Learning risk models for pancreatic cancer from electronic health records

Abstract

Pancreatic cancer is the third most lethal cancer in the U.S., causing an estimated 45,750 deaths in 2019. Of all treatments, surgical resection provides the best survival rate for pancreatic cancer. This is not feasible for the majority of pancreatic cancer patients, whose cancer is typically not diagnosed until the tumor is unresectable. Most symptoms of pancreatic cancer are typically subtle, which underscores the need for better risk modeling to predict a patient's chance of pancreatic cancer well before it would usually be diagnosed. We propose a series of novel models that apply standard machine learning techniques to Electronic Health Records (EHRs) to predict risk of pancreatic cancer in advance of cancer diagnosis. On the test dataset, two of our models achieved AUROCs of 0.85 (CI 0.81 - 0.90) and 0.79 (CI 0.77 - 0.82) with a 365-day lead time.