Risk prediction for estrogen receptor-specific breast cancers in two large prospective cohorts

Abstract

BACKGROUND: Few published breast cancer (BC) risk prediction models consider the heterogeneity of predictor variables between estrogen-receptor positive (ER+) and negative (ER-) tumors. Using data from two large cohorts, we examined whether modeling this heterogeneity could improve prediction. METHODS: We built two models, for ER+ (Model(ER+)) and ER- tumors (Model(ER-)), respectively, in 281,330 women (51% postmenopausal at recruitment) from the European Prospective Investigation into Cancer and Nutrition cohort. Discrimination (C-statistic) and calibration (the agreement between predicted and observed tumor risks) were assessed both internally and externally in 82,319 postmenopausal women from the Women's Health Initiative study. We performed decision curve analysis to compare Model(ER+) and the Gail model (Model(Gail)) regarding their applicability in risk assessment for chemoprevention. RESULTS: Parity, number of full-term pregnancies, age at first full-term pregnancy and body height were only associated with ER+ tumors. Menopausal status, age at menarche and at menopause, hormone replacement therapy, postmenopausal body mass index, and alcohol intake were homogeneously associated with ER+ and ER- tumors. Internal validation yielded a C-statistic of 0.64 for Model(ER+) and 0.59 for Model(ER-). External validation reduced the C-statistic of Model(ER+) (0.59) and Model(Gail) (0.57). In external evaluation of calibration, Model(ER+) outperformed the Model(Gail): the former led to a 9% overestimation of the risk of ER+ tumors, while the latter yielded a 22% underestimation of the overall BC risk. Compared with the treat-all strategy, Model(ER+) produced equal or higher net benefits irrespective of the benefit-to-harm ratio of chemoprevention, while Model(Gail) did not produce higher net benefits unless the benefit-to-harm ratio was below 50. The clinical applicability, i.e. the area defined by the net benefit curve and the treat-all and treat-none strategies, was 12.7 × 10(- 6) for Model(ER+) and 3.0 × 10(- 6) for Model(Gail). CONCLUSIONS: Modeling heterogeneous epidemiological risk factors might yield little improvement in BC risk prediction. Nevertheless, a model specifically predictive of ER+ tumor risk could be more applicable than an omnibus model in risk assessment for chemoprevention.