A transformer-based AI model demonstrated the ability to predict pancreatic cancer risk up to 5 years before diagnosis, according to findings from a study presented at the American College of Surgeons Clinical Congress 2026.  

Because universal screening for pancreatic cancer is not feasible and no population-level risk stratification tests are currently available, Mayo Clinic researchers sought to develop an AI model to identify patients at highest risk of developing the disease who might benefit from targeted screening.  

“We know that pancreatic cancer forms over five to seven years, but the things that a clinician or patient sees don’t happen until it’s too late,” explained study coauthor Cornelius Thiels, DO, MBA, FACS, a surgical oncologist at Mayo Clinic in Rochester, Minnesota.  

Study and Model Methods

The researchers identified adult patients from the Mayo Clinic Platform, including patients with pancreatic cancer and age- and sex-matched controls without cancer.  

They used longitudinal sequences of diagnostic histories coded using ICD-9 and ICD-10, spanning at least 1 year, along with routine laboratory test results, to train an AI model to predict patients’ 5-year risk of developing pancreatic cancer. The model was a transformer-based neural network incorporating a multihead attention mechanism.  

“We built this to be as generalizable, scalable, and easy to put into practice as possible,” said presenting author Chris Varghese, MBChB, a surgical data scientist at Mayo Clinic in Rochester. “If it’s shown to work, it could be used in almost any setting.” 

Performance measures included the area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC), calibration slope and intercept, positive predictive value, and negative predictive value.  

Study Findings

The study included 6,066 patients with pancreatic cancer and 36,396 controls. The AI model incorporated a median of 13 years of clinical history per patient (interquartile range = 7.5–19 years).  

The model achieved an AUROC of 0.853 (95% confidence interval [CI] = 0.847–0.858) and an AUPRC of 0.712 (95% CI = 0.706–0.717) for predicting pancreatic cancer risk 5 years before diagnosis.

The model demonstrated a calibration slope of 0.92 and an intercept of –0.030.  

At a predicted risk threshold of greater than 50%, the model achieved a positive predictive value of 88% (95% CI = 86%–90%). However, the best balance between positive predictive value and sensitivity was observed at a threshold of greater than 16%, with a positive predictive value of 63% and a negative predictive value of 94%.  

“Our model showed that a greater than 50% risk of pancreas cancer predicted by our model indicated an 88% likelihood of being diagnosed with pancreatic cancer in one year,” Dr. Varghese explained. 

The research team is now working to validate the model prospectively at Mayo Clinic. “We’re proving that we can move this from a retrospective research tool into our clinical environment and run it prospectively for validation,” said Dr. Thiels. He added that the researchers are also seeking to validate the model prospectively at a health system outside Mayo Clinic.  

“We are also working on developing more advanced machine learning architectures, which appear to improve the performance even more,” he added.

DISCLOSURES: The study authors reported no disclosures. 

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.