One versus Two Breast Density Measures to Predict 5- and 10-Year Breast Cancer Risk

This article is featured in Highlights of This Issue, p. 887

Breast density is one of strongest risk factors for breast cancer and provides important information for risk assessment. Breast Imaging Reporting and Data System (BI-RADS; ref. 1) breast density has been shown to improve the discriminatory accuracy in two breast cancer risk prediction models (2, 3). Interrater agreement of BI-RADS breast density assessment is moderate and has raised concern about using this subjective measure of density in risk prediction models to evaluate individual risk (4–8). Misclassification of BI-RADS categories could result in an under- or overestimation of breast cancer risk.

Our prior work (9) has shown that an increase in BI-RADS breast density category is associated with an increase in breast cancer risk, and a decrease in density with a decrease in risk, within 1 year of the most recent mammogram. Reductions in breast density after 12 to 20 months of tamoxifen therapy have been associated with reduced risk of breast cancer among high-risk women, and reduced risk of second breast cancer events in breast cancer survivors (10, 11).

We developed a 5- and 10-year breast cancer risk prediction model that uses the same risk factors as the Breast Cancer Surveillance Consortium (BCSC) risk model, but included two measures of BI-RADS breast density instead of one, to assess whether sequential density measures can improve risk prediction by providing a more precise density assessment and/or enable evaluation of changes in breast density that influence breast cancer risk.

Data were pooled from seven mammography registries that participate in the BCSC (http://breastscreening.cancer.gov; ref. 12). Registries collect data, including patient characteristics and clinical information, from community radiology facilities. Breast cancer diagnoses are obtained by linking women in the BCSC to pathology databases; regional Surveillance, Epidemiology, and End Results (SEER) programs; and state tumor registries with completeness of reporting estimated at >94.3% (13). Vital status is obtained through linkage to SEER registries, state tumor registries, and the state death tapes. Data are pooled at a Statistical Coordinating Center. Registries and the Coordinating Center have received Institutional Review Board approval for active or passive consenting processes or a waiver of consent to enroll participants, link data, and perform analyses. All procedures were Health Insurance Portability and Accountability Act compliant, and registries and the Coordinating Center received a Federal Certificate of Confidentiality and other protections for the identities of women, physicians, and facilities.

The study sample included women of ages 35 to 74 years who had at least two mammograms with nonmissing BI-RADS density between January 1994 through December 2010. We excluded women with a history of ductal carcinoma in situ (DCIS) or invasive breast cancer, breast implants, or mastectomy prior to the second mammogram. We selected a woman's earliest pair of screening or diagnostic mammograms for which the time between mammograms was ≥9 months and ≤4 years for a sample of mammograms on average 1 to 2 years apart, consistent with the recommended screening frequency in the United States (14), and for which family history of breast cancer, history of breast biopsy, and race/ethnicity were nonmissing at the second examination (see Supplementary Figure). Women diagnosed with DCIS or invasive breast cancer the 3 months following their second examination were also excluded. Our study population had similar distributions of age, race/ethnicity, breast density, and family history of breast cancer as the distributions among all women in the BCSC (data not shown).

Demographic and breast health history information were obtained on a self-administered questionnaire completed at each mammography examination. We obtained self-reported information on history of first-degree relatives (mother, sister, or daughter) with breast cancer, history of breast biopsy, and race/ethnicity (non-Hispanic white, non-Hispanic black, Hispanic, Asian/Native Hawaiian/Pacific Islander, Native American/Native Alaskan, or other/mixed race).

Radiologists categorized breast density at the time of clinical interpretation of the mammogram, as part of clinical practice, using American College of Radiology's BI-RADS breast density categories (1): (i) almost entirely fat, (ii) scattered fibroglandular densities, (iii) heterogeneously dense, or (iv) extremely dense. We classified women into 1 of 16 possible density combinations based on the densities assigned at the two examinations.

Women were considered to have breast cancer if diagnosed with invasive carcinoma during the follow-up period.

Risk factor frequency distributions were determined for women with and without breast cancer. We used the Cox proportional hazards regression to model time to invasive breast cancer for the same covariates included in the BCSC risk model (2): age at entry (linear and quadratic terms), race/ethnicity, history of first-degree relatives with breast cancer, and history of breast biopsy. We also included interaction terms between age at entry (linear) and BI-RADS density, first-degree relatives with breast cancer, and race/ethnicity and between age at entry (quadratic) and first-degree relatives with breast cancer. Age was modeled using linear and quadratic terms because breast cancer incidence increases nonlinearly with age.

We fit two models. In the first model, we used the BI-RADS density assessment from the woman's most recent mammogram. In the second model, we included both density assessments. All other covariates were assessed at a woman's most recent mammogram. Follow-up time started 3 months after the most recent mammogram. Women were censored at the time of death, diagnosis of DCIS, mastectomy, end of complete cancer follow-up by mammography registries, or 10 years after study entry. We assessed the proportional hazards assumption by calculating interval-specific hazard ratios (HR; i.e., 0–3 months, 3–6 months, 6 months–1 year, 1 year–2 years, etc.), which appeared consistent over time for each predictor variable.

We estimated and plotted the cumulative incidence of invasive cancer by the most recent BI-RADS density measure and by the 16 combinations of two BI-RADS density measures accounting for the competing risks of DCIS diagnosis, mastectomy, and death.

We developed absolute risk models to estimate the 5- and 10-year risk of invasive breast cancer based on one and two density measures in addition age, race/ethnicity, family history, and history of breast biopsy. We estimated the baseline breast cancer risk from the age- and race/ethnicity–specific incidence of invasive breast cancer from the SEER 18 registries (2000–2010), which provide the most representative data for the U.S. population (15). For each race/ethnicity group, we estimated age-specific incidence by fitting a third-order polynomial model to the SEER data. We used the methods described in Gail and colleagues (16) for translating the HRs and risk factor distributions into absolute risks. We estimated the age- and race/ethnicity–specific distributions needed to standardize HRs to be relative to average risk using data from a larger set of 4,610,085 mammograms from the BCSC. Specifically, we used logistic regression to model BI-RADS density, history of first-degree relatives with breast cancer, and history of breast biopsy as a function of age (linear and quadratic terms), race/ethnicity, and an interaction between race/ethnicity and age (linear term), among women of ages 40 years and older. We used the model-predicted probabilities to estimate the proportion of women in each age, race/ethnicity, and predictor category to adjust the HRs to be relative to average risk. The age- and race/ethnicity–specific competing risk of death for women was calculated using 2010 U.S. Vital Statistics (17). Age-specific mortality for each race/ethnicity group was estimated by fitting an exponential model to the all-cause mortality rates, adjusted for mortality due to breast cancer. The age- and race/ethnicity–specific competing risk of DCIS was estimated by fitting separate models by race/ethnicity with a third-order polynomial effect of age to SEER breast in situ rates. We applied the adjustments for whites to women of other/mixed race for whom insufficient data were available.

We assessed model calibration by calculating the ratio of the expected breast cancer rate from the absolute risk model (E) to the observed breast cancer rate (O) by age group, race/ethnicity, and individual risk factor distributions. We used the Kaplan–Meier estimator to estimate the observed rate within each subgroup to take into account censoring (18). We assumed that both the observed and the expected number of breast cancers follow a Poisson distribution and calculated the 95% confidence interval (CI) for E/O and 5- and 10-year risk of breast cancer as follows: (E/O) × exp (±1.96 × 1/sqrt(observed number of breast cancers)) and 5- and 10-year risk ± 1.96 × 100 × sqrt (expected number of breast cancers)/number of women in the subgroup, where the expected number of breast cancers is equal to the number of women in the subgroup × 5- and 10-year risk/100.

We used risk reclassification tables (19, 20) to compare the performance of the two-measure versus one-measure density models. Women were cross-classified on the basis of their risks estimated by the Breslow estimator of the 5-year survivor function from the two Cox models (21), using risk categories 0% to 1.66%, 1.67% to 3%, and >3%. We used the Kaplan–Meier estimator to estimate the number of breast cancer events and nonevents at 5 years within each cross-classified risk category, as in French and colleagues (18).

The discriminatory accuracy of the model was summarized using the area under the time–dependent receiver operating characteristic curve (AUC; ref. 22) at 5 and 10 years, with absolute risk as the marker. We performed 5-fold cross-validation to confirm the internal validity of the model (23, 24). AUCs were calculated separately for all women and for the subset of women whose BI-RADS density category changed from their previous to their most recent mammogram. All analyses were performed using SAS (version 9.3, SAS Institute, Cary, NC) and R (version 3.0.3, R Foundation for Statistical Computing, Vienna, Austria). The survivalROC package (25) was used to estimate time-dependent receiver operating characteristic curves. The cmprsk package (26) was used to estimate cumulative incidence functions for invasive breast cancer in the presence of competing risks.

We included 722,654 women of ages 35 to 74 years who underwent two mammograms with BI-RADS density measures on average 1.8 years (range, 0.75–4 years) apart; 13,715 developed invasive breast cancer during a mean of 6.6 years of follow-up (range, 1 day to 10 years). The overall 5- and 10-year cumulative incidences were 1.32% and 2.87%, respectively. Women with invasive breast cancer were more likely to be older, white, have a family history of breast cancer, and have heterogeneously or extremely dense breasts (Table 1). A total of 63.5% of women had the same BI-RADS density on two sequential examinations while 17.9% had an increase in breast density category and 18.6% had a decrease. The most common combinations of changing density categories were heterogeneously dense on the earlier examination and scattered fibroglandular on the most recent examination (10.0%) and scattered fibroglandular densities on the earlier examination and heterogeneously dense on the most recent examination (9.9%).

Figure 1A–D shows the 10-year cumulative incidence curves of invasive breast cancer for the combinations of BI-RADS density measures. Women with fatty breasts on their most recent mammogram had similar low risks of breast cancer regardless of previous density measures. Women with scattered fibroglandular densities on their most recent mammogram and prior density of heterogeneously dense (10% of screened women) or extremely dense (0.6% of screened women) were at higher risk of breast cancer than those women whose density remained scattered fibroglandular densities on both examinations. In contrast, women with scattered fibroglandular densities on their most recent mammogram and prior density of almost entirely fat (3.5% of screened women) were at lower risk of breast cancer than those women whose density remained scattered fibroglandular densities on both examinations. Women with heterogeneously dense breasts on their most recent mammogram and prior density of extremely dense (4.2% of screened women) were at higher risk of breast cancer than those women whose density remained heterogeneously dense. Conversely, women with heterogeneously dense breasts on their most recent mammogram and prior density of scattered fibroglandular densities (9.9% of screened women) were at lower risk of breast cancer than those women whose breast density remained heterogeneously dense. Women with extremely dense breasts on their most recent mammogram had a high risk of breast cancer regardless of previous density measures.

In general, the strength of the breast density association with breast cancer was greatest for women with extremely dense breasts on a recent or previous mammogram and lowest for women with almost entirely fatty breasts on a recent or previous mammogram relative to women who had scattered fibroglandular densities on two sequential mammograms (Table 2). The strength of the breast density association with breast cancer decreased significantly with increasing age; for example, the HR for extremely dense versus scattered fibroglandular tissue decreased from 2.1 (95% CI, 1.9–2.3) for women 35 to 49 to 1.3 (95% CI, 1.2–1.5) for women 65 years and older (data not shown). The strength of the family history association with breast cancer also tended to decrease with increasing age, for example, from HR of 1.8 (95% CI, 1.6–1.9) for women of ages 35 to 49 years to HR 1.5 (95% CI, 1.4–1.6) for women of ages 65 years and older, although this interaction was only borderline statistically significant (P values for tests of interaction with linear and quadratic age = 0.09 and 0.15, respectively).

The two-measure density 5-year risk model was well calibrated overall (Table 3). The two-measure density model discrimination, as measured by the AUC, was 0.640 (0.639 from 5-fold cross-validation) and was marginally greater than that of the one-measure density model AUC of 0.635 (0.635 from 5-fold cross-validation). Similar results were observed for the 10-year two- and one-measure density models; AUC, 0.628 and 0.622, respectively. Among women who changed density categories, discrimination for the two-measure density 5-year risk model was 0.641 (0.639 from 5-fold cross-validation and for the one-measure density model 0.630 (0.629 from 5-fold cross-validation). Calibration of the two-measure density 5-year risk model was reasonably accurate across risk factor subgroups (Table 3). The model was well calibrated across family history of breast cancer and history of breast biopsy. As expected, the predicted absolute risks of breast cancer were lower than the breast cancer rates observed in the BCSC in younger women and among Asian women. Also, predicted risks were lower than observed BCSC breast cancer rates for rare changes in density combinations such as extremely dense and fatty breasts and higher than observed for women with fatty breasts on both examinations.

Table 4 presents the predicted 5-year risk by breast density groups and subgroups defined by the presence or absence of other risk factors that can be used to determine how changes in density measures between examinations can affect risk among subgroups. For example, women with scattered fibroglandular densities on the most recent examination and heterogeneously dense breasts on the previous examination with a family history of breast cancer would have a 5-year risk of 1.62% if the most recent density measure were used to calculate risk, and 1.83% with the two-measure model. Of 18.6% (134,404 of 722,654) of women who decreased density categories, 15.4% (20,741 of 134,404) had a clinically meaningful increase in 5-year risk; women whose density decreased from heterogeneously or extremely dense to a lower category with a family history of breast cancer or history of breast biopsy had an increase in 5-year risk from <1.67% with the one-density model to ≥1.67% with the two-density model (bold numbers in Table 4). In contrast, a small percentage of women had a clinically meaningful change in risk if breast density increased; of the 17.9% (129,690 of 722,654) of women who increased density categories, 0.6% (751 of 129,690) decreased 5-year risk from ≥1.67% with the one-density model to <1.67% with the two-density model if the prior density was fatty and women had a family history of breast cancer or history of breast biopsy (italicized numbers in Table 4). Women with fatty breast density on their most recent mammogram and no other breast cancer risk factors had 5-year risks <1.67% regardless of previous breast density measures.

Among women who had a change in density categories between examinations, we calculated the proportion of women reclassified correctly (cancer cases to a higher risk category, noncases to a lower risk category) and the proportion of women reclassified incorrectly. Using a cutoff point of ≥1.67% to define high risk, there was a slight increase in the proportion of women diagnosed with breast cancer who had risk ≥1.67% (42.3% one-measure and 44.5% two-measures), a similar proportion not diagnosed with breast cancer who had risk ≥1.67% (25.6% for one-measure and 26.2% for two measures) and a less than 1% decrease in women not diagnosed with breast cancer who had risk <1.67% (74.5% one-measure and 73.8% two-measures). The positive predictive value increased slightly from 2.19% to 2.24%. The two-measure density model reclassified 6.6% of women to a risk ≥1.67% and 5.9% to a lower risk of <1.67% compared with the one-measure model.

The distributions of 5- and 10-year risk with the two-measure and one-measure density models are shown in Fig. 2A–D. Forty percent of women had a 5-year risk that was less than 1% with the two-measure model. Only 3% of women had a 5-year risk that was 3% or greater. The 10-year risk distribution similarly had concentrated numbers of low-risk women.

The BCSC 5-year risk model has improved discriminatory accuracy compared with models without a measure of breast density and has been validated in a large mammography cohort (2, 27). We evaluated whether two measures of BI-RADS breast density would further improve the discriminatory accuracy of the BCSC 5-year risk model and found that for most women, two measures would not change breast cancer risk sufficiently to affect clinical decisions. However, for the modest fraction of women (18.6%) who experience a decrease in breast density measures between two mammograms on average 1.8 years apart, the use of both measures offers some improvement in risk classification. In some cases, this improvement could affect clinical discussions about supplemental screening for women with dense breasts and discussions of chemoprevention for women with several risk factors.

High interval cancer rates have been reported for women with 5-year BCSC risk of ≥1.67% and extremely dense breasts or 5-year risk ≥2.50% and heterogeneously dense breasts (28) with recommendations that discussions of supplemental imaging be directed to these women who have the potential to benefit from alternative screening modalities. Women with extremely dense breasts on the most recent examination and almost entirely fatty on the previous examination with a family history of breast cancer or history of breast biopsy would have a 5-year breast cancer risk of 1.93% to 2.02% if the most recent measure were used to calculated risk (i.e., ≥1.67%) and 1.16% to 1.65% with the two-measure model (i.e., <1.67%). Similarly, women with heterogeneously dense breasts on the most recent examination and almost entirely fatty on the previous examination with a family history of breast cancer and history of breast biopsy would have 5-year breast cancer risk of 3.01% if the most recent measure were used to calculated risk (i.e., ≥2.50%) and 2.35% with a two-measure model (i.e., <2.50%). Thus, if the two-measure model were used, discussions of alternative screening strategies for women with dense breasts would be avoided for some women.

Accurately identifying women at high risk of breast cancer is important so that primary care providers can discuss primary prevention interventions (29, 30). The American Society of Clinical Oncology and National Comprehensive Cancer Network define elevated 5-year risk as ≥1.67 when considering primary prevention with selective estrogen receptor modulators (SERM), while the US Preventive Services Task Force (USPSTF) defines elevated 5-year risk as ≥3% (29, 30). Women with scattered fibroglandular densities on the most recent examination and heterogeneously dense breasts on the previous examination with a family history of breast cancer or history of breast biopsy would have a 5-year risk of 1.57% to 1.62% if the most recent density measure were used to calculate risk (i.e., <1.67%), and 1.79% to 1.83% with the two-measure model (i.e., ≥1.67%). Similarly, women with a family history of breast cancer, history of breast biopsy, and scattered fibroglandular densities on the most recent examination and extremely dense breasts on the previous examination would have a 5-year risk of 2.44% if the most recent density measure were used to calculate risk (i.e., <3%), and 3.06% with the two-measure model (i.e., ≥3%). Thus, the two-density model can be particularly informative in estimating risk among women with several risk factors to inform patient-provider discussions of prevention interventions.

There is mounting evidence that women with sustained levels of high breast density are at highest risk of breast cancer, those with low levels of breast density over time are at lowest risk and those that change density over time can increase or decrease their risk (9, 31). The IBIS-I has reported for women on tamoxifen who had a reduction in breast density of 10% or more, the risk of breast cancer was significantly reduced 52% relative to controls (10). A study reported improved survival among postmenopausal women with breast cancer with a decrease in breast density on adjuvant tamoxifen (32). Another recent study reported a decrease in risk of contralateral breast cancer among women who have a decrease in breast density within the first 2 years after breast cancer diagnosis (33). Our results are consistent with these studies. We found women who decrease from a higher density to fatty breasts or decrease from heterogeneously dense to scattered fibroglandular density instead of density remaining unchanged, have a reduction in breast cancer risk. Automated density measures incorporated into the clinical practice setting will be critical to providing reproducible density estimates for evaluating changes in density and risk over time.

Women may be reluctant to engage in risk reduction therapies unless they are confident that their breast cancer risk is sufficiently high in the near and/or long term (34). Commonly used breast cancer risk prediction models estimate 5-year risk of breast cancer (2, 16). The Tyrer–Cuzick breast cancer risk model estimates 10-year risk using a complex model that collects 12 personal risk factors and extensive family history information (35, 36), but has not been validated in an average risk population. Our one- and two-measure density models are simple and easy to use and provide moderately accurate estimates of 5- and 10-year risk that can be used in breast cancer prevention discussions to assist women in making clinical decisions.

A major strength of our study is the large number of women with two breast density measures and the large number of breast cancers. Despite this, some groups that changed density category were very small and the numbers of cancers in these groups were few, limiting our ability to estimate risk accurately in these groups. Misclassification and/or changes in breast density reporting could affect women changing density categories. However, studies have found no change in historical distributions of breast density categories from 1996 to 2009 (8, 37). Finally, for risk prediction modeling, it does not matter if an observed change in density is due to subjective differences in measurements or changes in a woman's breast composition.

In summary, for most women, a risk prediction model that includes two BI-RADS density measures provides minimal improvement over one recent measurement for predicting 5- and 10-year breast cancer risk. However, for those women whose prior BI-RADS density measure varies from the most recent measure, health care providers should take into account both recent and previous breast density measurements when calculating breast cancer risk because estimates could differ and impact clinical decision making.

No potential conflicts of interest were disclosed.

Conception and design: K. Kerlikowske, J.A. Tice, D.L. Miglioretti

Development of methodology: K. Kerlikowske, J.A. Tice, D.L. Miglioretti

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): K. Kerlikowske, B.L. Sprague, D.L. Miglioretti

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): K. Kerlikowske, C.C. Gard, B.L. Sprague, J.A. Tice, D.L. Miglioretti

Writing, review, and/or revision of the manuscript: K. Kerlikowske, C.C. Gard, B.L. Sprague, J.A. Tice, D.L. Miglioretti

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): K. Kerlikowske

Study supervision: K. Kerlikowske, D.L. Miglioretti

The authors thank the participating women, mammography facilities, and radiologists for the data they have provided for this study. A list of the BCSC investigators and procedures for requesting BCSC data for research purposes is provided at: http://breastscreening.cancer.gov/.

This work was supported by the National Cancer Institute–funded Breast Cancer Surveillance Consortium (P01 CA154292 and HHSN261201100031C to K. Kerlikowske, B.L. Sprague, and D.L. Miglioretti) and U54 CA163303 to B.L. Sprague. The collection of cancer data used in this study was supported, in part, by several state public health departments and cancer registries throughout the U.S. For a full description of these sources, please see: http://breastscreening.cancer.gov/work/acknowledgement.html.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.