expertise with that engineer with that. and that happens if you're in a health setting. Yeah. From a vendor perspective, I do think that, I see that is not, that model is not always used to have that subject matter expertise.
And now it's hard because in different areas, if you're in children's, if you're in substance use, if you're in cancer, right, it's. All very particular subject matter expertise where you have to have a human there to help train the model and know, and no one to tell it, yes, this is right, or yes, this is wrong.
So the model, so the model can continue learning. I think in healthcare it's hard because when you're looking at selling to a customer, oftentimes the customer is not the end user and the customer is most certainly not the patient. So your three steps removed from the customer. Customer might be the CTO of a hospital, but then you have the end users who are the clinicians, and then the ultimate end user who is a patient who is so far.
Absolutely. From the actual like building of these things. So it's really important to incorporate patients and doctors and not CMOs who haven't been, no offense, but who haven't been on a hospital floor for 20 years. Incorporate those folks into your, like governance structure into your, so that you get that feedback loop.
Mm. .
Ronald Razmi: Yeah. I'll also say there are different types of physicians. There are physicians who take part in developing this stuff. Yeah. Who are much more innovation oriented. Mm-hmm. . And then you develop it and you take it to the real world and 99.9% of physicians don't want to change their workflow, even though you, you included physicians in designing the.
It doesn't gain the traction that you thought it should get because mm-hmm. , the ones who participated are more innovation in centric. And then to add on top of it at Mayo which, you know, I did all of my medical training there. It's a fantasy world. Physicians who work at Mayo, the environment at.
It's so different than anywhere else. I mean, a lot of people don't realize, like physicians at Mayo can't wear white coats. They have to wear suits and tie every day of the week, seven days a week, you're not allowed to wear white coats. I showed up to clinic one time. I had been on call all night at the C C U taking care of patients.
Okay. I, I, I had patients waiting for me in the clinic. I threw on my. And went in to see my first patient and then a nurse walked in the room while I was with the patient and said, somebody's on the phone for you. . I walked out, it was my program director who said, the nurses are telling me you haven't shaved, you need to shave before you see patients.
Oh, wow. Wow. Mayo is a very , it's a really mayo is a very harsh environment. Wow. And physician behavior, how they practice medicine. Right. It's very different. So you can, which I think, I mean, well, that, that becomes a problem because if you develop it in that, and you gain traction. It may not translate, but
Megan Antonelli: it speaks to, I mean, Mayo sounds very unique, for certainly.
But you know, I mean, to the problem of developing the tools that make it, you know, frictionless. I mean, every specialty is different. Every hospital is different. Every physician, their age, their, that's, you know, their technology comfort level is gonna be different. So, You know, it's a complex problem.
Divya Pathak: Yeah. And actually, just to answer that, yes, Mayo's environment is very cult-like and they're very proud of it. Right, right. That's, that's the truth. But one thing I would speak about, you know, adoption of AI when it comes to physicians yes. You have a very innovative group of physicians passionate about technology.
I think the younger generation, and I'm not, this is not, not, not to really be biased, but I think the medical training now, Much more looking at advancement when it comes to informatics and ai. So that's really part of the training now versus 10 years, or even 15 years ago, how medical professionals were trained.
So that's number one. So there's more understanding of the use of technology in medical training in in Medi in Medicine. The second one is, I read an article I can't quote, which, what the, what the. It was very interesting article, which was talking about AI will not replace physicians, but AI will replace physicians who do not use ai.
So eventually that's where I see AI leading is right? Eventually more of the, as we move into the Gen Z world, , they'll be more adoption of ai. And it, it won't be an afterthought. It would be of how can technology be infused in the way I work? Yeah. Mm-hmm. . Yeah. And, you know, I don't wanna take the example of Metaverse.
It's really in some bad press now, but there was a nice, again, another YouTube video that Meta had published, which was talking about virtual surgical training. How can I teach medicine virtually and how can the surgeons operate on a. virtually. So that's really taking us to the red player, one ready player, one kind of environment where that's, that's the future.
That's why I see the future. But they, we are very, I'll call it, we are in an infancy state. Mm-hmm. when it comes to precision and preventative ways. Mm-hmm. , we're much more advanced in the other areas of administrative and scheduling and
Ronald Razmi: triaging. As a matter of fact, years ago we invested in a surgical training using vr.
It was at its infancy back in 2007, 18. It's still at, it's still
Elise Kohl: still. It's a very bad decision. . Yeah. Well and I think you bring up a good point cuz I think the hysterica comes around people being worried that a robot's gonna operate on me or that it's gonna replace all these jobs. We're in such a like workforce shortage anyways.
I'm personally, I'm not worried about that aspect cuz I think healthcare will always need a human touch. Right. But when you're accessing healthcare services and when you're writing down your notes and when you're sitting in the front desk and you're trying to triage and run a whole control center of where are my patients going and sending nurses to each room to check, are the rooms ready, are the rooms available?
That's a lot of that's a, like a lot of human work. That shouldn't be human work cuz that's not why. . That's not why providers got into the business of healthcare. They got into the business of healthcare to treat patients. So I really think we'll always need that human element, but it's things that automate the intake process, right?
That is a pain that I've worked in healthcare for over decades, and I still don't understand really how the system works from a patient perspective. So as. . You know, when you get into the nitty gritty, it's kind of boring. It's not as sexy to talk about. But those are the areas Yeah. Where we're seeing a lot of
Ronald Razmi: movement.
And actually right now you have companies like Lean Toss, q, Ventus, scheduling. Yeah. Who are helping with hospital operations. Mm-hmm. improving CER or scheduling, and so, And I think the use case that you mentioned earlier, like applying N L P to notes coding and so forth, that is the frontier that is gonna be the first applications.
And there's nothing wrong with that because it's gonna save a lot of work into workflows and create room for patient
Megan Antonelli: care. Mm. Right. And that's what it, you know, taking out the stuff that is in patient care, right? Automating the easy stuff, which we talked a little bit about getting rid of the stupid stuff, which I do love.
But, you know, and, and that's, you know, the pro promise of it is to be the augmenter, right? So, you know, chat, G B T, we've. certainly seen a lot about it on social. Some of us have watched their children write papers with it. And some of us have experimented with it. You know, and so I wanna open up the conversation cause I don't think anybody's an expert in it.
I, except you guys, of course, . But . But in terms of any questions we have, I, I, I mean, I've started to hear people's stories of, of using it, so. Okay. I don't know if you've used it, if you wanna talk about it a little bit. Yeah. Each of us can kind of share a story. I already. , my son wrote paper. So, but you, for you guys tell us and then I'll speak the Mayo
Divya Pathak: Clinic story.
Yes. . It's off, it's right off the press. Right. Chat g pt, since late last year, what was November 30th is when Chad g p t was released. Mm-hmm. . The around the use of Chad PT was huge and believe it or not, a lot of neo physicians started playing with it. And what happened? , I was getting a lot of inquiries.
So did my team around, can we use it? Hey, I'm finding this very educational. There was some compliance issues too, when you start using it for, in a medical organ, a medical, you know, organization. So we had to come up with a very soft policy right now to actually. Make sure that we provide guidance when it's, when you're using charge G p T in healthcare.
First it's not a HIPAA compliance service and you really passing any information, patient specific information to charge G P T servers is, is against the terms of use when it comes to hipaa. That's number one. The second is it's, it's a research preview version. As we all know what, they're still trying to improve their models as they get more feedback.
Being able, and it's, it's only dated back till 2021. It's the data that's used, the knowledge that's fed into those large language models are only, goes back only till 2021. It's off the internet. So you know that anything in the internet is as credible as it, as it can be, can be used for Wikipedia, for it's, yeah, it's, it's Wikipedia Plus plus.
Right. Anything I could write on my, you know, my blog. The third was that it doesn't have any medical knowledge. It's no nothing based on medical knowledge. So it's not looking at, you know, U M L S, it's not looking at any knowledge or terminologies. Mm-hmm. . So anything you, you see in it, it's probably some literature and you
Elise Kohl: can understand and
Megan Antonelli: like we have a tool we write you that you Right.
Use the tools
Divya Pathak: we gave you . And last but not the least is. It's so important for every organization, especially the healthcare I'm gonna speak to come up with policies, right? In use of charge G P T, it's still a very nascent field. Yes, there's lot of promise we can talk about. End number of potential applications and use of charge G P T in healthcare.
but I think we're really in this generative AI space where we are really learning the prom, we're learning what possibilities can happen, but there's no risk framework around it. There's no policies around it in an organization. So we've provided guidance to all physicians and educators at Mayo Clinic to use it.
We can't, we don't have a policy. We're still building a policy, but they've got to use it based on understanding the risks associated with it. Okay. Chat, G p T in healthcare education. Absolutely. Being able to summarize your literature. If you've got questions around literature, medical literature, absolutely.
It can be used in terms of education, but not really in support of patient care. Right, right.
Elise Kohl: Can I ask the audience, raise your hand if you've used chat. G P T. I can raise mine too. . Okay, so like ha, a little more than half. So I have to go on the record saying I'm not speaking from AW s perspective.
I'm speaking from my personal perspective on Chat GPT . I've used it right? And how I used it was I have a two year old and somehow we got into this routine of me making up a story about him on the farm. He's a Brooklyn kid that loves a farm . And after two weeks I started running outta story. So I went on chat G P T, and I said, tell me a story about my two-year-old who lives on the farm and loves pigs.
And it spit out a pretty good story. And I used it and I used that example to say, and going back to the three things I'm most excited about, administra AI in administrative activities, AI and health equity and AI in precision medicine, health equity. There's a national push for health literacy, right?
And to get content out there that can reach the patient. Don't have a medical background. Right. And getting content out that that's in different languages. If you ask a doctor or a nurse to write some form of patient facing material at a third grade level, it's really hard to do because you have all the acronyms, you have all that medical knowledge.
It's kind of cool to think that we can think about chat G B T and saying, Hey, write this for me. In a third grade level or in a second grade level. So that way we can speak to more patients that are out there to create more equitable access to healthcare services and help patients manage their own care and actually learn a little bit how to navigate the healthcare world.
Ronald Razmi: I had prepared a few slides when I didn't know what this panel was gonna be, . It was a possibility that I. Like give a talk. One of them was this guy who is a physician and went on Twitter and said look guys, for those who say there is no use case for chat GPT in healthcare, look, I asked the, to generate a letter for this patient to get this procedure.
It generated the letter and. Why it's justified and put the references on there. And he posted it on there and people were like, thumbs up and stuff. And then a few threads down. Some people started asking questions, you know, I can't find the reference. Yeah, it's true. And so, so it turns out the references looked very scientific.
Yeah. very correct. And they were completely made up. Yeah. So, , so Chat GPT is capable. Of very convincing bullshit , and, and it looks, it looks very authentic. It looks very authoritative.
Megan Antonelli: So what you're saying is that my son's paper was in his own voice .
Ronald Razmi: Well, , what I'm saying is, what I'm saying is Chat GPT cannot be used.
To generate unsupervised content? Yes, absolutely. Yes. That's sent to patients or providers looking up, oh, I have this patient here. What is the best treatment? You don't know what's, if what's it's given you is the best possible treatment or completely made up. Made up. It's no. And actually now that clinicians and healthcare people have played with, We have about a few weeks worth of track record, which is 50% of the time it's making things up.
Yeah. Or it's giving you content that's close to what you asked, but it's not relevant. Like you might ask something about diabetes type two. It gives you information about diabetes. Type one, you don't know it's about diabetes type one because it's stated very authoritatively. Mm-hmm. and it's very nicely format.
however, type two and type one are different. Mm-hmm. , you know, one uses insulin, one doesn't, one is obesity. So that's it cannot be used in however, what I think it could be very helpful is it can take you 50, 60% of the way there. It can generate the content and you can review it. Yeah. If it's, , you make it available to patients.
Right. If it's not, you edit it and it saves you time. Right.
Elise Kohl: Going back to, it needs that human element. Yes. Right. To look over. So can
Divya Pathak: always say AI is only to aid in clinical decision making. Yes. Yes. Cannot make decisions like ization. So the human feedback at the loop, human in the loop is, is, you know, it's, it's not right.
Megan Antonelli: Well, thank you all very much.