Combining Breast Cancer Risk Prediction Models

• Activity Number: 172 - Risk Prediction and Analysis
• Type: Contributed
• Date/Time: Monday, July 30, 2018 : 10:30 AM to 12:20 PM
• Sponsor: Section on Risk Analysis
• Author(s): Zoe Guan* and Danielle Braun and Lorenzo Trippa and Hajime Uno and Kevin S. Hughes and Giovanni Parmigiani
• Companies: and Harvard T. H. Chan School of Public Health and Harvard and Dana Farber Cancer Institute and Massachusetts General Hospital and Harvard T.H. Chan School of Public Health / Dana-Farber Cancer Institute
• Keywords: breast cancer; risk prediction; BRCAPRO; Gail; stacking

Abstract

There exist numerous breast cancer risk prediction models and different models often yield conflicting predictions for the same patient. It is important to integrate information from different models in order to provide clinicians with the most accurate risk predictions. Two of the most widely used breast cancer risk prediction models are BRCAPRO and Gail. BRCAPRO is a Mendelian model that estimates an individual's probability of carrying a BRCA1/BRCA2 mutation, as well as his or her future risk of breast cancer, based on family history and mutation prevalence and penetrance. Gail is a logistic regression model based mainly on non-genetic risk factors. We consider two approaches for combining these complementary models: 1) penetrance modification, in which we incorporate Gail relative risks into the penetrance functions used by BRCAPRO via a proportional hazards model, and 2) a variant of stacking, an ensemble method, in which we fit a logistic regression model with BRCAPRO and Gail predictions as input. We compare the performance of the combination models, BRCAPRO, Gail, and another widely used breast cancer risk prediction model, Tyrer-Cuzick, in real and simulated data.