Prediction Model for Detection of Sporadic Pancreatic Cancer (PRO-TECT) in a Population-Based Cohort Using Machine Learning and Further Validation in a Prospective Study

OBJECTIVES: There is currently no widely accepted approach to screening for pancreatic cancer (PC). We aimed to develop and validate a risk prediction model for PC across two health systems using electronic health records (EHR). METHODS: This retrospective cohort study consisted of patients 50-84 years of age meeting utilization criteria in 2008-2017 at Kaiser Permanente Southern California (KPSC, model training, internal validation) and the Veterans Affairs (VA, external validation). ‘Random survival forests’ models were built to identify the most relevant predictors from >500 variables and to predict PC within 18 months of cohort entry. A prospective study was then conducted in KPSC to assess feasibility of the model for real-time implementation. RESULTS: The KPSC cohort consisted of 1.8 million patients (mean age 61.6) with 1,792 PC cases. The estimated 18-month incidence rate of PC was 0.77 (95% CI 0.73-0.80)/1,000 person-years. The three models containing age, abdominal pain, weight change and two laboratory biomarkers (ALT change/HgA1c, rate of ALT change/HgA1c, or rate of ALT change/rate of HgA1c change) had comparable discrimination and calibration measures (c-index: mean=0.77, SD=0.01-0.02; calibration test: p-value 0.2-0.4, SD 0.2-0.3). The VA validation cohort consisted of 2.6 million patients (mean age 66.1) with an 18-month incidence rate of 1.27 (1.23-1.30). A total of 606 patients were screened in the prospective pilot study at KPSC with 9 patients (1.5%) diagnosed with a pancreatic or biliary cancer. CONCLUSIONS: Using widely available parameters in EHR, we developed a population-based parsimonious model for early detection of sporadic PC suitable for real-time application.