Prospective Breast Cancer Risk Prediction Model for Women Undergoing Screening Mammography

Abstract

The Gail breast cancer risk model was developed in the late 1980s with information from women undergoing annual mammography screening in the Breast Cancer Detection Demonstration Project ( 1 , 2 ) . The Gail model uses retrospective risk factor information that was collected in a case–control study embedded in the cohort study. Those data were coupled with cancer incidence rates obtained from the National Cancer Institute's (NCI) Surveillance, Epidemiology, and End Results (SEER) program to estimate the probability that a woman who undergoes annual screening will be diagnosed with invasive breast cancer within the next 5 years. The Gail model used risk factors known at the time (i.e., in 1989), including current age, age at menarche, age at birth of first child, number of first-degree relatives with a family history of breast cancer, and number of previous breast biopsy examinations. Race and atypical hyperplasia were added in a revised model ( 1 , 2 ) .

The Gail model has been validated ( 2 – 6 ) , although the predictive accuracy and calibration of the model are only fair ( 5 – 13 ) . Predictive accuracy is the ability to correctly predict the outcome for an individual woman, and calibration measures the predicted versus actual number of breast cancers in a cohort. The predictive accuracy of the model is highest in annually screened non-Hispanic white women and is lowest in women with different demographic characteristics (age, race, ethnicity, and nationality) than the population from which the model was developed ( 8 – 13 ) . Women with high Gail scores have been encouraged to be screened, to undergo genetic or biomarker evaluation, and to participate in intervention trials ( 14 – 20 ) . The Gail model is also used to estimate the number of expected breast cancers in a particular population ( 21 – 23 ) . Consequently, both predictive accuracy and calibration of the model are important.

Risk prediction models for breast cancer may be improved by the addition of recently identified risk factors, including breast density and use of hormone therapy. It may be desirable to include other routinely available risk factors and to have separate models for premenopausal and postmenopausal women. Recently, higher breast density has been shown to be strongly associated with breast cancer risk ( 24 – 31 ) . Other factors associated with breast cancer risk include use of hormone therapy, high body mass index (BMI), and the result of the previous mammographic examination ( 32 – 42 ) . Finally, the effects of race and ethnicity on risk models have not been well validated or explored ( 10 – 11 , 43 ) . For example, Asian women are assigned the risk of white women in the online Gail risk model ( http://www.cancer.gov/bcrisktool/ ), although Asian women have a markedly lower incidence rate of breast cancer than white women ( 44 ) .

To develop an enhanced breast cancer risk model requires a large cohort with measured risk factors and well-ascertained cancer status. The Breast Cancer Surveillance Consortium (BCSC) was established in 1994 by the NCI to assess community mammography practice and outcomes ( 45 ) , collect risk factors prospectively at the time of each screening mammogram, and ascertain cancer outcomes for all women. In this report, we used data from the BCSC to estimate the probability of a diagnosis of incident breast cancer (invasive breast cancer or ductal carcinoma in situ) after a screening mammogram. Although most breast cancers were detected shortly after the screening mammogram, we included breast cancers diagnosed within 1 year, a commonly recommended screening interval ( 46 ) . To estimate probabilities, we developed a model that includes traditional risk factors and adds others including race, ethnicity, breast density, BMI, use of hormone therapy, type of menopause, and previous mammographic result.

P ATIENTSAND M ETHODS

Study Population

Seven mammography registries of the BCSC ( http://breastscreening.cancer.gov ) participated in this study (Carolina Mammography Registry, Colorado Mammography Project, Group Health Cooperative's Breast Cancer Surveillance Project, New Hampshire Mammography Network, New Mexico Mammography Project, San Francisco Mammography Registry, and the Vermont Breast Cancer Surveillance System) ( 45 ) . The mammography registries are community based and located in geographic areas that broadly represent the United States ( 47 ) . Screening mammograms in the calendar period from January 1, 1996, through December 31, 2002 (or December 31, 2001, at one registry), from women aged 35–84 years were included. Women with previous breast cancer were excluded. Women with breast augmentation were also excluded because augmentation decreases breast cancer detection by mammography ( 48 ) . We also restricted our analysis to women who had a known previous mammogram (screening or diagnostic) or who self-reported a mammogram within the last 5 years, to eliminate women with long-term prevalent breast cancers. There were 2 392 998 eligible screening mammograms from women without previously diagnosed breast cancer who had had a prior mammogram in the preceding 5 years. Within 1 year of the screening mammogram, 11 638 women were diagnosed with breast cancer.

Each mammography registry actively links mammography records to cancer outcomes recorded in cancer registry or pathology data. De-identified mammography and outcome data were sent to the BCSC Statistical Coordinating Center for analysis. Procedures for maintaining confidentiality have been described previously ( 49 ) . All study activities were approved by the institutional review boards of the participating institutions.

Screening Mammograms and Risk Factors

Screening examinations had to be designated as bilateral screening by the radiology facility and needed to be done at least 9 months after any preceding breast imaging to ensure an accurate designation as a screening mammogram. Patient information was primarily obtained from self-report at the time of the screening mammogram. Questionnaires may differ across registries or calendar time, but all contain certain core elements (available at http://breastscreening.cancer.gov/elements.html#questionnaires ). Questions include information about birth date, race, ethnicity, education, time since last mammogram, personal or family history of breast cancer, prior breast procedures, age at menarche, age at the birth of first child, menopausal status, age at menopause, type of menopause, and current use of hormone therapy. Prior breast procedure included self-reported breast biopsy, fine needle aspiration, cyst aspiration, breast reconstruction, lumpectomy, or mastectomy, but women with previous breast cancer in either breast were excluded. Prior pathologic findings such as lobular carcinoma in situ or atypical hyperplasia were not available for all women and so were not included as predictors. Age at menarche was often not collected or not reported, but it was tested as a risk factor in women who reported it. Height and weight were only recently added to the questionnaire, and so BMI (weight in kilograms/[height in meters] ² ) was available primarily for recent examinations. Family history in this report refers to the number of first-degree relatives with breast cancer (classified as none, one, or two or more).

When results are presented without regard to menopausal status, they include screening mammograms from all women aged 35–84 years. For the analysis of premenopausal women, we included only women aged 35–54 years who stated that they were still having periods. Analyses of postmenopausal women included all women aged 55 years or older and women aged 45–54 years who reported their periods had stopped permanently or who were on hormone therapy. Women aged 35–54 years of unknown menopausal status were excluded from both pre- and postmenopausal models to avoid misclassification. Postmenopausal women were classified as having surgical menopause if the woman reported that surgery was responsible for the cessation of menstruation. It was not always possible to ascertain whether the woman had oophorectomy and/or hysterectomy.

We included only information known at the time of each mammogram even though additional information became known subsequent to the mammogram. Data were missing primarily because the item was not on the mammography registry's questionnaire and secondarily because the woman chose not to report this information. We included “unknown” as a category in most analyses.

In addition to self-reported data, breast density was also recorded at the time of the mammogram and was typically classified by use of the four categories in the Breast Imaging Reporting and Data System (BI-RADS) coding system: 1) almost entirely fat, 2) scattered fibroglandular densities, 3) heterogeneously dense, and 4) extremely dense ( 50 ) . This method is the predominant method for reporting breast density among radiologists, and efforts are under way to standardize its reporting ( 50 ) because use of the scale can vary among radiologists ( 51 , 52 ). If breast density was recorded using a different system or was not recorded at all, then it was classified as “unknown.”

All women had a prior mammogram in the last 5 years. If the previous mammogram was in our data, then the final interpretation of that previous mammogram was categorized as false positive (BI-RADS category of 0, 4, 5, or 3 with immediate follow-up) or true negative (BI-RADS category of 1, 2, or 3 without immediate follow-up) ( 50 ) . Women with previous true-positive or false-negative mammograms would have been excluded because of prior breast cancer.

Follow-up for Breast Cancer

Women were classified as having breast cancer if there was a breast cancer diagnosis within 1 year of the screening mammogram. We used 1 year because women were encouraged to be screened every 1–2 years, and the observation period was truncated by a new screening examination. Breast cancer cases were identified through linkage of each mammography registry to a cancer registry or pathology data. Both invasive carcinoma and ductal carcinoma in situ were included as breast cancer, although some analyses included invasive cancer only. Lobular carcinoma in situ was not included as a breast carcinoma, and women with this diagnosis continued to be screened.

Statistical Analysis

We evaluated the probability of a cancer diagnosis within 1 year of each screening mammogram by use of logistic regression analysis in SAS, Version 9.0 ( 53 ) , and Stata, Version 8 ( 54 ) . The primary goal was to find a model that best predicted a diagnosis of breast cancer separately in premenopausal and postmenopausal women with a minimal number of predictors. Because the number of observations was very large, small effects can be statistically significant at usual nominal levels such as .05. Consequently, we included only covariates that were statistically significant at a very stringent level of P less than .0001. All P values were from two-sided tests, when possible.

We did not include interaction terms to keep the model parsimonious, but a few interactions were tested as noted below. Because of the potential for overfitting, we used a 75% random sample (training sample) to generate and test different statistical models before determining a final model. We then tested the prediction in the remaining 25% (validation sample) with estimates derived from the model of the training data. We report the concordance value (i.e., the c statistic) for the validation sample by use of the estimates derived from the training sample. The c statistic ranges from 0.50 to 1.00, with a higher score indicating better prediction for an individual woman ( 55 ) . Approximate 95% confidence intervals (CIs) for the c statistic (or for the area under the curve) are derived from an empirical receiver-operator curve analysis, by treating the estimated score from the logistic regression model as a continuous marker of cancer risk ( 54 ) . The c statistic may also be stratified by age group to determine the remaining effect of the other predictors ( 56 ) . We also tested the goodness-of-fit of the data to the estimated model by use of the Hosmer–Lemeshow test ( 55 ) . After the training sample was used to guide model development and the validation sample was used to test model prediction, we consolidated the data to estimate the coefficients for the final prediction model. Thus, reported odds ratios, 95% confidence intervals, and P values were from the full dataset, although the training sample was used to choose predictors and the validation sample was used to assess prediction.

In these analyses, the screening mammogram is the unit of analysis because many of the associated characteristics (e.g., age, breast density, menopausal status, or prior mammogram result) can change over time. We considered several possible regression approaches to address the complexity introduced by a woman having several screening mammograms during the study period. Many women (38%) had only a single screening mammogram, 52% had two to four screening mammograms, and 10% had five or more screening mammograms. However, this cohort was an open cohort, so that not all women were under observation for the same length of time. Choosing only one mammogram per woman would underestimate the contribution of women who are regularly screened compared with women who may have a single screening mammogram over the study period. To parallel the data used in the Gail model, it was important to include all screening examinations from women undergoing annual or biannual screening. A diagnosis of breast cancer would make her ineligible for further screening examinations. Therefore, the only possible sequence for a woman would be a series of screening mammograms with a negative outcome (no cancer) followed by a last screening mammogram that could be either negative or positive (cancer diagnosis). Thus, all cancers were found on the last screening examination that a woman had. If only the last mammogram was chosen for each woman, then the cancer risk would be biased upward because all preceding negative examinations would be ignored.

We used several alternative models to address the possible correlation of mammograms within a woman, but the results were almost identical to the one used here, a standard logistic regression model for each mammogram. These alternative models include 1) a model of the sequence of screening examinations that uses a negative binomial model of time to first failure (cancer diagnosis), 2) a Poisson regression model, and 3) logistic regression that uses a generalized estimating equations approach for all mammograms from the same woman ( 57 ) . These alternative models gave very similar results because of the low outcome rate and the low correlation between examinations after adjustment for risk factors ( 58 ) . We used a standard logistic model because it allows estimation of the concordance between prediction and outcome (by the c statistic) and because the odds ratio is a good approximation to the risk ratio ( 59 ) .

We also compared our absolute incidence rates with those reported by the SEER program over the same calendar period ( 60 ) . The incidence rates and the percentage of all breast cancers that were invasive were computed by use of the SEER*STAT program ( 61 ) . Confidence intervals for rates were computed by use of the normal approximation of the binomial distribution. We reported incidence rates per 1000 examinations (or per 1000 women in SEER). We also show the percentage of women reporting a screening mammogram in the last 2 years according to data from the National Health Interview Survey (N. Breen for data from the 2003 National Health Interview Survey: personal communication).

Data Availability

Other investigators may wish to explore modification of risk factors by demographic factors or statistical issues, such as the effect of data imputation for missing values, or alternative estimation models. To aid in this exploration, these data will be available to others by following the link http://breastscreening.cancer.gov/rfdataset/ . The particular dataset used in this study was a large cross-classification of risk factors by cancer outcome.

R ESULTS

There were 2 884 197 screening mammograms, but we included only the 2 392 998 (83.0%) index screening mammograms from 1 007 600 women who had had a previous mammogram in the prior 5 years. Breast cancer was diagnosed within 1 year of a screening mammogram in 11 638 women, for an absolute rate of 4.86 breast cancers per 1000 screening mammograms (95% CI = 4.78 to 4.95). Most (75.7%) of these 11 638 breast cancers were diagnosed within 3 months of the screening mammogram.

Risk factors vary by menopausal status, so that separate models were fit for premenopausal and postmenopausal women. This procedure required excluding 7.6% of the mammograms from women aged 45–54 years with unknown menopausal status. The remaining mammograms were classified as premenopausal (n = 568 215; 25.7%) or postmenopausal (n = 1 642 824; 74.3%). There were 1726 breast cancers among premenopausal women, for an absolute rate of 3.04 per 1000 screening mammograms (95% CI = 2.89 to 3.18), and 9300 breast cancers among postmenopausal women, for an absolute rate of 5.66 per 1000 screening mammograms (95% CI = 5.55 to 5.78).

Distribution of Risk Factors

We next determined the distribution of demographic factors and cancer rates per 1000 screening mammograms by menopausal status ( Table 1 ). The cancer rates suggest that risk patterns are dissimilar for premenopausal and postmenopausal women. For example, premenopausal women had decreasing breast cancer rates with increasing BMI, but postmenopausal women had increasing cancer rates with increasing BMI ( Table 2 ).

To better describe the relationship between cancer risk and the joint effects of age and breast density, we computed the observed cancer risk without regard to menopausal status. In this analysis, the observed breast cancer rates (per 1000 screening mammograms) increased with both breast density and age ( Fig. 1 ).

Fig. 1.

Open in new tabDownload slide

A ) Observed breast cancer rates per 1000 screening mammograms by age and breast density. Density was classified by use of the Breast Imaging Reporting and Data System (BI-RADS) coding system: 1) almost entirely fat ( circles ), 2) scattered fibroglandular densities ( diamonds ), 3) heterogeneously dense ( squares ), and 4) extremely dense ( triangles ). Both invasive cancers and ductal carcinomas in situ are included. B and C ) Breast Cancer Surveillance Consortium (BCSC) estimates are based on a logistic regression model of 7577 invasive breast cancers diagnosed after 1 642 824 screening mammograms in postmenopausal women aged 45–84 years. B ) Estimated 5-year invasive breast cancer risk (percent of women) for high- and low-risk postmenopausal women with the BCSC model and the Gail model by age. The mean risk is the mean of the estimated risk in that age group. The high-risk estimate is based on a non-Hispanic white woman with a prior breast procedure, with a family history of breast cancer (one relative), with a body mass index (BMI) of 30–34 kg/m ² , with current use of hormone therapy, and who had natural menopause. She had a breast density of 3 on the BI-RADS scale, had her first child before the age of 30 years, and had a negative last mammogram, so that she was not at high risk on all factors. The low-risk woman had a breast density of 2 on the BI-RADS scale, a negative family history, a surgical menopause, and no current use of hormone therapy, but her other characteristics were otherwise identical to the high-risk woman. Ages were varied from 45 to 80 years, and the estimated 5-year risk was plotted. The Gail estimates assume that age at menarche was 12–13 years, that the family history of breast cancer was positive indicating one affected first-degree relative, and that a prior breast procedure was equivalent to a single benign breast biopsy examination. C ) Estimated 5-year invasive breast cancer risk for postmenopausal women from the BCSC model by breast density. The age-specific estimate was based on a non-Hispanic white woman with no prior breast procedure, a BMI of 30–34 kg/m ² , no current hormone therapy, natural menopause, birth of her first child before age 30 years, and the last mammogram being negative. The risk for that woman is shown if she has none or one first-degree relative with breast cancer. Density ranged over all four BI-RADS categories.

Model for Premenopausal Women

Factors that were not statistically significantly associated with breast cancer risk for premenopausal women when the training data were analyzed included race ( P = .40), Hispanic ethnicity ( P = .41), age at birth of first child ( P = .48), age at menarche ( P = .56), BMI ( P = .89), and education ( P = .13). The P value for the result of the previous mammogram ( P = .0018) did not meet our strict criterion. Four factors—age, breast density, family history of breast cancer, and a prior breast procedure—were all statistically significantly associated with breast cancer risk ( P <.0001). Application of the model to the validation data showed the model was well calibrated, with an observed cancer rate of 3.042 breast cancers per 1000 mammograms compared with an estimated rate of 3.037 breast cancers per 1000 mammograms ( Table 4 ). In addition, the c statistic yielded a value of 0.629 (95% CI = 0.603 to 0.656), and the Hosmer–Lemeshow test showed no lack of fit ( P = .91).

The model was then refit to the combined data (training and validation data together) ( Table 3 ). All factors that were statistically significant in the training set remained statistically significantly associated with breast cancer risk (at P <.0001). Risk increased with age (trend test, P <.0001), breast density (trend test, P <.0001), having a positive family history of breast cancer, and having had a prior breast procedure. The fitted model had a c statistic of 0.631 (95% CI = 0.618 to 0.644) ( Table 4 ). If stratified by age group, the c statistic was 0.600 (95% CI = 0.587 to 0.613) for the three remaining risk factors. If breast density was completely excluded from the model, the c statistic decreased to 0.607 (95% CI = 0.592 to 0.621). The predictors age and breast density individually had c statistics of 0.574 (95% CI = 0.562 to 0.585) and 0.565 (95% CI = 0.553 to 0.578), respectively, indicating almost equal ability to predict breast cancer. Finally, prediction of invasive cancer alone by the four risk factors had a c statistic of 0.633 (95% CI = 0.618 to 0.648), and thus, predictive ability differed little from the model for both invasive and ductal carcinoma in situ disease combined.

Model for Postmenopausal Women

Many risk factors jointly predicted the likelihood of a breast cancer diagnosis for postmenopausal women ( Table 5 ). Age at menarche was not associated with breast cancer risk ( P = .24), and the P value for educational level ( P = .0002) did not meet the pre-established statistical significance level. All factors shown in Table 5 were statistically significantly associated with breast cancer risk ( P <.0001). Application to the validation sample showed slight overestimation of the cancer rate, but the c statistic of 0.626 was similar to that in the training data (i.e., c statistic = 0.623), and the lack of fit test was not statistically significant ( P = .23) ( Table 6 ).

The model was then refit to the combined data ( Table 5 ), and risk was found to increase with age, BMI, and breast density (trend test P <.0001, for all variables). The risk among Asian/Pacific Islander women and Native-American/Alaskan Native women was lower than that among white women, but the risks among African American women and women of other or mixed race and of unknown race were similar to those among white women. Being of Hispanic ethnicity conferred lower risk compared with being of non-Hispanic ethnicity. Other factors associated with higher risk included later age at birth of first child (≥30 years) or being nulliparous, prior breast surgery, family history of breast cancer, natural menopause, current use of hormone therapy, or a false-positive result on the prior mammogram. The c statistic for the overall model was 0.624 (95% CI = 0.619 to 0.630), indicating moderate prediction ( Table 6 ). If stratified by age group, the c statistic was 0.599 (95% CI = 0.593 to 0.604), indicating that age contributes to the prediction but that the other factors also contribute substantially. When breast density was excluded, the c statistic decreased to 0.605 (95% CI = 0.600 to 0.611). Conversely, the c statistics for age and breast density alone were 0.571 (95% CI = 0.565 to 0.577) and 0.552 (95% CI = 0.547 to 0.558), respectively, so that both were highly predictive. If only invasive cancer was modeled, the same risk factors were identified, and the c statistic of 0.628 (95% CI = 0.622 to 0.634) was similar to that for all breast cancer.

Inspection of the highest and lowest deciles of risk showed that no risk factor automatically conferred high or low risk. The observed numbers of breast cancer cases for the two extreme deciles were 341 and 1766, which were close to the predicted numbers of 345 and 1787.

Some interaction terms were tested to determine whether there was sufficient improvement to justify more complicated models. Race did not interact with breast density ( P = .50) or BMI ( P = .95). There was a statistically significant interaction between breast density and age ( P <.0001), but this interaction increased the c statistic by only 0.002 at the expense of 28 additional covariates. There was also a strong interaction of BMI and current use of hormone therapy ( P <.0001). BMI was a strong predictor of breast cancer among women not currently taking hormone therapy ( P <.0001) but not among current users of hormone therapy ( P = .067). However, adding this interaction to the model only improved the predictive power (i.e., c statistic) from 0.624 to 0.626. In some additional analyses, combining pre- and postmenopausal data, BMI had a very strong interaction with menopausal status ( P <.0001), showing a possible protective effect of high BMI in premenopausal women and a deleterious effect in postmenopausal women.

To compare our BCSC model with the Gail model, we refit the BCSC model to predict invasive cancer only. We could not directly compare our model with the Gail model because of the difference in the time intervals for cancer ascertainment. However, we obtained crude estimates of a 5-year risk by using the formula Prob(invasive cancer in 5 years) ≈ 1 − [1 − Prob(invasive cancer in 1 year)] ⁵ (where Prob is probability) and averaged the computed 5-year probability estimates within each age group for all postmenopausal women ( Fig. 2 ). We also computed the risk for a high-risk woman who we defined as a non-Hispanic white woman with a prior breast procedure, one first-degree relative with breast cancer, and BMI of 30–34 kg/m ² and who is a current user of hormone therapy who had natural menopause. We assumed that this woman had a breast density of 3 on the BI-RADS scale, had her first child before the age of 30 years, and had a negative last mammogram, so that she was not at high risk on all factors. We also computed the risk for a low-risk woman by changing breast density to 2 on the BI-RADS scale, the family history to negative, the menopause type to surgical, and the hormone therapy use to negative. This woman was identical to the high-risk woman on all other characteristics. Age was varied from 45 to 80 years, and the estimated 5-year risk was plotted ( Fig. 1 ). Finally, we computed Gail estimates of the 5-year risk by assuming that age at menarche was 12–13 years and that a prior breast procedure was equivalent to a single benign breast biopsy examination. For the low-risk patient, the Gail model and the BCSC model provided very similar 5-year estimates ( Fig. 2 ). However, for the high-risk patient, the BCSC risk was much higher. This result may be due to risk factors not included in the Gail model, such as use of hormone therapy, BMI, and breast density. For invasive cancer only, inclusion of risk factors not in the Gail model increased the c statistic to 0.628 from 0.598.

Fig. 2.

Open in new tabDownload slide

Upper ) Comparison of breast cancer incidence per 1000 women from the Surveillance, Epidemiology, and End Results (SEER) program database with breast cancer rates per 1000 screening examinations by age. Using SEER public-use files, we computed breast cancer (invasive cancer and ductal carcinoma in situ) incidence from January 1, 1996, through December 31, 2002, by age and compared it with the observed incidence in the Breast Cancer Surveillance Consortium (BCSC) cohort of women undergoing mammography screening. For this analysis, we combined premenopausal and postmenopausal women. Lower ) Percentage of women reporting in the National Health Interview Survey that they underwent screening mammography in the last 2 years.

We compared SEER cancer incidence rates, which are based on both screened and unscreened women, with the BCSC rates, which are based on women undergoing screening. Use of SEER data allowed computation of cancer incidence (invasive and ductal carcinoma in situ) by age over the calendar period from January 1, 1996, through December 31, 2002, but did not allow separation by menopausal status. Consequently, we had to compare the BCSC and SEER rates by age without regard to menopausal status ( Fig. 1 ). For women younger than 75 years, the BCSC and SEER rates were roughly parallel, with the BCSC rate indicating that approximately one more breast cancer case per 1000 examinations was identified with BCSC data compared with SEER data. The percentage of breast cancer that was invasive was similar between BCSC and SEER women aged 50 years or older (data not shown). Figure 1 also shows the percentage of women in the National Health Interview Survey reporting a screening mammogram in the last 2 years by age (N. Breen: personal communication).

D ISCUSSION

We used prospectively collected risk-factor information to predict a breast cancer diagnosis after a screening mammogram. The models establish breast density as a highly clinically significant predictor of breast cancer risk that is almost as powerful a risk factor as age. The models also confirm most previously established risk factors as being additional independent predictors of breast cancer, particularly for postmenopausal women. Nonetheless, ability to accurately predict breast cancer at the individual level remains limited.

For premenopausal women, only four risk factors were statistically significantly associated with the risk of breast cancer: age, breast density, number of first-degree relatives with breast cancer, and a prior breast procedure. Having had a prior breast procedure was associated with an approximately 50% increase in risk even without knowledge of the type or result of the prior breast procedure. Breast density was strongly associated with the risk for breast cancer among women with extremely dense breasts, with the risk almost four times greater than that for a woman with breasts that are almost entirely fat, after adjustment for age. Vacek and Geller ( 62 ) found a slightly higher, but similar, risk of 4.6 (95% CI = 1.7 to 12.6). The effect of breast density may be underestimated in general because mammography has poorer sensitivity for dense breasts than for fatty breasts ( 63 ) . Despite the large sample size, many potential risk factors were not associated with a diagnosis of breast cancer in premenopausal women, consistent with other reports ( 64 ) . These risk factors include race, Hispanic ethnicity, education, age at menarche, age at the birth of the first child, and BMI.

Risk in postmenopausal women increased dramatically with age, as expected. Unlike what is observed in SEER program data, we observed no decrease in incidence for the most elderly women. Because screening decreases with older age, our screening cohort may become less representative of the general population as it ages ( 65 , 66 ) . Hispanic ethnicity was associated with a 30% reduction in risk, and being American Native or Pacific Islander was associated with a 44% reduction in risk, although it is always possible that this result can be partially attributed to lower rates of diagnostic follow-up that would result in underdiagnosis subsequent to a positive mammogram. Screened Asian women appeared to have lower incidence of breast cancer compared with white women, even though breast density is greater for Asian women ( 67 ) . Other factors associated with increased risk included higher BMI, late age at the birth of the first child or being nulliparous, prior breast procedure, family history of breast cancer, natural menopause, and use of hormone therapy. Breast density was strongly associated with breast cancer risk, with women with the highest breast density having a risk of 3.16 (95% CI = 2.72 to 3.66) compared with those with the lowest density even after adjustment for several factors including age and BMI. This value is lower than the hazard ratio of 3.9 (95% CI = 2.6 to 5.8) found by Vacek and Geller ( 62 ) . We confirmed that having had a false-positive mammogram at the last mammographic examination was associated with an increased risk for detecting breast cancer within a year of the current index mammogram. McCann et al. ( 35 ) found an odds ratio of 2.15 (95% CI = 1.55 to 2.98) for detecting cancer at the second round of screening in women who had a false-positive result on the first round, compared with women who were negative on the first screening round.

Recent studies show that increased breast cancer risk was associated with estrogen plus progestin hormone therapy but not with estrogen-only therapy ( 68 – 72 ) . We could model only the effect of current hormone therapy because earlier questionnaires did not distinguish among types of hormone therapy. Therefore, the association between estrogen plus progestin and breast cancer risk may have been underestimated in this study, and the association between taking unopposed estrogen therapy and breast cancer risk may be overestimated. We were unable to examine the association between parity and breast cancer risk because information on parity was unavailable ( 73 ) . Age at menopause has previously been demonstrated to be statistically significantly associated with the risk of breast cancer ( 74 – 77 ) , although it was not included in this analysis.

With the exception of age at menarche, the BCSC models support an association between some known factors and the risk for breast cancer and show that breast density is a salient addition to models to predict the risk of breast cancer. As a single predictor, breast density is almost as powerful as age. However, the c statistics show that breast density has only moderate predictive ability when added to other factors ( 6 , 19 ) . Tice et al. ( 31 ) found that the Gail model risk factors had a c statistic of 0.67, which was increased to 0.68 when breast density was added to the model. It is difficult to compare predictive ability across different models and populations. The moderate predictive power also suggests that major determinants of breast cancer risk remain to be discovered.

There are important similarities and differences between the BCSC and Gail populations ( 1 ). The Gail model was developed by use of data from Caucasian women who participated in a demonstration project of annual mammography screening population recruited from January 1, 1973, through December 31, 1975 ( 78 ) . The BCSC used modern screening mammography techniques, and the population was racially and ethnically diverse. Unlike the Gail model, our model used prospectively measured risk factors and internal estimation of the cancer rates, rather than SEER program rates. The BCSC incidence rates are slightly higher than the SEER rates. This fact is not surprising because women undergoing screening mammography are at higher risk for breast cancer ( 79 ) and because the SEER program combines screened and unscreened women. The similarity of the BCSC and SEER rates does provide further validation that the BCSC population is representative of the population of US women.

The high percentage of missing values in this report may be cause for concern. However, in most cases, the missing values are attributable to the question not being asked in the questionnaire or omitted by a particular mammography facility, rather than women choosing not to answer the question. Inclusion of a “Missing” category in analyses has the potential to bias the odds ratios for the observed nonmissing levels of that predictor ( 80 ) . For each risk factor, we eliminated women with missing information for that risk factor only and refit the model. The odds ratios did not change by more than 0.01 for any predictor, except breast density for which the odds ratios increased if women with unknown breast density were excluded. However, the ultimate goal is to predict breast cancer from a collection of risk factors, some of which may be unknown for some women. Therefore, it seems most appropriate to include “Missing” as a legitimate level in a prediction model.

There are some other limitations of the use of BCSC data for assessing risk. Typically, incidence was assessed on a cohort known to be cancer free at the beginning of the study, time t₀ , and the incidence rate was estimated on new cases accrued since t₀ divided by cumulative person-time. In this study, we used the index mammogram time as t₀ , so the outcome was highly associated with the starting point because most cancers are diagnosed within 3 months of the examination. We did include all cancers diagnosed within 1 year, including so-called “interval cancers,” for the following reasons. First, it is likely that the interval cancers were present at the time of the index mammogram but may not have been detectable mammographically. By including all cancers detected within 1 year of a screening mammogram, we avoid consideration of the sensitivity of mammography that is affected by many of the included risk factors. Second, it has been suggested that risk models can probably include cases obtained in the initial screen ( 81 ) . Our model is really a short-term prediction model for diagnosing a recently developed breast cancer within a year of a scheduled screening examination. Despite that, the risk factors are similar to those of the Gail model, which were based on a longer follow-up. As more longitudinal follow-up time accrues in the BCSC trial, direct modeling of the 5-year risk will become possible.

A recent conference sponsored by the NCI called for better risk prediction models that can be validated ( 82 ) . Such models can identify women at high risk who may benefit from more intensive screening with magnetic resonance imaging or with ultrasound or may benefit from prophylactic interventions, such as use of antiestrogen agents. In the absence of strong predictive accuracy, however, some argue that the models may be misleading, frightening, or even unethical ( 83 ) . At the population level, the models may identify some factors that are modifiable, such as BMI, use of hormone therapy, or potentially breast density ( 84 , 85 ) .

Clinically, it might be useful to offer a revised prediction of the likelihood of breast cancer after the result of the mammogram is known. For women with a positive mammogram, the use of covariates would refine assessment of the positive predictive value, the probability of breast cancer given a positive mammogram ( 86 ) . Similarly, women with a negative mammogram may still be at increased risk given certain risk factors. Analysis of positive and negative predictive values is beyond the scope of this article, but the combined use of mammographic findings and risk factor data has the potential to address these clinical questions ( 87 ) .

The risk factors in these models can be collected at the time of mammography screening. Dr. Steven Cummings (personal communication) of the San Francisco Coordinating Center (and California Pacific Medical Center Research Institute) has suggested integrating risk assessment, including breast density, into routine mammography. This step would expand the role of mammography facilities into centers for breast cancer prevention (personal communication). This practice could include evaluation of biomarkers that may aid in a final diagnosis of breast cancer ( 88 , 89 ) . Computation of the risk of breast cancer could assist recruitment to prevention interventions for women estimated to be at high risk. Finally, it would allow the mammography facility to tailor current and future screening to the actual risk of the patient.

References