Breast Cancer Surveillance Consortium (BCSC): A Research Infrastructure sponsored by the National Cancer Institute Breast Cancer Risk Models William Barlow, PhD Senior Biostatistician, Cancer Research and Biostatistics Research Professor, Dept. of Biostatistics
Gail model (Gail et al., JNCI, 1989) Claus model (several references) BRCAPRO (many references) Breast cancer risk models in 1994
Population-based screening mammography population Prospectively collected risk factor information coded in an uniform manner Huge longitudinal dataset that is increasing Can examine the very important role of breast density as a risk factor, available on almost all screening mammograms Advantage of BCSC data for estimating risk
Barlow et al., JNCI, 2006 (77 citations) Short-term prediction of a diagnosis of breast cancer Tice, et al., Annals Int Med, 2008 (17 citations) Long-term prediction of a diagnosis of breast cancer Simplified model to be used in clinical care Janes, Pepe, and Gu, Annals Int Med, 2009 (17 citations) Compares the change in risk prediction from the Gail model to the Tice model Statistical technique for evaluating how well a model improves on past models for identifying high risk individuals BCSC breast cancer risk models
Published in the same JNCI issue as the revised Gail model which also includes breast density Based on one million women and 2.4 million screening mammograms Published in the same JNCI issue as the revised Gail model which also includes breast density BCSC risk prediction model
Prospective collection of self-reported risk factors: demographic information (age, race, ethnicity) family history of breast cancer previous breast procedures menopausal status and use of hormone therapy body mass index (BMI) Breast density (BI-RADS scale) and mammographic assessment reported by the radiologist Outcome is a diagnosis of breast cancer within one year of the screening mammogram Invasive breast cancer DCIS BCSC risk factors & outcomes
Available online along with the risk program and documentation: Dataset has been downloaded 228 times What’s not included? Mammographic outcomes; location; individual-level data What’s missing? genetic markers in-depth family history assessment environmental factors dietary factors (other than BMI) exercise levels BCSC risk data
Prediction of a diagnosis of breast cancer within a year of a screening mammogram Extrapolated to 5-years for comparison to the Gail model, but not intended for long-term prediction Most Gail risk factors are included Adds breast density to the risk model Purpose of the risk model
Observed incidence by age and breast density
Risk factors for pre-menopausal and post-menopausal women are different Required that a factor be significant to the level before including it Goal was to keep the model as simple as possible Develop the model on 75% and validate it on the remaining 25% Good calibration in the validation sample (correct estimation of cancer rates) Good prediction of individual outcomes (c-statistic) Risk model factors and assessment
Risk Model for Premenopausal Women Risk Factor LevelsOR 95% CI Age in years (versus Age 35-39) (versus Age 35-39) Age – 1.56 Age – 2.41 Age – 3.02 Prior breast procedure (versus No procedure) (versus No procedure) Yes – 1.65 Unknown – 1.10 First degree family hx (versus No family hx) (versus No family hx) Yes, 1 relative – 1.74 Yes, 2 or more relatives – 3.36 Unknown – 1.07 Breast density (BI-RADS) 1: Almost entirely fat : Scattered fibroglandular densities – : Heterogeneously dense – : Extremely dense – 6.28 Unknown/different system – 5.24 c = 0.63
Postmenopausal Predictors (c=0.62) Age Race Hispanic ethnicity Body mass index (BMI) Age at first birth Family history of breast cancer Previous breast surgery Menopausal status Type of menopause Hormone therapy use Breast density Previous screening result (false positive/true negative)
Five-year risk of invasive breast cancer based on extrapolation of the model: High and low risk women
Risk model summary This particular model did somewhat better than the Gail model All risk models have difficult getting beyond a c-statistic of 0.65 Still cannot identify a group of women who are risk-free Risk models may identify high-risk groups, but still have difficulty with precise estimation of risk for a particular woman
Clinical implications Breast density was a known risk factor, but we did not appreciate its importance relative to other risk factors Can better estimate risk for prevention efforts Breast density may be an early marker of a reduction in risk
Next steps Augment the screening mammogram assessment with risk factor information to better estimate outcomes from positive or negative screening mammograms Use changes in risk factors to develop better models of long-term risk prediction Incorporate extensive family history and and SNPs into risk models with breast density (P01 application by BCSC)