Narrator - Dr. Abel 00:00 Welcome to HelixTalk, an educational podcast for healthcare students and providers, covering real life clinical pearls, professional pharmacy topics and drug therapy discussions. This podcast is Narrator - ? 00:12 provided by pharmacists and faculty members at Rosalind Franklin University, College of Pharmacy. Narrator - Dr. Abel 00:17 This podcast contains general information for educational purposes only. This is not professional advice and should not be used in lieu of obtaining advice from a qualified health care provider. Narrator - ? 00:27 And now on to the show. Dr. Sean Kane 00:31 Welcome to HelixTalk episode 181 I'm your co host, Dr. Kane, and I'm Dr. Patel, and the title of today's episode is from meds to machine learning how AI is and will revolutionize pharmacy practice. So Dr. Patel today, we're obviously talking about things like chat GPT and how AI and large language models or llms are already making an influence on how healthcare is delivered to patients, and what will or is the role of the pharmacist, given that we have this new technology out there. Dr. Khyati Patel 01:00 So aside from using a little bit chat GPT here and there, I have no experience on what happens behind the scenes and how AI thinks and, you know, generates responses. So I am here to learn along with our listeners here today, but to kind of set the stage. Dr. Kane like, why is it important for student pharmacists, or currently practicing pharmacists to learn about AI. Dr. Sean Kane 01:22 Well, Dr. Patel, if I told you that I had a tool, right, like a calculator, that could improve patient care, I think most people would be interested in being able to use that tool. And that's exactly what AI is. It's a way that is a copilot or a tool that can extend the services of a provider and make the quality of those services better. We're already seeing lots of healthcare systems integrating, AI companies, pharmacies. It's already happening, and it's going to happen more and more as the technology becomes more available. And it's already pretty inexpensive, it's going to become even less expensive as we get more competitors into the market. Dr. Khyati Patel 02:01 I guess pharmacists need to kind of understand, what is the language of AI, right? What is the acronym that your LLM, that you just Dr. Sean Kane 02:09 described earlier? Yeah, so LLM, or large language model, is the technology used to generate these very human like responses, and at a very basic level, I think, you know, especially if it's integrated into someone's workflow, it would be really useful to know how is it actually working under the hood, not mathematically, per se, but just the concept of, how does an LLM generate the text that you see? Because that actually informs understanding. What are some of the downsides to the LLM technology, or generative AI technology, and where is it getting the data from? You know, it looks really smart, but at the end of the day, it's using mathematical models to come up with a predictive, generative text that you see, right? Dr. Khyati Patel 02:50 And I think it's important then for pharmacists to see how this technology fits in and around their role in whatever setting that they're working in, exactly. Dr. Sean Kane 02:59 And it's not a question of if pharmacists and other healthcare providers are going to be using AI or be exposed to AI. It's when and how much and what is the the interaction between the human and the AI. So this is Pandora's box. It's already been opened, and it's really we're at a point where I have to figure out how to best use it in a safe way with good safeguards to make sure that we're doing the right thing for our patients. Dr. Khyati Patel 03:22 So Dr. Kane, if I were to go ask, you know, 15 people out there, I think two out of three, or, you know, most out of 15, would know what chat GPT is. But what's, what are some of the other similar tools, AI, tools that are available right now for the users? Dr. Sean Kane 03:38 Yeah, so chat GPT, because it was kind of first to mainstream media. That's why it's become really popular. And that was produced by a company called Open AI, and they have chat GPT 3.5 which is the one that's free and you can kind of use right now. And then they also have a premium version called GPT four. And Microsoft actually purchased open AI, and they are using chat GPT four and wrapping it in a product that they call copilot. So copilot exists as a web based interface, and Microsoft has also integrated it into things like Outlook and Word and other platforms that they already have. So that's like the open AI, chat GPT thing that I think most people are familiar with, but people may not be aware that there's actually a number of other models out there that, in many cases, are freely available. So for example, Google has one that used to be called Bard, but is now called Gemini. A company called anthropic has a model called Claude cohere. The company has a model called coral. And then there's actually open models where you can literally download it and run it on your computer. And those models, they're intended for developers, but really anyone could use them. So an example of that would be with meta or Facebook. They have a model called llama two. There's another product called Mistral AI, and there's literally hundreds of other open source. Source models that you can download and run on your own computer if you Dr. Khyati Patel 05:03 want to wow. So there's plethora of different AI resources and tools available out there, but when we focus more on the behind the scene, I should say a language like you mentioned, large language models. How exactly this program or the llms, work? Dr. Sean Kane 05:21 Well, the first step is that these models, they have a mathematical algorithm to basically take text and then learn from that text. So these models will train or learn from a huge, what's called corpus size. It's like the amount of data that it reads to understand how, in this case, the English language works, but it's obviously been produced for other languages as well, and just to put a number to it, most of these modern llms were trained on more than a trillion words. That's literally like, essentially the size of the public Internet, where everything that has ever been written and is publicly available. That is what all of these llms were trained on to understand the patterns of human communication in a written format. Dr. Khyati Patel 06:05 And when you say pattern, you're saying like, what comes after something else, right? So I'm sure there's there's algorithms, or there is methods of how you connect these different words together, yes. Dr. Sean Kane 06:21 So there's a mathematical model in terms of how the algorithm will learn and then how it then predicts text based on an input that the user provides. So it really honestly all starts at this very simple concept of what are called tokens. So a token is a word or a part of a word that the LLM, or these generative AI models will actually code as a number. So for example, the letter A is the number 32 so everything is a number to an LLM, and it has other words that it knows, like the word total that has a number of 2860 but then there are more complex words, like semaglutide would be an example where that's actually made up of four different tokens to make up the word semaglutide. So the flexibility here is that either words or parts of words are encoded as a number, and then those numbers are then encoded in a database that have relationships between those different tokens, and the learning process occurs where token A has a reference to token B, and the training happens in terms of how strong that reference is. So a good example would be the word apple in the data set. The LLM knows what is an apple. It knows that is whatever token number that corresponds to, and it looks around and figures out, what are words that are related to the word apple. So red and blue might be two words that it looks at, and the relationship between red apple is much stronger than the relationship to Blue apple in terms of the trillions of words that It read, it knows that red and Apple are more correlated with each other. So then the model will know when it has to come up with text. When it's talking about an apple, it's much more likely to talk about a red apple than a blue Apple, because it's been trained using all of these different words in terms of what is a typical pattern and the relationship between words. It has a whole mathematical model to do that Dr. Khyati Patel 08:16 and the way these tokens are connected to each other, it's probably called something else too, right? Yeah. Dr. Sean Kane 08:22 So these are called parameters. Let's say that you have 30,000 tokens and these data sets, that means that every token has a parameter or a link to every other token, and it becomes like a really big data set in terms of all of these different parameters. So when you download a model, you're downloading a small number of tokens. We'll call it 30,000 tokens, but you're downloading, like, literally a trillion of relationships between those 30,000 tokens, and the weight, or the importance of those different relationships is what makes a model a model. Wow. Dr. Khyati Patel 08:57 So this all seems really, really smart, and basically you're saying between the tokens and the parameters. Once these models are created, it will generate Dr. Sean Kane 09:07 text, yeah. So an example would be, if you provided an input to chat GPT or any LLM that says the most common brand name of Warfarin is, so it's going to take that, it's going to chop it up into tokens where you know the letter A is the number 32 so a bunch of tokens, and based on that sequence of tokens and which tokens are included, it's going to figure out what is the next most probable token. So in the case of the most common brand name of warfarin, is it's going to figure out. In this case, I put it into chat GPT, and it was a 56% probability that the next token should be C, O, U, like the word fragment, coup. And then, once it figures that out, then the next token that it knows is most appropriate with like a 99% probability is the letters M, A, D, so, COO, mud, and then the last one is n, I, N. So it predicts the word Coumadin as the most likely thing. But it actually predicts it using three different tokens. And it kind of looks like magic in terms of like, how would it know that? It just knows that things like warfarin, brand name, most common, all of those are inputs or parameters that help it understand well, the most likely next word or token is going to be Coumadin in this case. So it doesn't actually know that Warfarin and Coumadin are branded generic names, but it does know that there's a very strong relationship between those concepts of generic name and Warfarin and then this three token thing of Coumadin. Dr. Khyati Patel 10:38 So these models are smart by predicting what is the next best related word or token, not necessarily knowing the right Dr. Sean Kane 10:47 answer exactly. And a lot of people, as an analogy, talk about word completion on your phone when you're texting, where it kind of guesses what the next word should be. And it is like that. But what makes these newer, large language models, special is that they aren't just looking at the word previous. So like when you text, it just looks at what was the last word you said and what was the most likely next word. These algorithms look at the entirety of everything that you've provided it. So it knows that four words ago used the word generic name, and two words ago use the word warfarin, and it looks at everything. It's called attention, and it provides different weights of attention based on the Word and the entire sequence that you've provided to it to figure out what is the next best token. So it's not just the last word that it's looking at. It's the entirety and the sequence of it and the importance of those different words in relation to each other. Dr. Khyati Patel 11:39 And that's where the model becomes smarter because it's not just looking at one thing, it's looking at the entire text. Dr. Sean Kane 11:45 Yeah, and it looks like it's really smart. It's really good at predicting what is the next best token based on all of the tokens or words that have been given to it. And that's a really important concept to understand, is that it's not like it has a pharmacotherapy textbook and it may have read one and trained on it, but it doesn't know things. It just knows what is the relationship between those tokens and words to come up with? From a probability standpoint, the most appropriate next word or token in that sequence, right? Dr. Khyati Patel 12:14 And we will talk about some of the, you know, ups and downs of using this technology, and we'll learn that not everything that these models predict or provide as an answer are correct for that reason. But in general, you know, before we when we get there, when we think about these tools generally, what do? What type of general tasks that we can enlist these AI tools to help us with? Dr. Sean Kane 12:38 Yeah, and you know, certainly there's a lot of almost like personal tasks that it's really good at, like grocery lists and stuff like that. But if we just focus on tasks that are relevant to a healthcare provider for a second, one would be the creation of text. It's really good at being creative and creating text. So this could be a first draft of an email to a colleague about a difficult topic that you're trying to bring up, or how to explain a medical term or topic to a patient and very patient friendly language. Or maybe you want to make a handout about Coumadin or warfarin, and you want it to give you, like a starting point of, you know, some of the most important things at a third grade reading level. And it can do that. It's also really good at creativity, humor, puns, stuff like that, when asked to do so again, it's really creative. So if you want it in a certain format or a certain analogy to movies or whatever, you can do that, because it's really good at those creative tasks. Dr. Khyati Patel 13:34 And a little secret to our audience, the way we come up with our episode titles too, is enlisting some of this AI help so they can be a little funny and, you know, catchy Absolutely. Dr. Sean Kane 13:46 You know, another thing it's really good at is revising text. So in addition to creating text, it's good at revising the text. So this could be, let's say that you've made an email and the tone is not what you want it to be. It's either too serious, professional or not serious enough. You could literally give that email to a large language model and say, This is my email, but make it third grade reading level, less funny, more professional, or whatever you want it to be. It can also provide feedback to you. So you could give it your document or letter or whatever and say, like, look for grammatical errors or factual inaccuracies and tell me if you identify any of those, and it can absolutely Dr. Khyati Patel 14:22 do that. That's very neat. And I've also seen where people feed large amount of feedback or information and then kind of tell it to summarize, and it provides pretty concise, small summaries, absolutely. Dr. Sean Kane 14:35 So again, the response from the large language model is only going to be as good as the prompt that you give it. So let's say that you had a 10 page New England journal article, and you want it to identify the five most important points from that article you depend on which large language model you're using. Some of them allow you to upload a PDF file where it will then take the PDF file, turn it into text, turn it into tokens, and then summarize it. It could absolutely do that. You could ask for 100 word or a 500 word summary, or whatever it is. You have to tell it what you want, but it can absolutely summarize 1000s of words into hundreds of words, or even bullet points, if that's what Dr. Khyati Patel 15:12 you want. And then obviously, you know, you could ask questions, right? So it's good at providing answers. But I wouldn't, I wouldn't imagine these are complex answers. These are probably simple answers. Dr. Sean Kane 15:23 It depends. So for example, if you wanted to know all about warfarin, and you had very specific fact based questions about warfarin, Warfarin has been around a really long time. We have lots of data on the internet that's publicly available that it would have trained on to understand kind of the history of warfarin, where it came from, what its mechanism is, what are some drug interactions? It's pretty good at stuff like that, and it can do that without any context, meaning like you don't have to tell it go, look at the package insert. It just knows the quote, unquote facts based on the relationship between these words. What gets really interesting is that you can also feed it information, and it can use that as context. So as an example, let's say I had a 20 page hospital protocol, and I wanted to know if ABC happens, what should I do? Or does it discuss a given scenario? You can feed it that 20 page document and ask a question about that specific document, and the LLM has never read the document before. It wasn't trained on that, but it can look at all of the words in the document, develop the relationships within that document, and then come up with an answer based on whatever question you pose to it. So you can actually provide IT stuff that it's never seen before, and it can actually analyze that and then provide an answer based on whatever your Dr. Khyati Patel 16:38 prompt is. And that's pretty neat. What else can these models do for us? Dr. Sean Kane 16:43 So probably the last thing is data analysis. You can actually provide many of the large language models with data like literally an Excel document, and it can take that, it can analyze it for you. If you have a specific question. It can provide graphics, tables, whatever you need, figures, it can certainly do that. I would say that this is a less common use case, but maybe a more common example would be, I'm trying to do this thing in Excel. How do I do it? And you're describing the format of your Excel document, or maybe you're trying to use SPSS or R or some statistical analysis package, and you're trying to figure out, how do you do a thing? It's actually really good at programming oriented tasks, of telling you how to do a thing in a piece of software, or how to type the code to do something in a piece of software. Dr. Khyati Patel 17:32 That makes sense. So these are all the great things that the AIS can help with. What are some of the tasks that the llms are not good at completing Dr. Sean Kane 17:43 you know, one thing, especially the most popular platform, chat, GPT 3.5 is its knowledge base. And what it was trained on has a finite period of time. And they've kind of moved the goal post a little bit in terms of adding a little bit more timeline. But let's say the old threshold was 2021 so November, 2021 was the time where anything current event wise, that happened after that period, it didn't know about so if an article was published in 2022 or new guidelines came out in 2023 it just wouldn't know about that because it wasn't trained on that. It didn't develop those relationships in the model. So that historically was a bigger factor, especially if you're looking for time sensitive responses from the large language model. But at this point, many of the large language models are actually connected to the internet so they can go out and understand So a good example is I typed in, does Kate Middleton have cancer? And this is like a rumor happened weeks ago, right? And some of the llms knew that because they have access to news and the Internet and stuff like that, whereas others didn't, if they didn't have access to that newer data source. So it kind of depends on how the LLM is set up, but that historically, especially, has been a bigger factor, is becoming less of a factor. Dr. Khyati Patel 18:54 Okay, what about some of the complex tasks that require problem solving or involves multiple steps, like, do this first, if it doesn't work, do that. You know, kind of like algorithmic, yeah, Dr. Sean Kane 19:06 so certainly humans are better at that, and that can be an issue with these large language models, especially if the prompt is not very good. And there's like a whole science called prompt engineering, which basically involves making it's almost like programming. Like, how good is your prompt, and how can you make your prompt better to be more successful in whatever the output that you get? Generally speaking, all things being equal, complex tasks and multi step tasks that require problem solving tend to be more difficult for these large language models. Math can be that way. So like, if you ask it two plus two, it will know that. But if you ask a more complex thing, where it's almost this word problem with multiple steps, it may struggle with that. And there are some tricks, from a prompt engineering standpoint, to get it better at doing that. But generally, that's probably the biggest deficit for these is multi step complex problem solving tasks that it's not good at yet. Yeah. Dr. Khyati Patel 19:59 And then. Something tells me that, because this is machine learning and AI, it doesn't have common sense. Dr. Sean Kane 20:05 Is that true? Yeah, so common sense for sure. It isn't great at and the reason is, again, why it's important to know how llms work. It knows relationships between words and a sequence of words, how the relationships are. It doesn't intrinsically know that Coumadin and Warfarin are branded generic names. It doesn't intrinsically know ethical things or anything of that nature. So when it comes to like a common sense thing, it's going to struggle with that, because it just knows how to connect word a and word B together. And what that strength of relationship is. It doesn't actually know, no in the in the way that humans know things or have that critical thought process Dr. Khyati Patel 20:46 that is very interesting. And in addition to that, I imagine, like very specialized knowledge. I mean, we gave just an example of updated information with, you know, Kate Middleton's issue, but somebody who knows about the type of cancer she has, or the oncologist talking about the cancer or the process. Do you think that the AIS know that much of a detail about a particular topic generally? Dr. Sean Kane 21:11 No, for very specialized topics, it's not going to be as good for generally. And I put an asterisk there, and the reason is that remember something like chat GPT was trained on basically the internet, right? And many of these llms were trained on everything that's publicly available to be as conversational and human like as possible. But remember, there's a bunch of llms out there, so some of the models are specific for certain tasks. So let's say I don't know if this exists yet. It will at some point. There may be a pharmacy LLM that's really, really good at incredibly detailed, specialized drug knowledge, right? And for that kind of an LLM, it might be really, really, really good at everything there is to know about drugs, but it might be terrible at pop culture references, right? So the llms that are publicly available were meant to be as general as possible, and they're not going to be good at very specific tasks because they weren't trained for that, right? But if you have a trained LLM that is specific for a given task, it can excel at those really specialized tasks. Dr. Khyati Patel 22:12 But then you have to go for that specialized LLM exactly to begin with. And then I guess we kind of covered earlier, depending on the LLM, if they are not, you know, trained on new information, they're just trained on older information, then that extremely new information may not be available. Dr. Sean Kane 22:30 Yeah. And then, really, the last limitation that definitely was in the media when chat GPT first came out, and this has gotten better, is the concept of hallucinations for the AI to do. And it really comes down to the AI is trying really hard to answer your question, right? And it isn't necessarily trained, especially early on, at saying I don't know, or kind of giving a more neutral answer. It's trying really hard to give you the best predictive word in the sequence, right? So sometimes it would literally come up with completely factually inaccurate things, but be really confident at that. And that's called an AI hallucination. I would say that this is better. It can still certainly be a problem. And again, this kind of comes back a little bit to prompt engineering, where, literally, if you tell it, if you don't know the answer, please tell me and don't guess. Or what is the confidence of your answer that can actually improve the conversation that you have with AI, but it's still going to be an issue, because, again, it doesn't intrinsically know facts. It just knows the relationship between words that tries to come up with, from a probability standpoint, the most probable answer. Dr. Khyati Patel 23:31 And honestly admitting here, I've been hearing this AI two commercial on NPR, and they advertise that they are hallucination free. And for the past few times that I've heard it, I've been wondering, what do they mean by hallucination free? It's not a drug, it's a it's an AI. What hallucinations are they talking about? But now it makes Dr. Sean Kane 23:50 sense, and I also have heard that, and I'm unsure how they could make that claim of zero hallucinations. That's a pretty big claim that I feel like wouldn't stand up with a lot of rigor in terms of, you know, enough people testing the platform and things like that. Yeah, interesting. Dr. Khyati Patel 24:09 Well, we work with patient information day in, day out. Pharmacists do pharmacy students do as well. Obviously, these tools are not HIPAA compliant. Are there HIPAA compliant? Llms out there, Dr. Sean Kane 24:20 there are so absolutely anything that's commercially available that you log into have an account, especially if it's free, these are not FERPA for students or HIPAA for patients. Compliant, even if they say that they don't share the data to be HIPAA compliant, doesn't just mean that they don't share data or that people don't have access to it anytime that you transfer protected health information from yourself to a third party entity. There are very specific things that have to happen in terms of how the transfer of data happens, how it's stored on the other end. There's a lot to it, right? So it's not just, oh, they don't, they don't look at my prompts. That's not enough to be. Be HIPAA compliant, but obviously this has a huge medical potential here. So there are companies that have HIPAA compliant llms, some of which are even integrated into the electronic health record. So there are epic plugins, if you will, that are llms that are obviously HIPAA compliant, and whatever compliance epic has the electronic health record to allow these platforms to exist in the electronic health record themselves. So that would be the more typical example where a healthcare provider is going to use an LLM. It's not going to be that they'll go to a website, but it would be directly integrated into a health record. Dr. Khyati Patel 25:36 So particularly focusing on the practice of pharmacy, and what's the future of pharmacy? What are some of the, I should say, realistic example, or practical example of how the llms could help, Dr. Sean Kane 25:51 literally, if you do a Google search, right now, there's a ton of companies that are using AI for medical reasons, right? So not specific to pharmacy. And I'd say, like 90% of them, focus on reducing documentation and time for typically, physicians, but could be anyone in terms of generating progress notes or making that process more efficient. And a lot of the call to action is that we can give you your time back, to allow you to spend more time with your patients, the thing that you love, and spend less time writing notes, billing your patients, stuff like that. So related to that, you're going to see features like progress note summaries, where you know a patient may have been in the hospital 12 times in the last 12 months, and they've got like, 100 notes. These llms can, quote, unquote, read all of those 100 notes and then give you 100 word blurb about the patient summary that certainly you could go into specific notes if you wanted more detail, but it gives you kind of the cliff notes version of a bunch of notes that have been written about the patient. Dr. Khyati Patel 26:54 You know how many hours we can save doing this with chart reviews? Oh my goodness, yeah. Dr. Sean Kane 26:59 And remember, it's not just the progress. Note that it's reading. It can read meds, vitals, labs, prescription, fill, history, whatever's in the electronic health record. It can theoretically consume that data and then summarize it to you as a provider, Dr. Khyati Patel 27:14 interesting and I assume then it can also search more into, let's say a start of a drug in a patient's chart. I can tell you exactly when it was started, who was started by, what quantity the first time patient got it, you know, all of that stuff. Dr. Sean Kane 27:30 Yeah. So again, to your point, like, how many hours have I spent where I'm searching for when did a med start? Or who, who gave it the very, very first time. And you know, an epic as an example, it's smart enough to know, like, a pixel band is Eliquis. So like, kudos that. Like, there are some synonyms that it can search for, but I can't, in epic, say who first prescribed this patient's Eliquis and when was it? But with an LLM, you can do that because it understands the prompt, because you're asking the question, and it understands the context, which is the entirety of the patient chart. So instead of typing in a pixabane and kind of going through 100 or 1000 entries trying to look for the one progress note that has the thing that I'm looking for, you can just ask what you're looking for, and it can do that hard work for you, which is like a really powerful thing to think about, that you're human text asking a question, not just looking for a keyword, right? Dr. Khyati Patel 28:24 And I think earlier, you talked about scribe features for some of these llms, but I can imagine that they could also help you document some of the nodes in a standardized format, like soaps and Dr. Sean Kane 28:36 such, absolutely. So you know, right now it's very common for health systems to have templated progress notes. So, you know, the provider fills in like four sections or whatever, and everything else gets populated in. So like the vital signs get populated in, or the labs plop goes right into the progress note. But with the use of an LLM, it could provide a lot better context. So it could have those things, but it could also summarize the last four encounters for the patient, and that goes into the template, or even for today's visit that you have for a patient. There's a lot of companies that basically record the conversation between the healthcare provider and the patient, and then, using that conversation, it will summarize it in the template of the progress note, or the soap note, or whatever you're doing. Or, you know, right now, many physicians use what's called a Dragon software where it's text to speech, right? So you talk in a microphone and it types it out for you. Or a dictation software where you have it recorded, and then a human being will literally type it out for you. AI can do that. So either it could potentially summarize the conversation from patient to provider. Or a provider can literally just can freely say, like, Hey, I just met Mr. Jones. Mr. Jones came in with a higher blood pressure today, so we did XYZ for him. And it's kind of freely doing that, but then it formats that in a very standardized way that is in the format of a progress note. So even though the input that the LLM received was just kind of a chat, right, just talking to a colleague about this encounter that you just had with a patient Dr. Khyati Patel 30:07 that is very neat. And I think you mentioned earlier, too, that you know you could ask these llms to provide you information in a third grade reading level. So when it comes to patient education material, or I call it the after visit summaries at the end of the encounter to provide to the patient, I assume that they can also generate Dr. Sean Kane 30:25 those absolutely so it could be in patient friendly language. Obviously the provider needs to finish their note for it to be summarized. But what a great way to make something that is more accessible for patients and very personalized too, like it could be in the after visit. Summary, today, you met with Mr. Jones, and he would like you to increase your exercise from 30 minutes a week to 45 minutes a week. That's a very like specific, personalized thing that can go in there with very little effort from the human in terms of getting that to the patient. Dr. Khyati Patel 30:56 So then Dr. Kane, if I have a patient coming in speaking Tagalog, for example, language that you don't have, commonly an interpreter or translator available. Can these notes or patient AVS is can be described in their language? Yeah. Dr. Sean Kane 31:11 So all llms are really good at translation, and one of the powerful things about how they translate is, again, with that word prediction. Thing that we talked about, it understands the context of all of the words in the relationship of all of the words, not just like a single word. So if you've ever used like Google translate back in the day, where it would provide like, weird verb conjugations or use like maybe not the best word based on the context, llms have much better context awareness. So the translation tends to be better, and we'll get to it. Accuracy is really the primary concern of at some point. You rely on it so heavily if it makes a mistake like that's kind of on you, right? So as long as you accept that it's a starting point and it needs to be validated by someone, it would be a great use of translation as well. Dr. Khyati Patel 31:59 Yeah, speed ups the you know, process a little bit, yeah. What about some of the clinical decision support, especially when it comes to pharmacy related tasks, right? Like screening out drug interactions or looking for drugs that you know needs renal monitoring, for example? Yeah? Dr. Sean Kane 32:16 So again, we're kind of, like raising the temperature, and we'll get to the point where we start worrying about, like our jobs in the future, right? But absolutely so it can do clinical tasks, especially if it has a good training set where it understands and it's been trained on common scenarios and kind of what things to not worry about, what things to worry about, things like that. But it could certainly look for duplication of therapy. It could certainly look at proper dosing. It could look at even med compliance in terms of this patient has missed three of 12 predicted refills like that should be something that gets looked into. It. Could also look at meds that aren't there. So this patient had a reduced ejection fraction A month ago, and they're not on the four pillars of their heart failure regimen yet. So a lot of this comes down to the output of how what the LLM analyzes and gives to you is only good as good as the prompt that you provide it and the context. So if you provide it enough data, and you tell it, these are the things that are most important for you to accomplish when you're reviewing a patient chart, and it's renal adjustment, duplication of therapy and drug interaction, and you provide it specific details of that. It can get really good at doing that way, way, way better than what we see now with drug interaction checkers, which is, you know, almost if, then statements of if they're on a beta blocker and a calcium channel blocker, worry about hypotension and maybe bradycardia for non hydropyridine calcium channel blockers. And it's like a rule based system, right? So this is not specific rules. It's more great in terms of identifying that a drug interaction may or may not happen, and it has a gradient in terms of how important that interaction Dr. Khyati Patel 33:49 would be, right? And at some point, I tend to think that the human factor is still needed to absolutely to kind of take a whole stock of the patient. What situation are we looking at, right like, how, in the grand scheme of things, How concerning is this drug interaction? And then really, what actions needs to happen? Anything? Maybe monitoring, maybe switching the drug, maybe changing the dose, etc. And I think that's where the human factor is going Dr. Sean Kane 34:16 to come in totally and at this point, with the current level of large language models and the sophistication I would view the role for the LLM similar to drug interaction checkers. Now where, and ideally, it would be better. But right now drug interaction checkers, I absolutely need to use my clinical judgment, because it's going to false alert a lot in terms of these two drugs may interact, but in terms of my clinical knowledge of the patient you know as an example, let's say it's worried about two blood pressure meds, but the patient's blood pressure is 180 I don't care about that drug interaction, literally at all, and that's very common to have two drugs that work on the same indication, but drug interaction checkers don't know that. And. And when you use an LLM, it increases the sophistication. But at the end of the day, it still doesn't know all of those details, like what you're talking about, or is the patient having the side effect that we're worried about. When you have these two different therapies, it doesn't know if it's not in the chart written somewhere, or if you didn't talk about it with the patient, right? So absolutely, at this stage, it's still a tool that the pharmacist uses to help extend and to analyze data quicker. But at the end of the day, the buck still stops with the pharmacist. Dr. Khyati Patel 35:27 Yeah, that's good to know. And I think that's just one example of a particular, you know, clinical task that a pharmacist is doing. But when we think about system levels, pharmacy operations, how does the AI help here? Dr. Sean Kane 35:41 Yeah, so literally, any task that humans do could potentially be done by AI, especially as they get more sophisticated. So for example, let's say a new prescription comes in. The AI could be the first check on that. And it could say, you know, the quantity doesn't match the SIG or the day supply, something very simple, right? It could be the first pass where it is flagging areas of concern. It doesn't mean it's wrong or right, it just means that, based on its algorithm, it thinks that this might deserve more attention. So it's a way to highlight for the pharmacist or the technician to be more efficient. It could be something as simple as prior authorizations. So understanding when a prior authorization is needed, it could literally fill out the paperwork. If it knew how to do that, it could fill out the paperwork and even submit it. Or on the insurance company side, the LLM could be the first check to say whether all of the paperwork was properly submitted, if all the documentation was actually there, and if it's not, it could immediately be rejected, back to the pharmacy to ask for whatever is missing. So again, it's a way that it can highlight or be an extender for the pharmacist to help draw attention to areas that would normally need human attention make sure that a thing is fulfilled or done correctly. Dr. Khyati Patel 36:53 And if you're taking votes on enlisting AI for doing prior authorization, my vote is yes. Dr. Sean Kane 36:58 And then the other thing that I came across that I thought was really fascinating was the concept of clinical trial registration. So you know, right now, if you want to be part of a clinical trial, you have to have a provider who is enrolled in that clinical trial. And that can get tricky if your provider doesn't do clinical trials, but you might like to be a part of a clinical trial. So right now, you know it's possible to use the electronic health record to identify patients that should be part of a trial. And we actually saw that with a trial that was published through the VA system where the electronic health record prompted the provider to say, hey, we're looking at hydrochlorothiazide versus chlorthalidone. Can we enroll this patient? And that was a very rule based system where they had to meet certain criteria and then it flagged. But a lot of times, clinical trials have very complex inclusion and exclusion criteria. What a great way to use an LLM to basically analyze a patient without the provider having to go through all of this crazy documentation. And again, it wouldn't be the final check, but it would be a way to say this patient probably qualifies. Is this something that you know, Joe over in Wisconsin, can then call that patient and see if they'd like to be enrolled in this stage four lung cancer trial that nobody knew about in Georgia, but because Joe in Wisconsin knew about it from the LLM that connects the patient to that clinical trial, right? Dr. Khyati Patel 38:18 And that could be certainly benefiting the patients, right? So this is a useful help from Ai. What else can ai do for us? I assume, like, as we talked about earlier, you know, looking for a dose. So, like, medical information, I can imagine that it's really good at, Dr. Sean Kane 38:34 yeah, and, and right now, everyone is hesitant to use llms for medical information because of the risk of hallucinations and accuracy and things like that. But there will come a time, I don't know when it will be yet, where the LLM is more efficient and better than a human using Lexicomp or micro medics or whatever. What's going to happen is these llms will be trained on medical data, right? So not just a general LLM, but one that's specifically good at providing answers to healthcare providers about medical information. And it will have access to every article in PubMed that is freely available. It'll have access to the lexicomp database or every package insert, and when you ask it a question, it's going to have all of those data pieces and put them together and provide potentially a better answer than what a typical person would get when trying to answer on their own. We're not there yet, and many llms actually prohibit the use of medical information, so this would only be for healthcare providers. It would be very likely a subscription that a hospital or health system would pay for, but at some point we're going to get there, and again, it's going to make it easier for that healthcare provider to be more efficient and get the right answer for their patient. Dr. Khyati Patel 39:44 And I know this episode is not about the ethics and whatnot, but I like what you said about just now that this is not going to be like a freely available information where they're going to treat it as Doctor LLM, for example, and it's going to be only available to people who have knowledge. Healthcare providers, and so a patient who doesn't know much about medication or disease states cannot interpret information as provided by the LLM, right. So these are just way too many tasks. Dr. Kane, it seems like the possibilities with AI is endless, but as we are talking about how this is an evolving field. I'm thinking right now, and with the information that I just received, it seems like, boy, in like, 10 years, these AIs are going to be smarter than I am, and what's going to happen to my job. So really, if we kind of summarize that concern, what's next, what's coming up, Dr. Sean Kane 40:37 so what's coming up is the technology is going to get better and scarier, in quotation marks, in terms of the quality of what it can do. So the llms are only getting smarter and better at completing tasks. We're now seeing really good voice generation, where it can talk on the phone to a patient. We've already seen snippets of video generation where it can literally make a video using a prompt, and at some point that that video could be a person talking, so it would be almost like you're having a video conference with AI that looks and sounds like a human, but is AI, and at some point, the lines are going to blur and that it's a little bit scary, right? The good news, though, is, at this point, right? Ai definitely lacks features that human cognition has, right, in terms of problem solving skills and the ability to empathize with a patient or understand the full context of everything. And yeah, it'll get better at those tasks, but it's going to be a long time until it meets or exceeds the capability of the human brain to be able to do that. I also think that humans like human interaction, right? Like we're scared about AI for a reason. It might take some jobs and stuff like that, but more, we value that human interaction that we have, and we want to literally sit down with a human and talk to them, not a computer, right? I think that that will never be replaced, that we'll always value that relationship building that happens with a patient and a trusted healthcare provider, and then also, we haven't seen it yet, but I'm sure we will government regulation of this in terms of disclosure that you're using an AI disclosure of when it's appropriate to have a video based conference with an AI individual or not, things like that. At some point, this is going to get regulated to get regulated to the point where there are some safeguards on it to protect from some of them, the scarier scenarios that people are really Dr. Khyati Patel 42:27 worried about, right? So in a more comfortable state of mind, how would I view AI right now? Dr. Sean Kane 42:34 So I think the best thing to think about is copilot. It's a copilot, right? And Microsoft did a great job naming their version of chat, GPT as copilot, because it isn't replacing you, but you're using it as an extender or a tool to make your job more efficient, better and to provide better quality care, potentially especially for documentation stuff, which again, is like 90% of what companies are using it for right Now, to expedite documentation if it can help you do the tasks that you don't like to do and free up your time to do the tasks that you do like to do. That's a big win in terms of having more time to talk to patients and having more time to do the things that you love and develop relationships and talk about non pharmacologic therapy for your patients that maybe get shortened because you just don't have time. This helps with that. Dr. Khyati Patel 43:21 Yeah, and then kind of putting a positive spin right, like when informatics was getting big in pharmacy, that was an emerging field, but now we have so many informatics pharmacists, so thinking of that term, there will be a need for integration of AI and healthcare in pharmacy operations. And so the new graduates and the students can look forward to having those positions, who would be the leaders in implementing these technologies, absolutely. Dr. Sean Kane 43:47 And you know, for that to happen efficiently, colleges and schools of pharmacy need to figure out, how are we training students and educating students on proper use of these technologies, how they work and then prompt engineering is a great example where if we don't teach students about prompt engineering and how to effectively and efficiently use it, they're not going to get the most value out of these tools, right? So we're going to see lots more AI, and it's getting to the point where every graduate in every healthcare setting needs to have some degree of education about proper use. And we're definitely going to see new fields evolve because of this, something as simple as being the person to program the AI that does the task, that is a clinical task that requires a lot of complexity to it. It's going to happen at some point. Well, Dr. Patel, whether we left on a scary or uplifting note, I'm not exactly sure, but certainly the time will tell in terms of where this technology goes for the audience. If you'd like, you can take a look at our show notes at HelixTalk.com Again, this is episode 181 we also have a mailing list, so if you go to our website, HelixTalk.com you can sign up for our mailing list and get an email every time a new episode comes out. So with that, I'm Dr. Kane Dr. Khyati Patel 44:54 and I'm Dr. Patel, and as always, study hard. Narrator - Dr. Abel 44:58 If you enjoyed the show, please help. Us climb the iTunes rankings for medical podcasts by giving us a five star review in the iTunes Store. Search for HelixTalk and place your review there to Narrator - ? 45:09 suggest an episode or contact us. We're online at HelixTalk.com thank you for listening to this episode of HelixTalk. This is an educational production copyright Rosalind Franklin University of Medicine and Science.