AI may open a pathway to early detection of pancreatic cancer

But there’s one group of patients with a better prognosis — the few pancreatic cancer patients who are screened for the illness because they have a known genetic risk. If their cancer is detected early enough, up to 80 percent survive five years.

They’re the only ones known to be at risk, and they make up only 10 percent of pancreatic cancer patients. Surely there are more high-risk people who could benefit from an early warning, and Appelbaum wanted to find them. She thought artificial intelligence could help, and 6 years ago, she teamed up with the Computer Science & Artificial Intelligence Laboratory at MIT.

Late last year, the group reported that it had developed computer models capable of identifying three and a half times more high-risk people than current standards. The models combed through a vast trove of electronic health records to tease out patterns common among those who went on to develop pancreatic cancer, such as changes in blood chemistry or more frequent doctor’s visits.

Published in The Lancet last month, this work has the potential to enlarge the group of pancreatic cancer patients who can benefit from screening from 10 percent to 35 percent. The group hopes its model will eventually help detect risk for other hard-to-find cancers, like ovarian.

While promising, the approach still needs to be proven in the clinic.

The study is notable for the diversity of its population and the vast number of potential risk factors considered, but the model still misses a lot of patients, said Christie Jeon, an associate professor at Cedars-Sinai Department of Biomedical Sciences in Los Angeles, whose research focuses on early detection of cancer but who was not involved with the MIT project.

“This model gets us closer,” Jeon said, “but in itself it cannot be implemented right away.”

Appelbaum acknowledged that even if the models work as expected, 65 percent of patients who go on to develop pancreatic cancer will not benefit from early detection. “But currently, we’re missing 90 percent,” Appelbaum said. “So this is a big improvement. … This translates into saving thousands of lives.”

Appelbaum and the MIT team are not the first to use AI to predict pancreatic cancer. They’re not even the first group to do so in Boston. In May of last year, a team based at Harvard Medical School and Dana-Farber Cancer Institute reported similar results from an AI system working with data from Denmark and from the VA Health Care system.

“Both papers point out the advantage of using these AI methods, compared to standard screening based on family risk and genetic predisposition,” said Chris Sander, a faculty member in Systems Biology at Harvard Medical School, who led the Harvard research and wasn’t involved with the MIT study.

The MIT study, Sander said, “is particularly interesting because of the large number of real-world patient records they were able to access.” Also, he said, the researchers demonstrated that their predictions would hold true regardless of race or geographic location.

Sander said that his work, together with Appelbaum’s, fills him with optimism that AI can eventually help control pancreatic cancer, the third leading cause of cancer deaths even though it’s rare.

The pancreas, which aids digestion and regulates blood sugar, is a six-inch-long organ buried deep within the abdomen, difficult to feel on examination, and even difficult to see in imaging tests. When cancer arises, it does so sneakily: A precancerous lesion simmers quietly for years, and then, as if boiling over, it explodes into full-blown cancer and rapidly spreads.

People do have symptoms – they may develop diabetes or lose weight without trying or have stomach aches. But such symptoms are more likely to be caused by something other than a rare cancer.

“Most primary care physicians, since it’s a rare disease, wouldn’t think of that,” said Appelbaum, who no longer sees patients but used to work as a radiation oncologist in Israel. The ultimate goal is to use AI to give primary care doctors an alert when a patient has the precise mix of symptoms and factors that signal a high risk, she said.

A main component of the recent study was the group’s collaboration with TriNetX, a Cambridge-based company with a huge database of standardized and de-identified electronic medical records from health care organizations around the world. That enabled the researchers to include a much more diverse population than has been used in previous studies.

The AI-powered programs sorted through anonymous patient records from 55 health care organizations across the United States and looked at 35,000 people who had pancreatic cancer and 1.5 million who did not. The models tracked virtually every interaction these patients had with the health care system, focusing on 87 features derived from demographics, diagnoses, medications, and lab test results, to identify the mix of features that distinguish those who go on to develop pancreatic cancer.

Next, the researchers tested these findings in a simulation. They picked out the records of individuals who fell into this high-risk group before 2020, and then tracked whether they developed pancreatic cancer over the next 18 months. The model performed as expected, picking up three and a half times more at-risk people than the standard methods.

Now, the team is working on its next and arguably most critical challenge: To show that this model’s predictions hold true when looking into the future. Using the same TriNetX database, they have assigned “risk scores” to 6 million people, and they’re waiting to see how many and which ones develop pancreatic cancer. Because the records have been de-identified, the researchers don’t know and will never know who these people are.

At the same time, Beth Israel Deaconess is conducting a prospective study on its own patients. Those who are flagged by the model are asked to give blood and saliva samples, which eventually may help pinpoint additional signs of pancreatic cancer. The carefully worded request does not suggest that these patients are at higher risk, because at this point that remains uncertain.

Why not screen these people right now?

“This model is experimental, and I really don’t know how well it’s going to work,” Appelbaum said. “And the last thing I want to do is make a lot of people panic. … A lot of these people are false positives. They’re not ever going to develop cancer.”

Additionally, screening for pancreatic cancer is no small matter. Patients undergo yearly MRI scans or endoscopic ultrasounds, typically performed at an academic medical center. These are expensive, time-consuming procedures. It’s no picnic for those who are currently considered high risk because of their genetic predisposition. “They undergo these tests every year and are worried every year,” Appelbaum said, and yet 95 percent of them never get cancer.

Meanwhile, the MIT team has already adapted the model for kidney cancer, and is looking at deploying it on other hard-to-find cancers, such as ovarian cancer, said Martin Rinard a professor in MIT’s Department of Electrical Engineering and Computer Science and a senior author of the study.

“There’s an enormous scope here for deploying this across a wide range of cancers,” he said.

---

Felice J. Freyer can be reached at [email protected]. Follow her @felicejfreyer.