Aurora NCORP Pilots Natural Language Processing Tool to Find Patients with Pancreatic Cysts for EA2185

Woman talking with doctor
Trial Spotlight: Robert Ferris on the EA3132 Trial for Patients with Head and Neck Cancer
December 20, 2022
Evaluation Form
36 Member Networks and Programs Receive Commendations in 2022 for Outstanding Research Performance
December 20, 2022

Aurora NCORP Pilots Natural Language Processing Tool to Find Patients with Pancreatic Cysts for EA2185

Nurse at computer

Close-up of unrecognizable nurse sitting at desk and using computer while analyzing medical notes

The teams that lead cancer prevention and screening trials face one barrier to accrual not typically seen with treatment trials: lack of interaction with the eligible patient population. Because these trials involve participants who have not yet developed cancer, study teams rely almost entirely on referrals from colleagues in other disciplines.

A number of factors make this method inefficient, from lack of awareness among referring providers to communication breakdowns among the various teams. The EA2185 study team is examining a new approach in an attempt to bypass these issues and bolster accrual.

Clinical trial EA2185, Comparing the Clinical Impact of Pancreatic Cyst Surveillance Programs, is evaluating two common strategies for following patients with non-cancerous pancreatic cysts: the 2012 Fukuoka guidelines and the 2015 American Gastroenterological Association (AGA) evidence-based guidelines. The results may reveal an optimal surveillance strategy, helping to prevent the over- or under-treatment of patients. The study may also uncover findings that could predict the behavior of these cysts.

Recently, leadership at the ECOG-ACRIN Cancer Research Group (ECOG-ACRIN) and the National Cancer Institute (NCI), through interaction with sites, identified that a natural language processing (NLP) tool can help find eligible patients for the trial more efficiently. Many sites already have NLP software in place, primarily for business analytics. However, common NLP platforms such as mPower™ Clinical Analytics and Deep 6 AI are capable of using artificial intelligence powered NLP to analyze radiology reports for eligible participants.

ECOG-ACRIN and NCI now offer funding and support to sites to implement natural language processing for recruitment of patients to EA2185. Aurora NCORP is one of the first sites to utilize NLP software for EA2185. The site recently enrolled their first patient identified by the platform they use, mPower™ Clinical Analytics.

“In the future, techniques like this may be the way patients are identified for trial participation,” said Thomas J. Saphner, MD, ECOG-ACRIN principal investigator for the Aurora NCORP. “As cancer is divided into smaller groups defined by genetics, researchers may not be able to identify patients using the current means. NLP and more advanced techniques may be required—and, in fact, may become the only way to effectively identify patients for prevention and screening trials.”

In a recent interview, Kevin P. Morrow, MA, the clinical research coordinator leading EA2185 at Aurora, shared his experience with the tool.

How has your NLP program changed the way you identify and approach potential EA2815 participants at your site?

The nature of EA2185 is to enroll patients who have certain incidental findings on a scan. The software has changed the process completely—and for the better. Previously with this study, we experienced “hopeful enrollment.” I say “hopeful” because you hope that the conversations you have with your radiologists about the trial will lead them to notify you when they find an eligible patient. In reality, this only happens a small percentage of the time.

Even beyond this study, I have been trying to rack my brain for better ways to conduct medical record searches and find patients for trials like EA2185 with large sample sizes. These patients should seemingly be easy to find—but it turns out that they are not without a cohesive way to search medical and imaging records. Thankfully, after meeting with colleagues and radiologists at my site to discuss their NLP software capabilities, someone mentioned mPower™. After implementing the tool, the number of potential patients for the pancreatic cyst study went through the roof.

What was the implementation like?

I was not involved in the technical implementation—that was handled by our IT department. ECOG-ACRIN personnel connected our IT team with a consultant who helped to fine-tune the EA2185 search query. They developed this behind the scenes, and then let me know. I suspect the process was very smooth, because the time between my first notification that this system was up-and-coming and my introductory training call was very short. In the past, when I have tried to develop data sweep systems of our medical records the process has been slow and cumbersome.

How often do you use the tool?

I make a point, once or twice per week, to do a sweep for possible study participants. People are undergoing imaging all the time, so there is new data each week as more patients are scanned.

Interestingly, for EA2185, I’ve noticed that some of the eligible patients the tool identifies have not even been told they have a pancreatic cyst. In many cases, primary care orders a scan—for any number of reasons—the radiologist identifies a cyst, but then no one communicates that finding to the patient. This may be because these cysts are not considered very serious. Now, I can step in and ask the original care provider to take another look, and to consider their patient for our study.

Once you identify a possible participant, what happens next?

When I find a patient—and especially someone who may not have been told about their cyst—I first have a discussion with the succession of people who ordered the scan, read the scan, or spoke with the patient previously. Those physicians then follow up with the patient. If I just called the patient directly, they might be confused or uneasy. With something like mPower™, which can generate massive amounts of data, we take that information but then critically think about it—and think about the best ways to approach our patients.

How much time would you say mPower™ saves you compared to your previous method of identifying patients?

It saves me hours. The program streamlines the identification process and drastically reduces the amount of time spent manually reviewing scans. With mPower™, I just enter all the study criteria and quickly get results.

What else should sites know about NLP software?

The tool’s usefulness extends well beyond EA2185. For clinical trials to work across the board, you have to have people enrolling in them. Research shows that medical progress does not move as quickly as it should because of the low number of people who take part in clinical trials. Part of this comes down to finding eligible participants in the first place. Sites may have great trials open, but then they just have to wait and hope the right people walk into the clinic. That does not spur innovation at the rate that we need.

Any system like mPower™ that not only finds patients more expediently, but also finds a greater quantity of patients in a shorter amount of time, is a great benefit to scientific innovation as a whole—in addition, of course, to EA2185.

If you are interested in learning more about EA2185 or the natural language processing effort underway, please send an email to

Leave a Reply

Your email address will not be published. Required fields are marked *