A Major Stride in Pancreatic Cancer Treatment

How Artificial Intelligence (A.I.) is Showing Major Promise in Early Detection

Conversations and Curbsides - a Podcast between DoctorsJohn A Chabot, MD  is a Columbia professor of surgery, the Chief of GI/Endocrine Surgery at NewYork-Presbyterian Hospital, and is the executive director of the Pancreas Center at Columbia.

Dr. Chabot joined Dr. Hyesoo Lowe on an episode of the Columbia Surgery Podcast Series: Conversations & Curbsides. The two doctors discussed a major innovation in Pancreatic Cancer Care, namely the potential role of Artificial Intelligence (A.I.) in early detection.

The following is a transcription of the discussion, and is lightly edited for context and clarity.

Jump to a Section:


Dr. Hyesoo Lowe

We are so happy to have you with us today, Dr. Chabot. I'd love to focus our discussion today on some innovations in pancreas cancer, and so how about some hot takes on some new topics in pancreatic cancer?


Dr. John Chabot:

Sounds good. The hottest thing this year all around us is artificial intelligence. Turns out that artificial intelligence is going to have an impact on pancreatic cancer too.

A paper came out in May of this year where the investigators looked at just medical records, and in fact, just at the codes that are used to identify new problems that people have in their medical record.

It didn't even go into the granular part of the medical record. And by looking at the sequence of codes in the population of Denmark, they were able to come up with a set of conditions that predicted the development of pancreatic cancer about three years before people developed pancreatic cancer.

They then took that information and they applied it to the VA, Veterans Administration, electronic medical record, which is a very complete and thorough large medical record, and they were able to replicate the findings.

So just think about that. They were able to identify the top 3%, I think it was, of people who were at risk for developing pancreatic cancer. And then you can look at those people really carefully. So what does looking at them carefully mean?

We usually do MRI scans or CAT scans, and if we see a mass, we biopsy it, and that's pancreatic cancer. But if we are looking at people three years before they would normally be diagnosed, are they going to have something that we can see at that early time point?

Well, it turns out that another group of radiologists using artificial intelligence did something called radiomics in a large group of people who had developed pancreatic cancer. And they looked at CAT scans one, two, and three years before they developed. They were diagnosed with pancreatic cancer and they ran them through this AI algorithm. And it turns out the algorithm could identify the most subtle changes on CAT scans that no radiologist would ever see.

That's not something we can biopsy. It's a very subtle finding, but by putting these two things together, you get the sense that we are rapidly, rapidly moving toward a system of early diagnosis in pancreatic cancer, and that is what we've missed all along.

That's why pancreatic cancer is such a bad disease because we diagnose it late. We don't have screening tests, we don't have colonoscopy, we don't have mammography, we don't have pap smears, we don't have PSA, and now we might. So this could be an absolute sea change in our ability to cure people with pancreatic cancer.

Dr. Hyesoo Lowe:

Absolutely. It's interesting, the AI in the radiology world makes sense. Obviously, the cancer didn't come from nowhere, so it could have started smaller at a place where it was detectable or even sub clinically detectable on previous scans. So those patients, I guess, would've happened to have other scans for other reasons, and that's why they had the chance to look at them.


Dr. Hyesoo Lowe:

Very fascinating is the diagnosis codes. Can you explain just for a moment, what is a diagnosis code and how exactly was that used to measure the outcomes?

Dr. John Chabot:

Sure. It's obviously second nature to physicians, but there's no reason for the public to know about diagnosis codes.

Once a diagnosis is made for a given condition in a given patient, that is assigned a number, a code, and there's this giant book of diagnosis codes. And it's the way that doctors, hospitals, insurance companies, regulatory agencies, all of the business and regulatory side of medicine keeps track of what's going on in a patient, in a country, in a community. It's a way of codifying what we as physicians see and do.

Dr. Hyesoo Lowe:

Fair. So these are labels that are universally accepted as labels for a particular diagnosis that a person has, and that will carry over and allow us to understand all of the different diagnoses carried by a patient at any given time.

Are the codes that were relevant to the A.I. model the same risk factors that we usually talk about, in relation to pancreas cancer? If you have X, Y, Z, those might be related to the possibility of an increased risk of pancreatic cancer?

Dr. John Chabot:

Some of them, yes. Some of them, no.

For example, most of us know that somebody as an adult who develops diabetes who doesn't have risk factors for diabetes, really should be screened for pancreatic cancer. So new onset diabetes is one of those diagnosis codes.

But one of the unique things that this group did was they not only looked at the codes that were diagnosed in the three years prior to development of pancreatic cancer, but they looked at the sequencing of those things, and it's really in the sequence of what people come in and tell their doctor they're experiencing that gives this away. So it's really, really fascinating.

Dr. Hyesoo Lowe:

And there's probably a lot more power in those larger populations to talk about. Not only sequencing, but combinations of different various diagnosis codes.

Dr. John Chabot:

So add to that some of our traditional tests, there's a test called CA 19-9, which 90% of people with pancreatic cancer have an elevated CA 19-9.


Dr. Hyesoo Lowe:

When you say CA 19-9, can you explain what that is exactly?

Dr. John Chabot:

Sure. There are these things called biomarkers that we use as blood tests or urine tests or any other tests to ask the question, “What's going on in this person with regard to this disease?”

I think the best known of these sorts of biomarkers is something called PSA, prostate specific antigen, which is a blood test that's done to look for prostate cancer in a patient. Many, many, many people have heard about PSA.

CA 19-9 is a similar molecule that we look for in the blood for people who are being treated for pancreas cancer, usually it starts high and as they're being treated, it goes low and it correlates with the amount of cancer in the body. It's just a molecule. It's floating around in the bloodstream.

But if you screen the entire population with CA 19-9, you identify far more people who have nothing, no pancreatic cancer. And they get put through the mill in terms of their diagnostic evaluation, nothing turns up. We create all sorts of anxiety. So we really can't use CA 19-9 in the general population.

But if we could use AI to narrow the population, then apply something like CA 19-9, a known biomarker, as well as a CAT scan with radiomics, now we're really, really getting somewhere.

Dr. Hyesoo Lowe:

Fascinating stuff. So we're coming into a whole world of a lot of information that's going to be just ready to be analyzed in the near future. Exciting.


Dr. John Chabot:

Well, there are all sorts of issues for us to work through. There are privacy issues. There are, who owns the data, issues. Who owns the data in this EMR that covers a million patients, and are they going to share? So there are non-scientific issues for us to work through, but if we've all got the right motivation, that should all be manageable.


Don’t miss an update on news and content from Healthpoints, the blog from the Department of Surgery at Columbia