Obesity as risk factor for subtypes of breast cancer: results from a prospective cohort study

EPIC is a multi-center prospective cohort study with more than 500,000 participants across Europe. In Germany, 53,088 participants (30,270 women) in the age range between 35 and 65 years were recruited at the study centers in the cities of Heidelberg and Potsdam between 1994 and 1998 [7, 8]. At baseline, anthropometric measurements were carried out by trained personnel, and data on diet, physical activity, smoking, alcohol consumption, medication use, reproductive factors and socio-economic status were obtained [7].

Incident cases of breast cancer were either self-reported during follow-up or derived from cancer registries. Each case was validated by a study physician using the information given by the patient’s treating physicians and hospitals. Overall, 1095 cases of primary breast cancer had occurred until Dec 31st 2010, the closure date for the present analyses. After exclusion of prevalent cases of cancer (n = 1669), individuals lost to follow-up (n = 947), individuals with unclear breast cancer status (n = 23), individuals with missing covariate information (n = 181), and incident cases without tumor blocks (n = 438) from the EPIC-Germany cohort, the study population for the present analyses comprised 27,012 women (Additional file 1: Figure S1).

Formalin-fixed paraffin-embedded (FFPE) tumor tissue material was available for a total of 657 cases (60.0%). There were no significant statistical differences regarding age, reproductive factors and lifestyle factors between these cases and those for which no tumor blocks were available, even though there were slightly more in situ and grade I tumors in the latter group (Additional file 2: Table S1). A board-certified senior pathologist (E.H.) selected representative tumor areas to construct tissue microarrays (TMA) on a hematoxylin and eosin stained slide of each tumor block. A TMA machine (AlphaMetrix Biotech, Roedermark, Germany) was used to extract tandem 1 mm cylindrical core samples. IHC staining was carried out using antibodies routinely employed for diagnostic purposes (Additional file 2: Table S2) and an immunostaining device (DAKO, Techmate 500plus). All TMA slides were examined by at least one pathologist (E.H., M.K.) with special expertise in breast cancer pathology. In case of a discrepancy between the scores derived from the first and second core of the same patient, the pathologists re-examined both cores and made a final decision. Whenever TMA analysis did not yield a conclusive result for a marker, it was assigned a missing value (ER: 2.0%; PR: 2.7%; HER2: 1.7%; Ki67: 6.1%; Bcl-2: 4.1%; p53: 6.7%). Tumors were categorized as ER positive/negative and PR positive/negative using the Allred Score [9]. HER2 was determined according to staining pattern and intensity, and scored as negative (0 and 1+) or positive (2+ and 3+) [10]. Ki67 proliferation activity was scored by percentage of positive tumor nuclei (< 20%: low proliferative activity; ≥20%: high proliferative activity) [11]. Bcl-2 was scored as negative if less than 10% of the cells were positive and staining intensity was weak, otherwise Bcl-2 was scored as positive [12]. Cases with more than 10% of cells stained were rated p53 positive, the remaining cases were rated p53 negative, as in most previous studies using this antigen [13]. Categorization of subtypes was based on visual estimation counting at least 100 tumor cells.

Relationships between BMI at recruitment and breast cancer risk were evaluated separately among 1) women, who were pre- or perimenopausal at baseline 2) women, who were postmenopausal at baseline and used hormone therapy (HT), and 3) women, who were postmenopausal at baseline and did not use HT, as differential risk associations with BMI across these subgroups have been reported [14, 15]. Statistical analyses on breast cancer risk by tumor subtype were carried out using multivariable Cox proportional hazards regression analyses to estimate hazard ratios (HR) and 95% confidence intervals (CI) across tertiles of BMI (created based on data of the full cohort), with age as the underlying time scale. All models were adjusted for height (continuous), number of full-term pregnancies (continuous), educational level (university degree vs. no university degree), smoking status (never, former, current), and study center (Heidelberg, Potsdam). Analyses among pre- and perimenopausal women were further adjusted for current use of oral contraceptives. The inclusion of other potential confounders (alcohol consumption, breast feeding, age at menarche, age at first pregnancy) only marginally affected risk associations and were not included in final Cox regression models.

Linear trends were estimated by entering BMI as a continuous term into the same model rescaling HRs to reflect a 5 kg/m² increase. Observations were left-truncated and censored at end of follow-up, death, or cancer diagnosis, whichever occurred first. In order to assess patterns of IHC markers, unsupervised hierarchical clustering was used to group cancer cases according to the similarity / dissimilarity of the IHC staining results for ER, PR, HER2, Ki67, Bcl-2, and p53, as previously published [16, 17]. In addition to BMI, we evaluated waist circumference and hip circumference as anthropometric markers of obesity in relation to breast cancer risk. Heterogeneity in associations between anthropometric factors and breast cancer risk across subtypes was tested for using a competing risk framework, as proposed by Wang et al. [18]. As the evidence on associations between BMI and in situ breast tumors is not consistent [19, 20], we decided to exclude cases of in situ tumors in sensitivity analyses. All statistical analyses were carried out using SAS, version 9.4 (SAS Institute, Cary, NC, USA). For unsupervised hierarchical clustering and for the generation of a dendogram / heat map to visualize clusters of tumor markers we used the d3heatmap package in R [21].

The analytical cohort for the present analyses comprised 27,012 women at a median baseline age of 48.4 (range: 35.2–65.2) years, and a median BMI of 24.7 (see Table 1, and Additional file 1: Figure S1). Overall, 40.8% of the women were postmenopausal at baseline. Among the postmenopausal women, 46.0% reported to use HT. The average follow-up duration was 13.0 (±3.1) years. Median age at diagnosis among the 657 breast cancer cases was 60.2 (range: 38.9–78.6) years.

Tumor stages and grades at diagnosis were as follows; In situ: 7.0%, Stage I: 38.7%, Stage II: 41.0%, Stage III: 11.3%, Stage IV: 2.0%; Grade I: 12.4%, Grade II: 56.8%, Grade III: 30.8% (Additional file 2: Table S1). Of the invasive tumors, 70.5% were carcinoma of no special type (NST), 18.3% lobular carcinoma, and 11.1% other; of the in situ tumors, 67.4% were ductal carcinoma, 13.0% were lobular carcinoma, and 19.6% other (Additional file 2: Table S3). The proportions of subtypes indicating more favorable prognosis were 84.8% for ER+, 70.7% for PR+, 87.5% for HER2-, 83.1% for Ki67_low, 66.0% for Bcl-2+ and 80.1% for p53-. Frequencies of luminal A (ER+ and/or PR+, HER2- and Ki67_low), luminal B (ER+ and/or PR+, HER2- and Ki67_high), Her2+, and triple negative (ER-, PR-, and HER2-) tumors were 68.6, 8.4, 9.7, and 13.3%.

The results of the unsupervised hierarchical clustering of breast cancer cases according to IHC staining profiles are shown in Fig. 1. The three main clusters identified by hierarchical clustering can be characterized as follows: Cluster 1 (42.7% of all cases) contains tumors with a profile of individual markers indicative of low aggressiveness (all cases are ER+, PR+, HER2-, Ki67_low, Bcl-2+ and p53-). Cluster 2 (19.0% of all cases) contains ER- tumors and ER+ tumors that are Bcl-2 negative. Cluster 3 (38.3% of all cases) mainly contains ER+ tumors that, unlike the ER+ tumors in cluster 1, show at least one criterion pointing to higher aggressiveness (i.e. p53 positivity, Bcl-2 negativity, high Ki67 expression, or HER2 positivity).

Among postmenopausal non-users of HT, BMI was directly associated with higher overall breast cancer risk (HR per 5 kg/m²: 1.27 [95% CI: 1.07, 1.50], p = 0.005), while a significant inverse association was observed among HT users (HR: 0.80 [0.66, 0.98], p = 0.024) (Table 2). BMI was not significantly associated with overall breast cancer risk in pre- and perimenopausal women (HR: 0.98 [0.85, 1.12], p = 0.72).

Analyses stratified by tumor subtypes as derived from hierarchical clustering are shown in Table 3. Among postmenopausal non-users of HT, each 5 kg/m² increment of BMI was directly and significantly associated with the risk of less aggressive cluster 1 tumors, i.e. tumors that were ER+, PR+, HER2-, Ki67_low, Bcl-2+ and p53-, with a HR per 5 kg/m² of 1.44 [95% CI: 1.10, 1.90], p = 0.009). BMI was not associated with more aggressive cluster 2 and cluster 3 tumors (Table 3). Among HT-users, BMI was significantly associated with lower risk of less aggressive cluster 1 tumors (HR per 5 kg/m²: 0.68 [0.50, 0.94], p = 0.018); again, no significant associations with the risks of more aggressive cluster 2 and cluster 3 tumors were observed. While risk analyses per 5 kg/m² did not reveal significant associations between BMI and risks of any tumor subtype in pre- and perimenopausal women, it is of note that women in the highest BMI tertile showed a significantly lower risk of less aggressive cluster 1 tumors as compared to women in the lowest BMI tertile (HR_{Tertile3 vs. Tertile1}: 0.55 [0.33, 0.93]). Sensitivity analyses excluding in situ cases yielded similar highly similar results (Additional file 2: Table S4). Associations between BMI and risk of luminal A tumors were similar to those between BMI and risk of cluster 1 tumors (Additional file 2: Table S5); there were no significant associations with luminal B and triple negative tumors.

In analyses on breast tumor subtypes defined by individual markers, BMI was significantly positively associated with risk of ER+, PR+, HER2-, Ki67_low, Bcl-2+ and p53- tumors among postmenopausal non-users of HT (Additional file 2: Table S6). By contrast, no significant associations with ER-, PR-, HER2+, Ki67_high, Bcl-2- and p53+ tumors were observed. With respect to postmenopausal users of HT, Cox regression analyses showed significant inverse associations with risks of ER+, HER2-, Ki67_low, Bcl-2+ and p53- tumors, and a non-significant tendency for an inverse association with PR+ breast cancer (Additional file 2: Table S7). Again, there were no significant associations with risk of ER-, PR-, HER2+, Ki67_high, Bcl-2- and p53+ tumors. Among pre- and perimenopausal women, BMI was not significantly associated with risks of any tumor subtype defined by individual markers (Additional file 2: Table S8). The results on BMI and risks of tumor subtypes defined by individual markers were similar after exclusion of in situ cases (see Additional file 2: Table S9, Table S10, and Table S11).

The directions of associations with risk of tumor subtypes were highly similar when using waist and hip circumference as anthropometric indices of obesity instead of BMI, while the associations between waist-to-hip ratio and breast cancer risk were weaker and non-significant (data not shown). Risk associations among premenopausal women only were very similar as the presented associations among peri- and premenopausal women (data not shown). Importantly, no formal heterogeneity of associations between anthropometric factors and breast cancer risk across tumor subtypes, as either derived from hierarchical clustering or defined by individual IHC markers, was observed.