This week, Advarra launched a pilot of the Insights Accrual Prediction platform to customers, powered by Wendy’s initial model and collaboration with our Onsemble Community. This is the second issue of our two-part Q&A series with Dr. Wendy Tate, Director of Research Operations at Advarra. In Part One of our series, we discussed how Wendy’s background in laboratory, regulatory, and oncology contributed to her accrual prediction method, as well as hurdles she hit during her initial development of the tool. In Part Two, we dive deeper into arguments for and against predictive modeling and how this technology can be leveraged in the future.
Some have argued that predictive models only repeat current trends and don’t invoke change. What would you say to that response?
Wendy: To ensure the model can help address current trends while also identifying opportunities for improvement, it’s essential to update and add new information to this platform. One of the ways we are constantly evaluating the platform is to analyze model accuracy across varying timeframes, as well as making sure our data is up to date. For some organizations, their entire closed portfolio predicts much more accurately than a subset that contains more recent data. But at others, you see a shift in accrual patterns based on changes in investigators or new areas of research. Sometimes taking a smaller slice of data that’s more recent will predict more accurately because you’re witnessing those shifts with the changing tide. We’re already talking about how to best update the models so they’re reflecting the most current data and information. We see good model fit consistently from programs with at least 150 closed protocols. One new protocol should not change the entire model predictions with hundreds of closed protocols. So, new models need to be generated regularly, but not with every new protocol that has been closed, as that is unlikely to greatly shift the predictive ability of the model.
When is this model most successful? In what scenarios should it be applied?
Right now the model is set up for interventional treatment industry trials. We selected industry trials because they are normally not the unique scientific output of a given clinical research program. Academic medical centers, cancer centers, and other clinical research science organizations will run institutional and IITs for completely different reasons beyond achieving accrual goals. From a business perspective, industry trials at the site must meet their budget and accrual goals to avoid wasting resources that could be utilized on other studies. Running industry trials more efficiently allows your institution to dedicate more resources to IITs and other scientific integrity initiatives, or to trials like rare disease that you may have open but are unlikely to consistently accrue.
Thinking beyond current use cases, what’s next? What is the potential of this tool?
When analyzing frameworks around accrual, I looked at three broad concepts: participant aspects (such as access and interest), organizational aspects (research program and facility characteristics), and investigator aspects (past performance, desire to participate in research, area of expertise). This model most directly addresses organizations and investigators to answer the question “can we enroll?” In the future, I’d love to investigate ways to incorporate patient data to answer the question, “Who can we enroll?” It’s the combination of the two questions that could build something much broader to supply a strong pipeline for studies. It is worthwhile to explore how this methodology could help those planning accrual study-wide, whether a pharmaceutical sponsor or site running multi-center trials. It takes significant time, effort, and money to start up a site. Non-accruing sites require additional time and money to be invested to start up additional, unplanned sites. This delays the ability to assess study outcomes and move the investigation to the next stage. Application of this model could help streamline which sites have the predictive ability to accrue study participants, providing more accurate assessments of how many sites are needed. This could shave significant costs off site activation through more efficient and accurate planning.
So, you’ve analyzed different stakeholders this tool can serve. Have you thought of applying machine learning to challenges beyond accrual?
Beyond accrual, most literature discusses machine learning in post-activation scenarios. While machine learning is valuable, to me—a more “traditionally trained” laboratory scientist—it is a tool to help us determine the most effective way to solve a problem. Study activation is my passion and where I think we can move the needle to get things done more efficiently through numerous mechanisms, machine learning only being one of them. When I look at other study activation processes—how to shorten timelines, for example—I’m not sure predictive modeling is the best tool. It has much more to do with process navigation. In my opinion, there is a lot of potential for AI post-activation in the conduct of the study, and it will be interesting to see if we can find ways to assist our data safety monitoring boards, and dive into real-world evidence studies with machine learning. For me, working in analyzing research methods, machine learning is one tool in my kit to help assess what we can do to streamline clinical research methods while maintaining safety and efficacy.
Everybody in the clinical research enterprise agrees research needs to be done quicker and better, but we cannot wait for national change to be made. Sweeping changes across the entire clinical research enterprise are slow and would need significant assistance to implement. One of the reasons we want to investigate accrual prediction from a site level is because, even for a single site, it can lead to progress for your patients, your investigators, and your science. If you can create a tool that’s valuable for enough sites to use, you will influence industry-wide change from the ground up. In research, sites are the boots on the ground. We can’t do clinical research without sites. I’m looking to give tools to our sites to empower them and to improve the clinical research enterprise as a whole. I thoroughly look forward to hearing more from our pilot customers and continuing to develop the Insights Accrual Prediction platform to advance better research.