Abstract 4945: Absolute risk prediction models for pancreatic cancer

Abstract

Abstract Introduction: Pancreatic cancer is one of the leading causes of cancer deaths in Western countries and has a 5-year survival rate of less than 10% as a high proportion of patients are diagnosed with advanced disease. Efforts are under way to develop biomarkers and imaging technologies for early detection. Given the low incidence of pancreatic cancer in the general population, a targeted approach that screens only those at high risk may be most effective. Here we build and evaluate prediction models for pancreatic cancer, including established lifestyle, clinical, anthropometric and genetic risk factors as well as emerging risk biomarkers. Methods: We identified incident cases and age-matched controls from four large U.S. population-based cohort studies (Health Professionals Follow-up Study, Nurses' Health Study, Physicians' Health Study, and Women's Health Initiative Observational Study); cases were diagnosed from 1984 to 2010. Data on lifestyle, clinical, and anthropometric factors were obtained from questionnaires completed around the time of blood sample collection; all cases were diagnosed after blood collection. We used conditional logistic regression models to build a relative risk model including (1) alcohol use, body mass index (BMI), waist-to-hip ratio, height, physical activity, periodontal disease, and diabetes; (2) a weighted genetic risk score incorporating 17 common single-nucleotide polymorphisms associated with pancreatic cancer risk; and (3) circulating biomarkers, such as metabolites, proinsulin, leptin, and others. We calculated five-year absolute risks by combining this relative risk model with gender-specific SEER incidence and U.S. mortality rates. We evaluated model discrimination by calculating the area under the ROC curve (AUROC) using a 4-fold cross-validation. We assessed predictive performance by plotting the distribution of age-specific risk in the general population. Results: We identified 503 incident cases and 1,174 matched controls. The AUROC for the final model was 0.68, which is an improvement over previously published risk models that only included a subset of the lifestyle, clinical, anthropometric, and genetic risk factors considered here (AUROCs ranging from 0.58 to 0.61 [Klein et al., PLoS One 2013]). The final model identified 1.6% of the general population who had greater than or equal to 3 times the population average risk of pancreatic cancer; this represents a four-fold increase over previously published models. Conclusions: We derived risk prediction models for pancreatic cancer in the general population, incorporating clinical, genetic, and biomarker information. We found that the models improved discrimination over existing models and could identify a small subset of individuals at notably higher risk of pancreatic cancer. These models may be useful to identify a targeted high-risk subset of the general population who could benefit from screening for pancreatic cancer. Citation Format: Jihye Kim, Chen Yuan, Ana Babic, Ying Bao, Lauren K. Brais, Marisa W. Welch, Meir Stampfer, Edward L. Giovannucci, Howard D. Sesso, JoAnn E. Manson, Charles S. Fuchs, Brian M. Wolpin, Peter Kraft. Absolute risk prediction models for pancreatic cancer [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2018; 2018 Apr 14-18; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2018;78(13 Suppl):Abstract nr 4945.