Deep Learning Tool Predicts Risk for Pancreatic Cancer
NCI-funded researchers collaborated with scientists from Denmark to develop a deep learning tool for predicting risk for pancreatic cancer, an aggressive cancer that’s difficult to diagnose in the early stages of disease.
The researchers established the model using clinical data from 6 million patients in Denmark’s national health system and 3 million in the U.S. Veterans Health Administration. Using diagnosis disease codes and clinical histories, they trained their model to predict if pancreatic cancer would occur, and if so, within what timeframe.
The model was able to project when pancreatic cancer would appear with reasonable accuracy; that is, within 3, 6, 12, and 36 months. The model proved to be substantially more accurate than current population-wide estimates of disease.
According to the authors, their model allows them to “cast a wider net” than traditional surveillance programs, which typically focus on people with a known risk, such as those who have family histories of disease or known genetic markers.
The researchers needed to make some adjustments when applying the model to U.S. vs. Demark populations, to offset differences in how clinicians denoted diseases using clinical codes. Applying the model to more global data, perhaps through multi-institutional collaborations (such as federated learning), could help overcome those variations, they noted.
Adding other types of real-world data (e.g., imaging, genetic, and patient-provided wearable devices), which go beyond the simple disease codes used to establish the model, also will help improve this diagnostic tool, they added.
Dr. Chris Sander, Harvard Medical School, and Dr. Søren Brunak, University of Copenhagen, who supervised this work, said, “We found the model was able to learn not only from symptoms very closely associated with pancreatic cancer (e.g., unspecified jaundice, abdominal pain, and tumors in the digestive organs), but also from longer disease histories (e.g., type II diabetes, substance use disorders). This allowed us not only to predict if cancer would occur but also to determine a timeframe, giving us the ability to forecast disease as far as 24 to 36 months in advance (albeit with lower accuracy for longer time intervals).”
Early detection of pancreatic cancer in the general population could be critical in real-world clinical practice. A machine learning tool like this offers a low-cost approach for detecting cancer when treatment is most effective.