Risk prediction models for familial breast cancer

Women with a family history of breast cancer have an elevated risk of developing the disease. In clinical practice, the probability of developing breast cancer over a specified timeframe is frequently estimated using breast cancer risk prediction models. It is currently unclear which of the available models performs best in women with a breast cancer family history.

To identify, describe, and appraise breast cancer risk prediction models developed or validated in women with a family history of breast cancer, and to meta-analyse their performance in predicting breast cancer occurrence.

We searched MEDLINE, Embase, Cumulative Index to Nursing and Allied Health Literature, and the Institute of Scientific Information Web of Science to February 2022, with a targeted top-up search of MEDLINE to 19 December 2024 to capture additional validation studies of included models. We also screened reference lists of included studies.

We included studies that developed or validated a breast cancer risk prediction model(s) in women with a family history of breast cancer, if the model(s) in question included family history of breast cancer among its predictors.

We based our data extraction form on the CHARMS checklist. Two authors independently screened references, extracted data, and assessed risk of bias using the Prediction model Risk Of Bias ASsessment Tool (PROBAST). We summarised risk prediction models developed or validated in the target population. Where performance statistics were reported by at least four studies, we pooled model performance measures using random-effects meta-analyses. We assessed model performance using calibration (agreement between predicted risks and observed breast cancer occurrences) and discrimination (ability to distinguish between women who did and did not develop breast cancer). We did not apply GRADE because guidance for prognostic model reviews is not yet available.

We included 45 studies and listed 17 as 'awaiting classification'. We identified 12 externally validated models in the target population. We meta-analysed four models that had at least four external validation studies. Reporting of studies varied. Several did not adequately report follow-up time or handling of missing data. Most validation studies reported at least one performance measure (calibration or discrimination), though some did not report both together. Basic details, such as validated model versions, were often missing. Most studies had a high or unclear risk of bias based on PROBAST ratings, although concerns about applicability were generally low. We report results for four models which had data available from at least four external validation studies in the target population. Gail/Breast Cancer Risk Assessment Tool (BCRAT) Calibration: the pooled observed (O)/expected (E) ratio of the Gail model (combined versions) in the target population was 1.06 (95% confidence interval (CI) 0.91 to 1.25), indicating that the model is well calibrated in this population. The 95% prediction interval (PI) was 0.65 to 1.74. Discrimination: the pooled estimate for the C statistic of the Gail model (combined versions) in the target population was 0.61 (95% CI 0.57 to 0.66). The 95% PI was 0.47 to 0.74. Tyrer-Cuzick/International Breast Cancer Intervention Study (IBIS) Calibration: the pooled estimate for the O/E ratio of the Tyrer-Cuzick model in the target population was 0.86 (95% CI 0.74 to 0.98) (combined versions) and 0.90 (95% CI 0.76 to 1.06) (version 8), indicating that the model overpredicts breast cancer risk in this population. The 95% PI was 0.56 to 1.33 (combined versions) and 0.55 to 1.47 (version 8). Discrimination: the pooled estimate for the C statistic of the Tyrer-Cuzick model (combined versions) in the target population was 0.62 (95% CI 0.58 to 0.66). The 95% PI was 0.49 to 0.74. The pooled C statistic for version 8 of the Tyrer-Cuzick model was 0.64 (95% CI 0.58 to 0.71). The 95% PI was 0.46 to 0.79. Breast and Ovarian Analysis of Disease Incidence and Carrier Estimation Algorithm (BOADICEA) Calibration: the pooled estimate for the O/E ratio of BOADICEA (combined versions) in the target population was 0.98 (95% CI 0.90 to 1.17), indicating that the model is well calibrated in this population. The 95% PI was 0.89 to 1.09. Discrimination: the pooled estimate for the C statistic of BOADICEA (combined versions) in the target population was 0.65 (95% CI 0.58 to 0.71). The 95% PI was 0.44 to 0.81. BRCAPRO Calibration: the pooled estimate for the O/E ratio of the BRCAPRO model (combined versions) in the target population was 1.44 (95% CI 1.25 to 1.62), indicating that the model underpredicts breast cancer risk in this population. The 95% PI was 1.02 to 2.04. Discrimination: the pooled estimate for the C statistic of BRCAPRO (combined versions) in the target population was 0.64 (95% CI 0.54 to 0.73). The 95% PI was 0.37 to 0.84.

Our meta-analyses showed that the Gail (BCRAT) and BOADICEA models are well calibrated in women with a family history of breast cancer. The Tyrer-Cuzick (IBIS) model overpredicts risk, while BRCAPRO underpredicts risk in this population. In terms of discriminatory accuracy in the target population, no model was clearly superior. Tyrer-Cuzick version 8, BOADICEA, and BRCAPRO showed similar modest discrimination in our meta-analyses, which was slightly better than that of the Gail model. Considering both calibration and discrimination together, our findings suggest that the BOADICEA model is well calibrated in this population and shows similar (modest) discriminatory accuracy to Tyrer-Cuzick (version 8) and BRCAPRO, suggesting that it may be useful for patient management in the familial breast cancer risk setting. However, this cannot be interpreted as conclusive: we judged most included studies to have a high or unclear risk of bias; the number of validation studies included in the meta-analyses was small (≤ 10 for each model); and the contributing studies were heterogeneous in terms of prediction time horizons and case-mix. Room for improvement remains in terms of the discriminatory ability of existing breast cancer risk prediction models in women with a family history of breast cancer. Reporting of prognostic model studies is currently suboptimal.

Funded in part by a Health Research Board Cochrane Training Fellowship REGISTRATION: Protocol (2018) DOI: 10.1002/14651858.CD013185.