Study
PLCO
(Learn more about this study)
Project ID
PLCO-277
Initial CDAS Request Approval
May 30, 2017
Title
Towards Precision Prevention: Testing a Machine Learning Algorithm in Pancreatic Cancer
Summary
Pancreatic cancer is a highly lethal cancer, thus identifying individuals at high risk for poor pancreatic cancer outcomes is critical. In line with the Precision Medicine Initiative, we propose to comprehensively evaluate the effect of multilevel risk factors, from genetics to behaviors, on pancreatic cancer outcomes using a machine learning algorithm that allows for epistatic assessments and the identification of co-occurring groups of risk factors that could be indicative of novel pancreatic cancer disease pathways. In a nested case control study from the Prostate, Lung, Colorectal and Ovarian Cancer Trial (PLCO; ~350 cases and 1400 controls), we propose to use a learning classifier system algorithm (LCS) to develop a comprehensive prediction model for pancreatic cancer and to empirically identify co-occurring risk factor subgroups related to pancreatic cancer risk. We will compare our prediction model to existing risk prediction models in literature using area under the curve estimates and comparisons of predictive accuracy from machine learning. We will also determine the potential clinical utility of identified risk factor subgroups by: 1) evaluating the absolute risk of developing pancreatic cancer in each risk factor subgroup using the Surveillance, Epidemiology, and End Results data; 2) conducting survival analyses (Cox Proportional Hazards) to determine which risk factor subgroups are related to early age at onset, later stage at diagnosis, and survival. The LCS algorithm will be adapted to allow for repeated measures, and imputations for missing data will also be explored. This will be one of the first studies to test machine learning methods in a pancreatic cancer population-based study. We hypothesize that machine learning methods could provide new insights into disease etiology and could have implications for precision medicine, given we will be determining whether individual behaviors, molecular markers, and genetic markers, co-occur or cluster in similar disease pathways to affect pancreatic cancer outcomes. This proposal is meant to be hypothesis-generating and to address pancreatic cancer outcomes across the cancer control continuum, from prevention to to survival.
Aims
1. Develop a comprehensive prediction model for pancreatic cancer using a LCS that evaluates genetic, molecular, nutritional, and biomedical risk factors while also identifying relevant risk factor sub-groups during model building. Hypotheses: Risk factor findings from previous PLCO studies will be validated; new risk factor subgroups will be identified and characterized; Risk factor subgroups will include combinations of genetics/behaviors (as opposed to just genetic or behavioral subgroup silos) potentially indicative of undiscovered, joint pathways to pancreatic cancer.
2. Compare the machine learning prediction model to existing, population-based prediction models. Hypothesis: The new machine learning model could improve prediction.
3. Determine the potential clinical relevance of identified risk factor subgroups in Aim 1 through statistical comparisons examining absolute disease risk, age at diagnosis, survivorship and tumor stage. Hypotheses: Empirically identified risk factor subgroups could improve absolute disease risk population estimates; new risk factor subgroups with earlier age at diagnosis, later stage at diagnosis, and poorer survival will be identified and could provide insight into improved definitions of high risk populations.
Collaborators
Rachael Stolzenberg-Solomon, Ph.D, National Cancer Institute, Bethesda, MD;
Jason Moore, Ph.D., University of Pennsylvania, Philadelphia, PA
Ryan Urbanowicz, Ph.D. University of Pennsylvania, Philadelphia, PA
Karthik Devarajan, Ph.D. Fox Chase Cancer Center, Philadelphia, PA
Laufey Amundadottir, Ph.D., National Cancer Institute, Bethesda, MD
