A parsimonious 19-predictor electronic health record model predicted 3-year incident pancreatic cancer with an AUC of 0.75 and a hazard ratio of 7.63 for the highest risk percentile.
Can a parsimonious EHR-based Cox model accurately predict incident pancreatic ductal adenocarcinoma in adults?
A parsimonious EHR-based risk model demonstrates strong discrimination and generalizability for predicting 3-year incident pancreatic cancer risk across U.S. and UK cohorts.
Abstract Background Pancreatic ductal adenocarcinoma (PDAC) is projected to become the second leading cause of cancer death in the United States by 2030. Early detection of PDAC improves outcomes. However, screening is impractical in the general population due to low disease incidence. We previously developed a PDAC prediction model using machine learning in an institutional electronic health record (EHR) database. Here, we aimed to improve generalizability, interpretability, and parsimony by developing and validating a Cox model in a national EHR database. Methods We used Optum Labs DataWarehouse (OLDW), a U.S. EHR and claims database, to develop a Cox model predicting incident PDAC in adults ages ≥40 years in 23 health systems (training cohort; N = 4,836,428). Elastic net with 10-fold cross-validation selected from candidate risk predictors, which included demographics, diagnoses/symptoms measured by International Classification of Diseases (ICD) codes, and routine laboratory values. Performance was assessed by 3-year area under the receiver operating characteristic curve (AUC) and calibration slope and intercept in 31 distinct health systems (validation cohort; N = 5,607,398). Sensitivity analyses excluded adults 50 years, those with abdominal imaging in the prior year, and PDAC diagnosed in the first 6 months, and stratified participants by sex. International validation was performed in the UK Biobank (UKB) (N = 498,754). Results In the training cohort (mean age 60.4 years), 14,405 patients developed PDAC, with a crude incidence rate (IR) of 56 per 100,000 person-years (PY); in the validation cohort (mean age 60.2 years), 11,693 patients developed PDAC (IR, 55/100,000 PY). The final elastic net model included 19 predictors. Top predictors included chronic pancreatitis and other gastrointestinal conditions, prior cancers, type 2 diabetes, elevated aspartate aminotransferase, current smoking, and male sex. 3-year AUC was 0.75 in both the training and validation cohorts; discrimination was equivalent in males and females. 3-year calibration in the validation cohort was excellent. The hazard ratio of PDAC in the top percentile of predicted risk compared to the 45th-55th percentile was 7.63 (95% CI, 6.85-8.49) and NNS in the top percentile was 128 (95% CI, 117-141). Performance was similar after excluding patients with recent abdominal imaging, PDAC diagnosed in the 6 months following index, and patients under 50. In UKB (IR, 44/100,000 PY), AUC was 0.71 with acceptable calibration. Conclusions A parsimonious EHR-based PDAC risk model developed in diverse U.S. health systems demonstrated strong 3-year discrimination and good generalizability to cohorts in the United States and United Kingdom. A subsequent prospective validation study will assess the feasibility of EHR-driven PDAC case-finding. Citation Format: Lucas A. Mavromatis, Viktor Zlatanic, Emil Agarunov, Long Chen, Shenin A. Sanoba, Leora I. Horwitz, Narges Razavian, Anirban Maitra, Tamas A. Gonda, Morgan E. Grams. Development and validation of a parsimonious electronic health record model for pancreatic cancer risk stratification abstract. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 1378.
Mavromatis et al. (2026) studied this question. A parsimonious 19-predictor electronic health record model predicted 3-year incident pancreatic cancer with an AUC of 0.75 and a hazard ratio of 7.63 for the highest risk percentile.