Abstract
Purpose
There are no models for German women that predict absolute risk of invasive breast cancer (BC), i.e., the probability of developing BC over a prespecified time period, given a woman’s age and characteristics, while accounting for competing risks. We thus validated two absolute BC risk models (BCRAT, BCRmod) developed for US women in German women. BCRAT uses a woman’s medical, reproductive, and BC family history; BCRmod adds modifiable risk factors (body mass index, hormone replacement therapy and alcohol use).
Methods
We assessed model calibration by comparing observed BC numbers (O) to expected numbers (E) computed from BCRmod/BCRAT for German women enrolled in the prospective European Prospective Investigation into Cancer and Nutrition (EPIC), and after updating the models with German BC incidence/competing mortality rates. We also compared 1-year BC risk predicted for all German women using the German Health Interview and Examination Survey for Adults (DEGS) with overall German BC incidence. Discriminatory performance was quantified by the area under the receiver operator characteristics curve (AUC).
Results
Among 22,098 EPIC-Germany women aged 40+ years, 745 BCs occurred (median follow-up: 11.9 years). Both models had good calibration for total follow-up, EBCRmod/O = 1.08 (95% confidence interval: 0.95–1.21), and EBCRAT/O = 0.99(0.87–1.11), and over 5 years. Compared to German BC incidence rates, both models somewhat over-estimated 1-year risk for women aged 55+ and 70+ years. For total follow-up, AUCBCRmod = 0.61(0.58–0.63) and AUCBCRAT = 0.58(0.56–0.61), with similar values for 5-year follow-up.
Conclusion
US BC risk models showed adequate calibration in German women. Discriminatory performance was comparable to that in US women. These models thus could be applied for risk prediction in German women.
Keywords: Absolute risk, Breast cancer incidence, Model transportability
Background
Breast cancer is one of the most common malignancies among women worldwide, including Germany [1, 2]. In developed countries, incidence has been increasing, due to increasing prevalence of several important breast cancer risk factors, including obesity, later ages at first life birth, and lower number of births [3]. To inform public health policy decisions, it is important to quantify breast cancer risk in a given population; it is also desirable to identify women at highest breast cancer risk for targeting prevention efforts. No model has been designed and validated for predicting individualized risk of breast cancer for women in the general German population. We therefore assessed the ability of two models developed at the US National Cancer Institute (NCI) for women in the general US population to predict the absolute risk of developing breast cancer in German women.
The first model, the “Gail model” [4], publicly available in a modified version as the Breast Cancer Risk Assessment Tool (BCRAT; http://www.cancer.gov/bcrisktool) has been validated extensively [5-9]. This model predicts a woman’s risk of developing invasive breast cancer over a specified time period (5 years or until age 90, i.e., lifetime), given her age, reproductive history and family history, and her past history of biopsy and diagnosis of benign hyperplasia. The probability of developing breast cancer is computed by combining relative risk estimates for a woman’s risk factors with attributable risk estimated from the Breast Cancer Detection Demonstration Project, age-specific breast cancer incidence rates, and competing mortality rates from all other causes from the Surveillance, Epidemiology, and End Results (SEER) program for the years 1983 to 1987. The second, more recently developed absolute breast cancer risk model, BCRmod [10], incorporates most of the risk factors included in BCRAT, and additionally includes potentially modifiable risk factors, namely body mass index (BMI), hormone replacement therapy (HRT) use, and alcohol consumption. BCRmod showed good calibration in one independent validation study in US women and might have better calibration than BCRAT in populations in which the distribution of these modifiable factors differs from that in US women [10].
Relative risks for many breast cancer risk factors do not vary greatly between white women in the US and in Europe [11-16]. Breast cancer incidence standardized to the world standard population, aged 20 to 85+, is also similar in the US and Germany (142.4 in Germany, 141.4 for the US for all racial groups combined) [17]. It is therefore reasonable to evaluate how well US breast cancer risk models can predict risk in German women.
We used data from the prospective European Prospective Investigation into Cancer and Nutrition (EPIC)-Germany cohort to assess calibration (i.e., bias) and discriminatory accuracy of the BCRAT and BCRmod models. We further assessed calibration by comparing 1-year breast cancer risk predicted from the models for all German women aged 40–70 based on the German Health Interview and Examination Survey for Adults (DEGS), a representative population-based survey, with overall German BC incidence. To allow for possible population differences, we also assessed the models’ performance after updating them with German breast cancer incidence and competing mortality rates. If well calibrated, these models can be recommended for individual and public health applications in German women.
Methods
Study populations
Epic cohort
We used data on 27,934 women aged 20 to 70 years from the two German centers (Heidelberg, HD, and Potsdam, PD) of EPIC, the European Prospective Investigation into Cancer and Nutrition. Details on EPIC are given elsewhere [18, 19]. EPIC participants were enrolled from 1994 to 1998, and at baseline completed a self-administered questionnaire and gave a computer-guided in-person interview. Anthropometric measures were taken in a physical examination. Incident invasive breast cancers were ascertained through follow-up questionnaires and cancer registry linkages. All cancer diagnoses were verified with hospital and pathology records (through 12/2009 for HD; through 7/2010 for PD). Information on vital status was obtained from follow-up questionnaires and linkages to municipal population registries.
The follow-up questionnaires included questions on anthropometric measures, alcohol and tobacco consumption, and current medication use. Participants were also asked about operations or newly diagnosed diseases since last follow-up, and dates of new diagnoses, which were verified using medical records.
From the baseline questionnaire, we used age at menarche, menopausal status (pre-, post-, perimenopausal/unknown), number of children (0,1, 2+), age at first full-term pregnancy, current and former use of hormone replacement therapy (HRT) and oral contraceptives, and alcohol consumption. Body mass index (BMI) was calculated from standardized measurements of height and weight. For HD, HRT type was obtained at baseline; information on number of 1st degree relatives with breast cancer and a previous diagnosis of benign breast disease were only available from a follow-up questionnaire that was answered on average 7 years after recruitment. As very few women (< 1% among women aged 40+ in EPIC) had 2 + first degree relatives with breast cancer, we combined them with the 1 affected relative group for imputation and analysis. For PD, information on breast cancer family history was collected at baseline, and HRT type at the first follow-up (around 3 years after recruitment). Information on a previous diagnosis of benign breast disease was not collected for PD. Information on the number of breast biopsies and the diagnosis of atypical hyperplasia (predictors used in BCRAT) were not ascertained in EPIC-Germany. We thus set the number of biopsies to one for all women who reported a diagnosis of benign breast disease.
We used multiple imputation to complete missing data as described below in the statistical section.
German Health Interview and Examination Survey for Adults (DEGS)
To facilitate comparing estimated risks with German breast cancer incidence, we used data on 3,705 women participating in the German Health Interview and Examination Survey for Adults (DEGS), 2008–2011 [20]. Information on age at first live birth, type and duration of HRT use, past diagnosis of benign breast disease, and family history of breast cancer was not collected in DEGS. We thus imputed these variables based on information from EPIC, as described in the statistical section.
German incidence and mortality data
After the initial validation, we updated the risk models with age-specific breast cancer incidence rates and competing mortality rates for German women [21] (Supplemental Table S1). We used rates from 1999 for EPIC and from 2012 for DEGS. Breast cancer mortality was subtracted from the overall mortality rates to obtain competing mortality.
Statistical methods
BCRmod was originally developed to predict risk in women aged 50+. However, to inform breast cancer screening decisions among 40–50-year-old women using risk estimates [22], we extended the age range using SEER incidence and competing mortality rates for 40–49-year-old women.
Imputation of missing values
We imputed missing values with chained regression models using IVEware [23] for number of life births, age at first full-term pregnancy (years), and age at menarche (years), separately for women aged 20–50 years and 50+ years: type of HRT (E, EP, or other), duration of HRT use (years), benign breast disease (yes/no), family history of breast cancer (yes/no), and alcohol consumption [g/d; continuous with a point mass at zero, corresponding to the proportion of non-drinkers]. Missing age at menopause (years) was imputed for women aged 50+ if they were postmenopausal, and if they reported unknown menopausal or perimenopausal status and their imputed age at menopause was higher than their age at recruitment. HRT use had largest amount of missingness in EPIC, 42%, among postmenopausal women; percent missingness for all variables is shown in Supplemental Table S2.
The imputation models included the log-transformed follow-up time and BC status [24]. We first created 5 imputed data sets for HD and PD separately, then combined them and imputed information on benign breast disease (not collected in PD) for PD using all available predictors. To impute completely missing variables in DEGS, we combined DEGS with EPIC, and additionally included the sampling weights and cluster variables in the imputation models.
Imputed data sets were analyzed separately and then estimates and variances were combined using PROC MIANALYZE (Inc. SI. SAS 9.4. Cary, NC2011).
Association parameters
Relative risks (RRs) and 95% confidence intervals (CIs) for the association between breast cancer and the model variables were calculated using multivariable Cox proportional hazards models with baseline hazard rates stratified by center and by age at enrollment (< 50 years; ≥ 50 years) for women 40+ years (PROC PHREG, SAS 9.4). Follow-up started at age at enrollment and ended at BC diagnosis or censoring. Censoring events were the first of death, loss to follow-up, or administrative censoring (31 December 2009 for HD; 31 July 2010 for PD).
Assessment of model performance
To assess model calibration, i.e., bias in the predictions, we computed absolute BC risk estimates for each woman given her age and risk factors at baseline for two projection periods, for 5 years and over the whole follow-up period, using SAS programs available at https://dceg.cancer.gov/tools/risk-assessment/bcrasasmacro and https://dceg.cancer.gov/tools/risk-assessment/br_en_oc_ram, for BCRAT and BCRmod respectively. Expected number of events (E), obtained by summing the risk estimates, were compared to observed numbers of events (O), overall or in subgroups defined by covariates or risk deciles in the population based on E/O using χ2 tests or the Hosmer–Lemeshow test [25]. As a measure of discrimination, we calculated the area under the receiver operator characteristics curve (AUC), stratified by 5-year groups of age at enrollment. To obtain variances for the AUC, we directly applied Rubin’s rule to separate results from each imputed dataset.
For women in the DEGS survey, we estimated 1-year absolute BC risk from the models and calculated weighted average values E/O with CIs accounting for the sampling design and multiple imputation (PROC SURVEYMEANS, PROC MIANALYZE). We compared these weighted projected age-specific incidence estimates visually to age-specific German BC incidence rates.
Results
Model validation in the EPIC cohort
Table 1 shows the characteristics of women aged 40+ and 50+ years and the distribution of the model risk factors in EPIC and DEGS (percentages weighted to represent the German female population). Characteristics of all women for EPIC-Germany and by center are in Supplemental Tables S2 and S3, respectively. In EPIC, 48% of the women were in the normal weight category (BMI < 25), but only 37% of all German women based on DEGS. The majority of women were parous (86% in EPIC and 82% of all German women). In EPIC, 7% of women reported a family history of breast cancer, and 4% a diagnosis of benign breast disease. Supplemental Table S3 shows the distributions of risk factors after imputation. During a median follow-up of 11.9 years, 745 incident BC cases were identified in EPIC (423 for HD; 322 for PD).
Table 1.
Characteristics and risk factor distribution in women at least 40 or at least 50 years old in EPIC-Germany and DEGS with sample sizes shown for the survey, and weighted percentages representative for the German female population
| n (%), median (min; max) | EPIC-Germany, women aged 40+ |
EPIC-Germany, women aged 50+ |
DEGS women aged 40+ | DEGS women aged 50+ |
|---|---|---|---|---|
| Study participants | 22,098 (100%) | 12,894 (100%) | 2,727 (100%) | 2,031 (100%) |
| Incident breast cancer cases | 745 (3%) | 508 (4%) | ||
| Age at recruitment (years) | 52.61 (40.00; 70.15) | 57.61 (50.00; 70.15) | 51.87 (40; 75) | 58.46 (50; 75) |
| Total follow-up time (years) | 11.86 (0.58; 15.53) | 11.84 (0.58; 15.53) | – | – |
| Age at diagnosis (cases only, years) | 60.83 (42.43; 77.52) | 63.82 (50.76; 77.52) | – | – |
| BMI | ||||
| Missing | – | – | 24 (1%) | 21 (1%) |
| BMI < 25 | 10,588 (48%) | 5,174 (40%) | 1,011 (37%) | 623 (30%) |
| BMI 25 ≤ 30 | 7,559 (34%) | 4,968 (39%) | 928 (33%) | 752 (35%) |
| BMI 30 ≤ 35 | 2,879 (13%) | 2,038 (16%) | 500 (19%) | 411 (22%) |
| BMI 35+ | 1,072 (5%) | 714 (6%) | 264 (10%) | 224 (12%) |
| Age at menarche | ||||
| Missing | 54 (0%) | 41 (0%) | 2,727 (100%) | 2,031 (100%) |
| 14+ | 9,194 (42%) | 6,303 (49%) | – | – |
| 12–13 | 10,465 (47%) | 5,478 (42%) | – | – |
| < 12 | 2,385 (11%) | 1,072 (8%) | – | – |
| Benign breast diseaseb | ||||
| Missing | 12,996 (59%) | 7,744 (60%) | 2,727 (100%) | 2,031 (100%) |
| No | 8,166 (37%) | 4,630 (36%) | – | – |
| Yes | 936 (4%) | 520 (4%) | – | – |
| Family history of breast cancerb | ||||
| Missing | 3,866 (17%) | 2,503 (19%) | 2,727 (100%) | 2,031 (100%) |
| 0 | 16,690 (76%) | 9,493 (74%) | – | – |
| 1 | 1,475 (7%) | 851 (7%) | – | – |
| 2 | 67 (0%) | 47 (0%) | – | – |
| Alcohol use (lifetime, daily) | ||||
| Missing | 10 (0%) | 4 (0%) | 34 (1%) | 28 (2%) |
| None | 346 (2%) | 229 (2%) | 450 (19%) | 355 (20%) |
| < 14 g/d | 19,243 (87%) | 11,309 (88%) | 2,031 (72%) | 1,484 (71%) |
| ≥ 14 g/d | 2,499 (11%) | 1,352 (10%) | 212 (8%) | 164 (8%) |
| Ever full-term pregnancy | ||||
| Missing | 64 (0%) | 46 (0%) | 183 (8%) | 139 (8%) |
| No | 2,991 (14%) | 1,656 (13%) | 289 (11%) | 197 (10%) |
| Yes | 19,043 (86%) | 11,192 (87%) | 2,255 (82%) | 1,695 (82%) |
| Among parous women only | ||||
| Number of children | ||||
| 1 child | 5,628 (30%) | 3,169 (28%) | – | – |
| 2 children | 9,375 (49%) | 5,175 (46%) | – | – |
| 3+ | 4,040 (21%) | 2,848 (25%) | – | – |
| Age at first child | ||||
| Missing | 12 (0%) | 9 (0%) | 2,255 (100%) | 1,695 (100%) |
| < 25 | 11,326 (59%) | 6,723 (60%) | – | – |
| 25 ≤ 30 | 5,399 (28%) | 3,275 (29%) | – | – |
| 30+ | 2,306 (12%) | 1,185 (11%) | – | – |
| Premenopausal | 7,192 (33%) | 479 (4%) | 701 (29%) | 172 (9%) |
| Perimenopausal/unknown | 3,534 (16%) | 1,718 (13%) | 959 (33%) | 951 (47%) |
| Postmenopausal | 11,372 (51%) | 10,697 (83%) | 1,067 (38%) | 908 (45%) |
| Among postmenopausal women only | ||||
| Age at menopause | ||||
| Missing | 2,158 (19%) | 1,733 (16%) | 404 (38%) | 320 (36%) |
| < 49 | 4,484 (39%) | 4,234 (40%) | 199 (19%) | 151 (16%) |
| 49 ≤ 55 | 4,014 (35%) | 4,014 (38%) | 190 (18%) | 163 (18%) |
| 55+ | 716 (6%) | 716 (7%) | 274 (24%) | 274 (30%) |
| HRT use and typea | ||||
| Missing | 4,759 (42%) | 4,612 (43%) | 12 (1%) | 4 (0%) |
| 0 or other not-EP type | 3,887 (34%) | 3,654 (34%) | 723 (69%) | 594 (66%) |
| < 10 years EP | 2,330 (20%) | 2,048 (19%) | – | – |
| 10 + years EP | 396 (3%) | 383 (4%) | – | – |
| Other HRT | ||||
| Missing | 7,239 (64%) | 6,944 (65%) | 1,067 (100%) | 908 (100%) |
| No | 2,927 (26%) | 2,615 (24%) | – | – |
| Yes | 1,206 (11%) | 1,138 (11%) | – | – |
In Potsdam women were asked at recruitment if they used HRT and how long. No detailed information on type of HRT use was derived. In Heidelberg all current medication was recorded at recruitment, which allowed definition of HRT type
As of follow-up 2, approx 7 years after recruitment
Relative risk estimates for all the factors in the respective models are given in Table 2. In EPIC, there was no association of BC risk with BMI (RR = 1.0, Table 2), with nulliparity (RR = 1.03, 95% CI 0.82–1.28), age at first life birth (RR = 1.02, 95% CI 0.91–1.14), and age at menopause (RR = 1.02, 95% CI 0.86–1.21). The RR for estrogen plus progestin HRT use was similar to that in BCRmod, but associations were stronger for “other HRT use” (RR = 1.58 (1.22–2.04)) and BC family history (RR = 1.64, 95% CI 1.26–2.13). Relative risk estimates did not differ when estimated from models fit separately to women born before 1939 (median birth year) or 1939 and later (data not shown).
Table 2.
Multivariate relative risk estimates for breast cancer among women aged 40 + in the EPIC-Germany study for variables in BCRAT and BCRmod using imputed datasets
| BCRmod-risk factors | Relative risks (95% confidence intervals) |
|
|---|---|---|
| EPIC (40 +) | BCRmod | |
| Body mass index (BMI) | ||
| < 25 kg/m2 | 1.0 (referent) | 1.0 (referent) |
| Per category increase | 1.00 (0.92–1.10) | 1.11 (1.07–1.16) |
| Estrogen plus progestin | ||
| No HRT use | 1.0 (referent) | 1.0 (referent) |
| Per category increase | 1.50 (1.29–1.75) | 1.53 (1.37–1.70) |
| Other HRT use | ||
| No | 1.0 (referent) | 1.0 (referent) |
| Yes | 1.58 (1.22–2.04) | 1.07 (0.98–1.17) |
| Age at first live birth | ||
| < 25 years | 1.0 (referent) | 1.0 (referent) |
| Per category increase | 1.02 (0.91–1.15) | 1.14 (1.08–1.20) |
| Parity | ||
| 1+ children | 1.0 (referent) | 1.0 (referent) |
| Nulliparous | 1.03 (0.83–1.28) | 1.31 (1.13–1.52) |
| Age at menopause | ||
| < 50 | 1.0 (referent) | 1.0 (referent) |
| Per category increase | 1.03 (0.87–1.23) | 1.20 (1.12–1.29) |
| Premenopausal | 1.35 (0.97–1.88) | 1.54 (1.32–1.80) |
| Benign breast disease/biopsy | ||
| No | 1.0 (referent) | 1.0 (referent) |
| Yes | 1.76 (1.33–2.35) | 1.34 (1.24–1.44) |
| Family history of breast or ovarian cancer | ||
| No | 1.0 (referent) | 1.0 (referent) |
| Yes | 1.64 (1.26–2.13) | 1.34 (1.22–1.47) |
| Alcohol | ||
| 0 drinks per day | 1.0 (referent) | 1.0 (referent) |
| Per category increase | 1.14 (0.93–1.39) | 1.06 (1.00–1.12) |
| BCRAT-risk factors | Relative risks (95% confidence intervals) |
|
| EPIC (40 +) | BCRAT | |
| Age at menarche | ||
| 7–11 | 0.92 (0.71–1.19) | 1.207 |
| 12–13 | 0.99 (0.85–1.16) | 1.099 |
| 14+ | 1.0 (referent) | 1.0 (referent) |
| Age 1st FTP * BC-fam. history | ||
| < 20, and none | 1.0 (referent) | 1.0 (referent) |
| < 20, and (≥) 1 1st deg. relative with breast cancer | 1.95 (0.88–4.31) | 2.607 |
| ≥ 2 1st deg. relative with breast cancer | 6.789 | |
| 20–24, and none | 1.07 (0.80–1.44) | 1.24 |
| 20–24, and (≥) 1 1st deg. relative with breast cancer | 1.77 (1.10–2.86) | 2.681 |
| ≥ 2 1st deg. relative with breast cancer | 5.775 | |
| 25–29 or nullip., and none | 1.06 (0.79–1.43) | 1.548 |
| 25–29 or nullip., and (≥) 1 1st deg. relative with bc | 1.88 (1.18–3.00) | 2.756 |
| ≥ 2 1st deg. relative with breast cancer | 4.907 | |
| 30 + , and none | 1.20 (0.84–1.72) | 1.927 |
| 30 + , and (≥) 1 1st deg. relative with breast cancer | 1.19 (0.50–2.83) | 2.834 |
| ≥ 2 1st deg. relative with breast cancer | 4.169 | |
| History of benign breast disease, used as | ||
| (≥) 1 biopsy (in women under 50) | 1.85 (1.23–2.78) | 1.698 |
| ≥ 2 biopsies | 2.882 | |
| (≥) 1 biopsy (in women 50 +) | 1.80 (1.31–2.48) | 1.273 |
| ≥ 2 biopsies | 1.620 | |
Categories for variables fitted with a trend are given below the table. The baseline hazard for EPIC-Germany was stratified by center and by age </>50 years for BCRmod, and only by center for BCRAT
Models were fit with baseline stratified by center. All variables are coded so that the lowest risk category is the reference category. Categories for: BMI: < 25, 25– < 30, 30– < 35, and 35 + ; menopausal hormone therapy use: 0, 1–9, and 10 + years; estrogen plus progestin HRT use: 0, 1–9, and 10+ years; other/unknown HRT use: no, yes; age at first birth: < 25, 25–29, and 30+ years; age at menopause: < 50, 50–54, 55+years, missing; benign breast diseases/biopsy: no, yes, missing; alcohol consumption: 0, < 1, 1+ drink/day, age at menarche: 14 + , 12–13, 7–11
We were unable to estimate RRs corresponding to the BCRAT parameterization, as we did not have detailed information on family history. Surprisingly, even though we misclassified women with 2 or more affected relatives as having 1 affected relative, all RRs for family history were lower in EPIC for women with one affected first degree relative than in BCRAT. All other RRs were also attenuated in EPIC. The only exception were the RRs for number of biopsies, with RR = 1.85 (95% CI 1.23–2.78) and RR = 1.80 (95% CI 1.31–2.48) for women with ≥ 1 biopsy in the ≤ 50 year of age and > 50 years of age categories, respectively. This increased effect is likely due to misclassification of that variable in EPIC; as we did not have information on biopsies, we assigned all women with a history of benign breast disease to the ≥ 1 biopsy group, which misclassified all women with 2 or more biopsies (Table 2).
There was no significant lack of fit for the 5-year BCRmod estimates, with E/O = 1.05 (95% CI 0.80–1.30) for women aged 40–49 and 1.09 (95% CI 0.94–1.25) for women aged 50+ (Table 3). When US incidence and competing mortality rates were replaced by German rates, the model significantly underestimated true 5-year risk by about 20% (E/O = 0.86; 95% CI 0.66–1.07) and 0.78 (95% CI 0.67–0.89) for women aged 40–49 and 50+ years, respectively). For the total follow-up period, BCRmod underestimated true risk with E/O = 0.83 (95% CI 0.72–0.93) and 0.97 (95% CI 0.88–1.05) for women aged 40–49 and 50+, respectively, when US rates were used in the model. This bias was stronger and statistically significant when German rates were used in BCRmod (Table 3).
Table 3.
Summary calibration and discriminatory measures for BCRmod and BCRAT in the German EPIC cohort
| 5-year risk and follow-up |
Total follow-up |
||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Observed cases | E/O | 95% CI | AUC | 95% CI | Observed cases | E/O | 95% CI | AUC | 95% CI | ||
| BCRmod (US rates/including younger women) | All ages | 256 | 1.08 | (0.95–1.21) | 0.58 | (0.54–0.62) | 745 | 0.92 | (0.86–0.99) | 0.61 | (0.58–0.63) |
| 40–49 | 68 | 1.05 | (0.80–1.30) | 0.61 | (0.51–0.70) | 237 | 0.83 | (0.72–0.93) | 0.60 | (0.56–0.64) | |
| 50+ | 188 | 1.09 | (0.94–1.25) | 0.57 | (0.52–0.62) | 508 | 0.97 | (0.88–1.05) | 0.61 | (0.58–0.64) | |
| BCRmod with German rates from 1999 | All ages | 256 | 0.80 | (0.70–0.90) | 0.58 | (0.54–0.62) | 745 | 0.66 | (0.61–0.71) | 0.61 | (0.59–0.63) |
| 40–49 | 68 | 0.86 | (0.66–1.07) | 0.60 | (0.52–0.69) | 237 | 0.67 | (0.58–0.75) | 0.60 | (0.56–0.64) | |
| 50+ | 188 | 0.78 | (0.67–0.89) | 0.57 | (0.53–0.62) | 508 | 0.65 | (0.60–0.71) | 0.61 | (0.58–0.64) | |
| BCRAT (“Gail model”, US rates) | All ages | 256 | 0.99 | (0.87–1.11) | 0.56 | (0.52–0.60) | 745 | 0.84 | (0.78–0.90) | 0.58 | (0.56–0.61) |
| 40–49 | 68 | 1.20 | (0.91–1.48) | 0.60 | (0.53–0.68) | 237 | 0.89 | (0.78–1.00) | 0.61 | (0.57–0.65) | |
| 50+ | 188 | 0.96 | (0.83–1.10) | 0.55 | (0.50–0.60) | 508 | 0.86 | (0.78–0.93) | 0.58 | (0.55–0.61) | |
| BCRAT with German rates from 1999 | All ages | 256 | 0.86 | (0.76–0.97) | 0.56 | (0.52–0.60) | 745 | 0.72 | (0.66–0.77) | 0.58 | (0.56–0.60) |
| 40–49 | 68 | 1.10 | (0.84–1.36) | 0.60 | (0.52–0.68) | 237 | 0.84 | (0.73–0.95) | 0.61 | (0.57–0.65) | |
| 50+ | 188 | 0.83 | (0.71–0.94) | 0.55 | (0.50–0.60) | 508 | 0.69 | (0.63–0.75) | 0.58 | (0.55–0.61) | |
The AUC-statistic was computed as weighted average over 5-year age groups
The AUCs for BCRmod were 0.61 (95% CI 0.51–0.70) and 0.60 (95% CI 0.56–0.64) for 40–49-year-old women for 5-year risks and risk computed for the total follow-up time, respectively, and 0.57 (95% CI 0.52–0.62) and 0.61 (95% CI 0.58–0.64) for 50+-year-old women for 5-year and total follow-up time, respectively.
Similar patterns were observed for BCRAT. There was no significant lack of fit in 5-year risk with E/O = 1.20 (95% CI 0.91–1.48) and 0.96 (0.83–1.10) for women aged 40–49 and 50+, respectively. When German rates were used in BCRAT, true risk was significantly underestimated, with E/O = 0.83 (95% CI 0.71–0.94) women aged 50+, but still slightly overestimated for women aged 40–49 (E/O = 1.10 (95% CI 0.84–1.36)). When projections were made for the whole follow-up time, E/O = 0.89 (95% CI 0.78–1.00) and 0.86 (95% CI 0.78–0.93) for 40–49- and 50+-year-old women, with lower estimates when German rates were used in BCRAT.
The AUCs for BCRAT were 0.60 (95% CI 0.53–0.68) and 0.61 (95% CI 0.57–0.65) for 40–49-year-old women for 5-year risks and risk computed for the total follow-up time, respectively, and 0.55 (95% CI 0.50–0.60) and 0.58 (95% CI 0.55–0.61) for 50+-year-old women for 5-year and total follow-up time risk estimates.
Supplemental Table S4 shows the risk stratification of both models using cutoff values 1.66 and 2.5% and cross classified women based on the two risk models. These categories were partly based on the risk threshold 1.66% recommended by the American Society for Clinical Oncology (ASCO, [26]). BCRmod identified more women at elevated risk, including more cases than BCRAT. However, both models slightly underestimated risk in the highest category (mean BCRmod and BCRAT risks in the ≥ 2.5% risk category were 2.4, slightly lower than 2.5%).
Using German rates did not noticeably impact the discriminatory accuracy of either BCRmod or BCRAT.
Figure 1 shows calibration in risk deciles for 5-year risk estimates for the original BCRmod model, and after updating BCRmod with German rates, and from BCRAT and after updating BCRAT. Both models overestimated risk in the highest decile, albeit not significantly.
Fig. 1.
Calibration in risk deciles for 5-year risk estimates for the original BCRmod and BCRAT models and after updating with German incidence and mortality data from 1999 (-D), in German EPIC data for women ages 40 and older
Supplemental Table S5 gives calibration results for BCR-mod and after updating with German rates in subgroups defined by age categories and covariates for 5-year projections, and Supplemental Table S6 over the whole follow-up period. BCRmod significantly overestimated true risk in women age 60+ years, with E/O = 1.47 (95% CI 1.08–1.86) for 5 years and E/O = 1.23 (95% CI 1.04–1.43) for the whole follow-up period. BCRmod also significantly overestimated among women in the 30–35 BMI group based on 5-year projections, with E/O = 1.65 (95% CI 1.00–2.29), but not for overall follow-up (E/O = 1.15(95% CI 0.91–1.40)). There was no indication of lack of calibration for all other categories or covariates. Calibration was much poorer after updating the model with German rates, with BCRmod generally underestimating the observed number of events (Supplemental Tables 5 and 6).
Supplemental Table S7 gives calibration results for BCRAT and after updating BCRAT with German rates for 5-year projections. BCRAT significantly underestimated risk for women aged 55–60 (E/O = 0.76; 95% CI 0.59–0.92) and even more strongly when German rates were used (E/O = 0.67; 95% CI 0.52–0.82). For all other subgroups BCRAT estimates were not significantly biased. After updating with German rates, risk was underestimated for women with no family history of BC and those with age at menarche 14+. Similar patterns were observed when total follow-up time was considered (Supplemental Table S8). The underestimation of true BC risk after updating the rates was even more pronounced and significant for all but the 40–45 years age group.
Comparison of 1-year model predictions to breast cancer incidence in the German population
In another independent validation, we compared 1-year projected BCRmod and BCRAT risks for women in the DEGS survey, conducted 2008–2011, weighted back to the female German population with age-specific German BC incidence rates (Fig. 2a, b). We also compared the fit after updating the models with German incidence and competing mortality rates for 2012.
Fig. 2.
Model estimates of 1-year risk and age-specific breast cancer incidence in German women for BCRmod (a) and BCRAT (b)
BCRmod overestimated risk for women aged 55+, with the strongest discrepancy observed for women aged 70+ years (Fig. 2a). To address the possible impact of covariates on the overprediction, we set alcohol consumption, parity, BMI and family history to their lowest level of risk (Supplemental Figure S1). When women who consumed alcohol or had a BMI > 25 were assumed to be at the lowest level of risk, BCRmod was well calibrated among women aged 55–65, but now underpredicted among women < 50 years. There was no notable impact of family history or nulliparity on the predictions. BCRAT risks agreed well with the observed BC incidence, with some slight overestimation in women aged 70+ (Fig. 2b). Using German rates in both models did not improve predictions (Supplemental Figure S2, panels a and b).
Discussion
As relative risks for many BC risk factors vary little across populations from different developed countries [11-16], BC risk predictions models might be transportable across populations, possibly after accommodating different incidence rates used in the model, see e.g., [8]. We thus assessed the ability of two US absolute BC risk prediction models, BCRAT and BCRmod, to predict BC risk in German women. We used a prospective cohort, EPIC-Germany, and data from a population-representative survey, DEGS, for an incidence-based calibration assessment. We focused on women aged 40+ years, the age range relevant for screening and early detection. In EPIC, we found no noticeable deviation in the estimates predicted from the models from the numbers of BCs observed during 5-year and total follow-up. Based on DEGS data, we further compared projected BC risk for the whole German population of women, computed by weighted 1-year projections based on the DEGS risk factor distribution, to BC incidence in German women. Here, model predictions also agreed well with observed incidence. Extending the BCRmod model to women aged 40–49 gave adequate predictions for German women, which is important as it could be used in counseling for risk-based mammography screening of women below age 50, who are currently not included in the national screening program. In a direct comparison of BCRAT and BCRmod using a reclassification table based on two risk thresholds, again both models showed good calibration, but BCRmod slightly better classified women than BCRAT (Supplemental S4).
However, for both, the EPIC and incidence-based calibration assessments, the models significantly overestimated BC numbers in women aged 70+, likely reflecting different screening recommendations in the US and Germany. In Germany, women do not undergo routine mammographic screens after age 70, whereas in the US the recommendations are less clear, and many women over 70 years continue to be screened. A careful comparison of BC mortality in older women in the US and Germany could shed light on the potential of screening at more advanced age.
The underestimation of risk in older German women may also be due to a weaker association of risk with BMI in German women, as compared to the risk effect included in BCR-mod. However, the model does not account for the interaction of age with BMI on pre and postmenopausal breast cancer risk, which seems to specifically affect older women not using HRT [27]. As BCRmod was developed for women 50 years or older, this interaction was not considered in that model. As the original models showed adequate calibration in EPIC and DEGS, it is not surprising that calibration did not improve further after updating the models with German rates. Instead we saw a rather pronounced miscalibration especially for EPIC, where observed counts were much higher than expected counts when using German rates in the models. One possible explanation is that women in the two EPIC centers undergo more frequent screening than the general German female population, due to a “healthy participant” effect, supported by higher BC incidence and lower BC mortality rates in EPIC-Germany than observed in the general female German population (Supplemental Table S1). Because of the possibly limited representativeness of EPIC, we also used the nationally representative DEGS survey to assess calibration. However, we also did not find improvement of calibration in the incidence-based assessment of calibration when using German rates.
Both models had limited discriminatory accuracy, with AUC values comparable to those seen in earlier calibration efforts (AUC = 0.58 for BCRmod and AUC = 0.61 for BCRAT). Having demonstrated reasonable calibration for these models provides a good basis for further extension with independent risk factors like biomarkers, e.g., endogenous hormone levels measured from serum [28], or polygenic-risk scores [29, 30], to improve discrimination.
Limitations of our study include missing data on some model predictors, including previous diagnosis of atypical hyperplasia and number of biopsies (for BCRAT). For the biopsy variable, we partially overcame this problem by assigning one biopsy to all women who reported a diagnosis of benign breast disease, acknowledging that this leads to potential underestimation of true risk in the predictions. Missing data also could have reduced the discriminatory accuracy and likely led to somewhat underestimated AUC values. However, most variables had less than 17% missing values, and we used multiple imputation to enable using all the data for model validation. Missing model predictors in DEGS were imputed based on the BC risk factors in EPIC. Strengths of the study are the large cohort size, use of a population-representative survey, the high data quality and carefully accounting for the added uncertainty arising from imputations in all variance calculations.
Conclusion
In summary, we showed that absolute BC risk models developed for women in the general US population were well calibrated in German women, especially those younger than 70 years. The discriminatory accuracy of the models was modest, and similar to that seen in US-based validations [7, 10]. This limits their applicability for deciding who should not be screened, but can add information to age in recommending screening for younger women. Adequate calibration also allows us to recommend using these models for public health applications and individual counseling of German women aged 40–70.
Supplementary Material
Acknowledgments
We thank David Check for help formatting tables, and Mitchell Gail for helpful comments.
Funding
The present project was supported by the German Federal Ministry of Education and Research (BMBF), the German Cancer Research Center (DKFZ) and by the intramural research program of the NIH. None of the funders had any role in the study design; in collection, analysis, and interpretation of data; in the writing of the report; and in the decision to submit the manuscript for publication.
Footnotes
Electronic supplementary material The online version of this article (https://doi.org/10.1007/s10552-020-01272-6) contains supplementary material, which is available to authorized users.
Availability of data and material The data that support the findings of this study are not publicly available but can be obtained from the authors upon reasonable request and with permission of DKFZ, DIfE, and RKI (all in Germany), but some restrictions apply.
Conflict of interest The authors declare that they have no conflicts of interest.
Ethics approval All participants in the EPIC and DEGS studies provided written informed consent. The EPIC study was approved by ethics committees at both study centers (Potsdam: Ethics Committee of the Medical Association of the State of Brandenburg; Heidelberg: Ethics Committee of the Heidelberg University Hospital). The Ethics Committee of Charite Hospital Berlin approved of the DEGS-Study, and the federal commissioner for data protection held no objections.
References
- 1.Ferlay J, Colombet M, Soerjomataram I et al. (2018) Cancer incidence and mortality patterns in Europe: estimates for 40 countries and 25 major cancers in 2018. Eur J Cancer 103:356–387 [DOI] [PubMed] [Google Scholar]
- 2.Ferlay J, Soerjomataram I, Dikshit R et al. (2015) Cancer incidence and mortality worldwide: sources, methods and major patterns in GLOBOCAN 2012. Int J Cancer 136:E359–E386 [DOI] [PubMed] [Google Scholar]
- 3.Pfeiffer RM, Webb-Vargas Y, Wheeler W, Gail MH (2018) Proportion of U.S. trends in breast cancer incidence attributable to long-term changes in risk factor distributions. Cancer Epidemiol Biomark Prev 27:1214–1222 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Gail M, Brinton L, Byar D et al. (1989) Projecting individualized probabilities of developing breast cancer for white females who are being examined annually. J Natl Cancer Inst 81:1879–1886 [DOI] [PubMed] [Google Scholar]
- 5.Costantino JP, Gail MH, Pee D et al. (1999) Validation studies for models projecting the risk of invasive and total breast cancer incidence. J Natl Cancer Inst 91:1541–1548 [DOI] [PubMed] [Google Scholar]
- 6.Gail MH, Costantino JP, Pee D et al. (2007) Projecting individualized absolute invasive breast cancer risk in African American women. J Natl Cancer Inst 99:1782–1792 [DOI] [PubMed] [Google Scholar]
- 7.Rockhill B, Spiegelman D, Byrne C, Hunter DJ, Colditz GA et al. (2001) Validation of the Gail et al. model of breast cancer risk prediction and implications for chemoprevention. J Natl Cancer Inst 93:358–366 [DOI] [PubMed] [Google Scholar]
- 8.Pastor-Barriuso R, Ascunce N, Ederra M et al. (2013) Recalibration of the Gail model for predicting invasive breast cancer risk in Spanish women: a population-based cohort study. Breast Cancer Res Treat 138:249–259 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Nickson C, Procopio P, Velentzis LS et al. (2018) Prospective validation of the NCI Breast Cancer Risk Assessment Tool (Gail Model) on 40,000 Australian women. Breast Cancer Res 20:155. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Pfeiffer RM, Park Y, Kreimer AR et al. (2013) Risk prediction for breast, endometrial, and ovarian cancer in white women aged 50 y or older: derivation and validation from population-based cohort studies. PLoS Med 10:e1001492. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Petracci E, Decarli A, Schairer C et al. (2011) Risk factor modification and projections of absolute breast cancer risk. J Natl Cancer Inst 103:1037–1048 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Cummings SR, Tice JA, Bauer S et al. (2009) Prevention of breast cancer in postmenopausal women: approaches to estimating and reducing risk. J Natl Cancer Inst 101:384–398 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Boyle P, Mezzetti M, La Vecchia C, Franceschi S, Decarli A, Robertson C (2004) Contribution of three components to individual cancer risk predicting breast cancer risk in Italy. Eur J Cancer Prev 13:183–191 [DOI] [PubMed] [Google Scholar]
- 14.Maas P, Barrdahl M, Joshi AD et al. (2016) Breast cancer risk from modifiable and nonmodifiable risk factors among white women in the united states. JAMA Oncol 2:1295–1302 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Li K, Anderson G, Viallon V et al. (2018) Risk prediction for estrogen receptor-specific breast cancers in two large prospective cohorts. Breast Cancer Res 20:147. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Barnes BB, Steindorf K, Hein R, Flesch-Janys D, Chang-Claude J (2011) Population attributable risk of invasive postmenopausal breast cancer and breast cancer subtypes for modifiable and non-modifiable risk factors. Cancer Epidemiol 35:345–352 [DOI] [PubMed] [Google Scholar]
- 17.Ferlay JEM, Lam F, Colombet M, Mery L, Piñeros M, Znaor A, Soerjomataram I, Bray F (2018) Global cancer observatory: cancer today, 2018th edn. International Agency for Research on Cancer, Lyon [Google Scholar]
- 18.Boeing H, Wahrendorf J, Becker N (1999) EPIC-Germany—a source for studies into diet and risk of chronic diseases. European Investigation into Cancer and Nutrition. Ann Nutr Metab 43:195–204 [DOI] [PubMed] [Google Scholar]
- 19.Riboli E, Hunt KJ, Slimani N et al. (2002) European Prospective Investigation into Cancer and Nutrition (EPIC): study populations and data collection. Public Health Nutr 5:1113–1124 [DOI] [PubMed] [Google Scholar]
- 20.Scheidt-Nave C, Kamtsiuris P, Gosswald A et al. (2012) German health interview and examination survey for adults (DEGS)—design, objectives and implementation of the first data collection wave. BMC Public Health 12:730. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.(2017) Human Mortality Database. Berkeley, USA, and Germany: University of California, Berkeley (USA), and Max Planck Institute for Demographic Research (Germany) [Google Scholar]
- 22.Esserman LJ, Anton-Culver H, Borowsky A et al. (2017) The WISDOM Study: breaking the deadlock in the breast cancer screening debate. NPJ Breast Cancer. 3:34. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Raghunathan TE, Lepkowski JM, Van Hoewykand J, Solenberger P (2010) A multivariate technique for multiply imputing missing values using a sequence of regression models. Surv Methodol 27:85–95 [Google Scholar]
- 24.White IR, Royston P, Wood AM (2011) Multiple imputation using chained equations: issues and guidance for practice. Stat Med 30:377–399 [DOI] [PubMed] [Google Scholar]
- 25.Hosmer DW, Lemeshow SP (2000) Applied logistic regression, 2nd edn. Wiley, New York [Google Scholar]
- 26.Visvanathan K, Chlebowski RT, Hurley P et al. (2009) American society of clinical oncology clinical practice guideline update on the use of pharmacologic interventions including tamoxifen, raloxifene, and aromatase inhibition for breast cancer risk reduction. J Clin Oncol 27:3235–3258 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Ritte R, Lukanova A, Berrino F et al. (2012) Adiposity, hormone replacement therapy use and breast cancer risk by age and hormone receptor status: a large prospective cohort study. Breast Cancer Res 14:R76. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Husing A, Fortner RT, Kuhn T et al. (2017) Added value of serum hormone measurements in risk prediction models for breast cancer for women not using exogenous hormones: results from the EPIC cohort. Clin Cancer Res 23:4181–4189 [DOI] [PubMed] [Google Scholar]
- 29.Husing A, Canzian F, Beckmann L et al. (2012) Prediction of breast cancer risk by genetic risk factors, overall and by hormone receptor status. J Med Genet 49:601–608 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Mavaddat N, Pharoah PD, Michailidou K et al. (2015) Prediction of breast cancer risk based on profiling with common genetic variants. J Natl Cancer Inst. 10.1093/jnci/djv036 [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.