Risk Prediction Models For Breast Cancer Subtypes Defined By Hormonal Receptor Status In The European Prospective Investigation Into Cancer And Nutrition
Kuanrong LI , International Agency for Research on Cancer , France
ISABELLE R. 1
, FERRARI P. 1
1 International Agency for Research on Cancer
Purpose: Several breast cancer (BC) risk models have been developed since the initial work by Gail in the late eighties. The discriminatory power (C-statistic) of these models is around 60% using questionnaire-based risk factors and 70% after combination with genetic variants. In this study, we investigated whether the predictive power could be improved by accounting for BC heterogeneity in the European Prospective Investigation into Cancer and Nutrition (EPIC).
Methods: Prospective data from 301,550 women were analyzed. Preliminary subtype-specific models included age, menopausal status, age at menopause, age at menarche, full term pregnancy (FTP), number of FTP, age at first FTP, breast feeding, hormone replacement therapy, height, body mass index (BMI), and the interaction between menopausal status and BMI. First primary cancers of other sites and non-BC mortality were considered as competing events. The predictive power was evaluated with a five-fold cross-validation.
Results: During an average follow-up period of 15 years, 13,164 BC cases were identified (ER+: 7,295; ER-: 1,613; unknown: 4,256). FTP, number of FTP, age at first FTP, and height showed differential associations between ER+ and ER- tumors. The five-fold cross-validation showed an average C-statistic of 69% (95% confidence interval: 66%, 72%) for ER+ and 56% (50%, 62%) for ER- tumors, and calibration values (expected/observed BC) were 1.14 (1.10, 1.19) and 0.98 (0.90, 1.07), respectively. Adding alcohol consumption and a composite dietary score to our models slightly improved the discriminatory power for ER- tumors (C-statistic: 59%; 53%, 65%).
Conclusion: Our subtype-specific models yielded higher predictive power for ER+ tumors than the previous overall models, although the predictive power for ER- tumors remained limited. Data on BC-related biomarkers, i.e. specific hormones and sets of fatty acids, as well as genetic variants available in EPIC nested case-control studies, will be integrated to the models.
Funding sources: IARC fellowship program.