Breast cancer risk prediction in women aged 35–50 years: impact of including sex hormone concentrations in the Gail model

Abstract

BackgroundModels that accurately predict risk of breast cancer are needed to help younger women make decisions about when to begin screening. Premenopausal concentrations of circulating anti-Müllerian hormone (AMH), a biomarker of ovarian reserve, and testosterone have been positively associated with breast cancer risk in prospective studies. We assessed whether adding AMH and/or testosterone to the Gail model improves its prediction performance for women aged 35–50.MethodsIn a nested case-control study including ten prospective cohorts (1762 invasive cases/1890 matched controls) with pre-diagnostic serum/plasma samples, we estimated relative risks (RR) for the biomarkers and Gail risk factors using conditional logistic regression and random-effects meta-analysis. Absolute risk models were developed using these RR estimates, attributable risk fractions calculated using the distributions of the risk factors in the cases from the consortium, and population-based incidence and mortality rates. The area under the receiver operating characteristic curve (AUC) was used to compare the discriminatory accuracy of the models with and without biomarkers.ResultsThe AUC for invasive breast cancer including only the Gail risk factor variables was 55.3 (95% CI 53.4, 57.1). The AUC increased moderately with the addition of AMH (AUC 57.6, 95% CI 55.7, 59.5), testosterone (AUC 56.2, 95% CI 54.4, 58.1), or both (AUC 58.1, 95% CI 56.2, 59.9). The largest AUC improvement (4.0) was among women without a family history of breast cancer.ConclusionsAMH and testosterone moderately increase the discriminatory accuracy of the Gail model among women aged 35–50. We observed the largest AUC increase for women without a family history of breast cancer, the group that would benefit most from improved risk prediction because early screening is already recommended for women with a family history.Electronic supplementary materialThe online version of this article (10.1186/s13058-019-1126-z) contains supplementary material, which is available to authorized users.

Predicting breast cancer risk using personal health data and machine learning models

“…For the BCRAT [7] model and all machine learning models, we reported the testing data set AUC, otherwise known as the concordance statistic, and the Delong [39] 95% confidence interval of this statistic. We used AUC to evaluate our models because many previous breast cancer risk prediction papers used this metric [11–18]. We used a Delong [39] method to calculate the 95% confidence interval of each AUC because this method is a nonparametric approach that makes fewer assumptions than other approaches do.…”

“…These works implemented simple statistical models and incorporated inputs derived from costly and / or invasive procedures. These inputs included breast density [11–13], genetic single-nucleotide polymorphism [12–16], nipple aspirate fluid cytology [17], and / or hormone level data [12, 18]. All of these works made use of simple statistical models such as Cox proportional hazards regressions [11, 17], logistic regressions [12, 16], or Gail model implementations [13–15, 18].…”

Early detection of prostate gland and breast cancer risk based on routine check-up data using survival analysis trees for left-truncated and right-censored data

“…The performance of each of the methods tested, for different time ranges, is summarized in Figures 4A and 4B . We also marked the AUROC of Gail’s breast cancer risk estimation for 5 years horizon as reported in [13]. TVsuRF had the highest AUPR on every time interval, and the highest AUROC on all intervals except one (though differences were not statistically significant) for 730 days, where Gail’s score was best.…”

“…[A] Performance (AUROC mean±SD) of five prediction models for different time intervals. The grey dashed line represents the (time-independent) AUROC reported for Gail’s Risk factor model [13]. [B] AUPR.…”

“…Gail’s model [7,8], BRCAPRO [9], IBIS [10] and BOADICEA [11] for BC risk, and the Prostate Cancer Prevention Trial Risk Calculator (PCPTRC) for PGS risk [12]. Both models use a few clinical parameters and are not based on routine laboratory measurements, and their performance is relatively limited [13].…”