Detecting Disease Early Via AI in Black Pancreatic Cancer Populations

AI may open pathways to locate pancreatic cancer earlier and in higher-risk patient subgroups, according to Debiao Li, PhD, and Stephen Pandol, MD.

CancerNetwork® spoke with Debiao Li, PhD, director of the Biomedical Imaging Research Institute and professor of Biomedical Sciences and Imaging at Cedars-Sinai; and Stephen Pandol, MD, director of the Basic and Translational Pancreas Research program at Cedars-Sinai, regarding their artificial intelligence (AI) tool and current plans to investigate why there are disproportionately high rates of pancreatic cancer in Black patients.

“The incidence of pancreatic cancer among the Black population is at least 50% higher than the incidence of other racial groups. Furthermore, research has shown that Black patients have the lowest survival rate,” Li stated in a news release on this study.¹

This study began following Li and Pandol’s initial creation of an imaging tool that utilizes AI to identify and predict pancreatic cancer through CT scans at earlier stages in 2022.² They emphasized that the tool could detect very subtle changes in the tissue of the pancreas, which can be predictive of future development of pancreatic cancer. Moreover, the AI tool can scour the electronic medical record, which has data from a great number of patients, to potentially locate biomarkers that can predict pancreatic cancer.

A major hope with this study is that, once a patient is identified to be at risk of pancreatic cancer, they can start receiving preventive measures that can greatly mitigate or eliminate their risk of disease. Pancreatic cancer is difficult to treat at a late stage, according to Pandol, and this tool creates more opportunities to catch the disease in earlier stages.

CancerNetwork: What was the purpose of this study—analyzing pancreatic cancer risk assessment—in Black patients specifically?

Pandol: One of the problems we have with pancreatic cancer is that the diagnosis is made late. A general statistic is that if you [make] a diagnosis early—that is, [if] the cancer is localized to the pancreas and it’s small in size—you can have, with a surgical treatment [plus] chemotherapy, an outcome such that 5-year survival can approach 50%. However, if the disease is diagnosed late, especially with metastasis of the disease outside the organ, where surgical therapy is not possible and chemotherapy is not that effective, the outcomes [reach] only a couple percentage points, at best, at 5 years.

There are many things that need to be done, but the big idea in the field is that we need to improve early diagnosis [to a point] where we have an opportunity to have good outcomes. Our project is designed around finding ways to improve early diagnosis…[Some] of the major ways we can improve diagnosis are [through] blood biomarkers, urine biomarkers, or imaging biomarkers. This study is about imaging biomarkers.

Li: We started talking about 6 or 7 years ago, and Pandol mentioned the importance of early detection and cancer risk prediction. When patients are diagnosed with pancreatic cancer, the survival rate is very low: 10% to 15% at most within 5 years. He raised the issue of whether imaging can help early detection and better predict the risk before patients have the cancer a few years ahead of time. That’s when we started to work together and use imaging. When patients go to the emergency department for whatever reason—for abdominal pain or other symptoms—they usually get an abdominal CT scan, but the majority of the CT scans are normal, so the patient is sent home, and nothing’s done.

But we [sometimes] find out since then that some of the images, even though they were normal at the time [they were taken], have some features and subtle changes in the tissue that could have predicted cancer down the road. That’s when we started to work together and accumulate data from Cedars-Sinai Medical Center.…Just last year, Pandol [formed] a connection at the University of Illinois Chicago. He had a special interest in the African American population and a long history of working with this group of patients, so we wondered whether there were unique features in this group, African American [patients], that are different from other populations [like those who are] White or Asian, for example….We found out, indeed, there are different image patterns or features that [Black patients] could have. We need to work on this specific population to develop an accurate prediction model.

Pandol: One of the things that we know is that African American patients have a higher rate of pancreatic cancer and poorer outcomes, so we need to appropriately address that population. We need to address them specifically in terms of early diagnosis….We know from the literature that patients would have CT scans for a variety of reasons, and then they would show up with pancreatic cancer months or years later. Going back and looking at some of the previous CT scans, they sometimes showed [features] that could be associated with pancreatic cancer that could have been missed, but in many cases, they there were no findings in them. We thought [that] maybe we could [make] something better than a radiologist’s eyes [to] determine whether there’s risk or if there is early cancer.

Why do Black patients have an increased risk of pancreatic cancer? What are some of the genetic, socioeconomic, and lifestyle differences that could lead to an increased pancreatic cancer risk?

Pandol: We don’t know the reasons for the increased risk. There’s no genetic information I’m aware of. There could be lifestyle factors, but those haven’t been extensively studied. In terms of lifestyle or genetic risks, we’re not certain that there are available determinants. The outcome issue is a little more complicated because it could have to do with access [to care] as well.

Ultimately, studies need to sort out whether outcomes are poorer in that population because of access or a different kind of disease than what other populations have. Those are important questions that need to be addressed. The next point is, can we contribute to that with our study? Our hope is that if we can improve early diagnosis in that population, as well as other populations, we can start to address parts of those questions.

When we develop our method and apply it in real time to populations—this would include African American populations as well as other populations—it will be able to identify those patients at risk. The at-risk population measurements will include the imaging data and risk factors [like] diabetes, obesity, smoking, etc. These factors will go into a model of risk. We envision that we’ll be able to identify patients at high risk and put them into prevention trials. That is the ultimate hope of what our methodology can add. The information about smoking, obesity, diabetes, etc. is also available in the medical records associated with the imaging, so we’ll be able to put those models together in different populations and then actually design intervention studies in patients with high risks. That’s where we’re going with all of this.

Was it foreseeable that, even with the AI tool, additional studies would be needed to focus on specific patient subgroups?

Li: AI is still relatively new in the domain of cancer risk prediction. The challenge is that anytime you use AI, you need a large amount of data for a model to be reliable and accurate. There’s still a long way to go in terms of the clinical testing before we can start to use this and make a prediction in a clinical setting. That’s something we are planning to do….This study only gave us [approximately] 1000 cases because pancreatic cancer is rare. Also, we need to do a prospective study. So far, [what] we have done is retrospective; the data has been collected, and the cancer has [already] occurred.

We’re going to collect a huge number of cases by involving more centers, and [we’re going to] perform more rigorous testing. We want to make sure that the model developed within our medical center can be applied nationwide or worldwide, in all different centers. That generalizability is also [an] issue [because] we need to make sure that the model is universal. There are technical advances that allow us to do that. We’re working on this…to make the model more robust. Also, in the meantime, we need to collect a lot more data to validate the model and prospectively study it as well.

How does AI’s recent advancement make a study like this possible today as opposed to conducting it several years ago without an electronic medical record to utilize?

Pandol: The electronic medical record, with all the clinical data in it, is also an essential part of this. That’s because all the clinical information that can be used in prediction models is attached or associated with the images that we get. That’s how we put it all together.

Li: AI is becoming much more powerful. [It allows one] pick up subtle changes that the human eye couldn’t detect. The cancer tissue has subtle changes that human eyes couldn’t identify. Computers have become so powerful over the last 5 years that they can [detect these changes]. Even though it still takes a little bit of time to run through all the cases because there’s such a tremendous need for computation….there are a lot of new methods for AI to make it more robust and more efficient.

What else would you like people to know about this study?

Li: This particular study [focuses on] the African American population. As we collect more data, we’ll be doing this for other groups as well, [like] Latino patients and others….Secondly, we are going beyond pancreatic cancer. We’re going to develop a similar model for liver cancer and for cirrhosis, pancreatitis, and diabetes. [We are focusing on] all kinds of diseases, including cancer. We’re hoping to use the imaging that is clinically available as a window to look into the health of patients with all kinds of disease and to develop a system that predicts many different types of disease so that patients can be aware of them and take action early. Another angle that we are looking at is using imaging as a measure of biological age. For example, some patients have accelerated aging in their organs, which could also be a risk factor for disease in the future. The bottom line is that we want to use clinical imaging as a window to predict risk and hopefully prevent disease in the future.

Pandol: One of the things that we talk about is that imaging markers can be another measure of prevention. For example, it would be similar to blood pressure, cholesterol, or glucose monitoring that are typically done in preventive medicine. Can there be imaging markers that can be used for prevention as well? That’s the holy grail here.

The other point to make is that pancreatic cancer, in the big picture, is difficult to treat at a late stage. Our overall goal is to have these prediction methods, like AI imaging, connected to other medical information [sources] to predict patients at high risk of disease and then institute interceptive or preventive therapy in the same way we think about using blood pressure medicines for blood pressure or cholesterol medications for high cholesterols.

Li: We hope that this could be a measure similar to [monitoring] cholesterol and glucose. Since imaging can also be done non-invasively, patients can see if they have a high risk, and if they [change] their lifestyle [and] take [medication], we can come back like we do for any physical exam with blood, and we can do an imaging exam to see whether their risk was reduced after their intervention. Those are exciting possibilities that that we hope this imaging research can help.

Pandol: Those are real possibilities. [One example] we talked about recently is that the amount of fat in the tissue in the pancreas [or] other organs like the liver is a predictor of some of these diseases, including cancer. If we institute therapy that affects fat in the tissue, we could affect outcomes. But there are also other preventive medicines that we have shown in preclinical models that can prevent the cancer or slow it down. Those are all possibilities, and developing this tool will allow us to proceed with another aspect of preventive medicine that wasn’t envisioned before.

References

1. Using AI to intercept pancreatic cancer in black patients. Cedars-Sinai. November 26, 2024. Accessed December 5, 2024. https://tinyurl.com/8rc87bxe
2. Qureshi TA, Gaddam S, Wachsman AM, et al. Predicting pancreatic ductal adenocarcinoma using artificial intelligence analysis of pre-diagnostic computed tomography images. Cancer Biomark. 2022;33(2):211-217. doi:10.3233/CBM-210273