Pre-natal exposures and breast tissue composition: findings from a British pre-birth cohort of young women and a systematic review

The Avon Longitudinal Study of Parents and Children (ALSPAC) is a prospective pre-birth cohort of 14,775 children born in Avon, England (representing 72 % of the eligible population [7]), between 1 April 1991 and 31 December 1992 [7, 8]. For this study, young, nulliparous women born from singleton pregnancies who regularly participated in follow-up surveys and had never been diagnosed with cancer or a hormone-related disease were invited to attend an MRI examination of their breasts at the University of Bristol Clinical Research and Imaging Centre between June 2011 and November 2014. Women who had contraindications for MRI (e.g., pregnancy, metal implants) were excluded. Of the 2530 potentially eligible women invited, 500 (19.8 %) attended. The low response rate reflected the highly demanding nature of the study, as well as relocation away from the study area (e.g., to attend university). However, participants were similar to potentially eligible women who did not participate in relation to socio-demographic factors and body size measurements (e.g., mean birthweight and BMI at age 16 years were 3390.9 g [SD 21.6 g] and 21.2 kg/m² [SD 0.2 kg/m²], respectively, amongst women who participated, and 3397.4 g [SD 11.4 g] and 21.5 kg/m² [SD 0.1 kg/m²], respectively, amongst those who did not). Mothers of participants provided access to their mammograms taken as part of the U.K. national screening programme if they were in the targeted age group (50–70 years).

The study received approval from all relevant ethics committees (listed below in the Ethics approval and consent to participate subsection). Participants provided written informed consent.

Information on maternal, in utero and birth size variables was collected from self-administered maternal questionnaires at enrolment in early gestation, throughout pregnancy and shortly after delivery, supplemented by obstetric and paediatric records [9]. During their MRI examinations, participants completed a short questionnaire on menstruation-related variables, and anthropometric measurements were taken. The mother’s BMI, parity and menopausal status closest to the time of mammography were obtained through face-to-face clinical assessments and self-completed questionnaires. The study website contains details of all the data that are available through a fully searchable data dictionary [9].

Young women underwent an examination using a 3-T Siemens MAGNETOM Skyra MRI system (Siemens Healthcare, Erlangen, Germany) with a breast coil that surrounded both breasts and with the women in prone position. For each woman, three sets of images through both breasts were obtained: (1) T1-weighted VIBE 3-D images (approximately 176 images per woman) with a voxel size of 0.76 × 0.76 × 0.90 mm³, (2) T2-weighted transaxial images (approximately 40 images per woman) with in-plane resolution of 0.85 × 0.85 mm² and slice thickness of 4 mm, and (3) sagittal Dixon images (between 37 and 44 per woman) with in-plane resolution 0.74 × 0.74 mm² and slice thickness of 7.7 mm. Fully automated algorithms were developed to estimate breast volume using both T1-weighted and T2-weighted images and perform fat/water segmentation on T2-weighted images, whilst semi-automated breast and fat/water segmentation methods were developed for the Dixon images (details provided in Additional file 1: Methods 1). These algorithms yielded left-right average estimates of volumes (in cubic centimetres) of breast, water and fat (the latter two correspond to mammographic dense and non-dense tissues, respectively), as well as percent water. Percent water has been shown to be highly correlated with mammographic percent density of the same women [10‐12]. In comparisons in a random sample of 200 participants, we found little difference in breast measurements across the different MRI images (Additional file 1: Methods 1), and results from T1-weighted and T2-weighted images are presented here. Valid breast parameters were obtained for 491 of the 500 participants who underwent the MRI examination.

Processed digital mammographic images were successfully retrieved from screening centres for 175 mothers. Left and right craniocaudal images were read using the Cumulus semi-automated area-based method [13, 14] to estimate average breast, non-dense and dense areas (in square centimetres), and percent density (mammographic percent density). Cumulus density readings of processed images are strong predictors of breast cancer risk [15]. Readings were performed by a single observer (IdSS) who was blind to the women’s characteristics (within-observer intra-class correlation 0.92).

Linear regression models were fitted to examine associations between participants’ breast tissue parameters and maternal, in utero and birth size variables. Breast tissue parameters were first log-transformed to achieve near-normal distributions. To improve interpretability, exponentiated estimated regression parameters are reported; these represent the relative percent change (RC) in breast measurements associated with a unit increase in the exposure of interest. Continuous exposure measurements were standardised and, where appropriate, grouped into relevant categories or using quartiles as cut-off points. Exposure effects were adjusted for (1) age, BMI, phase of menstrual cycle (luteal, follicular and irregular period), hormonal contraceptive use at the time of MRI (as described in Table 1), and, when investigating the role of maternal mammographic density measurements, also for maternal age and BMI at mammography; and (2) further adjusted for other maternal, in utero or birth size variables as specified in the tables and figures. For simplicity, variables (1) and (2) will be referred to hereafter as minimally and mutually adjusted effects, respectively.

Mediation analyses were performed to investigate separately whether the effect of maternal exposures on the participants’ percent water were mediated via birth size, and whether the effect of birth size was mediated through adult height and BMI [16]. Linear regression models were fitted to the percent water, the exposure and each mediator in turn, with relevant confounders included and interactions between each exposure-mediator pair investigated. Results are presented in terms of direct (i.e., not mediated) and indirect (i.e., mediated) effects and expressed as percent changes, with 95 % CIs for the indirect effects obtained by bootstrapping [17].

Sensitivity analyses were conducted using (1) an alternative method to estimate percent water on Dixon images (Additional file 1: Methods 1) and (2) multiple imputation to deal with missing exposure and confounder data under the missing-at-random assumption [18] to obtain results based on all participants with valid MRI breast parameters (n = 491). Imputation by chained equations method was used, including all the exposures, confounding factors and outcomes involved in the analysis. The models described above were fitted to each of 20 imputed datasets, and overall estimates were obtained using Rubin’s rules [19].

Data analysis was conducted using STATA version 14 software (StataCorp, College Station, TX, USA). All tests of significance were two-sided.

The original protocol and methodology of the review are given in Additional file 1: Methods 2 and 3. Briefly, a search for studies published between 1 January 1970 and 25 September 2015 was conducted in PubMed using the search terms detailed in Additional file 1: Methods 3. Paper screening and data extraction were completed independently by two reviewers (RD, IdSS). The quality of the eligible papers was assessed by developing a standardised quality score (ranging from 0 [lowest quality] to 59 [highest quality]) based on 15 individual parameters reflecting the potential for selection bias, measurement error and confounding (Additional file 1: Methods 3).

Estimates of association between pre-natal exposures and breast density, as ascertained by mammography or an alternative approach, were extracted from each study and reported graphically using forest plots, whenever appropriate. To summarise the results in terms of linear trends, we first estimated linear effects across categorical exposures using study-specific weighted regression across the reported regression coefficients (with weights proportional to their SE). Derived study-specific linear trend coefficients and study-specific linear effects of continuous exposures were then summarised using random effects meta-analysis. Between-study heterogeneity was assessed using the I ² statistic. To examine potential sources of heterogeneity, study-specific trend coefficient estimates were stratified according to variables defined a priori (i.e., menopausal status at mammography, source of pre-natal exposure data, breast density assessment method). Funnel plots and Egger’s test were used to assess publication bias.

Table 1 presents the distributions of maternal, in utero and birth size characteristics of the participants. The mean age of the participants at the time of MRI was 21.5 years. Relative to mothers for whom mammograms were not available, those with mammograms were, as expected given the age group targeted by the U.K. national breast screening programme, older (31.7 vs. 28.8 years) and more likely to have older children (55.6 % vs. 49.3 %) at the participants’ birth. There were no differences, however, in the participants’ characteristics according to whether a mammogram could be retrieved for their mothers (data not shown). Participants’ percent water was inversely associated with BMI at the time of MRI, but not with age (reflecting their rather narrow age range) or menstrual phase/contraceptive use (Additional file 2: Table S1). Mothers’ mammographic percent density was inversely associated with both their age and their BMI at the time of mammography, but not with parity or menopausal status (Additional file 2: Table S1).

In univariate analyses, percent water in daughters was positively correlated with maternal mammographic percent density (r = 0.23; p = 0.003; n = 175), mirroring positive daughter-mother correlations in the amounts of dense tissue (r = 0.18; p = 0.014) (Additional file 3: Figure S1). Both maternal height and mammographic percent density were positively associated with the participants’ percent water in minimally (Fig. 1d) and mutually (Table 2) adjusted analyses. One SD increment in maternal height and mammographic percent density were associated, respectively, with a 2 % (RC 1.02; 95 % CI 1.00–1.04) and a 6 % (RC 1.06; 95 % CI 1.01–1.10) increase in percent water (Table 2), reflecting mainly increases in water volume (Fig. 1c). No associations with other maternal characteristics or in utero exposures were observed (Table 2).

Weight, length, head circumference and ponderal index at birth were all associated with percent water in minimally adjusted analysis (Fig. 2b), reflecting similar positive associations with water volume, with the exception on ponderal index, (Fig. 2d). A 1-SD increment in weight (473 g), length (2.3 cm), head circumference (1.2 cm) and ponderal index (4.1 g/cm³) at birth was associated with a minimally adjusted RCs of 1.03 (95 % CI 1.02–1.05), 1.02 (1.00–1.03), 1.02 (1.01–1.04) and 1.01 (1.00–1.03), respectively. For birthweight, this corresponded to an absolute 5.45 % (95 % CI 1.05–9.85) difference in minimally adjusted mean percent water between the extreme categories of the birthweight distribution (i.e., ≥4.0 vs. <2.5 kg) (Additional file 3: Figure S2). Gestational age was not associated with percent water (Additional file 2: Table S2); indeed, further adjustment for this variable did not materially affect the magnitude of the birth size-percent water association (Table 3).

Weight and length at birth were correlated with each other (r = 0.67, p < 0.0001), and both were correlated with head circumference (r = 0.71, p < 0.0001, and r = 0.50, p < 0.0001, respectively) and ponderal index (r = 0.29, p = 0.001, and r = 0.46, p = 0.001, respectively). Both length and head circumference reflect linear (skeletal) growth, but in mutually adjusted analysis only the association of the latter with percent water persisted (Table 3). Ponderal index reflects adiposity, while birthweight is a function of both linear growth and adiposity; however, only the birthweight-percent water association persisted when head circumference or ponderal index was included in the model (Table 3).

Similar findings were observed in multiple imputation analyses or when the breast tissue measurements were estimated using an alternative method to measure percent water on Dixon images (Tables 2 and 3 and Additional file 2: Table S2).

The association between maternal height and the participants’ percent water was partly mediated by birthweight, with about half of its total effect (mutually adjusted total effect RC = 1.01; 95 % CI 1.00–1.01) being attributable to its influence on birthweight and the effect of birthweight on percent water (indirect effect 1.01; 95 % CI 1.00–1.01) (Fig. 3a). For the birthweight-percent water association (mutually adjusted total effect 1.03; 95 % CI 1.02–1.05), there appears to be some evidence of protective mediation through the adult height of the participants (indirect effect RC = 0.99; 95 % CI 0.99–1.00). In contrast, there was no evidence of mediation of the effect of maternal mammographic percent density on the participants’ percent water through birthweight, nor was there evidence of the effect of birthweight on percent water via BMI at the time of MRI (Fig. 3b and d).

In the systematic search, we identified 208 abstracts, of which 12 were eligible (Additional file 3: Figure S3). Of these, most were conducted with post-menopausal women, although four included or were restricted to pre-menopausal participants [20‐23]. Two studies included women in young adulthood [12, 24], but these contributed data only to the mothers-daughters breast density analysis (Additional file 2: Table S3 and S4). These two studies [12, 24] were also the only ones not to use mammography to assess participants’ breast density. All but two studies adjusted for age, BMI and, if appropriate, also menopausal status at breast density assessment [21, 25].

Nine studies investigated associations between birth size measurements and adult breast density. The meta-analysis of their study-specific trend estimates, including those derived from the present study, revealed positive trends in the relative amount of fibroglandular tissue in the breast, as assessed by mammographic percent density or percent water, with birth size, albeit with high between-study heterogeneity (Fig. 4a and Table 4). Analysis by potential source of heterogeneity showed that this positive trend was stronger among studies assigned a high overall quality score, and in particular among those that relied on less error-prone birth size data from hospital records and more objective computer-assisted density measurements (Table 4). The positive trend in percent breast density with birth size was found in analyses restricted to study-specific estimates derived from pre- or post-menopausal women only, albeit more marked for the latter (Table 4). Similar results were observed when estimates derived from the present study were excluded. There was evidence of publication bias (p<0.001), which disappeared (p = 0.928) when the Tamimi study [26] was excluded (Additional file 3: Figure S4).

Seven studies reported on associations between other pre-natal exposures with adult breast density (Additional file 2: Table S4). Meta-analyses of study-specific estimates derived from five studies, including the present one, revealed a positive trend between maternal age and relative amount of fibroglandular tissue in the breast, which was stronger in analyses restricted to post-menopausal women, with relatively low between study-heterogeneity (Fig. 4b and Table 4). There was some indication of publication bias (p = 0.032) (Additional file 3: Figure S4). Meta-analyses were not possible for other pre-natal exposures, owing to the small number of studies and differences in the way these variables were measured or analysed (Additional file 2: Table S4), but qualitative assessment of the evidence did not reveal any consistent associations of percent breast density with maternal parity (based on n = 4 studies, including the present one), maternal smoking (n = 3) or alcohol intake (n = 2) during pregnancy, gestational age (n = 5), or placental weight (n = 2).

Strengths of the present study include the unique pre-birth cohort design with a wide range of prospectively collected pre-natal exposure data. Breast tissue measurements were collected from ionising radiation-free MRI examinations, making this the first study to examine the effect of pre-natal influences on breast tissue composition in young adulthood, prior to changes induced by pregnancies and breastfeeding. Objective (fully automated and, hence, observer-independent) volumetric breast tissue composition measurements were taken using a previously developed and evaluated approach [32].

The response rate was low (approximately 20 %), although comparable to a similar MRI breast study [12], but there was no evidence that the participants constituted a biased sample. Data were missing for some variables, but analyses of complete records and imputed datasets produced similar findings.

Meta-analysis of study-specific estimates derived from all eligible studies identified in the systematic review, together with those derived from the present study, revealed a significant positive trend between birth size and percent breast density, as assessed by mammographic percent density or percent water, albeit with marked between-study heterogeneity. Positive associations between birthweight and percent breast density were reported in studies based on computer-assisted methods to assess breast density [22, 26], but not in those that relied on radiologist-assessed categorical measurements (e.g., Wolfe’s [33] or Boyd’s [34] categories, American College of Radiology Breast Imaging Reporting and Data System (BI-RADS) [35]). Inconsistencies are likely due to the relatively small effect of birth size, which cannot be captured by categorical density classifications (Additional file 3: Figure S2). Analyses stratified by source-of-exposure data showed a significant positive trend between hospital-recorded birthweight measurements and percent breast density, but not in studies that used parental or adult self-reports, consistent with findings derived from a pooled analysis of birth size and breast cancer studies [1].

Birth size is a strong predictor of later physical development, with both weight and length at birth being associated with childhood growth, age at menarche [36] and adult body size [37]. Thus, the observed association between birth size and percent breast density may be mediated by childhood and adolescent growth trajectories [26]. Mutually adjusted analysis revealed that birthweight had the strongest independent association with percent water, suggesting that both linear growth and adiposity may affect breast tissue composition later in life. Furthermore, the association of birthweight with percent water did not appear to be mediated primarily via BMI at the time of MRI examination, suggesting that the effect is mainly independent of childhood and adolescent growth. There was some indication of a protective mediation effect of adult height in the association of birthweight with percent water, which may potentially reflect interactions between birth size, post-natal catch-up growth and pubertal development on percent water. Overall, dependent on the strong assumption of no unmeasured confounding, the mediation analysis provides strong evidence of a causal association between birthweight and breast tissue composition.

The observed associations with birth size and percent water mostly reflect associations with water volume, indicating that the amount of fibroglandular tissue in the breast may be set in utero. These findings parallel the effect of birth size on breast cancer risk and provide further support for the hypothesis that the intrauterine environment may play a role in determining both breast tissue composition and breast cancer risk [38, 39]. The pool of breast-specific stem cells, whose size is determined in utero [40], may be a critical factor linking pre-natal exposures to breast tissue composition and breast cancer risk in later life. In utero levels of growth factors (e.g., insulin-like growth factors) and hormones (e.g., sex hormones) are thought to act as mitogens, influencing both the pool size of breast-specific stem cells and possibly birthweight [41], with the former likely to be strongly correlated with the amount of fibroglandular tissue present in the fully developed breast [42].

Maternal height and mammographic density were also positively related to percent water, supporting previous evidence that percent breast density is a highly heritable trait. Twin studies have estimated that an additive genetic model explains 53–60 % of the variance in percent breast density [43, 44]. In the present study, the observed correlation between the participants’ percent water and their mothers’ mammographic percent density was r = 0.23 (Additional file 3: Figure S1), consistent with previous studies of mothers and daughters (r = 0.25) [12] and dizygotic twin pairs (r = 0.27) [45]. Maternal adult height is strongly correlated with daughters’ adult height (in this study, r = 0.49), and greater adult height is associated with a higher amount of fibroglandular tissue in the breast [46, 47]. Furthermore, adult height is positively associated with breast cancer risk, and genetic variants and biological pathways affecting adult height play an important role in the aetiology of breast cancer [30]. Maternal height modifies the effect of pregnancy hormones on birthweight [48], and thus it may also affect the risk of breast cancer in daughters through this mechanism.