Commentary on Terry et al. , 10-Year Performance of Four Models of Breast Cancer Risk: A Validation Study. Lancet Oncol. 2019;20(4):504-17

Terry and colleagues [1] provided much needed information on the performance of four breast cancer risk prediction models in a high-risk population, the Breast Cancer Prospective Family Study Cohort (ProF-SC). Women in ProF-SC were recruited from high-risk clinics or were relatives of women with breast cancer in registries [2–4]. Terry and colleagues studied a subset of 15,732 women who were aged 20 to 70 years old and without previous breast cancer. They were followed for a median of 11.1 years to estimate 10-year risk of developing breast cancer. Eighty-two percent had at least one first-degree relative with breast cancer and 6.83% carried a BRCA1 or BRCA2 mutation (hereafter called BRCA mutation). For comparison, no more than 15% of women in the U.S. have an affected relative [5], and the prevalence of BRCA mutation carriers in the general population is only about 0.32% [6]. I estimated [7] that the annual breast cancer incidence rate in ProF-SC was 508/105, which is 3.04 times greater than the U.S. National Cancer Institute’s Surveillance, Epidemiology and End Results Program (SEER) rate for white women (from 2011 to 2015), adjusted to the ProF-SC age distribution. Thus, ProF-SC is a high-risk cohort, ideally suited to assess the performance of risk models like BRCAPRO [8], BOADICEA [6] and IBIS [9], which include an autosomal dominant component of genetic risk and are widely used in genetic counselling. ProF-SC provides much more information on the performance of these models than all previous studies. Terry and colleagues also evaluated the Breast Cancer Risk Assessment Tool (BCRAT) [10,11] that was designed for the general U.S. population. BCRAT warns that “This tool cannot be accurately used for women carrying a breast-cancer-producing mutation in BRCA1 or BRCA2” (https://bcrisktool.cancer.gov/). In fact, BCRAT recommends BOADICEA for such women. There are important differences among these models. BCRAT is empirical and does not invoke a genetic model. The only information required is age, age at menarche, age at first live birth, number of previous breast biopsies (0, 1, ≥2), presence of atypical hyperplasia on any biopsy, number of affected first-degree relatives (0, 1, ≥2), and race/ethnicity. The Open Access

performance of four breast cancer risk prediction models in a high-risk population, the Breast Cancer Prospective Family Study Cohort (ProF-SC).
Women in ProF-SC were recruited from high-risk clinics or were relatives of women with breast cancer in registries [2][3][4]. Terry and colleagues studied a subset of 15,732 women who were aged 20 to 70 years old and without previous breast cancer. They were followed for a median of 11.1 years to estimate 10-year risk of developing breast cancer. Eighty-two percent had at least one first-degree relative with breast cancer and 6.83% carried a BRCA1 or BRCA2 mutation (hereafter called BRCA mutation). For comparison, no more than 15% of women in the U.S. have an affected relative [5], and the prevalence of BRCA mutation carriers in the general population is only about 0.32% [6]. I estimated [7] that the annual breast cancer incidence rate in ProF-SC was 508/10 5 , which is 3.04 times greater than the U.S. National Cancer Institute's Surveillance, Epidemiology and End Results Program (SEER) rate for white women (from 2011 to 2015), adjusted to the ProF-SC age distribution. Thus, ProF-SC is a high-risk cohort, ideally suited to assess the performance of risk models like BRCAPRO [8], BOADICEA [6] and IBIS [9], which include an autosomal dominant component of genetic risk and are widely used in genetic counselling. ProF-SC provides much more information on the performance of these models than all previous studies.
Terry and colleagues also evaluated the Breast Cancer Risk Assessment Tool (BCRAT) [10,11] that was designed for the general U.S. population.
BCRAT warns that "This tool cannot be accurately used for women carrying a breast-cancer-producing mutation in BRCA1 or BRCA2" (https://bcrisktool.cancer.gov/). In fact, BCRAT recommends BOADICEA for such women.
There are important differences among these models. BCRAT is empirical and does not invoke a genetic model. The only information required is age, age at menarche, age at first live birth, number of previous breast biopsies (0, 1, ≥2), presence of atypical hyperplasia on any biopsy, number of affected first-degree relatives (0, 1, ≥2), and race/ethnicity. The

Open Access
Received: 08 August 2019  women who were not tested for BRCA mutations and whose affected relatives were not tested. I call these women "non-tested." The breast cancer incidence rate was higher in the non-tested women than in the tested women, probably because some of the non-tested women were BRCA-positive. I assumed that those who died, developed breast cancer or for non-tested "BRCA-negative" women. If the non-tested group had had the same incidence rate as the tested BRCA-negative group, the incidence rate in the entire "BRCA-negative" group would have been reduced by a factor of (0.368 0.0392 0.632 0.0392) / (0.368 0.0467 0.632 0.0392) 0.899 with a corresponding reduction in breast cancers observed. Thus, the E/O ratios in Table 1 for all BRCA-negative women would be increased by a factor of 1/0.899 = 1.112, as shown in the row labeled "All adjusted" in Table 1. Similar adjustments are given separately for BRCA-negative women <50 and ≥50 years old with respective divisors 0.967 and 0.939 (Table 1). These adjustments suggest that BOADICEA, IBIS and BCRAT may overestimate risk slightly in women aged ≥50 years who tested negative or whose affected relatives tested negative. In the subset of 7737 BRCA-negative women with at least one affected second-degree relative, BOADICEA and IBIS underestimated risk if only data on first-degree relatives were used.
There is reason to believe that high concordances of BRCAPRO, BOADICEA and IBIS in ProF-SC reflect their incorporation of BRCA mutation data, rather than their extensive use of pedigree data. Their concordances were higher in younger women (Table 1), which is consistent with the higher relative risks from BRCA mutations in younger women [6]. In the BRCA-negative group, their concordances were little different from that of BCRAT, which uses only number of affected firstdegree relatives. ProF-SC offers guidance for practice. BOADICEA and IBIS are well calibrated and take advantage of information on mutation status to improve discriminatory ability in high-risk populations like ProF-SC.
BRCAPRO underestimated risk substantially in this population.
Surprisingly, the simpler BRCAT performed reasonably well in BRCAnegative women.
Although well calibrated risk models provide useful information for clinical decisions and some public health applications, such as designing chemoprevention trials [15], higher concordances (well above 0.8) are needed for deciding who should not be screened or for achieving an appreciable population benefit from preventive interventions with adverse effects [16]. Previous theoretical calculations [16,17] indicated that adding mammographic density and single nucleotide polymorphism (SNP) scores might achieve a concordance near 0.7. IBIS recently introduced these factors [15], as did BOADICEA [18]. Assuming that risk had a lognormal distribution, I calculated from figures 1c and 2c in [18] that the modified BOADICEA had a concordance near 0.7 in the general population without BRCA measurements. Models like these combined with safer preventive interventions could reduce population risk. To achieve these benefits, safer interventions are needed, and the calibration of new risk models needs to be assessed with prospective studies such as that by Terry and colleagues.