The good, bad and uglyof using AIfor QA RA Compliance

If you are located outside of the European Union, the United Kingdom and or Switzerland, then you need an authorized representative. So, I have a good news. You have find it with EasyMedicalDivision. And if you are also in need of an importer in Europe and in Switzerland, then contact us definitely at info@ EasyMedicalDivision.com. And I'm sure we can help you. Welcome to the Medical Device Med Easy Podcast. I am Munir Alazuzzi, a medical device expert specialized on quality and regulatory aspects. My mission is to help you learn how to place a compliant medical device on the market. For that, I share with you my experience and the one of others on this podcast. Are you ready for your dose of regulation and standards today? Okay, so let the show begin. Here is Munir Alazuzzi from EasyMedicalDivision.com. And today, we'll try to explain to you more about artificial intelligence used in QA area compliance. So, mainly we are all now exposed to artificial intelligence. And the idea of today is, is it good or bad or ugly to use it mainly within our quality and regulatory affairs compliance and use it for our own benefit and maybe create a document that will be sent to authorities like that. So, this is our discussion. So, for that, I have with me an RG Kedziora from Estanda. So, RG, welcome to the Medical Device Med Easy Podcast. Hey, Munir. I'm glad to be here today. Looking forward to the conversation that it is good, bad, and ugly. Let's look at that. But first, before we begin to start with my question, can you make a small introduction of yourself? Yeah, my name is RG Kedziora. I'm one of the two co-founders of Estanda Solutions and Digital Health-focused software development data and AI consulting company here in the US. We focus very much on the front end of healthcare, of what I think of improving patient health and wellness, where we work with startups and medical device companies, pharma, in developing new, better ways of improving health very often, having to go into clinical trial or when they go into the liberal world, go through the FDA regulatory process. So, it makes it a fascinating journey for us taking on new challenges, new ways of improving life of our fellow humans, I like to say. Exactly. And I suppose that we understand that you are also using AI for compliance or those kind of things. We use AI more and more every day. I make a huge advocate of it. If you are not impressed by what AI can do, then you're not really using it. Exactly. So try a little harder. So, my first question is mainly about also this topic about AI. As you said, now we are all exposed to that. So, are you seeing or hearing maybe you are identifying some big increase of use of AI also with QA error compliance? So, is this something that is now becoming like normal or it's something that is starting to be new or something in this discipline? Yeah, I think we're, it is new. The technology is new. We're all still trying to figure out how to use it in the most effective fashion. But we're very quickly moving into it's becoming the norm. Like that you need to use this now to improve your productivity and efficiency. And that's really what this technology is about. Internally, we started in low risk areas and you see this a lot. How can we use this in low risk areas, particularly training? So instead of just reading documents or then we can create quizzes with those documents to assess the learning abilities or create different modalities of that information. There's a platform out there that Google's notebook LLM like it will read a document and create a podcast for you where two people are talking about that content. It's fascinating. So now it's, it's easier for us to create multiple modalities to help people learn and understand how to use those documents. And then you can, as you train systems with and about your process, you can have it asked questions to better understand and figure out how to do things. It can create documents for you. So yeah, it's huge boom for creativity, productivity, efficiency. And it's really about freeing people up to then do higher value tasks. But I suppose here the danger or if you are not using that correctly is also the fact that you should have the know I suppose I mean you tell me but I suppose you should have still the knowledge of what you are doing with AI for career compliance. It's not like you are completely new to the business like a startup company that is starting fresh and they are just using AI to create their quite management system or the create their technical documentation. If they have no clue many what how it looks like normally. Yes, I am a huge advocate of it and that's the good side of the as I am the beneficial side. The bad and potentially ugly side of it is it does make up stuff. It's not very good at saying I don't know it's there and inherently to try and please you so if it doesn't know the answer to something it's going to have a tendency to create something. So you have to think of it like an intelligent intern perhaps with a PhD because it is incredibly intelligent and capable of answering questions but it makes stuff up so you do have to review everything that it is producing and we're seeing this across multiple industries where people are generating reports. You go in there and you look at it and you are like wait a minute that reference doesn't exist. So yes, you do absolutely 100% have to cross check what it's producing. And as you said if you're new to the industry it can do a lot for education and helping you get up to speak but you can't rely on it because you're going to look really silly if you go to the FDA with some documents that just don't comply with what they're looking for or bad reference within that documentation. So hugely helpful that you do have to be cautious. As you mentioned you mentioned not book a land that I'm also using for from Google there is chat GPT, Jimmy night, Claude, Grogg there is I mean there is a lot that are coming so is there a tour within your knowledge is there any specific tools that mainly quality and real true affairs people should maybe use because it's more accurate or no for the moment there is nothing really like that that is existing. And generally accepted large language models that chat to be these Gemini's they they bounce around you know as they each release a new model they all start getting a little bit better than the other as we're recording this chat GPT 5 just came out. Now a little bit better than the other ones chat GPT I think because it was the first to the market has the most usage you know the most widely accepted but you could use any of them for this capability but the key if you can in here have the opportunity and there's numerous companies that are doing this is if you can access or use a purpose built model something that we specifically trained on the idea of Q a R a in compliance such that it will be. More accurate that company that's putting that model out will have tested it and making sure and put guard rails in place to help prevent the idea of those hallucinations so yeah I think for that learning the experimentation productivity you can use some of the general purpose ones. If you have access to one that's more specific you're going to be better off and as you talked about companies because I don't know for the moment if companies are. Or authorizing you to use chat GPT or use those those AI so is there still I mean I don't know now what are there maybe you know about that about the privacy policy or something like that from those tools so is something that is safe that we can provide some. Private data or confidential data because I need to get some information or it's something that know you should really avoid that because can be a problem in the future. Meet Alex he needs to create an electronic instruction for use. With Easy I view he uploads the document and in just second gets a QR code. The code is printed on the packaging and chips with the products. The user receive it, scan it and the I view appears fast and always up to date. Simple, fast, paperless, easy I view the smart way to monitor I views. Yeah so you can go to these organizations and create your own private version of these your own models of these two better protect your data so I would not put in. proprietary information into a general chat chat because you're not entirely sure what's going to happen. I think that's low risk from the perspective of things. There's discussions around if I'm Farma A and I put in some proprietary information, is Farma B going to go and find that information and understand our trade secrets and everything. I think that's a little far-fetched. I mean, the realm of possibilities, yes, it's in the realm of possibilities. But the key is that you can spin up your own large language model to ensure yourself and make sure that it is protected and that doesn't happen. So there are guardrails and protections that you can put in place on a personal perspective. Yeah, I don't put in my health information. That's identified as me. I don't say, "Hey, my name is RJ Kessier and I'm a 50 plus year old. I'm not doing that." But I do ask questions. So it's also because recently we heard about the platform called WeTransfer that says that all the data that you will be providing to us can be used for training our AI model, etc. So it means that if I transfer some data through WeTransfer, so those files can be included and then used to train their model. So I mean, they are open with that. And it's hidden somewhere inside the terms of use of privacy policy. So it's why asking those kind of things to be careful of that. Yes, absolutely. We all hate reading the fine print, but with these technologies, you want to read the fine print and you can even search for those words and what they're going to do with your data and are they going to use it for training purposes? Some do, some do. So yeah, I would mean towards ones that are going to protect your data and not use it for training purposes. So now if I am in a project in a quality, regulatory project, etc. and I want to use maybe charge GPT or any AI for my work. So what are the, what is the methodology that maybe you will be suggesting to people to reuse that? Because maybe you have some people that say, where should I start? What should I do? How it will be working? Should I start at the beginning and then it will give me everything or should I come at different stages? So what exactly is the, maybe one methodology that you would be suggesting to people for using that? Yeah, you have to think of these large language models, the charge GPT as extremely knowledgeable interns. But when you think of that intern that you're going to use, they can provide value, but you need to give them really specific instructions. So when you are using a large language model, you want to break things down into individual text and have it focus on a specific thing. You can ask it to create a project plan for you to create a quality assurance plan or my test plan. You can have it. Do that. But if you don't give it the necessary context, just as if you didn't give an intern the necessary context, what it's going to create isn't going to be as valuable. So the more content you can give it, content and context, and examples of what you wanted to output, the better off you're going to be, the better job it is going to be. So think of it as you're generating requirements, you're reviewing documents, creating test cases, test data. These are all things that the AI can do for you to help you be more efficient. We've done a lot of work in the blood glucose field and it's like, well, create me a blood glucose trace for 24 hours of a patient with diabetes and it generates a trace. It was fairly, I didn't give it a lot of information. It was very generic. While now include the idea that they're eating three meals a day and you generate that trace. Now you see that trace of data that it goes, blood glucose starts going up after a meal. It's like, okay, make the lunchtime meal carb heavy. You can see it goes up a little more and then you can add in factors like patients taking this medication and it adjusts that. Is it going to be as accurate as a human? No, but you were able to generate this data very quickly, very efficiently and it's that 80%. Then you want to make sure that you take your expertise to get it to that last 20% to make sure your test and all your edge cases and things like that. That's where this technology excels. Here, when we are in the projects, then we can say that you still need to be in control of your projects. For example, you are in the design and development for example, blood glucose device. Mainly, as you mentioned, you make a project plan but you still have to be in control of what is provided and maybe update it and readjust and rediscover it. We hear a lot of people now using chat GPT for everything. Like, "Oh, watch. Should I take this decision or that decision? Should I move here or there? How to do this?" I mean, it's start to be crazy when we see some things on social media that are why they are using that. It's less to be crazy. Here, the idea is also less being controlled of what we are doing and use that as an assistant, as you mentioned. Yeah. I'd like to word assistant. It's a partner, companion. It's there to support you and help you. Yeah, I really like thinking of it as that intern. Like, that you need to give it the instructions and the better you can do at those instructions to better the output that you're going to get. And still verify what they are giving back to you. Oh, yes. You're not giving you an intern a task and they can do great work. We've had lots of interns over the years that can do great work. You still got to double check it. Exactly. And here now about the new, and you talked a bit about that before because many giving contacts, giving information, etc. We hear now about a new discipline which is called prompt engineering. So can you explain maybe that for people what is prompt engineering because maybe we'll have more and more people that have like this qualification within their CV like I'm a prompt engineer or something like that. Yes. I am a prompt engineer. It's fascinating as this technology came to like that term quickly became special. It's like I have this special skill of prompt engineering. And it is. There is a trick. There's some magic. You need experience with it to do a better job in prompt engineering. It's just a fancy word to you having a conversation with the AI. This is how we interact with them. It's having a conversation. It's like you and I, you can go into chat with you and talk about this podcast. You can say, okay, here's my topic. Tell me about AI, UA compliance and the impact of AI. You can have that conversation with it. But to get the most out of it, that's where this idea of prompt engineering came in. It's like how can you make those prompts your conversation better? How can you make the AI produce better results be more efficient to produce something that that's meaningful? It's fascinating as I walk around at different conferences and talk to people and some people start playing with it. I am not really impressed. Well, you're not doing that prompt engineering. You're not thinking of it as like the intern that you have to help it along the way. And just from that perspective, talk to it and ask it to respond like Bill Gates or Bill Murray or Bill Clinton. And you will get three distinct responses from the technology just by telling it to act like one of those individuals. And so the first and biggest key I think in prompt engineering is for any exercise you're doing, even a simple prompt, tell it to act like an expert in whatever field or thing you're working in QA and regulatory compliance act like an FDA reviewer. And it will improve its response. It is really that simple act like a and then do that. The second thing I have talked about a lot about it is the context. The more context you can give it, the better understanding as we generate requirements as part of brain storming exercises or test cases. It's like act like an expert in ISO 1345. And it changes its response to in creating that and how it writes things out for you if you're using user stories. It's like, okay, write a user story. And if you can provide an example of prior projects you worked on, it's going to do even better. And then that's also the examples of the output. Here's what I want my output to look like. For your, as you're generating test data and you might need it in a specific format, it can do that for you. But I think ultimately, prompt engineering and the use of this technology is the first technology that you can use to learn how to use it better. So if you pick up a calculator, you can use a calculator to do incredible math. But you can't use a calculator to really teach yourself math. You need some external source with these AI's and prompt engineering very frequently. If I'm developing some complex prompt, I'm asking it to help me create a better prompt. Yeah, I agree. I'm like, yeah, engineering for me. And very frequently it's like, hey, chat to ET, ask me questions to better understand this task. And it's not simple. I'm using it to be better. So you have this cycle going through. And it's also when you do ask questions, if you just say ask me questions, they're probably through like 10 questions out here. So you can be like ask me questions one at a time. And as you respond to those questions, it then is learning more and more about you. One key thing is, as we talk about this and answering these questions and that in-term sort of, yeah, the idea, where you do work with an in-turn, they're going to learn and retain this over time. The LLM's large language of knowledge, while the Chinese, they're getting better retaining this information about you, but don't typically retain in long-term. So you do have to keep telling it this information again. Because you told your in-turn that you are ISO 13485 and trained them in what that means two days from now and you're using chat to ET, you need to remind it every time you need to say, hey, we are ISO 13485. This is how I want you to respond. So I'm just, you know, important elements to be able to do. Exactly. And I think, yeah, so many of those technologies have some good, if I can say so, to help you. We talk maybe about some of the bad and ugly, but do you have some warnings or some things to be careful of this or be careful of that so that they have really the full guide on how to use that? Yeah. And we touched on the one idea, the hallucinations. The problem with the challenge with this technology is that not just that it hallucinates or makes up facts, it sounds very authoritative. It sounds like it's right. When you're sitting there reading, you're like, oh, yeah. That's true. That out. And so you have to be cautious of not lower your guard and be like, don't get lazy with it. So if you start using it, you know, a couple of months go by, it's like, okay, it's done really good. It hasn't produced any bad information in the last two weeks. I'm not going to check this. That's where you're going to get caught. So it's good for brainstorming, you know, producing documents, summarizing information, but double check everything. And above all, I think don't relax. This is some of the things that we're seeing in the current research and this idea of critical thinking that it's breaking down our ability to think critically. There was an recent article on, they were in Europe somewhere, but they were doing assessments of medical images. And it's like these doctors who were very good at assessing medical images after using an AI system for a couple of weeks, when that AI system was taken away from them, their ability to assess those images had dropped down a little bit. So yes, you definitely have to be aware of that and keep that critical thinking going because if not, there's little sneaky hallucinations can come in there. So you do want to be careful of those. And maybe the other. Sorry, the other aspect of that. Many way to have said now is important is also the fact that I heard that also for maybe for children that can maybe grow with this kind of thing. So there will maybe have this easy way to create things or do things. But when you remove that from them, they will not be able to create anything like to create a text. For example, we talk about students. Sometimes now students are using also a judge to create their own text. And if you remove that from them, ask them to create that themselves. And they are not able to do anything because they are not, they were not trained or used to do that. Yeah, we do. It's just I use the analogy to calculator. It's like when calculators were becoming available, everybody panic, the education system, panic, and was like, oh, math is done. We're never going to teach math again because we have calculators. We now know that's not true. We all still teach and learn math is very valuable. And you can use a calculator. We're going through this now. This technology is in its infancy. And we're still trying to learn and figure out how to use it to our advantage, how to incorporate it into education because it is powerful. But you know, my daughter, we just got her first job at a college and she's a seventh grade English teacher. She's going to have an interesting challenge. But I think teachers a lot will see the evidence of students using it because it does have that tendency to create sort of common content, use common words. So it's interesting. When I use the system now to generate content that I then always check, I do have this like list of words that new chat GPT frequently uses. It's like, well, don't use these words. You know, one of the newer things is the double hyphen. Yeah, exactly. All of a sudden, yeah, they started the areas that started using double hyphen. So now I just have a little caveat in my instructions with chat GPT is like, don't use double hyphens. And so it's not in there and I don't have to worry about that. But yeah, it's useful for brainstorming. The other thing that you do have to be aware of is how this was trained, how these systems were trained. And I say that because they have a tendency to emphasize our biases. Yeah, true. I heard that. Yeah. And it's because that's where this data is coming from. From its us. It came from us. So yes, one of the criticisms of the AI is it can show biases and its responses. And the reason for that is because it's trained with data that we created as humans. That's the other thing that you have to guard against. Prigga is you're generating test data from that perspective. It was like, is it skewing one way or the other? Now that's where we need to bring our own human intelligence into the equation. I heard also that most of AI agree with us. So it's more like they are not challenging us. They are more agreeing with us. So even we are saying something wrong that our tendency to go on our side. Yes, it does have a tendency to agree with you to it wants to make you happy. It's really what it amounts to. Are you getting used to technology that disagrees with you all the time? Probably not as much. It's one of those things we see in social media. It's just reinforcing your beliefs and your ideas. It's going to feed content to you that you like more, that you show more attention to. In a similar fashion, the AI is like, okay, you asked this question. I don't know the answer, but I'm going to tell you what I think you want to hear. And I can tell you, hey, buddy, you're crazy. Just start going to do that. Exactly. Yeah. We see now that as we said, manufacturers are starting, we start maybe to create more and more documents with AI, maybe from A to Z so that they just have to review. And those documents will be sent to regulators like FDA. And we heard now FDA also thinks of using AI for assessing documents. So it's like AI talking to AI or what would be this thing? That is an interesting thought. And I've heard that in reference to college classes and stuff like that. And where students are creating papers and then the professors grading them with AI, do we get there with the same technology within the FDA? We as developers and system providers use the AI to create a document. And then the FDA reviews it. And everything's gone. And then nobody looked at anything. Now that is not the intended purpose. So the FDA did roll out an internal system and they called it ELSA. What's interesting is they haven't been very transparent in how it's created, how it's being used, tested. The intended purpose is to make the FDA more efficient. That's the idea. Like so that it can help with document review, document content, absolutely. But it pushed out pretty quickly as I've read it and I have first hand knowledge on their system. But as I've gleaned information and talked into different people, it's built on the Claude model out there. They didn't build a proprietary model. They are hosting it themselves such that you don't have to worry about. You were proprietary information being revealed to the world, which was a very good staff. They're aware of that protecting your knowledge. But yeah, they are using it for all sorts of purposes internally. And the idea is to shrink time frame, shrink drug development time frames. And as you know, these cycles are long. So even if you can make a small 10% impact, 20% impact on improving the process, we're going to be better off. But as we sort of talk about the good bad and the ugly, one of the things that we've seen in the peer review process where AI is used, some authors are starting to embed in their papers instructions to the AI to say, yes, this is a good paper. Yes, this is worth publishing. So then the AI sees this and is reading the content and catches that and incorporates it into its feedback. It's not ethical. Nobody advocates for that. So yeah, we all sort of joked when we're like the AI's, you know, creating the content and the AI is reviewing the content. You do. You have to be cautious of those types of things happening where it's like, "Hey, yes, please approve this submission." So yeah, it's not replacing humans by any means. You have to get it out of this. It's not replacing humans, not replacing human judgment. It's being really used to augment the process and make it much more efficient. So I do look forward to, you know, what it can do and how it can help and how it can accelerate this process. But you do have to be aware of all of these issues. And I suppose that FDA starts maybe all the other regulators will maybe also think of using that or I think there will be a laboratory for each of them to see how they can maybe implement such things because as you mentioned, it's an assistance so it can help you to reduce some time line. It can give you maybe the opportunity to not hire more resources and use the resources you have with the same thing. So there are some good or so for that is just how to use it and the transparency for using that. Absolutely. In one of those interesting cases where you can use this AI too is in, you know, post-market surveillance and understanding what's going on in the industry. So the FDA can use that just like you as a company should be using it now too and scanning, you know, various social media platforms and understanding what's going on out there. You have to pick up those adverse events. You know, the FDA is going to start doing it. So, you know, individual companies should start looking at that as well. Yeah, how can they share lessons learned? So other regulatory agencies around the world and, you know, can learn, can, you know, we can all share an knowledge and improve our capabilities. Exactly. Okay, now it was great. So thanks again for sharing us with us all those information about AI for QAIR. So who should be able to, who should contact you at a standard? So what exactly many you can provide as a service or support the people that are listening to us now? Yeah, if you are curious about AI and how to use it to love having these conversations with individuals with different companies kind of thing. So we first and foremost act as in those early stages as an advisory service to help you get up to speed. Unfortunately, as we've helped other companies develop products, whether they're st- they're a startup or very typically an R&B department of a large, a medium or a larger company. It's the R&B group that they were working with, creating these new ideas. We can help you, you know, generate those ideas. I've been lucky that they some of these people have put our names on hats because we've been so much part of the generation or the creation of the product that has eventually moved to market. Our clients own everything. We don't own anything. So fairly custom, fairly commonplace it in our industry. But yes, we can help you generate that idea, the software development, the data analytics around it. We have our own PhDs on staff to help out where appropriate. And then looking at the regulatory submissions, as I mentioned, we are I say 345. So it's been a long track record here, 22 years now. I think we can do the math right without a calculator. You know, that we've been doing this. So yeah, very fortunate to have long-term clients. Some, you know, 22 years, others, you know, 15, 12, 10, you know, long-term clients. So really, really enjoy helping others. That's what it's all about. Exactly. How can we work on these projects and help others help improve, you know, humanity? Exactly. So where we can follow, can they follow, can they follow, with you on LinkedIn or on your website? LinkedIn is, yes, LinkedIn is the best place. It is RJ, Kedzerah. You know, but if I'm a Richard Kedzerah, it finds me as well. But yeah, LinkedIn is the best. I'm always out there trying to share content about what we're doing, what I'm learning, what we're seeing in the industry to, again, just help others as we can. Exactly. Okay, so three, thank you very much, RJ. So thanks for all the information. It was really great. Thanks for the conversation also, but a lot of love about potential possibilities that will come in futures. So we'll see if it happens or not. And yeah, so for people that are listening to menu, you can go also on the show notes and check many, all the information about Kestanda or RG Kedzerah. So that you can maybe contact me directly on LinkedIn. Okay, RJ, so thank you very much and wish you a nice day. Thank you very much. Thanks for listening. So if you'd like to see this episode, please provide a review on the platform where you are listening to it. And also don't forget to share with your colleagues. Thank you very much.