AI tool helps identify pancreatic cancer risk up to 3 years before diagnosis
Click Here to Manage Email Alerts
Researchers at Harvard Medical School and University of Copenhagen developed an artificial intelligence tool with the potential to identify people at greatest risk for pancreatic cancer up to 3 years before diagnosis.
The tool — which uses patients’ medical records to determine elevated risk for the disease — is based on data from 6 million patients in Denmark. It underwent subsequent evaluation in collaboration with VA Boston Healthcare System, Dana-Farber Cancer Institute and Harvard T.H. Chan School of Public Health.
A paper published in Nature Medicine suggested that artificial intelligence (AI)-based population screening may be useful for identifying high-risk individuals and could help accelerate detection of a malignancy often diagnosed at advanced stages.
“Only 12% of patients are alive after 5 years, and it’s pretty much the same all over the world,” study co-senior investigator Søren Brunak, PhD, professor of disease systems biology and director of research at University of Copenhagen’s Novo Nordisk Foundation Center for Protein Research, told Healio. “It also is on the rise — for not very well-known reasons. So, the unmet need in this domain is quite significant.”
Healio spoke with Brunak about how the AI tool works, how it has performed to date, and the potential implications the approach may have for pancreatic cancer detection and outcomes.
Healio: How did you develop this AI tool?
Brunak: My group at University of Copenhagen initiated this collaboration with Chris Sander, PhD, and the Sander lab at Harvard Medical School, but my group has been working for around 10 years on what we call disease trajectories that are used as data input to the tool. Instead of just saying you have one diagnosis and one disease, we are looking at longitudinal, temporal patterns. In some sense, we try to redefine diseases as trajectories that go over several diagnoses. Fortunately, in Denmark and other Nordic countries — and, of course, in the United States — we have data that go back quite a long time.
For example, for someone like me, there will be diagnostic history in the Danish databases from the past 45 years. Everything that happened to me in the hospital system for the past 45 years is recorded. We can essentially take that trajectory and plug it into a deep learning algorithm to see whether we can discriminate individuals with pancreatic cancer from those who have not yet developed pancreatic cancer. These disease trajectories are quite effective, because they implicitly combine the germline risk — which is the genetic risk — with lifestyle factors. It is a combination of lifestyle-provoked disease and inherited disease, and that’s why it has a lot of predictive potential.
We designed several versions of the AI model and trained them on the health records of 6.2 million patients from Denmark’s national health system over 41 years. We then trained the AI models to discriminate.
Healio: How has it performed and what factors are predictive of pancreatic cancer?
Brunak: We looked at whether it would pick up many of the known risk factors because, when we have trained an algorithm, we can do this explainability exercise through which we try to look at the parameters in a neural network that trigger it toward coming up with a diagnosis of pancreatic cancer. It did pick up on many of the known risk factors, and that of course is a good sign. It has learned from these data and combined the risk factors in a nonlinear way. During the training, the AI algorithm identified patterns suggestive of pancreatic cancer risk based on disease trajectories. It identified diagnoses such as gallstones, anemia, type 2 diabetes and other gastrointestinal conditions as predictive of risk for pancreatic cancer.
Healio: What are the next steps in research?
Brunak: This is a hospital data-based exercise and, of course, many of the signs of this disease are more general practice-based. For example, a patient might have stomach problems or constipation for which they will go see their general practitioner. These aren’t necessarily symptoms that would lead to a person becoming an inpatient at a hospital. We are working on incorporating more general practice trajectories into the algorithm. We are also working to reduce false positives, although it already performs quite well.
For more information:
Søren Brunak, PhD, can be reached at Novo Nordisk Foundation Center for Protein Research, University of Copenhagen, Blegdamsvej 3B, DK-2200, Copenhagen; email: soren.brunak@cpr.ku.dk.
Collapse
Click Here to Manage Email Alerts