Risk factors for breast cancer

Risk estimation

Where there is no dominant family history, risk estimation is based on large epidemiological studies, which give 1.5- to 3-fold increased risks with family history of a single affected relative.^42,43 Clinicians must be careful to differentiate between lifetime and age-specific risks. Some studies quote ninefold or greater risk associated with bilateral disease in a mother or if the woman herself has proliferative breast disease. However, if these at-risk individuals are followed up for many years, the risk returns towards normal levels.⁵³ Clearly, if one uses these risks and multiplies them on a lifetime incidence of 1 in 8–12, some women will apparently have a > 100% chance of having the disease. The risks do not multiply and may not even add. Perhaps the best way to assess risk is to take the strongest risk factor, which in our case is nearly always family history. If risk is assessed on this alone, minor adjustments can be made for other factors. It is arguable whether or not these other factors will have a large effect on an 80% penetrant gene other than to speed up or delay the onset of breast cancer. Therefore, we can only really assume an effect on the non-hereditary element of the risk. Although studies do point to an increase in risk in family history cases associated with some factors, these may just represent an earlier expression of the gene. Generally, therefore, we will arrive at risks between 40% and 8–10%, although lower risks are occasionally given. Higher risks are applicable only when a woman at 40% genetic risk is shown to have a germline mutation, to have inherited a high-risk allele or to have proliferative breast disease.

Several methods based on currently known risk factors have been devised to predict risk of breast cancer in the clinic and in the general population.⁵⁵ Some depend on family history alone (e.g. the Claus and Ford risk evaluation models) and others depend on hormonal and reproductive factors in addition to family history (e.g. the Gail and TC models). Outside risk assessment clinics, where most women have sufficiently strong family histories to have a probability of harbouring mutations in BRCA1, BRCA2 and TP53 (tumour protein p53) genes, it is probable that models in which as many risk factors as possible are combined may be preferable. After all, only 10% of breast cancer occurs in the context of a first-degree family history of breast cancer. The Gail model accurately predicted the number of cancers in the Nurses’ Health Study,⁵⁶ but in our FHC, the TC model, which depends on the extent of family history and several endocrine factors, showed a better prediction than those that used fewer risk factors.⁵⁷ Our own clinical manual assessment was, nonetheless, as good as TC and significantly better than the other computer-based models.⁵⁷ Although these models have reasonably good predictive power for the number of cancer cases likely to be seen in a population, they have low discriminatory accuracy, in that they cannot positively identify which particular woman will develop breast cancer.^55,57 This is not surprising, given that most of the inherited component of at least 27% from twin studies^44,45 would not be identified from family history and the risk model cannot predict who has and who has not inherited any genetic factors in a particular family. At present, most of the known non-family history risk factors are not included in risk models. In particular, perhaps the greatest factor apart from age, MD,⁵⁸ is not yet included. Further studies are in progress to determine whether or not the inclusion of additional factors into existing models, such as MD, weight gain³¹ and serum steroid hormone measurements,⁵⁹ will improve prediction. These are not straightforward additions, as there may be significant interactions between risk factors. Although breast density is an independent risk factor for BRCA1 and BRCA2 cancer risk,⁶⁰ the density itself may be heritable and not increase risk in a similar way to the context of family history of breast cancer alone.

Women who are at increased risk for breast cancer can be identified on the basis of their individual risk factors. However, this approach does not permit the combination of multiple risk factors or the calculation of a woman’s lifetime probability of breast cancer. In breast cancer FHCs in Europe, the Claus tables⁴³ were frequently employed to assess lifetime risk and risk in a given decade (Table 1). However, these did not take into account risk modification from hormonal and reproductive factors or from having many unaffected female relatives.

TABLE 1

Original Claus tables for women with a single affected first-degree relative

Therefore, multivariate risk models have been introduced. These models allow the determination of a woman’s composite RR for breast cancer as well as her cumulative lifetime risk adjusted for several risk factors. Such models, therefore, provide an individualised breast cancer risk assessment, which is an essential component of the risk–benefit analysis from which decisions regarding the implementation of frequent surveillance, chemoprevention or risk-reducing surgery can be made.

In 2003, we published a study designed to compare the predictive value of the Gail, Claus, Ford and TC risk assessment models using a cohort of 3151 women attending the Family History Evaluation and Screening Programme.⁵⁷ The study used a read-out from the Cyrillic computer program [version 3.0; www.exetersoftware.com/cat/cyrillic/cyrillic.html (accessed 30 March 2004)].⁶¹ These computerised models were also compared with risk assessment undertaken by clinicians based on Claus tables with adjustment for other risk factors, particularly hormonal factors (the Manual model). This showed that although the TC and Manual models were accurate at predicting breast cancer risk over time, the Claus, Gail and Ford models substantially underestimated risk (Table 2). Our 2003 report⁵⁷ remains the only large-scale effort to validate risk assessment models in a family history setting. We now have more than fourfold follow-up time and over sixfold the number of prospective cancers to validate risk prediction models. We are also in a position to validate the BOADICEA model for the first time.

TABLE 2

Known risk factors and their incorporation into existing risk models

We have now expanded our follow-up in the FHC to nearly three times the size and twice the average years. To reassess the Manual model, we have updated the Claus tables to reflect modern breast cancer incidence in the UK. We show that the Manual model is accurate and convenient for use in the busy clinic, but that TC and BOADICEA also appear to predict cancers accurately.

Methods

Study tools

Manual risk calculation (the Manual model) was used to calculate risk in the clinic. The Manual model uses the Claus tables⁴³ and curves^52,62,63 to calculate a heterozygote and lifetime risk (which includes a population risk element). Families fulfilling Breast Cancer Linkage Consortium (BCLC) criteria or with a proven BRCA1 or BRCA2 mutation⁶⁴ were given risks based on the penetrance and 10-year risks from our regional penetrance estimates.⁵ All other women were given lifetime risks, with 10-year risks calculated from the equivalent figure given in the Claus tables after a clinician’s modification for hormonal and reproductive factors. Those women with average risk factors, such as menarche aged 12–13 years and first pregnancy aged 23–27 years, had no alteration to the genetic risk-based assessment. However, those with unfavourable factors, such as menarche aged < 12 years and late first pregnancy and nulliparity, would go up a risk category, say from 1 in 4 to 1 in 3 lifetime risk, and those with favourable factors of a late menarche and early first pregnancy would drop to 1 in 5, equivalent to a 20–30% alteration. More significant alterations would occur with a very early menopause which, aged 40 years, could halve the risk. A more detailed review of our Manual model can be found elsewhere.^52,62,63 The output of these lifetime risks was used to calculate the expected numbers of breast cancers.

The original Claus tables were based on invasive breast cancer incidence rates for the 1970s and 1980s in the USA.⁴³ We therefore updated these to reflect modern incidence rates for the 1990s and 2000s in the UK (Table 3). The increase in breast cancer incidence in recent times was discussed in NICE guidance,² in which 2001 incidence rates from the Office for National Statistics² predicted a risk to age 80 years for the average woman as 10.7%, whereas Claus tables predict a lower risk for women with a single first-degree relative affected > 60 years of age (see Table 2). Minor upwards adjustments were made only to values with a risk to age 80 years of < 25% to reflect the increase in risk in the last 20 years (see Table 3). The 3% 10-year risk for women in their forties with a single first-degree relative affected aged 30–39 is equivalent to the version 6 TC read-out for a typical pedigree (Figure 1).

TABLE 3

Adjusted Claus tables for women with a single affected first-degree relative

FIGURE 1

Tyrer–Cuzick read-out giving 10-year risk aged 40 years for a woman with a single first-degree relative with breast cancer.

Risk models

Gail model

• Absolute risks of 1–20 years (including and excluding competing mortality) were calculated using the code available from www.cancer.gov/bcrisktool.
• Ethnicity was taken to be ‘white’ for all probands, as 98% were of white origin in the Family History Risk Study (FH-Risk).
• Very little information was available about benign disease, and no information was available on the number of biopsies. In total, 23 of the probands had a previous hyperplasia diagnosis. These were all taken to have one previous biopsy.
• A summary of the other risk factors (age, number of first-degree relatives, age at menarche and age at first child) is in Table 4.

TABLE 4

Some characteristics of the cohort

Tyrer–Cuzick model

• Absolute risks of 1–20 years (no competing mortality), based on version 6.
• The family history information used included age, diagnosis and age of breast cancer, bilateral breast cancer and ovarian cancer for the following relatives:
  • mother
  • sisters and half-sisters
  • daughters
  • aunt, maternal and paternal
  • grandmother, maternal and paternal
  • nieces (from brother or sister)
  • cousins (from uncles and aunts).
• Affected father and brother status was also included.
• No BRCA test results were used, but they will be incorporated later.
• Similar to the Gail model, there was very little information on benign disease.
• In addition to the Gail model factors, information on height and weight was used (Table 5).

TABLE 5

Height and weight distributions in cohort

BOADICEA model

The BOADICEA model used the same family history information as TC; 10- to 20-year risks were generated. Unfortunately, BOADICEA risks could be generated only for individuals from families with only one proband (Table 6). Therefore, output was possible on only 6717 women.

TABLE 6

Number of probands per family and prospective cancers in the probands

Results

Manual model

The analysis of the whole study population was carried out on data from 8824 women [age range 14.5–81.3 years; interquartile range (IQR) 33.9–46.0 years; median 39.5 years at entry] (Table 7). Forty prevalent cancers (15 carcinomas in situ) were excluded (rate 40/8824 = 4.5 per 1000). The mean time of follow-up was 9.71 years (range 0.10–26.00 years). During follow-up, 406 first incident breast cancers were diagnosed (372 invasive, 34 carcinomas in situ) in 85,726.9 woman-years, giving an incidence of 4.7 cancers per 1000 women per year (95% CI 4.3 to 5.2) (Table 8). The expected incidence based on 9.71 years of follow-up for a woman aged 39.5 years would be 1.5 per 1000 in the local region based on NWCIS rates.

TABLE 7

Baseline characteristics of study population

TABLE 8

Manual risk prediction expected and observed breast cancers by decade

Eighteen female carriers contributed risk in their twenties, 81 contributed risk in their thirties, 37 contributed risk in their forties and 26, 18 and 8, respectively, contributed risk in the subsequent three decades. In total, eight breast cancers occurred, with 8.40 expected [observed vs. expected (O : E) 0.99, 95% CI 0.43 to 1.96] using rates from our regional penetrance estimates.⁶⁶ Of the 406 breast cancers that occurred, 203 (50%) had no upwards or downwards variation in their genetic-based risk based on reproductive risk factors by the study clinicians (DGE and AH), while 68 (17%) had a downwards change and 135 (33%) had an upwards change in risk. Of women without breast cancer, 42% had no variation in their genetic-based risk based on reproductive risk factors, while 39% had a downwards change and 19% had an upwards change.

The expected against observed counts for all breast cancers are shown in Table 9 divided into risk categories as defined by NICE guidelines.² These O : E ratios are close to 1.0 across all risk categories, especially for women in their forties. A further breakdown of risk for the forties group shows good discrimination around the 3% 10-year threshold for annual mammography set by NICE. The 12–16% lifetime risk category, which predominantly includes a single first-degree relative aged in their forties with breast cancer, has a 10-year risk of 2.75%, below the 3% threshold, whereas the 17–19% category, which predominantly includes women with a single first-degree relative aged in their thirties with breast cancer, has a 10-year risk of 3.1% above the threshold. The only risk and age category that statistically differed from 1.0 was the near population (average) risk group in their fifties, in whom there was a significant underestimate of the breast cancers that occurred (O : E 1.61, 95% CI 1.06 to 2.34).

TABLE 9

Observed and expected number of cancers from the TC and Gail 10-year cumulative risk bands

Computer models

Follow-up was for a median of 10.2 years (IQR 5.0–15.5 years; range 0.1–26.0 years). A total of 9527 women had assessable follow-up and 416 breast cancers occurred in 97,958 years of follow-up. 382.9 cancers were expected by O : E ratio 1.09 (95% CI 0.98 to 1.20) with TC and 412.5 were expected by O : E ratio 1.01 (95% CI 0.91 to 1.11) with Gail. However, although Gail accurately predicted the number of cancers, its calibration was substantially worse than TC. The observed and expected risks are tabulated in Table 9, based on 10-year absolute risk groups from TC and Gail. Gail significantly overestimates risk in those with risks in the lowest three categories (0–3% 10-year) risk and significantly overestimates risk in the highest risk category (> 12% 10-year risk). In contrast, TC significantly underestimated risk only in the lowest risk category (0–1% 10-year risk). Figure 2 shows the distribution of 10-year risk at baseline from the Gail model (includes competing risks) and the TC model (does not).

FIGURE 2

Baseline risks. (a) Histogram of 10-year risk distribution in this cohort at baseline from TC and Gail models; and (b) individual risk prediction comparison (Spearman’s correlation coefficient 0.79).

Figure 3 shows observed risk from Kaplan–Meier estimates by 10-year risk quintile. There is overlap of the quintiles with Gail, but good separation with TC.

FIGURE 3

Observed risk by 10-year risk at baseline quintile (Q1 = bottom quintile, Q5 = top quintile). (a) Gail; and (b) TC.

Figure 4 shows the same, but with emphasis on the spread of risk between the upper and lower quintiles of TC and Gail.

FIGURE 4

Comparison of the spread of observed risk from the upper and lower quintile groups at baseline. Q, quintile.

Figure 5 again shows the upper and lower quintiles, but adds the expected risk from those groups based on the two models. There is much better overlap in the TC plot than in Gail.

FIGURE 5

Comparison of observed and expected risk through follow-up time from the upper and lower quintile 10-year risk groups from the (a) TC and (b) Gail risk models.

The overall performance of the models is compared in Table 10. The log-likelihood of the predictions was used as a primary performance measure to compare models. Using this performance measure, TC performed much better than Gail: on a likelihood radio chi-squared scale the difference was 98.3.

TABLE 10

Performance summary of the Gail and TC models

Assessment of BOADICEA

The BOADICEA model could be assessed only in the families with only one proband. This limited the analysis to 6268 single-proband families. One hundred and ninety-seven breast cancers occurred in 70,293 years of follow-up. BOADICEA predicted 203.59 breast cancers, with good discrimination across the risk groups (Table 11).

TABLE 11

Number, observed and expected breast cancers assessed by BOADICEA

Discussion

Accurate individualised breast cancer risk assessment is essential for the provision of risk–benefit analysis prior to the initiation of any surveillance or preventative interventions. We have extended our original study⁵⁷ with a longer follow-up and greater number of incident cancers (sixfold power) to confirm that manual risk assessment and the TC model remain accurate tools for women with a family history of breast cancer. We also provide updated Claus tables for current use in the UK. The expected-to-observed ratios with the Manual model are all very close to 1.0, particularly for women in their forties, for whom discrimination between moderate and near population risk is even more important for risk stratification in order to implement annual mammography. NICE² has set a 3% 10-year risk aged 40 years as the threshold for annual mammography in England and Wales, and the Manual model accurately predicts this based on the adjusted Claus tables. The TC model (see Figure 1), Claus model (as accessed on version CancerGene6⁶⁷) and BOADICEA^68,69 also predict this accurately, with a 10-year risk of 3.1%, 3% and 3%, respectively, compared with the 3.1% observed for women who typically had this type of pedigree. However, when the same pedigree is input to Gail and BRCAPRO (FORD) (as accessed on CancerGene version 6; The University of Texas Southwestern Medical Center at Dallas and the BayesMendel Group, www4.utsouthwestern.edu/breasthealth/cagene/), the 10-year risks were < 3.0%, at 2.7% and 1.7%, respectively. Previously, we showed that the Gail, Claus and Ford models all significantly underestimated risk.⁵⁷ In particular, the Gail, Claus and Ford models all underestimated risk in women with a single first-degree relative affected with breast cancer, while the TC and the Manual models were both accurate in this subgroup. The new outputs from the Claus model, in particular, suggest that modifications have been made to this to provide data similar to those in the tables and this might now provide a more accurate assessment. Our modifications to the Claus tables provide more accurate assessment for current breast cancer risks.

When risks in other age groups are assessed, only the very high-risk group exceeds the 10-year 3% risk from age 30 years and the group as a whole falls short of the 8% 10-year risk previously deemed as the threshold for MRI screening in the thirties.^2,70 Indeed, when gene carriers are excluded, the 10-year risk drops to 5.7% in those at very high risk (lifetime 40–50%) of breast cancer. The American College of Surgeons recommends MRI screening in women with a lifetime risk of breast cancer as low as 20–25%.⁷¹ The incidence rates in this risk group are 20–25% of those deemed cost-effective by NICE at 8% 10-year risk in the thirties and 20% 10-year risk in the forties.² Thus, American College of Surgeons screening for moderate-risk women based on the cost-effectiveness analysis is likely to cost in excess of US$120,000 per quality-adjusted life-year (QALY).⁷⁰ In fact, NICE has now dropped the breast cancer risk threshold in favour of a simple mutation-based eligibility, as work from the MARIBS study showed that those testing negative for BRCA1 and BRCA2 dropped below the breast cancer risk thresholds.⁷² This was possible as NICE dropped the genetic testing threshold to a 10% likelihood of identifying a mutation and included unaffected women for the first time in the eligibility.²

The question of which risk model to use in the family history setting is a problem, as the models often give quite different read-outs.^73,74 To our knowledge there has been only one other assessment of breast cancer risk prediction in the family history setting, this time in North America, which again showed that TC substantially outperformed the Gail model.⁷⁵ In this study, the Gail model, as in our original series,⁵⁷ predicted cancer rates half of those that actually occurred. This is, perhaps, not surprising, as Gail does not include second-degree relatives or the age at breast cancer onset in first-degree relatives. Nonetheless, in the current update Gail predicted the number of breast cancers accurately, but not the proportions in each risk category. Changes have been made to the model since 2003 to increase the rates expected with a family history, but its discrimination compared with TC remains poor. Surprisingly, our interrogation of the Claus model showed that a maternal grandmother with breast cancer was ignored but a maternal aunt was included in assessing risk. This may be because the original Claus tables did not have a read-out for grandmothers.⁴³ Using the Manual model, we would include a grandmother in the same way that we would include an aunt in the tables.^52,62,63 It is likely, therefore, that the Claus model would predict lower incidence than the Manual model use of tables where there is an affected maternal grandmother. Nearly 1500 of our women had an affected maternal grandmother. As yet, there has been no formal validation of the BOADICEA model for its breast cancer risk output. We did consider using the Claus ‘extended’ model to incorporate ovarian and other cancers, but this would have meant reassessing > 8000 women.⁷⁶

Although the core risk assessment in the present study was adjusted Claus tables, there were upwards or downwards alterations owing to non-familial risk factors in > 50% of women. This was usually by only one or less than one order in the odds ratios (ORs) that we usually quote for women. For instance, an upwards variation from 25% or 1 in 4 lifetime risk usually was to either 3 in 10 (30%) or 1 in 3 (33%). It is of note that the predominant variation in those with cancer was an upwards variation, whereas in those without cancer it was downwards. This demonstrates the added value of adding hormonal risk factor information, which is currently not incorporated into the Claus model, BRCAPRO (Ford) or BOADICEA.

There are some potential weaknesses in the present study. Although screening is likely to increase the incidence of breast cancer through early detection, most of this effect will be offset by excluding the prevalent cases. Moreover, the Gail and TC models have been derived from a screened population. The subanalysis carried out on each risk group does reduce power to discriminate between models and, thus, for some estimates CIs are wide. We have not compared the Manual model directly with the computer models in this present study, other than in specific examples. The primary aim was to determine whether or not adjustments to the Claus tables to reflect modern incidence rates of breast cancer provided an accurate risk prediction. The longer mean follow-up data and the large study size also mitigate this potential weakness. We have included both invasive breast cancers and in situ disease. Exclusion of in situ disease would not have substantially affected the results; the E : O ratio would move from 0.95 to 1.03 with the Manual model. It would also have made the TC model more accurate. The other main issue is the alteration in risk in the Manual model from the ‘genetic prediction’ based on hormonal and reproductive factors. This study is, therefore, not a pure output of modified Claus tables and the modification by two experienced clinicians (DGE and AH) may not be entirely reproducible. Given that screening is likely to be recommended in women at moderate and high risk, we believe that our Manual model and the TC models are the most appropriate in the family history setting. The Manual model can be used quickly and gives similar read-out to the TC, which provides the most useful and accurate tool for clinicians involved in breast cancer risk assessment.⁵⁷

Conclusions

Manual risk prediction which adds hormonal and reproductive factors remains an accurate approach to breast cancer risk estimation when used in conjunction with adjusted Claus tables provided here. This approach would be reasonable as a back-up to use of validated tools such as TC or as a stand-alone in a busy clinic and where incidence rates of breast cancer are similar, such as North America, northern Europe and Australasia.

Statistics

Patients were grouped by genetic status and by ovarian cancer type (invasive epithelial or borderline). Person-years at risk analyses were performed to assess expected ovarian cancers in the family history population using data from the NWCIS. The expected numbers of cases for each genetic status (BRCA1, BRCA2, BRCA negative and BRCA untested) were calculated. Observed-to-expected ratios were assessed for statistical significance using the common method from Clayton and Hills based on the Poisson assumption. Confidence limits for RR were based on the Byar’s approximation of the exact Poisson distribution.

Results

We have assessed risk of ovarian cancer in 8005 women from 895 families from time of referral without ovarian cancer to our FHC. A total of 1613 women from breast cancer families that had tested negative for BRCA1/BRCA2 were followed for a total of 17,589 years between 0.04 and 25 years and checked against a cancer registry for ovarian cancer incidence. Data were censored at ovarian cancer diagnosis, oophorectomy or death. During follow-up, only one invasive epithelial ovarian cancer occurred, although two borderline tumours were diagnosed. Expected rates for this cohort from cancer registry data were 2.99 epithelial ovarian tumours, with 2.74 for invasive cancer and 0.25 for borderline tumours. The RR of developing invasive ovarian cancer in this group was 0.37 (95% CI 0.01 to 2.03), compared with the general population. The upper confidence limit for invasive RR was 2.03 and for borderline tumours was 28.90 (Table 12). The 351 women who were from BRCA2 breast cancer families had a total follow-up of 3230.47 years between 0.02 and 23.72 years. These years of follow-up to ovarian cancer, oophorectomy or death showed no borderline tumours, but five invasive epithelial tumours were recorded. Expected rates for women from BRCA2 positive families were 0.319 ovarian tumours, including 0.300 invasive cancers and 0.032 borderline tumours (see Table 12). The RR was 16.67 (95% CI 5.41 to 38.89) (see Table 12) for invasive cancers in BRCA2.

TABLE 12

Observed vs. expected ovarian cancers in BRCA1, BRCA2 and BRCA-negative families

Thirteen cancers were identified in the cohort of 310 women whose families had tested positive for BRCA1 in 1981.60 years of follow-up, of which all 13 were invasive epithelial tumours and none were borderline cancers. The expected rate of invasive tumours was 0.26, which, as expected, was higher than the expected rate of borderline cancers (0.03). The RR for the invasive cancers was 50.00, although with wide CIs (95% CI 26.62 to 85.50) (see Table 12).

From families untested for BRCA1/BRCA2, 5731 women had 58,904 years of follow-up and 14 ovarian tumours were diagnosed. There were nine epithelial ovarian cancers, two malignant germ cell tumours and three borderline tumours. Using the same average invasive ovarian tumour rates of 0.154 per 1000 as in the untested cohort, we would have expected 9.07 invasive ovarian cancers and, using a similar analysis, 0.84 borderline tumours. The RR for invasive cancers was calculated as 0.99 (95% CI 0.45 to 1.88), increasing to RR 3.57 for borderline tumours (95% CI 0.72 to 10.44) (see Table 12).

Discussion

The present study from our FH-Risk cohort has shown that in a large cohort of women from breast cancer families testing negative for BRCA1/BRCA2, there was no increased risk of ovarian cancer. There was a non-significant (95% CI 0.97 to 28.90) increased risk of borderline tumour, and an apparently non-significant reduced risk of invasive epithelial ovarian cancer, with only one cancer occurring when 2.74 were expected. This study provides strong evidence to support the counselling of women whose breast cancer-only family tests negative for BRCA1/BRCA2 that they are not at increased risk of ovarian cancer, although indications are that counselling in ovarian cancer risk does need to continue in women whose families have tested BRCA1/BRCA2 positive. Statistically significant increased risk was seen in both BRCA1 and BRCA2 families: RR 50.0 (95% CI 26.62 to 85.50) and RR 16.67 (95% CI 5.41 to 38.89), respectively.

Our study contained only 167 women who had developed breast cancer and who personally had tested negative for BRCA1/BRCA2, but support that these women may also be reassured comes from a large Swedish study⁷⁸ that found that their founder mutation, BRCA1 3171 ins5, explains the excess of ovarian cancer after breast cancer in 2600 women in their region. The authors estimated that BRCA1 gene mutations were associated with about 80–85% of the estimated 63 excess cases of ovarian cancer diagnosed after breast cancer.

The only other prospective study mainly in unaffected women showed that, during 2534 woman-years of follow-up, one case of ovarian cancer was diagnosed, when 0.66 was expected (standard incidence ratio = 1.52, 95% CI 0.02 to 8.46).⁷⁹ This study used questionnaires to families and did not verify diagnoses against a cancer registry as we have done. Our study also has over five times the follow-up. Nonetheless, both studies show no overall evidence of any increased risk.

Although data from the BCLC estimated that close to 100% of families with two or more ovarian cancers in addition to breast cancer (at least 2 < 60 years) had mutations in BRCA1/BRCA2,⁶⁴ recent results have shown that three of eight (37.5%) of the mutations in RAD51D (RAD51 paralog D) were in families with two or more ovarian cancers that fulfilled BCLC criteria.⁸⁰ However, the frequency of RAD51C (RAD51 paralog C) and RAD51D mutations was only 1.3%⁸¹ and 0.9%,⁸⁰ respectively, in breast ovary kindreds negative for pathogenic mutations in BRCA1 and BRCA2. Furthermore, although the initial study on RAD51C⁸¹ suggested that mutations might be high risk for both breast and ovarian cancer, the RAD51D study estimated that the risk was high only for ovarian cancer, with a non-significant increased breast cancer risk of less than twofold (1.37, 95% CI 0.92 to 2.05; p = 0.64)).⁸⁰ Neither study found mutations in breast cancer-only kindreds (0/737), supporting the lack of a strong link with breast cancer.^80,81

If BRCA1/BRCA2 mutations account for the great majority of the inherited link between breast and ovarian cancer, the main factor affecting the ability to confidently exclude risk of ovarian cancer in families testing negative is the sensitivity of the testing. Tests on most cancer-predisposing genes are limited, in that they do not screen the intronic areas outside the intron–exon boundaries, and nor do they screen for positional effects of mutations in other genes that can affect genes at a distance, such as EPCAM (epithelial cell adhesion molecule) mutations and MSH2 (MutS homolog 1).⁸² There are relatively few papers that adequately assess the sensitivity of BRCA1/BRCA2 mutation testing. Simply using a panel of found mutations and assessing different screening tests does not address the overall sensitivity.⁸³ The tests can be assessed only against a gold standard, such as gene sequencing, which, in any case, does not screen the introns. It is first necessary to identify families with a very high a priori probability of BRCA1/BRCA2 involvement, such as breast/ovarian families fulfilling BCLC criteria or such families with a Manchester score⁸⁴ of ≥ 40. We have previously shown that sequencing plus multiplex ligation dependent probe amplification (MLPA) identified mutations in 58 out of 65 (89%) families with both breast and ovarian cancer and a Manchester score of ≥ 40.⁸⁵ Breast cancer phenocopies can reduce the sensitivity of tests, as around 6% of tests of breast cancers in families with mutations are mutation negative;⁸⁶ thus, true sensitivity could be closer to 95%. Even if one takes the lower sensitivity estimate, this would reduce the excess risk of ovarian cancer in a breast cancer-only family by ninefold. The true likelihood of a missed mutation can be estimated from our testing of 2009 breast cancer-only families, of whom only 240 (11.9%: 100 BRCA1; 140 BRCA2) had mutations identified by sequencing plus MLPA. Allowing for a Bayesian calculation, no more than 1.5% of these breast cancer-only families would have had a missed mutation.

There are some potential limitations to the present study. Not all participants remained in active follow-up, and a small proportion may have moved out of the north-west of England and an ovarian cancer could have been missed. Despite the large numbers of women, the confidence limits are still quite wide and include a twofold RR of invasive ovarian cancer. Nevertheless, the data are now strongly supportive of informing women from breast cancer-only families who test negative for BRCA1/BRCA2 that their risks of ovarian cancer are not increased.

In conclusion, this, the largest prospective follow-up of a BRCA-negative cohort, has demonstrated that there is no evidence of an increased risk of invasive ovarian cancer in families who have tested negative for BRCA1/BRCA2.

Recruitment

A total of 320 cases and 980 controls were recruited between November 2010 and September 2012 to the MD study. A further series of cases and controls were available who had blood DNA but no assessable mammogram for a cancer. All cases have been matched to three controls, with two extra controls recruited. Samples of DNA have been obtained on 426 women affected with breast cancer and 1275 controls. Direct consent for FH-Risk was obtained for 320 cases and 980 controls (Figure 6). A further 38 who had been affected with breast cancer and had subsequently died having previously supplied a blood sample were also included.

FIGURE 6

Uptake to the FH-Risk study over a 2-year period. (a) Cumulative uptake to FH-Risk – cases; and (b) cumulative uptake to FH-Risk – controls.

Methods

Cases and controls were matched on a 1 : 3 basis, with age at assessable mammogram being the matching criterion. When sufficient controls were unavailable, matching was carried out on a 1 : 1 basis. Controls did not have cancer at last follow-up and cases had to have been diagnosed after an assessable mammogram or have an assessable mammogram at date of cancer diagnosis and no cancer in the contralateral breast. Controls were also matched for type of mammogram, either digital or analogue. The questionnaire data were updated to include clothes size and alcohol intake as well as ensuring that data were held on age at menarche and age at first full-term pregnancy. The TC model was used to incorporate an adjustment for MD using VASs and two digital methods [Volpara™ (version 1.4.5; Volpara Solutions, Wellington, New Zealand) and Quantra™ (version 2.0; Hologic, Inc., Marlborough, MA, USA)]. For breast cancer cases with no assessable mammogram but in whom blood DNA was available, controls were also recruited for the DNA SNP study in Chapter 3.

Results

The SNP results are in Chapter 3. The FH-Risk case–control study was also used to assess incorporation of MD into risk assessment. The study was affected by the policy of destroying analogue mammograms after 3 years and by the fact that no raw data from digital mammograms were saved. Therefore, only 320 cases had an assessable mammogram. The controls were matched for age and type of mammogram but, owing to hospital policy of destroying mammograms, the analogue controls were from a much later era (see Table 13).

TABLE 13

Demographic characteristics for family history analogue case–control analysis

Density results: analogue (film) mammograms

Table 13 shows the demographic characteristics of women with film mammograms. The majority were parous and had a body mass index (BMI) in the normal range; mean BMI was approximately 24.5 kg/m². There were very few current or previous users of HRT (< 10% of cancer cases and < 5% of controls) based on data available at the time of referral; however, the average age was approximately 43 years and the majority of women had not yet been through the menopause at the time of referral. Fifty-six per cent of cancers were in the left breast, and 44% in the right breast. The year of mammogram ranged from 1988 to 2007, with more of the control subjects than the case subjects having recent images. Matching on age alone appeared to be adequate in the analogue film group, as there were no significant differences between cases and controls for other important parameters such as BMI, parity, HRT use and menopausal status.

Table 14 shows the analysis of breast density measures made in women with film mammograms. The results show no association between MD, as measured by VAS, and risk of developing breast cancer. The only significant association with risk was in women with Cumulus percentage density of 41–50% with an OR of 4.16 (95% CI 1.32 to 13.06) compared with those in the lowest quartile (< 27% density). Those with percentage density of ≥ 50% also had an increased risk of breast cancer (OR 3.19, 95% CI 0.99 to 10.31), which approached statistical significance. Adjustment for breast area increased the ORs slightly, but there were no changes in statistical significance.

TABLE 14

Numbers of cases and controls in quartiles of density measures and ORs for the risk of developing breast cancer from univariate logistic regression, and after adjustment for breast area. Data are for analogue (film) mammograms

Density results: digital mammograms

Table 15 shows the demographic characteristics of women with FFDM mammograms. More case subjects were premenopausal (88% cases; 66% controls); this difference was not statistically significant. There was, however, a significant difference between BMI categories for the cases and controls. The cases had a lower mean BMI (24.2 kg/m²) than the controls (26.5 kg/m²), although it should be noted that BMI was not recorded for 22% of controls and 14% of cases. Fifty-six per cent of the cancers were in the right breast and 44% were in the left breast. There were very few current or previous users of HRT (< 5%) based on data available at the time of referral; however, the average age was approximately 46 years. The year of mammogram ranged from 2006 to 2012, with more of the control subjects than the case subjects having recent images.

TABLE 15

Demographic characteristics for family history digital case–control analysis

Table 16 shows the analysis of breast density measures made in women with digital mammograms. The results show no significant association between MD and risk of developing breast cancer for density measured by Cumulus (both percentage density and dense area), but a significant association for density measured by VAS following adjustment for menopausal status, breast area and BMI. The ORs for Cumulus tended to be higher after adjusting for BMI, menopausal status and breast area; however, none was statistically significant.

TABLE 16

Numbers of cases and controls in quartiles of density measures and ORs for the risk of developing breast cancer from univariate logistic regression, and after adjustment for breast area, BMI and menopausal status. Data are for FFDM mammograms

Discussion

In women with film mammograms, Cumulus was the only method that showed significant association with breast cancer risk. Use of Cumulus for mammograms acquired in this way is well established, and the inclusion of prior mammograms as well as contralateral breast of diagnostic mammograms provides a similar methodology to published validation of Cumulus for breast cancer risk.^26,87,88 In a study in which mammograms (contralateral breast for cancers, and matched controls) were assigned to density classes, the RR for women in the highest density group associated with radiologist allocation was higher than computer-based allocation using interactive thresholding.²⁶ However, in this Family History analogue mammography substudy, VAS assessments were not related to cancer risk; this may be due, in part, to interobserver variability dominating any genuine effect in a small evaluation. The failure to match on date of mammogram is also an important shortcoming in this study. It is of interest that in the density analysis of the larger PROCAS screening study (see Chapter 4), and in the subgroup of family history cases with digital screening mammograms which were matched with controls on a 1 : 3 basis, VAS provided the most predictive MD method. Visual assessment recorded on VAS, following adjustment for BMI, menopausal status and Cumulus breast area, was the only method significantly associated with an increased risk of breast cancer in the digital group. For the other methods, the ORs tended to be higher after adjusting for menopausal status and breast area, but did not show statistical significance. For the VAS measures, the CIs for the OR were very wide, reflecting the limited number of cases. Significance was achieved only following the inclusion of breast area as measured by Cumulus in the model.

These results are consistent with previously published work showing that the association between breast cancer and MD is higher in an asymptomatic population.³³ In this substudy we were unable to assess volumetric breast density; in analogue mammograms, the Manchester Stepwedge method can be applied prospectively only, and in the digital group only seven cases/controls were amenable to processing by Quantra and Volpara (with GE Healthcare images and raw data).

The small number of cases for which both Cumulus and VAS were available is a limitation of this substudy. Furthermore, Cumulus was performed on a single mammographic image (MLO), whereas VAS assessment used the mean across both views [MLO and craniocaudal (CC)]; ideally, we would have liked to have had Cumulus for both mammographic views for the contralateral breast. A further limitation is that the covariate information was collected at the time of referral, which was very often not the same time as the mammogram was acquired. Of all FH-Risk mammograms (both digital and film), approximately 25% were taken at the same time as referral to the clinic, 5% were taken within 1 year of referral and the remaining 70% were taken over 1 year after referral. For those with a larger interval, BMI, HRT use and menopausal status may be inaccurate.