Percent mammographic density (PMD) is a strong marker of breast cancer risk. It may be a correlate of the rate of breast tissue aging, as proposed by Pike to explain breast cancer age-incidence. We examined longitudinal changes in PMD in 645 breast screening attendees in London, UK, in which each had between 2 and 5 screens spanning 3–12 years at ages 50–65 years and compare these to Pike's model. Within-woman PMD declined during these ages, with a slowing rate of decline. Annual rates of decline were 1.4% (95% confidence interval: 1.2–1.6), 0.7% (0.6–0.9) and 0.1% (–0.2 to 0.4) at ages 50, 57 and 64. Dense area declined similarly, but the absolute magnitude of the rate of increase of nondense area was almost double that of dense area. PMD dropped by 2.4% (1.4–3.4) on menopausal transition and increased by 2.4% (1.4–3.5) with the use of hormone therapy. Higher body mass index, greater parity and being Afro-Caribbean or South Asian ethnicities were associated with lower PMD, but did not affect rate of change of PMD at these ages. Within-woman rank correlation of PMD was 0.80 for readings taken 9 years apart. Effects of menopause and parity and the lack of effect of menarche on age-specific PMD at these ages are consistent with the predicted determinants in Pike's model. A high degree of tracking of PMD indicates that at ages 50–65 years high-risk women could be identified by a single early screen at age older than 50.
Percent mammographic density (PMD), a measure of the proportion of radiodense fibroglandular tissue in the breast, is one of the strongest quantitative markers of breast cancer risk.1 Many breast cancer risk factors, such as parity, age at first birth and use of hormone therapy (HT), are determinants of PMD because of its possible intermediary causal role. However, PMD and breast cancer risk demonstrate opposite trends with age; risk increases whereas PMD decreases as women age, the latter observation having been made in both cross-sectional and longitudinal studies.2–4 To explain this apparent anomaly, Boyd et al. noted that the decline of PMD with age parallels that of the rate of breast tissue aging in a model proposed by Pike to explain the age-incidence curve of breast cancer.5, 6 In this model, the age-specific rate of breast tissue aging is high between menarche and age at first birth, drops slightly after pregnancies and considerably after the menopause. Breast cancer incidence rates are related to breast tissue age (i.e., the cumulative rate of breast tissue aging). However, there are few longitudinal studies of PMD to investigate the PMD—breast tissue age analogy.
A greater understanding of the nature and determinants of the profile of PMD and of changes in PMD within a woman over her life-course would help to identify periods of change, the processes that might account for them and factors that influence them. Reductions in PMD have been shown to be associated with reduced breast cancer, although not consistently.4, 7–9 If PMD is causally related to breast cancer, factors that reduce PMD within women and also lead to a lowering of breast cancer risk might provide new opportunities for primary prevention strategies. Risk reduction through this pathway (if it exists) also provides unique possibilities for monitoring the effectiveness of interventions. To date, only a few longitudinal studies have investigated a range of risk factors in relation to quantitatively measured within-woman changes in PMD,2, 3, 10 but none explored whether PMD tracks (maintenance of relative ranks) over time. Long-term tracking would enable earlier identification of women at high risk and a greater window within which risk-lowering strategies, if considered appropriate, could be implemented.
Herein, in a retrospective longitudinal study of PMD at ages 50–65 years, we investigate (i) the average profile of breast density during this age range, (ii) determinants of mean levels of density and of changes in density, (iii) the degree to which PMD tracks over a period of up to 9 years and (iv) whether breast density profiles concur with those suggested by Pike's model for the rate of breast tissue aging at these ages. This is the first longitudinal study of PMD in the UK National Breast Screening Programme, a free population-based National Screening Programme, providing the opportunity to investigate within-woman changes in PMD, in this case for 3 ethnic groups with different breast cancer risks. Uniquely, we perform a detailed comparison of the longitudinal profile of PMD with Pike's model of breast tissue aging.
Material and Methods
As previously described in detail, we conducted a retrospective study of PMD amongst Caucasian, Afro-Caribbean and South Asian women who were regular attendees at the Central and East London Breast Screening Service, United Kingdom.11 At the time of this study, this National Health Service Breast Screening Programme invited all women registered with a general practitioner in central and east London to attend mammographic breast screening every 3 years from ages 50–65 years. Women who attended their second or later breast screen in 2004 were identified, and an ethnically stratified random sample was invited to participate in 2005–2006 (conducted separately to screening so as not to interfere with screening uptake). Ethnicity information was extracted from a form self-completed by the woman at the time of screening.
Participation involved consenting to mammogram digitization and self-completing a lifestyle questionnaire with questions on menarche (years), reproduction (parity, age at each birth), breast feeding (duration for each child), height and weight from which body mass index (BMI in kg/m2) was calculated, weight changes in the past 3 years from which BMI 3 years previously was calculated, ethnicity and oral contraceptive (OC) use during the woman's 20s, 30s and 40s, family history of breast cancer and personal history of benign breast disease. Questionnaires were mailed to women at their homes and self-completed, but for some South Asian women an interviewer-administered telephone interview was conducted to help increase response rates which were initially extremely low. Final response rates were 58.5, 40.5 and 28.1% in Caucasian, Afro-Caribbean and South Asians. The study was approved by the East London and The City Local Research Ethics Committee.
By design all women had been screened at least twice, in 2001 and 2004, and some older women had additionally been screened in ∼1991, 1994 and 1997. At the time of screening, women completed a form that included information on whether they were currently taking hormonal therapy (HT), i.e., HT data are screen-specific. Mediolateral oblique (MLO) views were taken at every screen and additionally craniocaudal (CC) views at the prevalent screening round (at age 50–52) and in 2004 (due to a change in screening policy). All available screen-films (none were full-field digital) were retrieved and digitized on an Array 2905 digitizer at optical density 0–4.0, 16-bit depth and 75-micron resolution (Array Corporation Europe, Roden, The Netherlands). PMD was measured using Cumulus, an area-based threshold method by a single observer (V.A.M.), providing measures of the area of dense, nondense and total breast tissues, as well as PMD (100 × dense area/total breast area). All films (i.e., across all women, all views and all screens) were randomly sorted into batches of 200. Personal identifiers and dates were masked for the whole batch prior to reading. All left and right MLOs for every screen were read as well as all CC views from the 2004 screen, a total of 5277 films. Reliability, the ratio of between-film to total variance (i.e., between-film and within-film, the latter based on 2 independent readings of the same film for 100 films) was 0.90 for PMD.
The study design provided repeated measures of PMD over time (2–5 screening rounds per woman), with multiple readings at each screen (left and right breast, 1 or 2 views of each). The view and screen-specific left–right average PMD was calculated for each woman and used as the unit of analysis. This data structure, with correlated PMD between views and between screens within the same woman, was analyzed using 3-level hierarchical normal errors regression models with woman at level 3, at level 2 her screening round (1–5) and at level 1 the view taken (CC or MLO). We first modeled the basic trajectory of PMD with age using linear and quadratic fixed effects (random effects or treating age in 2-year bands did not provide a better fit). Thereafter, the influence of time-varying factors (varying between screens) HT and menopausal status on PMD were investigated and finally that of time-invarying factors (such as ethnicity and reproductive factors). All variables were included using indicator dummy variables for each category except for current BMI, which was modeled using both linear and quadratic terms when it was not the primary variable of interest. “Current” BMI was calculated from height and weight reported at the time of questionnaire completion (2005–2006) to refer to the 2004 screen and BMI 3 years prior to that to refer to all previous screens. We investigated influence on the rate of change of PMD by including and testing an interaction term of the factor with the linear age term. As the PMD age profile was quadratic in nature, the age at which PMD ceased to decline (agerate=0) was estimated by equating the first derivative of the age profile to zero (agerate=0 = –0.5 (linear/quadratic age parameters). All models were adjusted for view.
Similar analyses were carried out for dense area, nondense area and total projected area as for PMD. For all variables, a square root transformation was taken and effects were translated back to the original scale, referring to the effect for Caucasian women with median density of 20%, dense area of 27 cm2, nondense 108 cm2 and total breast area of 135 cm2.
Tracking was assessed using the Spearman rank correlation coefficient. Canalization, the tendency of a measure to remain in the same relative region over 3, 6 and 9 years was assessed using a weighted kappa statistic of agreement between deciles of the first density measurement and density at subsequent 3 yearly screens. The weights used were 1, 0.89, 0.78, 0.67, 0.56, 0.44, 0.33, 0.22, 0.11 and 0 (derived from 1 – |i – j|/(k – 1), where k = number of quantiles (i.e., 10), i and j are the deciles at different screens). Tracking and canalization were assessed using the MLO average, as this was the only view performed at every screen. Finally, logistic regression models were used to investigate the odds ratio of being in the highest PMD quintile at a woman's 2nd, 3rd and 4th screens (separately) for women who were in the top 4th and 5th PMD quintiles at their first screen, compared to those in quintiles 1–3.
To compare the age profile of PMD with that predicted by the model for the rate of breast tissue aging at these ages, we used the extended Pike's model created by Rosner and Colditz that incorporates ages at further and late pregnancies in addition to factors in the original model, i.e., ages at menarche, first birth and menopause.12 Parameter estimates from this model were based on breast cancer incidence rates in the Nurses' Health Study. Model predicted breast cancer incidence rates (log scale) were calculated for a reference woman with menarche at 12 years, a single full-term pregnancy at age 20 and menopause at age 50, and for 5 further women who differed from the reference woman by a single factor (factors not mentioned do not differ): (i) menarche at 15, (ii) nulliparous, (iii) parity 1 but age at first birth at 30, (iv) parity 3, with pregnancies at ages 20, 25 and 30 and (v) menopause at age 55. Appendix provides the formula for the predictions (taken from Rosner and Colditz12). Graphs of predicted rates of breast tissue aging were visually compared with corresponding age profiles of PMD.
Stata version 10 was used to perform all analyses, using the “xtmixed” command to fit multilevel models.
Characteristics of women
Six hundred forty-five women participated in the study, 267 Caucasian, 213 Afro-Caribbean and 165 South Asian women. Ethnic diversity resulted in a wide range in the distribution of risk factors (Table 1). Overall 88% of women were parous, and a relatively large percentage (11%) had had at least 6 full-term pregnancies. Among parous women, one-third had their first child at or under 20 years, 12% over age 30 and 70% had breast fed their children. Smoking prevalence was low (8%). Mean BMI at interview was higher than that 3 years previously as the majority of women reported a mean weight gain of 1.5 kg [standard deviation (SD) 6.4 kg].
Table 1. Characteristics of 645 participating women
Mean age at earliest screen was 52.3 years (SD 2.4). Films were available for between 2 and 5 screening rounds for each woman, with an average of 3 rounds each spanning a mean 5.9 years (SD 2.2). Eighty-one and 31 women had 4 and 5 screening rounds each, respectively. Most screens were taken when women were already postmenopausal, but 358 of 1930 screens were conducted when women were premenopausal, that is for 233 women (34.5%) mammograms were available from screens both when the woman was pre- and later postmenopausal. Thirty-one percent of women were taking HT at at least 1 screen, predominantly at younger ages.
Within-women changes in density
Table 2 shows the raw data for mean age, mammographic features and their within-woman rates of change for consecutive screens in 5 groups of women with between 2 and 5 screening rounds (the large number of women with 3 screens were split into 2 according to age at first screen so that any differences in the PMD profiles would not be concealed). Within each of the 5 groups, average PMD and dense area declined between screens whereas nondense and total breast areas increased. There was a greater drop in PMD between earlier than between later successive screens, e.g., in 83 women with 4 screening rounds each, mean rate of decline of a woman's PMD was 1.2% per year between 1st and 2nd screens (i.e., between mean ages 52.3–55.1), but the same women had a mean rate of decline of 0.2% per year between their 3rd and 4th screens (about 6 years later, at mean ages 57.9–60.9 years). PMD declines were mirrored on the whole by declines in the dense area and increases in nondense area, however increases in the latter were at least twice as large in magnitude and so the total breast area also increased between consecutive screens.
Table 2. Unadjusted mean (SD) of mammographic measures and their annual rates of change in sets of women with repeat screens
Trajectory of density
As suggested by the slowing rate of decline in PMD between consecutive screens, there was strong evidence of a nonlinear association of PMD with increasing age (p < 0.001 for quadratic age term). Mean within-woman PMD decreased by 1.7% (absolute reduction in PMD) [95% confidence interval (CI) 1.9–1.5] from age 50 to 51, whereas by the end of the age range from 64 to 65 by only –0.2% (95% CI –0.5 to 0.1). However, some of this decline was due to a tendency for increasing BMI and menopausal transition both of which decrease density. After adjusting for these and all other factors in Table 3, annual within-woman rate of decline in PMD, independent of these factors, was 1.4% (95% CI 1.2–1.6) at age 50 reducing almost to zero by age 65. Decline in PMD was mirrored by a nonlinear annual decline in dense area, of 1.6 cm2 (95% CI 1.2–1.9) at age 50 and almost no change by age 65. As expected, the opposite trend was observed for nondense area; however, its increases were of a much larger magnitude than the decreases in dense area, with an annual increase of 3.4 cm2 at age 50 and 1.3 cm2 at age 64 (Table 3). Based on this nonlinear trajectory, the predicted age at which density stabilizes is 65.6 years (95% CI 62.1–69.0), although this value is just beyond the limit of the age range for which we have data.
Table 3. Determinants of mean levels and rates of change of mammographic density (mutually adjusted estimates)
The time-varying factors menopausal status and use of HT were associated with PMD (Table 3). The menopausal transition was modeled in 2 ways—a sudden change at the reported age at menopause or a gradual decline by additionally incorporating a time-since-menopause variable but as the latter model did not explain more variation, menopausal status was included as a screen-specific binary indicator (i.e., modeled as a sudden change). Menopause was associated with a 2.4% (1.4–3.4) drop in PMD, arising from a decline in dense area of 3.3 cm2 and an increase in the nondense area of 4.2 cm2 (Table 3), but with no evidence of a change in breast area (data not shown). Of similar magnitude but in the opposite direction, HT use increased PMD and dense area and decreased nondense area.
The effect of time-invarying factors which differ between women on mean levels of PMD and on its rate of change are also shown in Table 3. Women with higher BMI, and of Afro-Caribbean and South Asian ethnicity (compared with being Caucasian) had statistically significant lower PMD and greater parity was consistent with lower PMD although was not statistically significant (after mutual adjustment). However, none of these factors affected the rate of decline of PMD or of dense area, with the exception of ethnicity for which White woman had a slower rate of decline in dense area than South Asian or Afro-Caribbean women. Figures 1a–1f display the model fitted trajectories of PMD predicted including interaction terms for each factor with age. Menarche, which was not found to affect percent density, was included (Fig. 1c) for comparison with Pike's model
Comparison with Pike's model of breast tissue aging
Figures 1g–1k show the age-specific rate of change of log breast cancer incidence (akin to the rate of breast tissue aging) as predicted by Pike's model for 2 groups of women, who differ by age at menopause (48 or 53, Fig. 1g), age at first birth (20 or 30, Fig. 1h), menarche (12 or 15 years, Fig. 1i), nulliparity (parity vs. nulliparity, Fig. 1j) and parity (1 or 3, Fig. 1k). At ages 50–65, corresponding to our data as shown in Figures 1a–1f, rates of breast tissue aging differ by menopausal status (Fig. 1g) and parity (Figs. 1j and 1k) only—more parous women have a lower rate of change of log breast cancer incidence and postmenopausal women have a lower rate, both of which are mirrored in Figures 1a and 1d for PMD. According to Pike's model, the rate of breast tissue aging at ages 50–65 years does not differ between women who had different ages at menarche (Fig. 1i), coinciding with no effect of menarche on PMD (Fig. 1c), or age at first birth (Fig. 1h) however PMD at these ages differed by this latter factor (Fig. 1b).
Tracking of density
Spearman rank correlation coefficients demonstrate a high degree of tracking, with a value of 0.87 for films taken 3 years apart (Table 4). Rank correlation declined slightly with increasing time, but even after 12 years, it was 0.73 (based on 32 women). Correlations were higher for total breast area than for dense area, showing that it is changing ranks of dense area that account for differences in PMD ranks between screens. In accordance, kappa statistics and percent agreement show high agreement of deciles. As HT is a time-varying factor that influences density, restricting analyses to women who did not take HT resulted in higher rank correlations (0.89 for women not on HT at either screen compared with 0.77 in women who were taking HT at either of 2 screens 6 years apart). Finally, being in the top quintile of PMD compared with being in quintiles 1–3 was associated with a 60-fold increased odds of still being in that quintile 3 years later, and a 25-fold increased odds 9 years later. Similarly, women in quintile 4 had greatly increased odds of being in the top quintile 3, 6 and 9 years later (Table 4).
Table 4. Tracking of mammographic density (unadjusted) over time
We found that women's PMD declined between ages 50 and 65 years, with an annual age-related rate of decline of about 1.4% in her early 50s dropping rapidly to almost zero by age 65. This trend is consistent although of a larger magnitude than in a longitudinal study in Hawaii which found a decline of 6% over 10 years, also with a slowing rate of decline with increasing age.2 Other estimates of annual rates of decline at similar ages are 1% for premenopausal women in their late 40s,13 1.14% in pre- and perimenopausal women in the Study of Women's Health Across the Nation (SWAN)3 and ∼1.6% in the control arm of a tamoxifen trial (ages 45–60 years at baseline).14 The predicted age at which density ceased to decline was 65 years, at the upper limit of the included age range, broadly in agreement with data from 2 other studies.2, 15
We found that mean levels of breast density are affected by both immediate and distal factors, but that its rate of change at ages 50–65 years is affected only by immediate factors, i.e., use of HT and menopause. Credibility of these findings are supported by their striking similarity with those from the Minnesota Breast Cancer Family Study.10
In our study, menopause was associated with a reduction in PMD of ∼2.4%, independent of the effects of changes in age and BMI, consistent with previous longitudinal studies10, 13 but not with an Australian study that found that the decrease in PMD was due to an increase in nondense area with no change in dense area.16 We did not have information on type of menopause and surgical menopause may be associated with a greater drop than a natural transition.10 In the opposite direction, HT increased PMD by over 2%, although this may be underestimated as HT type was not known and estrogen plus progestin therapy increases density particularly.17 These time-varying factors cannot occur without a concomitant change in age, thus these PMD changes are on top of age-related declines.
This is one of the few studies that has analyzed area components of PMD. Notably, although HT and menopause affected PMD without having an effect on the total breast area, age-related density changes were associated with an increase in breast area. PMD declined within women partly as a result of a decrease in dense tissue but, to a greater extent, as a result of an increase in nondense area which was increasing at twice the rate, after controlling for BMI. The impact of within-woman changes in BMI on PMD could not be adequately assessed in our study due to the lack of accurate screen-specific BMI data. Thus, if BMI increased with increasing age to a greater degree than our estimates, residual confounding by this suboptimal measure might have led to positive confounding of the age trend, and BMI-adjusted declines may be lower. It would be interesting to explore whether such a large increase in nondense area and in total area occurs in populations that do not undergo age-related increases in BMI to this extent. However, our results are based on area-based PMD and the degree to which they reflect changes in breast volume is less clear due to differential compression by density, but the need for a larger bra size after the menopause in another study suggests a real increase in breast volume.18 Volumetric methods of measurement made on repeated screens may help to more precisely quantify the changes.
A decline in dense area coupled with an increase in nondense area on menopausal transition and with increasing age is likely to reflect the process of lobular involution occurring at this age. There is a reduction in the number and size of lobules as a result of the substantial drop in estrogen levels on ovarian shutdown, a process which is itself associated with a reduction in breast cancer risk.19 Epithelial proliferation rates decline, fibroglandular tissue recedes and the stromal component declines resulting in a drop in PMD. Increases in the nondense area (and thus a decrease in PMD) may result from weight gain or from weight redistribution, which increase fatty tissue in the breast possibly without influencing the amount of fibroglandular tissue.
Breast density tracked for up to 9 or even 12 years at ages 50–65 years. This observation helps explain why, at these ages, breast density remains a predictor of breast cancer risk 10 years or more after density assessment.20 Such tracking suggests that differences between women and the factors causing them arise prior to this age (in agreement with the finding that distal factors did not affect rates of PMD decline). Tracking at these ages may explain inconsistencies found for the effect of change in PMD on breast cancer risk, as 2 studies primarily in postmenopausal women found no association,2, 4 but Kerlikowske et al. found that a decrease in BIRADS category in younger women was associated with a decrease in breast cancer risk.7 If PMD ranking does not change greatly at older ages, repeated measures of density do not provide extra statistical information (other than to reduce noise) to aid breast cancer prediction. Furthermore, tracking may be a useful feature for screening and prevention. In the context of the UK NHS Breast Screening Programme, women with high density can be identified at younger ages (at age 50 in this study), and possibly benefit from risk-lowering interventions or from appropriate screening modalities and shorter screening intervals.21
Our study benefited from a large number of screening rounds and accurate estimation of density through the use of the left–right average PMD as measurement error in the assessment of density would have weakened any signal of tracking. Reading films in a random order across all woman and screens was also essential to assess tracking so that measurement errors were unlikely to be correlated. Some features of the study design were less than ideal. Mammograms were taken 3 years apart so the nature of the decline in density during the perimenopausal period or on starting and stopping HT is not precisely captured. All exposure data were self-reported thus the accuracy, particularly of distal factors such as age at menarche, may be suboptimal. Additionally, participation rates were low, particularly in the South Asian ethnic group, and participating women may not be representative of the general population, e.g., if participants are more likely to be of higher socioeconomic groups that are likely to have higher density. However, as the 2 ethnic groups with lowest response rates had the lowest density, ethnic differences should be real, and it is unlikely that particular patterns of PMD change would have influenced response.
The similar effects of menarche, menopause and parity on PMD as those predicted by Pike's model suggest that a woman's age-specific PMD partially correlates with her age-specific rate of breast tissue aging. However in Pike's model, a woman's rate of breast tissue aging is constant once she is postmenopausal, whereas our and other data suggest that PMD demonstrates a gradual decline at these ages. The lag period of up to 10 years before PMD stabilizes suggests that PMD may represent the average rate of breast tissue aging over the past 10 years. Differences in PMD by age at first birth, that do not exist in the modeled rate of breast tissue aging at these ages (they are present earlier) may also be due to PMD representing an average rate of aging. This interpretation would be consistent with the gradual, rather than sudden, change in PMD with cessation of HT22 or commencement of tamoxifen treatment.14 Declining MRI-determined breast percent water measured cross-sectionally in women aged 15–30 years would also be consistent with PMD being an average of recent rate of breast tissue aging.23 Finally, as remarked by Rosner and Colditz,12 their model only considered reproductive factors, thus important breast cancer risk factors such as HT use and BMI were not considered and may affect breast tissue aging rates in postmenopausal women.
Some characteristics of PMD are suggestive of it being a marker of breast tissue aging. Further longitudinal studies of PMD at younger ages are needed to explore the meaning of PMD during this susceptible period.
V.A.M. was supported by a Cancer Research UK Graduate Training Fellowship. The authors thank the women who participated in this study, and staff at the Central and East London Breast Screening Service for their help and cooperation. They also thank the reviewers whose insightful comments and suggestions greatly improved this manuscript.
Model for log breast cancer incidence at age t:
where t0 is the age at menarche, ti, age at the ith pregnancy; M = 1 if postmenopausal, 0 otherwise; tm, age at menopause; b1 = 1 if parous, 0 if nulliparous; t*, minimum of age and age at menopause tm. Best fit to the Nurses' Health Study provided parameter estimates of , , , , , and