A systematic review of prediction models for risk of breast cancer

Abstract

Background: Predicting the risk of breast cancer (BC) can help with early diagnosis, treatment, and prevention strategies. Existing BC risk prediction models typically utilize demographic, genetic, and/or imaging-derived variables. This systematic review summarizes all developed models for BC risk in general and high-risk populations, and their discriminatory ability and calibration. Methods: MEDLINE and Embase were searched for studies developing and/or validating models estimating risk of developing BC in women and/or men. After removal of duplicates, 9,511 titles and abstracts were screened, yielding 360 full-text reviews. The current review focuses on all studies which developed a new model to predict BC risk in general and high-risk populations. Results: A total of 107 studies developing new models for BC risk prediction were included in this review. Sample sizes ranged from 535 to 2,392,998 for cohort studies, and from 133 cases with 113 controls to 95,075 cases with 75,017 controls for case–control studies. Areas under the receiver-operating characteristic curves (AUCs) ranged from 0.51 to 0.96. For studies reporting calibration using observed/expected events (O/E) ratio (n = 8), the range was 0.84 to 1.10. External validation was reported in 18 studies. Conclusions: Most BC risk models were developed in Caucasian populations, and their predictive ability and quality varied. Models including both demographic and genetic or imaging/biopsy data performed better than models based on demographic variables alone. However, their performance was not further improved by the addition of multiple types of such data.