ACS: Transformer-Based Model Can ID Pancreatic Cancer Risk

TUESDAY, Sept. 29, 2026 (HealthDay News) — A transformer-based model can predict personalized risk for pancreatic cancer five years prior to diagnosis, according to a study presented at the annual meeting of the American College of Surgeons, held from Sept. 26 to 29 in Washington, D.C.

Chris Varghese, M.B.Ch.B., from the Mayo Clinic in Rochester, Minnesota, and colleagues trained a transformer model to risk-stratify populations for targeted pancreatic cancer screening. The five-year pancreatic cancer risk was predicted by inputting coded diagnostic histories (one year minimum) into a custom transformer-based neural network with a multihead attention mechanism.

Data were included for 6,066 individuals with pancreatic cancer and 36,396 controls, with a median of 13 years of clinical histories. The researchers found that when predicting pancreatic cancer five years prior to diagnosis, the mean area under the receiver operating characteristic curve was 0.853 and the area under the precision-recall curve was 0.712. The calibration curve slope was 0.92 and intercept was −0.030, indicating that the model was well calibrated. The positive predictive value (PPV) was 88 percent at a screening threshold of >50 percent risk. At a screening threshold of >16 percent, the best balance between PPV and sensitivity was achieved with a PPV of 63 percent and negative predictive value of 94 percent.

“We built this to be as generalizable, scalable, and easy to put into practice as possible,” Varghese said in a statement. “If it”s shown to work, it could be used in almost any setting.”

Press Release

More Information