Pharmacy Times interviewed Emmanuel Walckenaer, CEO of Yseop, on the value of understanding the potential of artificial intelligence (AI) in medical writing. Walckenaer expresses that medical writers should receive education and training to feel comfortable adopting it into their work.
About the Expert
Emmanuel Walckenaer has worked as the general manager and senior vice president of Sierra Wireless’s Cloud & Connectivity Services business unit and the senior vice president at Gemalto. He previously held a variety of positions with Esso within IT and business development. Since 2017, Walckenaer has served as the CEO of Yseop, where he leads the company’s growth, bringing the benefits of automation and NLG to enterprise companies worldwide. He has over 25 years of international experience in the high-tech service and business development industries.
Pharmacy Times: How can AI be used in medical writing, and how is it currently used by medical professionals?
Emmanuel Walckenaer: We are aiming to help the medical writers to write all the documents required by the FDA, the EDCTP…So, we’ve got the clinical study report (CSR), the patient narrative, you’ve got a whole bunch of documents that the medical writers have to write. These are based on the clinical data that data statisticians, build tables—statistical tables—which reflects what happened during the trial, and it's the job of the medical writers to actually analyze these tables and write all these reports. It's our job to do content generation leverage AI algorithm—AI models—to actually transform data into high quality narratives, high quality text. So really, we do content generation for a highly regulated industry, that's all we do. We're helping the medical writers to work more efficiently, and eventually, accelerate the time to market for new drug introduction.
Pharmacy Times: A recent study published in Frontiers of Medicine shows a low rate of awareness among medical professional, but a high rate of support. What does this mean for AI being efficiently applied to medical writing by those in the field?
Walckenaer: I would not say AI is widely used, but I think they are more used to deploy AI models for controllers, for risk assessment, and so on. In the pharma industry—for good reason—for what is at stake, you cannot allow any mistakes. There’s a little bit more conservatism because they have to do their job in a perfect way.
The first company to introduce AI to this population, we did that 74 years ago, and this was seen as almost impossible. Due to the complexity of the analysis, due to the perfection you would have to match, the expectation was super high. There's no doubt that this population is both expecting for something new to happen because their job is really tedious. In some studies, they have dozens and dozens of statistical tables to analyze, and they do that manually. So, it's very, very tedious. There's a bit of skepticism, but a little bit of hope that if this works, this will be great. And the good news is, it works.
Pharmacy Times: How is AI currently being used by medical professionals and how might that current use be advanced?
Walckenaer: Today we are working with half a dozen of big pharma companies. I’m proud because this is deployed in production at scale, and we have hundreds of medical writers using it on a daily basis. It's not proof of concept (POC) anymore, or pilots, this is really used on a daily basis by hundreds of medical writers, which is great. We got a ton of feedback, and we are constantly improving our solution, our model. We're working on an application where the medical writers, paragraph by paragraph, can actually finetune with a few clicks what is written. As an example, when you are describing the adverse events (AEs)… By default, we may present the top 3 or 4 AEs, and the writer might say…he has a little slider, “I want more detail,” and it automatically will generate all the AEs. And he may say, “No, no, no, no, I want something more condensed…” so he would actually change the parameters and boom, it automatically would do so. So, it's really a dialogue. I think this is very important for the adoption. If you just produce a draft, and it's a black box, it doesn't know where it comes from, it doesn't have the data. People may say, “My God, should I trust the machine? Of course not.” This is a built-in application, however, learn step by step how to build applications, so that we can actually build this trust step by step.
Pharmacy Times: Can you define the terms “large language models (LLMs)” and “generative AI”? How are they used?
Walckenaer: LLMs, and this language model in general, look at data transformers. They are transforming data into texts, they may transform text into text... They may summarize the text, they may expand the text, they may explain data. You can say, “Okay, tell me what's new in these tables.” They’re really transformers, they can transform text into images sewn into videos, this is really transformation. That’s how we use them. We use different types of models, depending on the type of transformation you need to do when you build a report for medical writers, some transformations are pure text to text. For example, I have in the protocol a big, long description of the study and I want a short version, a condensed version, so we use the transformation to condense it. Sometimes you need to change the tense, condense it, and change the tense. The protocols will do that, we are going to do this. And in the CSR, in the documents who are presenting the results, you are saying we have done that. So, you do some condensing of the text, summarizing the text, and changing the tense. This type of transformation, which of course, will save a lot of time for medical writers, they don't have to do it themselves. It’s automatic, I would say.
Other transformations are data to text, which are way more subtle: you have the statistical tables, you have got the results of your logical trial, and you have to write the analysis. So, we use another type of model we call symbolic AI, or the type of model you can do that perfectly without isolation or omission, which are a little bit, I would say, a disease embedded in in LLMs. So, we were using the right model—and there are hundreds of them—and our job is to pick, train, sometimes develop our own model, so that we have a little bit of all the types of transformation we have to do, and then we apply that for the right transformation.
This is super important, we’re not choosing to use a third-party platform for security reasons. We're dealing with clinical data, which is the most important asset of our company, so for security, the way we are using them has to be 100% secure. Our customers…they say, “We don't want you to train your model with our data.” So, we're using what we call ‘synthetic data’, we're using data from the public domain to pre-train our model, and then with few short learnings to really do the last mile for a specific customer. But security, reliability, cost performance, and the quality of the text you’re rating…this is where you need to have the right model and the right training. This is a couple. If you have a wrong model and an excellent training, you know, it could be okay. Excellent model, poor training could be okay, but what we are looking for is the best model and the best training so that you have perfection.
Pharmacy Times: How might AI help pharmacists in ways that can allow them to pursue critical thinking tasks (e.g., publishing)?
Walckenaer: We're going to provide a draft of some sections of this type of application. We definitely could help summarize some other text, to help you write some portion of it, but this will be the tedious sections or the most repetitive ones, and we could help there. But you still need the pharmacists to step back and say, “Alright, I've got this raw material, but this is my value add.” Our job is to help, whether it's medical writers or the pharmacists, so that they can have all the tedious parts of the job done in a trusted way. I want them to forget that there’s a machine. I want them to be completely…I want to free their mind. I don't want them to be obsessed, “Oh my god, is this correct?” Or you know, “Is this stuff okay?” No, I just want them to say, “Okay, this is a first draft, [I] can trust that.” Now, on top of that, what do you want to add? And this would be the true value add. I want to free their mind, I want to free their time.
Pharmacy Times: How might the perception be on drawbacks surrounding AI, and what is your perspective on it?
Walckenaer: I think AI in general…I don't know the exact number, but I would say for 10 POC, there is 1 solution which is actually put into production, and this is the essence of machine learning. You are trying to find something, sometimes you don’t find something, you don't add a lot of value. What is great with natural language generation or document automation, we provided immediate tangible value. You don’t know how many minutes, how many hours it manually takes to do a report, and you can measure the gain with AI. It's as simple as that. If you used to take 3 weeks, and with AI, it takes you 1 week, this is a gain.
For the documents…we're doing the most automation, it goes up to, I would say, 90% to 95%. So, it's not any more manual, tedious work…you generate the text, and you can have the time to review, ensure it's consistent. I think we are in this process, I'm very confident that this is going to completely change the way the pharma companies are working right now. I'm pretty sure this is going to dramatically change the way pharmacists are going to work. You know, their job is not going to disappear, not at all, their job is going to change. I think training and education are going to be very important so that they really understand these new tools and use them—you know, now it's, you know, everybody's using PowerPoint or Excel or whatever—there will be new tools to help them to do their job faster and better. But training, education, there’s not too much mystery around AI, this pretty simple. You know, there are some limitations, but there are some massive benefits, and we are going to spend more and more time to build some learning sessions, training sessions and so on. Because I think yes, the job will change. Yes, this creates anxiety, but there should not be too much worry. It’s just a new way of working, and frankly, a much more enjoyable way.
