Development and evaluation of a method to assess breast cancer risk using a longitudinal history of mammographic density: a cohort study

Mammographic density is one of the strongest known risk factors for breast cancer [1]. Mammographic density is commonly measured using the Breast Imaging Reporting and Data System (BI-RADS) four-category scale, summarised as (A) almost entirely fat, (B) scattered fibroglandular, (C) heterogeneously dense or (D) extremely dense breasts [2]. Including BI-RADS density in breast cancer risk models improves the accuracy of individual risk assessment [3]. Providing accurate estimation of a woman’s breast cancer risk could aid decisions regarding supplemental screening, risk-reducing strategies based on diet and exercise, preventive therapy and risk-stratified screening.

Most studies assessing the effect of mammographic density on breast cancer risk are based on a single measurement. Whilst some studies measure density at more than one time point, the majority of studies assess change in density between mammograms taken at two time points [4, 5]. Moreover, the focus of many of these studies is on the association between density change and breast cancer risk, rather than the predictive ability of multiple serial measurements of density. A recent systematic review [4] identified just one study that had evaluated the predictive performance of using more than one serial density measurement for risk of subsequent breast cancer [6]. This US cohort study of ~ 700,000 women explored using two density values over two years to assess invasive breast cancer risk and observed a small improvement in risk discrimination (concordance index 0.640 vs. 0.635) [6]. However, using a woman’s entire history of breast density measurements might be more informative for breast cancer risk estimation than using only one or two time points. This is partly because mammographic density is a dynamic trait that decreases most strongly with increasing body mass index (BMI) [7, 8], age [9, 10] and during the menopause [10, 11]. It also decreases in response to endocrine treatment [12‐15] and increases with current hormone replacement therapy use [16‐18]. There is also variation in radiologists’ visual interpretation of mammographic density due to inter- and intra-reader variability [19], and thus, there is a need for measurements of breast density to be more reliable than they are at present [20]. Using a woman’s history of previous density measurements should help to reduce such measurement error. In addition, when using a breast density history with more than two records, it may be important to include information on mammograms arbitrarily spaced through time, i.e. not necessarily only two mammograms approximately one or two years apart, but previous work does not appear to have considered this.

The aim of this cohort study was to develop a method to assess breast cancer risk using a longitudinal history of breast density and to assess the predictive value of this approach in comparison with breast density measured at a single time point.

The cohort included 132,439 women with a median follow-up of 5.2 yr (interquartile range (IQR) 2.4–11.1 yr) and maximum follow-up of 19 yr. The median time between mammograms was 1.8 yr (IQR 1.0–2.0 yr) and the median number of mammograms per woman was 3 (IQR 2–6), with 32,010 women (24.2%) having a baseline mammogram only. The number of mammograms was similar across different ages at baseline and throughout the follow-up. Women were of Asian (9%), Black (4%), Mixed (3%), Other (3%), Unknown (2%) or White (80%) race, and Hispanic ethnicity was reported in 5% of women. In total, 2704 women (2.0%) were diagnosed with invasive breast cancer during follow-up. Summary statistics on age, BMI and breast density at baseline are shown in Table 1.

In the linear mixed model for continuous longitudinal density (Additional file 1: Table S1), an unstructured covariance matrix was chosen because it provided a substantially better model fit than an independent structure (ΔLR-χ²(1) = 2473.3). A quartic polynomial term for age and an interaction between age and BMI were included because likelihood-ratio statistics and graphical plots of predicted density by age and BMI identified an improved model fit. There was also evidence of a nonlinear relationship between longitudinal density and breast cancer risk (Fig. 1), with improvement in model fit when including a quadratic term (ΔLR-χ²(1) = 15.0), but not a cubic term (ΔLR-χ²(1) = 1.1).

All breast density methods were strongly associated with risk (Table 2), but most statistical information was in the continuous longitudinal density measure. It added 23% more statistical information to an age- and BMI-only model than BI-RADS density (ΔLR-χ²(2) = 379.6 vs. ΔLR-χ²(3) = 307.7, respectively). This corresponded to a marginal increase in the mean c-index over all of the follow-up: 0.64 for continuous longitudinal density vs. 0.63 for BI-RADS density (mean difference in c-indices = 0.008, 95% CI 0.003–0.012) (Fig. 2). The eight-category version of longitudinal density was inferior to the continuous measure (ΔLR-χ²(7) = 375.1), but aids interpretation of the hazard ratios associated with the continuous measure (Table 2). The densest of the eight categories (‘Extremely dense’ II) had threefold greater risk (HR = 3.0, 95% CI 2.5–3.7) than the reference (‘Scattered’ I), with a sixfold greater risk when comparing densest and fattiest breasts (HR = 6.3, 95% CI 4.7–8.7, relative to ‘Fatty’ I). This compared with a fourfold increased risk between densest and fattiest BI-RADS density (Extremely dense HR = 4.3, 95% CI 3.4–5.5, relative to Fatty).

In a secondary analysis restricted to women with at least three mammograms (n = 76,313 (58% of cohort), 2169 cancers after third mammogram), the gain in statistical information when using the continuous longitudinal density measure was 35% more than when using BI-RADS density (ΔLR-χ²(2) = 251.2 vs. ΔLR-χ²(3) = 186.7, Table 3), with respective mean c-indices of 0.63 and 0.62 (mean difference in c-indices = 0.010, 95% CI 0.005–0.015). Using eight-category longitudinal density (ΔLR-χ²(7) = 246.3), the risk gradient between densest and fattiest breasts was greater for the extremes of the eight-category longitudinal density (HR = 4.8, 95% CI 3.5–6.8) than BI-RADS density (HR = 3.5, 95% CI 2.7–4.5). Using breast density measures at/to the third mammogram and not updating further through time yielded similar results to the main analysis with breast density as a time-varying covariate (Table 3).

Cumulative distribution plots for relative risk associated with the second and third mammogram suggested that longitudinal density is more stable through time than BI-RADS density: 11% of women had an estimated relative risk outside of the stable range (relative risk was considered to be stable between 4/5 and 5/4) when using longitudinal density, compared with 28% when using BI-RADS density (Fig. 3).

Exploratory analysis including density trajectories showed only a small improvement on the continuous longitudinal density model (ΔLR-χ²(1) = 4.0, p = 0.046), with a 1% gain in statistical information. There was little evidence of an interaction between continuous longitudinal density and age, BMI or race (age at baseline interaction, HR (per 5 yr) = 0.99, 95% CI 0.94–1.03; BMI interaction, HR (per 5 kg/m²) = 1.08, 95% CI 0.99–1.17; race interaction, HR (non-White relative to White) = 0.99, 95% CI 0.81–1.21). Longitudinal density performed better than BI-RADS density in both premenopausal (n = 87,156, 1581 invasive breast cancers) and postmenopausal (n = 45,283, 1123 invasive breast cancers) women (ΔLR-χ² was 23% and 21% greater for continuous longitudinal density than BI-RADS density, respectively), and for both oestrogen receptor positive (n = 2175) and negative (n = 418) subtypes (excluding 111 invasive breast cancers with unknown oestrogen receptor status, n = 132,328 in both the oestrogen receptor positive and negative models, ΔLR-χ² was 24% and 19% greater for continuous longitudinal density than BI-RADS density, respectively; heterogeneity p-values: linear = 0.816, quadratic = 0.772).

In sensitivity analyses, results were robust to excluding mammograms taken up to 6 months before an incident breast cancer (all women, follow-up from baseline mammogram, ΔLR-χ² was 23% greater for continuous longitudinal density than BI-RADS density). We also found similar results when removing mammograms with imputed BMI (n = 129,748, 2668 invasive breast cancers, follow-up from updated baseline mammogram, ΔLR-χ² was 30% greater for continuous longitudinal density than BI-RADS density). Furthermore, continuous longitudinal density outperformed baseline BI-RADS density (ΔLR-χ²(2) = 379.6 vs. ΔLR-χ²(3) = 296.2, respectively, resulting in a ΔLR-χ² that was 28% greater for continuous longitudinal density than baseline BI-RADS density; baseline BI-RADS density was adjusted for age at baseline and baseline BMI). This suggests that a greater gain in aetiological information can be obtained by including multiple historical measures of mammographic density instead of a single measure that reflects a woman’s ‘peak’ available breast density. Four-category longitudinal breast density also performed slightly better than BI-RADS density (ΔLR-χ²(3) = 312.3 vs. ΔLR-χ²(3) = 307.7), indicating that the improvement in statistical ability when using longitudinal density is explained by more than just longitudinal density being measured continuously compared with categorical BI-RADS. Finally, in a post hoc sensitivity analysis, we tested whether the superior performance of longitudinal breast density compared with BI-RADS density could be due to the incorporation of an interaction term between age and BMI in the calculation of the longitudinal breast density measure. However, a proportional-hazards model including a single (time-varying) BI-RADS density measure, age at baseline, BMI (time-varying) and an age–BMI interaction did not perform as well as the proportional-hazards model including longitudinal density (time-varying; including a quadratic term), age at baseline and time-varying BMI (ΔLR-χ²(4) = 308.8 vs. ΔLR-χ²(2) = 379.6).

Using a woman’s longitudinal history of breast density could improve breast cancer risk estimation in women attending mammography screening over using a single measurement alone. Our proposed longitudinal breast density measure provided a gain in statistical information that was 23% more than that provided by BI-RADS density. Women in the highest longitudinal density category had a sixfold greater risk of developing breast cancer than women in the lowest category, compared with a fourfold difference when using BI-RADS density. This greater spread of risk with eight-category longitudinal density (which demonstrates the greater spread of risk that is achievable when using the continuous longitudinal density measure) suggests a potential for improved risk stratification with the continuous longitudinal density measure. We consider there to be two main benefits of using repeated mammograms and deriving a longitudinal breast density measure. Firstly, the longitudinal breast density measure is likely to be reflecting a woman’s long-term average exposure to breast density and therefore aligning with Boyd et al.’s model of breast tissue ageing which suggests that cumulative rate of exposure to hormones, and consequently breast density, is positively associated with breast cancer risk [27, 28]. Secondly, since the longitudinal breast density measure appears to be more stable over time than a single measure of BI-RADS density, it is likely that it reduces variation in estimates of an individual’s visual assessment of breast density and therefore also breast cancer risk between screens. This should prevent fluctuations in the classification of women into different risk categories through time. We also note that the improvement in discrimination when using longitudinal breast density was relatively small, suggesting that the benefit of the measure is more so in its ability to improve the reliability and accuracy of a woman’s breast density measurements and personalised estimate of breast cancer risk than in its ability to differentiate between women who will, and will not, go on to develop breast cancer. However, given that the longitudinal density measure is likely to reduce the amount of noise in data, its utilisation should result in an improvement in the performance of established breast cancer risk models that currently use BI-RADS density.

These results support previous findings that suggest an improvement in predictive ability of breast cancer risk estimation when using density values from more than one time point [6]. Kerlikowske et al. assessed BI-RADS density in a screening cohort of ~ 700,000 women from the BCSC, where a two-measure density score was developed combining first and last BI-RADS density measures taken on average 1.8 years apart [6]. They found an improvement in the BCSC 5-year risk model when using their two-measure score compared with a one-measure score, whereby the area under the receiver operating curve (AUC) increased by ~ 0.005. Our study (which includes some of the women in Kerlikowske et al.’s study) also found a slight improvement in discriminatory accuracy when using multiple density measures (compared with a single density measure, the concordance index increased by ~ 0.008). Several other studies have also made use of more than one serial density value [4, 5], with most suggesting an association between change in density and breast cancer risk [5, 31‐41], demonstrating the benefit of including more than one time point for breast cancer risk estimation. However, these association studies did not focus on evaluating the predictive ability of using serial density measurements to estimate future risk of breast cancer.

There is a question of whether the longitudinal breast density measure improves accuracy or whether it provides aetiological information. In our view, the main statistical advantage of longitudinal density is to reduce measurement error and improve reliability and accuracy by functioning as a shrinkage estimator making use of multiple data points. This was apparent from our results because the predictive ability of longitudinal density improved by a greater amount when it was included in the proportional-hazards survival analysis as a continuous variable than as a categorical variable. This result also highlights the advantage of using finer-grained descriptions of a woman’s history of breast density. In addition, we initially hypothesised that including information on the rate of change of each woman’s density might be informative (analysis including random slopes in the model). However, our data suggested only a small effect on predictive ability when including each woman’s individual slope. Whilst this finding indicates a high level of density tracking (which is consistent with earlier work [42, 43]), it differs from the results of a recent study which reported a better model fit when using both current mammographic density and individual slope compared with current mammographic density alone [44]. A possible explanation for the contrasting findings with our study is that different density measures were used. [BI-RADS density was used in our study, and STRATUS (an automated area-based measure of absolute density) was used in Illipse et al.] It would be worthwhile investigating the longitudinal breast density method using different measures of mammographic density such as the aforementioned. Another consideration of the longitudinal breast density measure is that, whilst the method is clearly most useful when a woman has more than one density measure, it may also improve reliability of breast density measurements even with just a single density score. This is because it borrows strength from other women of the same age and body mass index, and shrinks the estimates accordingly. For example, one might expect a 40-yr-old woman with healthy weight and BI-RADS density D to have truly denser breasts than a 70-yr-old woman with obesity and BI-RADS density D; and this would be reflected in the proposed algorithm.

The major strength of this study is the use of a large cohort with repeated measures of breast density through time, and the development and predictive evaluation of a new method to assess serial density, particularly using more than two measures arbitrarily spaced through time. This might make longitudinal density a particularly useful tool for clinical practice where women can have several mammograms taken at any point in time. We additionally observed improvements when assessing women with at least three prior mammograms, which demonstrates an increasing value of the method with more data. Therefore, the benefit of using longitudinal density to assess breast cancer risk in terms of reducing measurement error and preventing fluctuations in the classification of women into different risk categories through time may be improved further by updating the measure at each screening examination to provide more density values. Another advantage of using longitudinal breast density is that it would help to identify women with persistently dense breasts who may be eligible for supplemental screening. This is an important piece of information that cannot be discerned from a single BI-RADS density measure. In addition, the beneficial consequence of having greater stability when using longitudinal breast density compared with BI-RADS density is likely to have implications for breast cancer screening programmes which employ a risk-based approach. For instance, according to the UK’s National Institute for Health and Care Excellence guidelines and previous studies assessing the application of these guidelines to population risk assessment, breast cancer risk thresholds of ≥ 8% and 5–8% absolute 10-year risk may be used to classify women into high- and moderate-risk groups, respectively [3, 45, 46]. Therefore, in the UK context, if a group of high-risk women with an 8–10% absolute 10-year risk of breast cancer (based on an established risk model) were to experience a 20% reduction in relative risk (i.e. relative risk = 4/5), then 17% of these women would be reclassified as moderate risk if using BI-RADS density, compared with only 6% if using longitudinal breast density (under the assumption that the distribution in Fig. 3 holds for comprehensive risk assessment within this group). Such a large difference in reclassification would have important ramifications for women who are truly at a high risk of breast cancer and may benefit from enhanced screening, but would be ineligible for it due to omission of information from previous mammograms. Furthermore, in view of the recent U.S. Food and Drug Administration regulation updates requiring patient mammography reports to indicate whether they have ‘dense’ or ‘non-dense’ breast tissue [47], and given that longitudinal density is more stable over time than a single density measure, the use of the longitudinal density measure could help to mitigate confusion amongst patients if their density values were to vary between screening visits. (This is especially relevant for women whose BI-RADS density oscillates between B and C.)

Limitations of the study include the statistical models used. In particular, BI-RADS density categories were modelled as quantitative integer values to crudely approximate a linear association between density and age and BMI in the linear mixed model. There is some justification for this because the relative hazard associated with BI-RADS density adjusted for age and BMI is approximately 1, 2, 3, 4 (cf. comparing relative differences in Table 2). However, alternative models, such as joint models, might better fit the data and improve predictive ability [48]. Furthermore, the linear mixed model used to develop longitudinal density was adjusted for age and BMI only. Adjusting longitudinal density for additional confounders such as hormone therapy use, benign breast disease or reproductive factors might improve its approximation; although we note that these adjusting factors may be more relevant when predicting breast cancer risk than breast density. Moreover, in a recent association study examining the relationship between change in volumetric percentage density and subsequent breast cancer risk, each breast was considered independently in the statistical model [49]. It is possible that modelling each breast independently could provide additional aetiological information. This approach was not possible in our analysis, however, because the density values at each time point were measured as a combined assessment of the four breast view combinations available to the interpreting radiologist at the screening examination. Another limitation is that the algorithm for calculating longitudinal density is not straightforward to implement using a simple rule. However, it is less complicated to implement in computer systems than fully automated measures of breast density, and the algorithm to evaluate this new measure is fully available in Supplementary Methods, Additional file 1. Finally, since the focus of this study was on evaluating the predictive performance of longitudinal breast density above that of age and BMI, rather than in the context of established risk models, we did not adjust for additional breast cancer risk factors. Future work will aim to validate the longitudinal breast density method in independent datasets, as well as assess the value of incorporating the measure into established breast cancer risk models such as the IBIS/Tyrer–Cuzick algorithm [50]. Ultimately, by adapting such risk models to include a woman’s historical information, dynamic risk prediction could be made possible. Patient data could be updated regularly, and accurate and reliable estimates of breast cancer risk (that take into account all of the woman’s longitudinal information) could be provided at each new screening visit.

We conclude that estimating mammographic density using a woman’s history of breast density is likely to be more reliable than using the most recent observation only, which may lead to more reliable and accurate estimates of individual breast cancer risk. Longitudinal breast density has the potential to improve personal breast cancer risk estimation in women attending mammography screening.