TRANSCRIPT
Welcome And What Matters Today
Deborah Borfitz
Hello and welcome to the Scope of Things podcast, a no-nonsense look at the promise and problems of clinical research based on a sweep of the latest news and emerging trends in the field, and what I think is worthy of your 30 or so minutes of time. I'm Deborah Borfitz, Senior Science Writer for Clinical Research News, which means I spend a lot of time with my ear to the ground on your behalf, and a lot of hours every week speaking to top experts from around the world. Please consider making this your trusted go-to channel for staying current on things that matter, whether they give us hope or cause for pause. In a few minutes, I'll be speaking with Mike Sullivan, who heads up IT Globally for Development Operations at Bristol Myers Squib about what AI in clinical trials is apt to look like in 2030. The first, the latest news, including building the capacity to collaborate with patients, a new playbook for pediatric clinical trial reporting, lithium for slowing verbal memory decline in older adults, sex-specific effects of dementia, an open source database for scaling cancer trials, and a proposed 10-year trial of anti-obesity drugs for preventing obesity-related cancers. Academic researchers in Finland examined the question of what public and patient engagement means in routine research practice to learn the focus on power relations between researchers and patients are varied rather than fixed arrangements between them, as they are often portrayed in the literature. Tensions between tokenistic involvement and co-creation, institutional structures and everyday work, and conflict and reflexivity were explored. Recognizing when and how power relations dynamically operate, they say, could help build the capacity to collaborate with patients in research. At the Hospital for Sick Kids in Toronto, researchers led the development of a new playbook for pediatric clinical trial reporting, which introduces new guidelines for ensuring inclusions of details that young people and their families consider important. Internationally, 42 young people aged 10 to 21 contributed to the development of trial reporting items through a series of workshops held in Canada, England, France, Scotland, and Spain. A pair of explanation papers on trial protocols and final trial reports simultaneously published, offering tips on how the new recommendations can be applied in practice. A two-year exploratory clinical trial at the University of Pittsburgh suggests that low-dose oral lithium may slow verbal memory decline in older adults with mild cognitive impairment, particularly among those with evidence of amyloid beta, one of the hallmark biomarkers of Alzheimer's disease. The treatment was also found to be safe and well tolerated in the study population. The trial launched nearly a decade ago before blood-based tests for Alzheimer's pathology were available, so participants were enrolled based on clinical symptoms alone, and only a subset turned out to be amyloid positive. The research team now hopes to conduct a larger, more definitive clinical trial informed by the pilot study's findings using blood-based biomarkers to identify individuals most likely to benefit to determine whether lithium can meaningfully delay cognitive and neurodegenerative changes associated with Alzheimer's disease. Mayo clinic researchers have opened an entirely new direction for understanding sex-specific differences in the burden of dementia with their study finding Alzheimer's-related brain changes progress up to 20 times faster when women also have abnormal levels of alpha-sinuclein, the Parkinson's-related protein. The same pattern was not observed in men. The discovery emerged from an analysis of data from 415 participants in the Alzheimer's disease neuroimaging initiative and could aid the design of more targeted clinical trials and ultimately more personalized treatment strategies. Next steps include examining whether the sex-specific effects also appear in patients with dementia with Lewy bodies, where alpha senuclein is the primary driver rather than a coexisting pathology. Johns Hopkins researchers have created an open source database known as AstroID to organize and integrate diverse large-scale cancer research data and thereby facilitate biomarker discovery as well as enable more efficient, scalable clinical studies. Investigators anywhere can now more easily study multiple types of cancer data in one setting for research on clinical outcomes, or the data can be merged and queried with a variety of scientific correlates. The resource organizes clinical and related blood and tissue specimen information in six tiers, including information from the patient, diagnosis, clinical events, specimens, and details about how those are processed by the lab. The tiers were translated to a query-oriented platform. At Johns Hopkins Medicine, Astro ID has been deployed for 16 different patient groups with multiple tumor types, but the structure could be adapted to any disease process. And finally, at a major scientific conference in Turkey this May, a global panel of experts will be putting out the call for a 10-year trial with 5,000 participants to establish the efficacy of new anti-obesity drugs and preventing obesity-related cancers. The obesity epidemic is predicted to cause a surge in 13 obesity-related cancers. Since the trial would be too expensive to adequately power with 50,000 participants, the experts are instead proposing a one-to-one randomized trial enrolling a smaller number of individuals with overweight or obesity and a cancer precursor condition, such as Barit's esophagus, endometrial hyperplasia or calodic pulps. The intervention group will receive a GLP1 or dual receptor agonist drug with a behavioral weight loss intervention, and the control group will receive the behavioral weight loss intervention only. Other studies will be needed to determine whether it is the weight loss, a specific effect of obesity drugs, or both that could reduce cancer risk. As a reminder, links to the articles, studies, and press releases referenced in this month's news segment can be found in the show notes. It is now time to bring in today's guest, Mike Sullivan, for a conversation about what's achievable with AI between now and 2030 to help solve the clinical insight latency problem that has long plagued the industry. Welcome to the show, Mike.
Mike Sullivan
Hi, Deb. Thanks for having me. Great to be here.
The Real Blockers Fear And Data
Deborah Borfitz
So happy you could join us. At the recent Scope event in Orlando, you made a presentation on this very topic. And one of the major takeaways was how creating value with AI depends on redesigning how clinical operations teams work, which is, of course, also the key challenge. You've been in this business a good while now, Mike. So let's start with what you see as the big work redesign issues that are most likely, if unwittingly, to hold up progress.
Mike Sullivan
Yeah, Deb, that's that's a great question. And I think that can be summed up in probably two macro ways. There's a bit of fear and there's a bit of data that are acting as a bit of rate limiters to how AI can really deliver the promise of the future for clinical operations in our industry. So if I think about, for example, the fear piece, right? I was in meetings recently where I was trying to help some colleagues, some industry peers think through how we can fundamentally change, not just augment incrementally, but fundamentally change the way we execute clinical trials in our industry. And immediately the conversation goes to, well, this requires certain regulatory consideration, or there this will violate potential compliance concerns, or this will run afoul of our SOPs. And while those are important things, we absolutely need to abide by the law and follow compliance and be prepared for regulatory considerations. The fear is preventing folks from really rethinking how process should be completely torn down and rebuilt in the face of AI. The second thing I mentioned was data. Data, let's face it, data is the modern currency, right? Maybe two, three hundred years ago it was beaver pelts, and then it became, you know, gold bullion and so on and so forth.
Mike Sullivan
But today it's data. The better your data, the more data you have, the more accessible it is, the higher quality it is, the better insights you're going to get. So if you can lower the fear, increase the overall data quality consumption and turning it into insights, we have an opportunity here to really change how you know clinical development can work. And I'll say fundamentally, underpinning all that is we really can't bolt on AI to existing processes. These existing processes would be no different than, say, a car that you see roaming down the street with tires that are clearly a bit too large for the vehicle. Now, it looks great and it works. It certainly drives forward and it gets you where you're going, but it's probably not the most efficient. Well, that's kind of what happens when you bolt AI onto processes today. It really will work. And it may be cool to say we're using AI, but it likely won't be most efficient. What people really need to do is consider how do we completely clean sheet the process. Think about what it would mean to sit at the table with, say, experts, SMEs, and I don't mean just C-suite or senior executives, but SMEs who really do the work, who can think openly, sit at the table with AI experts and allow the process, a new process to unfold. For example, perhaps the experts, the SMEs, say, I can do process one and two, but then the AI experts say, you know what, I can do process two, but then the SMEs can do three and four. And now AI can do steps five and six to close it out. And then humans apply judgment throughout the way. This is how this should work going forward. The the old tenet of building a process and then allowing technology to support it was great, but I think those days are over now, Deb.
Pillar One Autonomous Agent Workflows
Deborah Borfitz
Yeah, got it, got it, got it. I know you built your whole talk around these pillars that you cleverly named, you know, these four pillars of what AI and clinical operations could look like, you know, inside of five years, which I guess is as far ahead as we want to predict, given all the fast pace of change. And this first pillar you named was autonomous clinical workflow. And that sounds intuitive enough, but but can you elaborate a bit on how AI agents might create, you know, these push-button workflows you talk about, you talked about and the role of humans in the loop, of course.
Mike Sullivan
Yeah, so when we talk about agents in general, right, it's important to understand that agents can and probably should be involved in most things we do. We'll assume the audience here has a reasonable understanding of what agents mean in the modern world of AI. But when we think about the potential value of really changing and getting to clinical workflows that are no longer linear or manual or reactive, right, when we talk about autonomously orchestrated agents that can automate clinical workflows, we're really talking about agents that can plan multi-step workflows, execute across systems, and really escalate or elevate decisions to humans just in time. So think about the notion or an aspiration of push-button clinical workflows. I think you might have mentioned this. So push-button clinical workflow. Some might think, well, yeah, Mike, we're kind of doing that today. We can create startup documents with AI. That's wonderful. That's a good example. And oh yeah, we can help digitize protocol, which we'll probably talk about very shortly. We can help digitize a protocol with some agents, and that's great too.
Mike Sullivan
But the real value, the real change that we can make across our industry is thinking about how agents can elevate all of that to be executed in one click. And only at the just the most critical decision-making moments are humans involved to intervene and say, yes, this is worth pursuing or proceeding on. No, we should go back and rethink this. Agents go rethink it and bring it back when you're done. Yes, let's continue on and proceed on to the next deeper part of the process where perhaps we create other documents, start up sites, enroll patients, et cetera. So the point here is we can really bring agents, not just at the micro or the level three or level four process steps, but bring agents up to execute all of that. It's a bit of an orchestration layer. This is the fundamental principle of agentic architectures. For those of you who are more technically savvy out there, that's the fundamental principle. You have an orchestration agent that helps manage what agents are doing. But this is where we get into the real step change in the industry, is with you know getting to these autonomous workflows that can run nearly the whole process by itself. That should be an aspiration, but not without human intervention, making decisions along the way and resolving AI ambiguity.
Pillar Two Machine-Readable Protocols
Deborah Borfitz
Gotcha, gotcha, gotcha. Okay, pillar two. This is where you got into, and I think you mentioned a minute ago, adaptive machine readable protocols, which I believe is an advanced form of digital protocols, involves more tailoring to specific audiences and ensuring data flows to all the necessary places. And that sounds great. Okay, how do we make that happen?
Mike Sullivan
Yeah, so good question. So this this has been probably one of the hottest topics over the last several years. Certainly for those scope attendees who are listening to this. This has been a key session over the last few years. If you think about what this means, right, let's let's define that for a second. So, digital protocol means different things to different people, right? In this context, we'll call the digital protocol, which is really digitizing or making electronic the content and the decisions and the design concepts that go into the development of a protocol. And we really want to empower clinical scientists, clinical child physicians as early in that process, perhaps even as far back as a clinical development plan, to digitize their thoughts, to get their design concepts onto some technology that then allows it to very quickly give them insight into patient burden, site burden, potential outcomes, but give them real-time feedback. So we are simplifying protocols so sites can actually execute them, simplifying protocols so patients have a better experience in the end. This is where we get into the continuously simulated protocols as well, where we can use protocol content now, right? Notice I'm not saying document, but I'm saying content data, aka data. We can use that now to simulate what might be the outcomes with these endpoints, right? And now let's take that insight and have it inform study design before we ever get into burdening patients and sites with an overly complex protocol. Now, there's a few key, you know, AI concepts that are driving this, this shift. You know, first, if you think about multimodal AI, and and this is very simply, and this feels very natural, I think, today to most people, but multimodal AI is really just the ability for AI to consume text, structured data, images, video, and and turn it into something that we can now consume, whether it's like via chat or other sort of models or or agents.
Mike Sullivan
But there's other pieces like neuro symbolic AI. And let's demystify that for a second. The neuropiece of AI is, for example, we'll use a cat. I can train AI with a million images of a cat, you know, typical household cat. And then if I show AI another image, it says, Yeah, you know what, that looks like a cat. But if I ask it, why do you say that's a cat, AI isn't going to be able to provide a meaningful response. It won't be able to provide evidence. It'll just say, well, it looks like all the other cats you showed me. That's not very high quality. But now, if you go to the symbolic part of neuro symbolic AI, the symbolic part is really rules and logic. So now let's take that same AI model and train it on the fact that does it have pointy ears, does it have whiskers? Does it say meow? Now, if you combine that, those rules and logic with the neuroside, all the million pictures, suddenly now AI can say, this is a cat and here's why. Now take that into a clinical setting. And much of this is to some degree been occurring over the last several years.
Mike Sullivan
But, you know, take diagnostic images of, say, lung cancer, right? There have been plenty of studies that show that AI has been able to identify potential cancerous tumors ahead of, or at least on par with, actual humans who are well-trained and high-quality physicians who do this for a living. So the ability to combine all of that, rules and logic, with the imaging, with the multimodal AI, suddenly now you bring all that into developing a protocol, simulating it end-to-end, you really can start to predict potential outcomes and more importantly, potential impacts on patients well before patients have to experience it. So this is a really critical area. I guess the bottom line on this one, Deb, I know I'm probably a little wordy on this, but the protocol becomes software. It's really just data. The protocol document, think of it just as a dashboard, a simple visual representation. But really, the protocol as data, that's where the real value comes in. The amount of work you can automate through agents, through AI, is tremendous.
Subscribe And Share Topic Ideas
Pillar Three Digital Twins For Burden
Announcement
Are you enjoying the conversation? We'd love to hear from you. Please subscribe to the podcast and give us a rating. It helps other people find and join the conversation. If you've got speaker or topic ideas, we'd love to hear those too. You can send them in a podcast review.
Deborah Borfitz
Yeah, and I want to I want to pivot immediately back to something you said just a moment ago. And this gets to pillar three about, you know, predicting the experiences of sites and patients and potentially the use of digital twins where sponsors are forecasting what's going to happen site to site. And these sounds like changes that study participants would welcome, well welcome. But but sites, maybe not so much. Please share how you see this playing out and if some resistance from sites might be expected.
Mike Sullivan
Yeah. So so dev, that this is a good one because it it depends where this goes. So what does that mean? So let's talk about sites for a second, and then we'll we'll focus on patients.
Deborah Borfitz
Okay.
Mike Sullivan
So I so sites, I think would be very happy. And I know some sites do a a version of what I'm about to describe today, which is, you know, as a site, if a sponsor sends a protocol, the site is gonna assess, all right, can I do this? Do I have all the right, you know, equipment and staff, and you know, can I handle the enrollment requirements and so on and so forth? So the site is naturally doing some assessment. I know I've talked to many sites. Some sites already have some, I'll say, algorithm loosely, whether it's AI, automation, or just a really terrific team, but they have something in place today to assess their own feasibility to execute the study and to really assess is this something that is going to be really complicated for them to do, even if they can execute. Right. And then similarly, sponsors are beginning. You know, I know every sponsor will say they've done something along this line, but sponsors are really beginning to look at that complexity and understand the impact on sites that the complexity of their protocols will have. But when we start to get into digital twins, we now have an ability, say as a sponsor, to pre really create a preview of what that impact could be and make adjustments before we even have to burden a site with, hey, do you think this is too complex? Can you run this?
Mike Sullivan
And then similarly, sites also could use digital twins of their own environment to simply look at you know, potential capacity constraints, what might an influx of patients, maybe it's a car T site, and trying to understand the implications there on the influx of patients and the timing and the duration they need to remain on site to complete treatment and assessments. There's so much that can be done with a digital twin to preemptively understand the impact both on resources, cost, and time. And these all translate to patient benefit and patient experience in the end. And so this is why I think this is a critical item. Now, to your earlier point in the question, why might this be controversial for sites? Well, one way to really help sponsors better learn how to make this a better experience for sites and patients is through this notion of federated privacy-preserving learning. So let's face it, AI models, they have to be trained. And you train models on data and the experiences that data has across different parts of the organization and across systems. But sponsors, well, we have a lot of data as sponsors, they don't have all the data. Imagine, and this is where the federated part of federated privacy preserving learning comes in. The federated part is imagine now if sponsors could take a model and allow it to run in the site's environment. And the site says, okay, we're willing to accept this. And the model runs overnight and it learns from all the site activities from that day. What patients came in, what were the responses, how did all the diagnostics and the tests go? What was the treatment like? What were the patient reported outcomes?
Mike Sullivan
It learns from all the work that was completed and then sends home to the sponsor all of those new insights and connections that it observed at the site's in the site's data. Now, the privacy preserving piece of that is we would certainly not take. IP, it would not take you know HIPAA protected information, certainly respect country regulations in terms of cross-d cross-border transfer. So there's there's a lot to be considered in this. And this is where I wouldn't blame sites for saying, I'm not going to let you as a sponsor run algorithms in my environment somewhat unattended. So while it's a very powerful concept, I think it might take a little time for us to figure out how to effectively train what I would say is perhaps maybe global industry models, global AI models, to benefit from the insights of sites globally, all toward the benefit of patients and bringing medicines to those patients. So it it could be a sticky wicket, as some might say, but I have hope that we will figure it out as an industry. And the more training we can do, the more data sharing we can do, the more we're all going to learn from it more quickly.
Pillar Four Zero-Latency Data Quality
Deborah Borfitz
Yeah, whatever feels like normal today will not be what feels normal tomorrow. So all things are possible. I want to now get to pillar four, which may be the clinical ops nirvana zero latency data and continuous quality, where data quality issues will self-identify and self-correct. Is this really achievable by 2030?
Mike Sullivan
Well, so Deb, I of course I'm the one to say yes, right? Okay. But but yeah, but here's why, right? So as an industry, for years, I mean, go go back for some listeners, go back to the days before EDCs and there was just paper, right? The industry has for decades tried to evolve past manual entry and tried to uplift data quality, right? And and I'm really talking about the the ecosystem that is patient visiting a site, site, you know, treating a patient, getting important data back to sponsors, labs also processing data, getting that back to sponsors, sponsors getting data over to health authorities, right? There's just a a super highway of data that has to occur. And it's been occurring for decades, just in different ways. Gotcha. But but imagine now though, right, to get to that analysis ready at the moment of creation and and self-correcting, right? We've been adding edit checks to CRF pages for years now, if not decades. Right. What I'm talking about is beyond those, I'll say more traditional approaches, which we all use. Let's face it, those traditional approaches are still there. But to make 2030 real, there are plenty of vendors and a lot of sponsors. I mean, I talk to my peers across the industry quite frequently. We're all doing something to try and reduce that burden on sites, to uplift the quality. There are a number of vendors that are trying to just seamlessly make data from site go right to, say, a clinical data warehouse where statistical analysis can be performed.
Mike Sullivan
So, but how do you do that? Yes, we could transfer data at light speed today. That's not novel. But the ability to ensure that that data is normalized, standardized, synthesized, analyzed, transferred, and ultimately turned into insight, right? That's the steps. And put them in any sequence you like, but those are the steps that really need to and will compress into a simple push button with AI helping to bring the intelligence at each of those steps to break down the formerly manual or highly batch-oriented processing that has historically plagued our industry. So while I'm not giving you a line-by-line playbook on how to make your data self-correct, the fact is any one of us can quickly go on, you know, pick your LLM of choice and ask it the question, and it'll give you 15 ways to go do it. The point here is we all need to agree to do it at an industry level. We don't all need to get in a forum and agree and shake hands. But if each one of us says, boy, this is important to me, and then a site does that, and then a lab does that, and then some of our clinical trial vendors do that, suddenly you start to hit critical mass of people demonstrating just how we can affect change without 50 sponsors and 200 sites all agreeing in some large contractual consortium that this is the way we'll do this. Those forums, those consortiums, those are all great. They do wonderful things. But what I'm saying is we don't need to wait for that any longer. AI now enables us to do these things in a much more, much faster and with much less overhead than we used to have in the past.
Deborah Borfitz
Well put, Mike. I expected nothing less. I'd like to end with a question, though, more squarely focused on the human in the loop, because I know the widespread perception is that AI means job losses. Can we level set here what kind of job losses can be expected and what new jobs might be created? And as you see it today, anyway, how might the two balance out?
Mike Sullivan
Yeah, Deb, this is a good question. And and this is the the scary one that seems to get all the the headlines in in the in the media. Sadly, they only talk about the the the job side of it, and there's very few news outlets that talk about the job creation side. There's been plenty of studies. There was a World Economic Forum report that suggested about 83 million jobs could be displaced, but but also 97 million new roles could be created. Now, look, those are really big numbers, and I may be dating myself here, but if you go back to say the inception of the internet way back when, I recall very early in my career people talking about oh, jobs are going to be lost, the people are being displaced. But think about the amazing number of industries, the fantastic businesses that have been created and the wonderful new skills that entire two generations of people have had to learn and develop since the inception of the internet. This is what we're on the cusp of right now with the modern AI. There's so many other people who are our fantastic experts and PhDs in this topic. I certainly will leave the rest of the pontification to them. But the fact is, there will be tremendous number of roles that evolve because of where AI is headed. Yes, some people could be impacted, but I think the key here is that there was actually there was a good reference. There was a book. I'm not gonna remember the title of it, but back in the 60s. No, I wasn't there in the 60s when it was first released.
Mike Sullivan
But I'm now I am trying to date myself. So but the the point is here is that there was a good book, and it was recently brought to my attention. There was an interesting quote, and that book at the time was talking about the future, and it said something like this. It's not an exact quote, but the essence was humans have to do this learn, unlearn, and relearn. And that was relevant then and it is absolutely relevant today. Whatever we learned, fantastic. We now, it's time, especially in light of AI, we have to unlearn what we have have learned and now relearn what it means to do our job in the future. Not to do our exact role, but how would we do our job, leveraging our skills, our experience, our training in the future with AI? So it's not about resisting, it's about reskilling. And that's really the critical piece here. And if I think about, you know, what what what does this mean then, say in the clinical operations setting where where humans ultimately go in the face of this sort of reskilling, right? Things like ethical oversight, exception handling, strategic design, system governance, those are gonna continue and evolve to be really front and center. But one in particular stands out for me, and that's really relationship leadership, right?
Mike Sullivan
Now, I'm sure many people will say, Oh, I've been managing my stakeholders, I've been leading my team, I've been leading my C-suite, whatever your role is. But in the face of AI, this becomes more around ensuring alignment, ensuring strategic direction, ensuring outcomes and value are clear. So when you deploy agents and push the button to execute clinical workflows, that it is actually aligned and in support with corporate objectives or functional goals and outcomes. This is I think a key piece that people are overlooking. Everyone wants to quantify the impact of AI on will I have a job or not. But this is a case where it's not about hands on keyboard, it's about elevating away from your hands and into human judgment. Someone once said to me, you get paid from the neck up. And that's never been more prevalent than right now. You get paid from the neck up. What are you thinking? How are you applying judgment? How are you applying human reasoning in the face of all the AI insight, actionable insight that AI is able to provide you? And that's really where humans need to move toward. We don't need to continue to defend our ability to touch keyboards. We need to defend our ability to think and reason logically, rationally, and well informed with data and insight.
Deborah Borfitz
Wow, Mike, I like this picture. It's hopeful, it's sensible, it's practical, it's promising, and if embraced, could foster a culture of innovation like the industry has never seen. Let's hope, right? Thanks. Thanks for sharing your pearls of wisdom on the near future realities with AI. I'll have to check back in with you come 2030 and see how you did with your predictions.
Mike Sullivan
I better make a poster and hang this on my wall so I stick to my objectives and how comfortable. You're gonna hold me accountable, Deb.
Closing Thanks And Where To Find More
Deborah Borfitz
I will. And as always, a big thank you to everyone out there for listening in. If you're not subscribed to this podcast yet, please consider going to Apple Podcasts and doing so right now so you don't miss your monthly dose of news and perspectives. You'll be hard pressed to find anywhere else. And if you're up for it, I'd also be so very grateful if you'd leave a rating and review while you're there. For more straight talk on studies involving humans, visit Clinical Research News Online.com. And if you're a clinical research professional, we hope also to see you at our next Scope Conference, where we make things happen. Bye for now.