UNIVERSITY OF WISCONSIN SYSTEM
Breast cancer is a disease for which personalized risk prediction could vastly improve care. Our currently funded R01 has enabled us to develop an algorithm called view learning that integrates predictive data from a complex multi-relational mammography database by automatically defining new fields which optimizes a Mammography Bayesian network (MBN) that estimates breast cancer risk. Recently, genome-wide association studies (GWAS) and large population based studies have recognized that genetic and epidemiologic risk factors profoundly influence breast cancer risk. This supplement proposal will test whether incorporating Single Nucleotide Pleomorphisms (SNPs) and rich epidemiologic data into our MBN using view learning can improve breast cancer risk prediction for women undergoing mammography.
Our team has recently published several of our R01 milestones which show that our MBN can outperform clinical radiologists on a large retrospective dataset and that our algorithms discover new predictive variable not previously established. For this supplement, we have the following goals which clearly fit within the scope of our peer-reviewed and approved R01 activities: first, to hire a Research Program/Database Manager to construct a multi-relational dataset incorporating patient-specific mammography, epidemiologic and SNP data and second, to hire a Research Assistant to focus on development and testing of view learning that incorporates these new variables.
We aim to prove the following hypotheses: view learning can 1) use mammography, genetic and epidemiologic data to improve our MBN risk prediction accuracy and 2) discover new valuable combinations of predictors (called predicate invention). We will employ a retrospective case control design including patients from two valuable databases, the Marshfield Clinic Personalized Medicine Research Project and the University of Wisconsin Breast Cancer Epi/Biobank, which contain 367 and 404 breast cancer cases and age matched controls respectively. By integrating epidemiologic and genetic information into our multi-relational database and algorithms, the new team members we plan to hire will expand our capacity to improve risk prediction and accelerate the tempo of discovery of novel concepts to optimize and personalize breast cancer diagnosis.