IQVIA Holdings Inc. (IQV) Earnings Call Transcript & Summary

Hello and welcome to our webinar day, AI for field medical nodes and how this drives true insights for medical affairs. Today's speakers are Reinhard Berkels, Head of Medical Affairs Strategy; and Hywel Evans, Director of NLP, Medical Affairs and commercialization. You can find their full bios on the right of your screen. This webinar is being recorded and will be e-mailed to you within 24 hours after the event. At the bottom left of your screen, you'll find resources relating to this webinar. Time permitting will follow the presentation with a Q&A session. So please submit any questions you have using the Q&A box on the left of the screen. We'd really appreciate it if you could take a few moments to fill out the survey on the bottom right to your screen so we can improve our webinars in the future and understand a little more about our audience. Hywel, over to you.

Thanks, Melanie. So great to be here today talking about this very interesting subject. Today, Reinhard and I will talk a bit about 2 aspects of Insights generation within medical affairs. But firstly, Reinhard will talk about the context, which is vital when thinking about Insights, how to generate impactful Insights within medical affairs? And then I'll talk a bit about how we apply different technologies, including AI and NLP technologies to get value from MSL nodes data. Before we get started and before Reinhard introduce himself more fully, let's have a quick poll. So the question is, do you have an end-to-end Insights process mapped out within your team today. So I'll take a few moments to consider that. And then we'll take a look at the results when we've got some responses. We've got a couple coming in. I'll just wait on a couple of seconds. We got little bit of a flurry. All right, we got just about half the attendees providing an answer there. More coming in. All right. Let's take a look at the results that we got so far. Interesting, we're about a little more in the Yes. We've got the approaching 60% in the yes and around 40% with a no. Okay, fantastic. Well, we'll refer back to that as we go through the presentation. But right now, I'll hand over to Reinhard to introduce himself and also talk to us about the context in medical.

Thanks, Hywel, for the nice introduction. I'm Reinhard Berkels. I'm leading our Medical Affairs Strategy and I spent the last 18 years in various roles in Medical Affairs. And I just wanted to kind of set the scene a bit about Medical Affairs and Insights, the importance and so on. So how I see Medical Affairs is kind of the bridge between R&D and commercial. It's deeply embedded on the left-hand side, giving input into the TPP then owning the scientific communication orchestrating the real evidence generation, engaging with the KOLs and of course about scientific -- owning the scientific communication. But then also on the right-hand side being a part of the brand team, being deeply embedded in the brand team as a true strategic partner with business acumen. And in order to do that, Insights are absolutely crucial. And what you can see here is a bit -- I wanted to start first of all with what is an Insight because there is a kind of an inflation about the word Insight, everything is an Insight, which is not true because 90% -- probably 95% of what you see is called an Insight is actually only information. So it's the why behind the what. So to give you an example, a physician, if you were in an interaction with the physician, a physician would say, I treat this subpopulation of my patients with the drug class A and subpopulation with drug class B. That is what. That is the information. What you would like to know and to understand is the underlying reason because then if you understand that, for example, the subpopulation, A is elderly people, and there might be, let's say, interactions and so on. That's the kind of the reasoning why the patients are being treated in this manner. And that's what you constantly need to do to dig for the why behind the what and the Medical Affairs Professional Society also states that medical Insights are a cornerstone of the Medical Affairs value proposition. And the MSLs, as you will see later with Hywel, play obviously, a critical role because they are interacting with the top experts at a daily base. And yes. And so the Medical Affairs Professional Society has done a survey. And I think the results kind of showcases exactly the same as our poll here because that survey found out that half of the organizations had the impression that the Insights and, let's say, and the informations are very fragmented across the organization and not really integrated. So there's no holistic process around it. There are different data lakes and I will come to that in a minute. But before I come to the different data lakes, who is responsible for gathering Insights. I mentioned MSLs and Hywel will also refer to that later on. But it's not only MSLs. It could also be medical information. It could be a great source of Insights if you ask for the why behind the what and somebody, a patient or a physician calls, if you really understand why they are calling, then that can be a great source of information and also Insight. But at the end of the day, it's literally the whole organization who can gather information and insights and that's also the sales force can be a great information and insight source. But everybody who is interacting with external HCPs and doing literature research and so on. Everybody can contribute with information and insights and especially if the insights are actionable. So where do I find Insights. There is -- as I just said, there are different sources where you can get insights. You can get insights from KOL interactions; from digital opinion leader, you can do social media listening, you can listen for sentiment there. Medical information I just mentioned is -- could be a really good source. Then advisory boards, that's a classical piece for Medical Affairs, there you gather information and then you try to share it inside the organization, medical education and online medical platforms. Then, of course, patients if you -- Medical Affairs talks to patients or patient organizations, they can be a great source because understanding the patient journey and unmet needs is absolutely critical. And then, of course, published sources in the public domain and of course, also congresses where you can interact with physicians, but also listen and see what the competitors do. Having said that, where can you apply Insights? And I think they drive a whole lot of different things across the whole life cycle of the molecule. But starting very early on even in the clinical development, when you look at the proof of concept, understanding the unmet needs of the patients is critical, the patient journey, that's where insights are absolutely key in order also to drive the TPP of the molecule. Then it is about understanding how is the disease awareness. What is the patient journey. And then in the prelaunch phase, it's about understanding the data gaps, understand the unmet need. And also, as I just referred to at the very beginning, treatment behavior and why patients are treated in a certain manner. What is behind certain treatment algorithms and then, of course, medical education, which ones -- which are, let's say, data gets, which -- where is the disease awareness lacking, what do I need -- what do I need to do in order to address that. First of all, I need to understand that. And then further on also understanding the need of real evidence, which ones are strategically important and critical and need to be generated. Again, medical education and of course, also positioning and the strategy around the molecule. Having said that, there are different kinds of Insights. There are, of course, those spontaneous Insights I was just referring to when you're having a conversation with an external expert and you understand the why behind the what that you have been discussing. That's a spontaneous Insights. But then there is also something that is called Key Intelligence Themes or Key Intelligence Questions, which is I think a more structured manner. That's when the brand team or also the Medical Affairs team sits down and reflects on what are key questions. What do we really want -- what do we need to know. And then the MSL can go out there into the field and in a conversational manner tries to dig into these questions during an interaction and then brings that back and collects that and that then being collected internally, so then you can answer some key questions in a more structured manner or you could even go more into the qualitative -- quantitative piece and do something like primary market research with these questions. And in order to do that, you will be able to together Insights in, let's say, in different manners. And then of course, there are also, as I said, different data sources. The question is then how do I make sense of that because that's one of the key questions of Medical Affairs in general. We've got tons of data. We've got different data lakes and how to make sense of that. And now I'll hand over to Hywel who'll tell us a bit more how we might do that and how technology can help.

Thanks, Reinhard. So we'll have a deep dive now in terms of specifically MSL nodes, but as a bridge from what Reinhard was talking about. This slide talks to some of the data sources that Reinhard had mentioned and some of the different teams and also some of the areas in which those Insights land. But what we're talking about now in this section is really this middle part. It's about how do we enable that process of taking data that we capture and taking data that's available externally or from other sources and helping turn that data into a useful form so that we can have actionable Insights. And that could be in the form of understanding disease in patient journeys. It could be in the form of strategy development, improving engagements as well as care pathways. And so really, there's a huge potential to become data-driven. But in seeking to become data-driven, that process of bringing in data and turning it into usable form and that is what we'll talk a bit more about now. So at the beginning, we asked a question around the overall process. And this is the second and last poll for today, which is about asking, are you capturing MSL field nodes in a digital form for collection and analysis. So please take a moment to complete the poll. While you're doing that, one of the things that we found is that there is a varying degree of data capture in play. So this poll will hopefully shine some light on what that's number is at least for the audience today. And this is really an important question on a couple of fronts. One, of course, it tells you whether it's possible to do some of the things that we're about to talk about. But it also speaks to some of the maturity around data capture and also some of the acceptance of digitizing this sort of information and then applying analytics to it, which is certainly increasing, and we see that, but it's not universal yet, that's for sure. Okay. We have a good number of responses. I will just skip forward here. Interesting, interesting. So 68% do capture MSL nodes in a digital form and around 32% do not. I think that's very much in line with what we're seeing. And I think sometimes we see even those who do capture in the digital form do analysis to a varying degree on top of that. So there's still a lot of potential to look at ways to use technology and take advantage of this very rich data. That's also data that you have some degree of influencer control over as well. So when we think about technology for dealing with MSL nodes, it's a family of technology around NLP, or Natural Language Processing. And the reason that we use this technology is really because MSL nodes are often richer than the structured data captured during or after an interaction. And that's true of many data sources that Medical Affairs deals with. And that's because things like scientific literature, congress abstracts, social media and patient forums as well as, of course, the internal nodes we're talking about today. These all have one thing in common, which is that they are rich and then they are unstructured. So that rich, unstructured data needs Natural Language Processing in order to deal with it at scale. So on the top left-hand side, you'll see different data sources, and we connect to those. And those can be in any form documents, there could be snippets of texts from databases, et cetera. We then apply this technology to extract information from within that text. And that can take a form of themes or topics, specific words, disease terms, treatments, therapies, drug names, all sorts of different information locked away within that text. And then once we've extracted that information, we can then visualize it or push it back to a database to be used in other ways. In terms of the type of questions or areas that we address, at the bottom of the slide, you have the challenge areas. And so the top 3 are really the most common that we see in Medical Affairs. That's the use of therapy areas specifically, science and research from sources like MEDLINE, PubMed, ClinicalTrials.gov, also Preprints. And then the focus of today's discussion is the interactions data. So data that is generated as you interact with key stakeholders. Those could be the broader HCP community. They could be KOLs and DOLs, as Reinhard mentioned earlier. They could be the digitalization of output through [ output ]. So there's lots of different potential there on interactions. Then the third pillar is online social and events. So what's happening out there. And those are quite structured information like congress abstracts. And we have database of those in the form of the [indiscernible] database who joined the IQVIA family last summer. And so that's a really rich data source in terms of emerging science and linking that science to key opinion leaders. And then there's social media and news, not necessarily new as a concept, social listening and news feeds have been around for a while. But the application of deep text mining approaches that really helps to unlock further value from resources. There's 2 more mentioned on the slide, Labeling Intelligence, which is primarily a regulatory use case, but for sure, the Medical Affairs teams are interested in looking and understanding that label information as well as other [indiscernible] sources. And then Reinhard also mentioned at the end of his section that in terms of surveys and market research, there can actually be a role to re-mine that data. So you can -- you can be very structured about how you ask questions, but you can also open up questions in a way that is an open-ended format, and you can mine those results. So to talk a little bit about where NLP fits in, now let's get into some concrete examples. So when considering MSL nodes, we are looking often for known topics or areas of content that are key to our day-to-day operations as well as achieving our strategic goals. And so for a particular therapy area, a particular drug, discontinuation and switching might be and is often a topic of interest. So what the process does is it ingests the entire node and then the linguistics runs through that node in order to extract the key information. The first piece of information here is that, yes, the MSL node contains information about this continuation. Second piece of information that we get out is the details around which drug has been switched to which drug. And so the direction of that switch, which drugs are involved, those are all useful pieces of information. And this helps us zoom in on specific nodes that are related to this topic. Each node can have multiple topics associated with it. So for example, we have 5 different topics that we're looking for. And a node is very rich and may contain references to all 5, we will tag that node as being relevant to all 5 topics and where possible, we would extract as much information about each topic as we have done here. And then at some point, then the user will then find themselves wanting to also read the nodes themselves. And I think that's a very natural process. This is not about replacing the expert review of the Medical Affairs team. There's nobody best placed to extract true Insights from MSL nodes than the Medical Affairs team working closely with their colleagues in the field medical team because they understand therapy area best and have the detail. But often having to manually do that work on a regular basis and review every single node can become cumbersome. And so this technology assistance can help to enhance the data and make it more usable. The second example that we've got is around cost. And I really put this on because it's a very simple example, but essentially, we can codify any topic that's of interest. So we have a series of topics that we commonly look for. We'll come on to that in the next slide. But we can also look for topics that may or may not have extra detail. And so we're building up a patchwork of as much information about the nodes as possible without forcing any particular rules to be met around the depth of quality there because really, at this point, we're trying to find valuable and useful data points from within this amount of text. We've got some questions that have come in. I might address 1 or 2 of those as we go on the next slide. So like I said, there's no limit to what can we look forward in the data, but there are common areas for teams to consider. So here are some of the areas that we commonly look for. The blue are primarily medical and treatment focused. And these are usually on the list or at least most of them are on the list of key areas for understanding. We also have slightly more patient-focused as well as more physician-focused data points that we look for. So for example, the name of common trials and studies that may be being discussed. They're also looking for things like standard of care. And then from a patient perspective, anything that can help us understand if there are such determinants or practical reasons why something happened and they may explain, let's say, a switching event that may be important and maybe not generalizable. So maybe there are special cases that we need to take into account that maybe shouldn't influence our overall decision making. In other situations, there may be a critical mass of those examples, and we may have a case that we need to address here in terms of education and awareness of evidence that sort of thing. I can see a question here around [ pretext ] as many times limited due to the compliance risk. How do you influence the organization to [ set up pretext nodes ]. Yes, it's a great question. And it really talks to the poll results. Reinhard, if you got any sort of [indiscernible] on this one

Yes, I do because that's obviously, it's an age old topic, and that's often why in a lot of organizations, the salespeople do not have pretext sales. Medical needs to be well-trained in order to, let's say, first of all, understand what the information is, what the insight is and then really trying to write down insight and also being aware, okay, what are the compliance rules and what is about, for example, of adverse events and so on that should not be in those pretext fields. So there's a couple of things where you need to have some proper training. But I think this is very valuable. And as Hywel said, NLP can help majorly to, let's say, aggregate the data to a meaningful chunk so that then the expert can have a look at it.

Brilliant. Thank you, Reinhard. Absolutely. And on projects, we see the discussion go even beyond that of should we capture this type of information, should we capture more, but also into how do we train field teams to really ask the right questions and enrich their own data as they capture it. So it's a fantastic opportunity, but obviously, care is needed. I think this suites the slide very well. We've had another question. How important is [ taxonomy ] for the effectiveness and efficacy of NLP solutions and overall insights process. I'll probably take that one, at least for the NLP part. A lot of our methods are heavily reliant and rightly so on very rich ontologies. And these are large databases, we have over 6 million entries in our database of terms and terminologies, and that helps us do a couple of things. The first thing it helps us do is to be able to create results quickly. You're not looking for examples and building up lists as you go. A lot of that information is already within ontologies such as NCI, MeSH, MedDRA and many others. We maintain those ourselves as well, and we create enhanced ontologies that gives us that control over that very rich source. So the first piece is we are reliant on those, and we use them at a good effect to get those resolved quickly. In terms of the taxonomy of a specific therapy area and for a specific team, we would take a slightly different approach. And what we would do is we would work with the experts within those teams to really understand what are the terms and terminologies that are important to get exactly the right result for our therapy area. And so we would create custom taxonomies or customer ontologies in those situations. And then for other situations, we may actually turn to machine learning models as opposed to the ontology and rules-based approaches because it may be that there is more ambiguous information that lends itself very well to being modeled as a document overall. So for example, if a series of documents refers to a challenge patients have rather than capturing every single challenge a patient has, you let the model work out that there's likely to be a patient-driven problem or scenario in terms of, let's say, being able to access care. And so we can mix -- we can blend our methods to use trained models where we train the model on the data and then the more targeted rules-based models that use ontologies. So a lot of different options there.

Okay. And I think that's exactly one of the questions here. How much is it's like an rule-based? And what -- how much is it is, let's say, true, let's say, artificial intelligence. Can you comment a bit on that one? And I think we discussed it the other day, because usually, you have to tell the machine to what to look for. And it's not like that an AI would come up, oh, this is a great insight, which I just mined out of your data. This is -- would be rather unusual [ ways ].

Yes, yes. No, it's a good point. I think we very much take a blend of methods approach to these techniques. And so we would choose the most appropriate method for the challenge. In terms of supervised and unsupervised, we used unsupervised learning to do clustering in order to surface naturally occurring topics within the data. Some of those topics would be very familiar. And you would say, yes, I would expect to see those topics being surfaced. And then there are other topics that perhaps we weren't expecting. And that method is really powerful on both of those fronts. So we use unsupervised where we need to. Then the supervised approach, we can train models or we can create rules and use ontologies, really whatever is going to get us to the best result. So that blend of method helps us do that. And now, of course, with the advent of GPT, we are really looking at how do we use these methods in a compliant and appropriate way for this type of data. And so we're taking baby steps in that space because this type of data is sensitive. And there are considerations around, obviously, compliance, but also engaging the stakeholders and understanding what exposure this data is getting to outside technologies, et cetera. And so we see real great potential for generative methods like GPT in different ways. Sometimes to do this kind of content analysis or topic analysis. But other times, it may be to summarize the data so that we can expose the summaries to a broader set of users and not expose the underlying details to that wider user group. So it's a very exciting space, and we are going to be working on incorporating these large language methods as well but in an appropriate time frame and in a way that we can really be very clear on any risks and exposure within that space. Okay. I'm going to skip ahead to the next one. So one of the key kind of outputs here is the visualization. So the visualization allows you to aggregate the analysis. Sorry, aggregate the data coming from NLP in such a way that users can see how many nodes have been written? How do those nodes breakdown by disease area? Or how do they break down by topic? And how do they break down by things like conference names, study names, et cetera. So many different dimensions that we can use to analyze this information. And then that can help with understanding overall -- are we seeing the level of conversation about our key topics that we would expect? Are we seeing a level of conversation about topics we don't expect and therefore, we need to adapt our content, our medical communications or our education or guidelines to the field medical team. Do we need to do something as a result of perhaps these new topics going in. And so aggregate analysis can be very useful for deciding on a new direction. Also, you can track these things over time. So you can start to understand is something changing. Are the nature of the conversations that we're having with physicians, those interactions, are they changing over time? And how are they changing? And what do we need to do to respond. So aggregate analysis is very important. The other half of this is then being able to tag the raw data and filter an immediately find specific nodes that are of interest and moving those at the detailed online level. So these 2 methods are complementary, and we often blend output, as you'll see in the next slide. So here, we have some aggregate analysis around topics. And this helps again to sort of [indiscernible] what's happening, who's talking about what. But then we then include a drill down that really takes you down to that note level with all the extra tags applied that help users zoom in on 1, 2, 5, maybe 10 nodes that are of interest for a detailed review. And then they can zoom out again, if you like, and see how those nodes might be influencing the trend lines that we're seeing all the key terms that have been surfaced. This slide is really just a summary of some of the benefits of why apply technology to MSL nodes. The first is, of course, efficiency. So the aim here is to reduce the manual work to review all the nodes and also to be able to quickly find specific subsets or individual nodes using the cleaned and categorized data. So really, that aim is to increase efficiency. The second is around insights. So being able to surface insights in a truly data-driven way and really have less bias towards the priorities or the experiences of viewers can be helpful. We can also evolve our medical and evidence strategies in response to the stakeholder needs. So really connecting the insights to the needs of the very people we're trying to engage. Then in terms of the interactions themselves, and we've talked a bit about that already around how does the field team really continue to find ways to have effective engagements with physicians. And how do we understand the way the data is captured, how do we influence that and introduce training, for example, that will help us in a compliant way, capture richer data. And then, of course, responsiveness. So by codifying these topic or topic analysis, we have that IP in a piece of methodology. So we have it in a model or we have in the script. And by having that information codified, we can run it again and again. So if we want to check every node as it comes through and tag it and see if there's something of interest in it, we can do that far more frequently than is commonly done today. And that, in turn, allows us to be more responsive to changing needs. So if we see the emergence of a new topic or we see the increase in the discussion or a decrease in the discussion around a topic that we're already focused on, we can react to that and we can start to interrogate that data and start to formulate response. So we're often asked what have we learned by doing this. And really, there's a couple of things that we've learned. The return on automation investment is an interesting one. Essentially, the more effort you put into automation -- of course, the more -- the richer the output is from rerunning those processes over time. However, there are diminishing returns. And so we don't see this, as I mentioned already, as a replacement for expert review. This is really about complementing the experts by giving them cleaned and tagged and augmented data to help them do their jobs more efficiently. And therefore, we would always advise that there is a point in time where you're over-fitting, you're creating too many rules to take care of too many small scenarios. And really, that can be seen through the nodes that are extracted either of you or any way. And so we would help to balance the return on automation investment by creating -- by deciding on a cutoff point mutually agreed where we see less benefit for investing anymore. And what we should not do is run the process over a few cycles and really see how that's flowing through to the insights process. The second learning is around volumes and frequency. If you only have a small number of nodes and they're relatively infrequently captured, then this is probably not the right solution. However, as soon as you do see those volumes increase, and you want to see that frequency of both the capture and the analysis increase, you really -- turning to technology is very valuable and gives you that nimbleness and that ability to run this process far more good. And I think we've covered #3 already really with the answer to the previous -- one of the previous questions. We have the ability to use different models, and those models will suit different questions from zero-shot approaches and large language models for those interested are -- you interested in this space, like GPT through training, machine learning models, using annotated data where we may work with your teams to actually annotate data in order to run that modeling. And all of those rules-based approaches, the capture the language, sentence structures, the expert terminology that's needed. All of these things are available to us, and we are often looking for ways to get as much value from the data as possible using these technics. And then finally, collaboration and adoption. It is no point developing these things in a vacuum. Key benefits of doing some of these projects is the involvement of the team themselves and looking at the data with fresh eyes with a view to automating some of the insight processes on top really helps to, a, get good results. and b, then when the process is up and running, the team feels that they were part of that process, part of tuning the models to get the right results, et cetera and therefore that's absolute adoption. I think we've covered this, so I won't dwell on this too much, but essentially, there is a key distinction between known topics, topics you're really looking for that are important to you, codifying those and then those data-driven topics. And both play a role in this work. One is really about aligning what you're finding in the data to what your key priorities are and your understanding all the particular therapy area is. The second is about saying, I don't know what I don't know. We need to be able to surface any new information quickly so that we can be responsive and ensure that we're really listening to the physicians that we're interacting with. We've just put in here an example of a tweet and some of the rapid ways that we can start to classify this type of information. So we focus so far on MSL nodes. In this case, this is a tweet from social media, and we've scrambled it a little bit just for compliance purposes. But essentially, we're able to very quickly classify this type of document using the latest techniques and then applying a layer of judgment on top, some heuristics on top to see if we think that the content is relevant. And the way that we understand that content to be relevant in this case is, is this something that is from a patient's perspective? Is it talking about disease? Is it talking about treatment? Is it talking about caregivers or doctor's perspective? And so we run these different topics against the data, and then we get some scores back. And then using rules like tweet length, we can then focus on content that's likely to actually be saying something rather than to be retreating something else or approving or disapproving of some other content. So we can use things like tweet length, we can use the combination of the presence of these different things to help us classify the content. And then that relevant content can be pushed to a detailed review or can be used as examples to go alongside trending and visualization. Okay. So thank you very much for listening, and we will now take some questions.

Just while we're waiting for that. What is your personal opinion? How long does it take until we do not need rule-based technology until [ AI ] -- is an expert in the therapeutic area, understand the treatment algorithms is able to mine for, let's say, 2 data gaps or something like that.

I think the promise of the large language models that we see today is huge. And so that space is moving so quickly that I would say, the first place to go for us is unstructured data will become large language models for sure. Rules based, however, is very important for specific extractions and also for normalizing the data. So within our ontologies, for example, we have synonyms for drugs and we can really sort of be highly specific about any transformation or cleaning of the data that is needed. And so I think there will be a role for rules-based NLP for the long run, for sure. I just think we now have yet another powerful tool in the toolkit that will allow us to interrogate the data.