Assessment of breast cancer risk prediction models in a high-risk population.

Author(s): Chun J,  Schnahel F,  Ogunyemi O

Journal: Cancer Res

Date: 2009 Jan 15

Major Program(s) or Research Group(s): BGCRG

PubMed ID:

PMC ID: not available

Abstract: Background: A variety of risk assessment models have been developed to quantify a woman's risk for developing breast cancer. Although the Gail model (GM) is the most widely utilized model, there are limitations. The Tyrer-Cuzick model (TCM), which has been validated in the United Kingdom (UK), addresses these limitations but its performance in other populations is unclear. The purpose of this study was to evaluate the predictive ability of the TCM in a cohort of high-risk women from the New York metropolitan area and compare the results to the GM.Methods: The Women At Risk (WAR) Registry provided the study population. Due to the age limitation of the Gail model, we excluded women who were under 35 and over 80 years of age. Calculation of lifetime Gail scores included the following: age, race, age at first menses, age at first live birth, number of first degree relatives with breast cancer, number of previous breast biopsies, and atypical hyperplasia. These variables were also included in the lifetime Tyrer-Cuzick scores as well as information on menopausal status, use of hormone replacement therapy, LCIS, and age at diagnosis for relatives with ovarian and breast cancer. The Receiver Operating Characteristic (ROC) and estimated area under the ROC curve (AUC) with 95% confidence intervals (CI) were used to assess prediction accuracy of both models.Results: Out of a total study population of 1523 women, 82 (5%) developed breast cancer during a median follow-up of 56 months. The median age of women without breast cancer was 49 years, and the median age of women who developed breast cancer was 54 years. The AUC and 95% CI for the Gail model was 0.547 (0.479 to 0.615) and 0.501 (0.433 to 0.569) for the Tyrer-Cuzick model. Our results indicate poor discriminatory accuracy for both models, which performed no better than pure chance (0.5).