Using AI, Researchers Develop New Tumor Marker for Pancreatic Cancer

It can be difficult for physicians to assess how the most common and most aggressive pancreatic cancer responds to treatment.

That’s because the cancer, called pancreatic ductal adenocarcinoma, forms fibrous tumors that often don’t change in size, even if they are responding well to chemotherapy.

Instead of relying on images of tumors, physicians judge treatment success by measuring the change in a patient’s levels of CA19-9, a protein expressed by the cancer that circulates in the blood.

But nearly 30 percent of pancreatic cancer patients don’t secrete this tumor marker at elevated levels, and 10 percent don’t secrete it at all.

At the Medical College of Wisconsin (MCW), researchers set out to develop a new pancreatic cancer biomarker that could be used across all patients.

Led by Anai Kothari, MD, MS, assistant professor of surgical oncology, the team used artificial intelligence and machine learning to determine whether other healthcare data routinely measured in these patients could be used to create a new biomarker.

By combining 38 common blood tests that every patient undergoes – including measures of white blood cell counts and liver function – the team developed a new measurement of treatment response.

That number, called e19-9, accurately reflected treatment response among patients who did not produce CA19-9, the team found. Their research was published in JAMA Surgery.

“A lot of recent excitement in healthcare AI research centers on predicting risk, but the output is usually a ‘yes’ or ‘no,’” Dr. Kothari says. “What is different about e19-9 is it more closely fits how cancer treatment is delivered, where decisions are dynamic and we have to be equipped to monitor changes in real-time.”

Training an AI Model to Measure Cancer Treatment Results

The problem of CA19-9 had stymied Sam Thalji, MD, as a surgical resident at MCW. When he connected with Dr. Kothari, an expert in AI and machine learning, they realized that this was a problem perfectly suited for AI.

“AI models are really good at making sophisticated statistical inferences,” says Dr. Thalji, now a surgical oncology fellow. “We wondered: could we give an AI model the lab work of patients who have an elevated level of CA19-9 and see if it can make a new tumor marker for everybody?”

The team trained the AI model using patient data from 2010 to 2022, which included data from more than 3,200 pancreatic cancer patients who had elevated CA19-9. The model was trained to predict the expected CA19-9 levels among patients with pancreatic cancer.

“MCW is one of only few places in the world that has this robust clinical outcomes database for people with pancreatic cancer,” Dr. Thalji says. “That information allowed us to do this high-quality research.”

That database included information from common blood tests that patients had taken, including complete blood count with differentials (commonly known as a CBC), basic metabolic panels, and liver function tests.

“These are all typical lab draws you would get if you were receiving treatment for cancer,” Dr. Thalji says.

The team found that these 38 lab results combined created a new overall measurement – which they named e19-9 – that correlated with how well patients had responded to treatment.

“We designed this to work with data that’s already routinely collected, so there is no need for additional testing,” Dr. Kothari says. “The model output is a single number that behaves like CA19-9, so you can track how it changes over time.”

If the number starts at 200 and drops down to 50, for example, that change is associated with a lower risk of the patient’s cancer metastasizing. This information is important for treatment, since physicians need to quickly adapt therapies if they are not working.

The e19-9 model was then applied to the data of 121 patients who did not have elevated CA19-9, and the research team found that it could accurately predict certain outcomes that CA19-9 had been used to measure.

Specifically, the change in e19-9 over time predicted whether patients completed therapy and whether their cancer had metastasized. The trajectory of the number was also associated with overall survival rates.

A Potential Tool to Assess Treatment for Other Cancers

Next, the team plans to put its tool in a head-to-head competition with CA19-9 among a new patient database to see if it performs even better than the blood biomarker.

“It doesn’t need to replace CA19-9. It can just become a tool that physicians use to make decisions,” Dr. Thalji says. “The important thing right now is to continue to test it to ensure that it is accurate.”

Once accuracy is validated in more patient sets, the team aims to create a tool within electronic medical records that any physician can use.

And because the model relies on common blood measurements, it could potentially be used as a measurement tool among other cancers, as well.

“Our hypothesis is that e19-9 is picking up on overall physiological changes that happen in a person when their cancer is worsening or improving,” Dr. Kothari says. “This could potentially be used by oncologists across cancers to drive treatment decisions.”