Prediction and clinical utility of a contralateral breast cancer risk model

Breast cancer (BC) is a major burden for women’s health [1]. Survival has improved substantially over the past half century due to earlier detection and advanced treatment modalities, for example in the Netherlands, 10-year survival of a first primary BC improved from 40% in 1961–1970 to 79% in 2006–2010 [2]. Consequently, increasing numbers of BC survivors are at risk to develop a new primary tumor in the opposite (contralateral) breast, with subsequent treatment and potentially less favorable prognosis [3]. BC survivors are more likely to develop contralateral breast cancer (CBC) compared to healthy women to develop a first primary BC [4].

Women at elevated CBC risk have been identified to be BRCA1/2 and CHEK2 c.1100del mutation carriers and to have a BC family history, particularly a family history of bilateral BC [5‐10]. For BRCA1/2 mutation carriers, in whom CBC risk is high, contralateral preventive mastectomy (CPM) is often offered [11]. However, the average risk of CBC among all first BC survivors is still relatively low, with an incidence of ~ 0.4% per year [12‐14]. Despite this, in recent years, CPM frequency has increased among women in whom CBC risk is low [15]. For these reasons, there is an urgent need for improved individualized prediction of CBC risk, both to facilitate shared decision-making of physicians and women regarding treatment and prevention strategies for those at high CBC risk and to avoid unnecessary CPM or surveillance mammography after first primary BC when CBC risk is low.

To our knowledge, only one specific CBC risk prediction model (CBCrisk) has been developed to date. CBCrisk used data on 1921 CBC cases and 5763 matched controls with validation in two independent US studies containing a mix of invasive and in situ BC [16, 17]. Moreover, the level of prediction performance measures such as calibration and discrimination needed for a CBC risk prediction to be clinically useful have not yet been addressed [18].

Our aim was twofold: first, to develop and validate a CBC risk prediction model using a large international series of individual patient data including 132,756 patients with a first primary invasive BC between 1990 and 2013 from multiple studies in Europe, USA, and Australia with 4682 incident CBCs, and second, to evaluate the potential clinical utility of the model to support decision-making.

A total of 132,756 invasive primary BC women diagnosed between 1990 and 2013, with 4682 CBC events, from 20 studies, were used to derive the model for CBC risk (Additional file 1: Table S2). Median follow-up time was 8.8 years, and CBC cumulative incidences at 5 and 10 years were 2.1% and 4.1%, respectively. Details of the studies and patient, tumor, and treatment characteristics are provided in Additional file 1: Table S6. The multivariable model with estimates for all included factors is shown in Table 1 and Additional file 3. BRCA1/2 germline mutation status, family history, and systemic adjuvant treatment showed the strongest associations with CBC risk.

The prediction performance of the main model (PredictCBC, version 1A) based on the leave-one-study-out cross-validation method is shown in Fig. 1. The AUC at 5 years was 0.63 (95% confidence interval (CI): 0.58–0.67; 95% prediction interval (PI): 0.52–0.74)); the AUC at 10 years was also 0.63 (95% CI: 0.59–0.66; 95% PI: 0.53–0.72). Calibrations showed some indications of overestimation of risk. The calibration-in-the-large was − 0.13 (95% CI: -0.66–0.40; 95% PI: -1.62–1.37). The calibration slope was 0.90 (95% CI: 0.79–1.02; 95% PI: 0.73–1.08) in the cross-validation. Calibration plots are provided in Additional file 2: Figure S2 and S3.

The nomogram representing a graphical tool for estimating the CBC cumulative incidence at 5 and 10 years based on our model and the estimated baseline of the Dutch Cancer Registry is shown in Fig. 2. In the nomogram, the categories of each factor are assigned a score using the topmost “Points” scale, then all scores are summed up to obtain the “Total points”, which relate to the cumulative incidence of CBC. The formulae of the models (PredictCBC-1A and 1B) providing the predicted cumulative incidence are given in Additional file 1: Formula to estimate the CBC risk and formula to estimate CBC risk in patients not tested for BRCA.

The DCAs for preventive contralateral mastectomy showed the potential clinical utility of PredictCBC-1A between thresholds of 4–10% 10-year CBC risk for BRCA1/2 mutation carriers and non-carriers (Table 2 and Additional file 3). For example, if we find it acceptable that one in 10 patients for whom a CPM is recommended develops a CBC, a risk threshold of 10% may be used to define high and low risk BRCA1/2 mutation carriers based on the absolute 10-year CBC risk prediction estimated by the model. Compared with a strategy recommending CPM to all carriers of a mutation in BRCA1/2, this strategy avoids 161 CPMs per 1000 patients. In contrast, almost no non BRCA1/2 mutation carriers reach the 10% threshold (the general BC population, Fig. 3). The decision curves provide a comprehensive overview of the net benefit for a range of harm-benefit thresholds at 10-year CBC risk (Fig. 4).

Decision curves for CBC risk at 5 year and the corresponding clinical utility are provided in Additional file 2: Figure S4 and Additional file 1: Table S7, respectively.

We also derived a risk prediction model (PredictCBC, version 1B) omitting BRCA status to provide CBC risk estimates for first BC patients not tested for BRCA1/2 mutations. This model has slightly lower prediction performance; AUC at 5 and 10 years was both 0.59 (at 5 years: 95% CI: 0.54–0.63, 95% PI: 0.46–0.71; at 10 years: 95% CI: 0.56–0.62, 95% PI: 0.52–0.66), calibration-in-the-large was − 0.17 (95% CI: -0.72–0.38; 95% PI: -1.70–1.36), and calibration slope was 0.81 (95% CI 0.63–0.99; 95% PI: 0.50–1.12) (Additional file 1: Results of the prediction model without BRCA mutation). Details of development, validation, and clinical utility are provided in Additional file 1: Tables S8–S10 and Figure S5–S10.

In a sensitivity analysis (see Additional file 1: Assessment of limited information of CPM), we studied the impact of CPM on our results using two studies, in which CPM information was (almost) completely available. The lack of CPM information on cumulative incidence estimation hardly affected the results of our analyses (Additional file 2: Figure S11).

Using established risk factors for CBC which are currently available in clinical practice, we developed PredictCBC, which can be used to calculate 5- and 10-year absolute CBC risk. The risk prediction model includes carriership of BRCA1/2 mutations, an important determinant of CBC risk in the decision-making process [6].

The calibration of the model was reasonable and discrimination moderate within the range of other tools commonly used for routing counseling and decision-making in clinical oncology for primary BC risk [56‐59]. As expected, the prediction accuracy was lower when we omitted the BRCA mutation carrier status although the prevalence of BRCA mutations among BC patients is quite low (2–4%) [60, 61].

In the breast cancer population, CBC is a relatively uncommon event (~ 0.4% per year) and difficult to predict. Therefore, physicians should carefully consider which patients should consider CPM using a prediction model [62]. The current clinical recommendations of CPM are essentially based on the presence of a mutation in the BRCA1/2 genes. Based on the risk distribution defined by the current model (Fig. 3), this is a reasonable approach: essentially no non-carrier women reach a 10% risk 10-year threshold. However, more than 50% of carriers do not reach this threshold either, suggesting that a significant proportion of BRCA1/2 carriers might be spared CPM. Contralateral surveillance mammography may also be avoided although detection and knowledge of recurrences may be necessary for better defined individualized follow-up and patient-tailored treatment strategies [63, 64].

CBC risk patterns and factors were identified previously in a large population-based study with 10,944 CBC of 212,630 patients from the Surveillance, Epidemiology and End Results (SEER) database diagnosed from 1990 to 2013 [65]. However, SEER does not include details of endocrine treatment and chemotherapy, therapies administrated to reduce recurrences and CBCs [13, 66]. Furthermore, in this study, the model was not validated or evaluated based on prediction accuracy, nor was a tool provided. Another study provided general guidelines for CPM by calculating the lifetime risk of CBC based on a published systematic review of age at first BC, BRCA1/2 gene mutation, family history of BC, ER status, ductal carcinoma in situ, and oophorectomy [34, 67]. However, the authors specified that the calculation of the CBC lifetime risk should be considered only as a guide for helping clinicians to stratify patients into risk categories rather than a precise tool for the objective assessment of the risk.

Only one other prediction model (CBCrisk) has been developed and validated using data of 1921 CBC cases and 5763 matched controls [16]. External validation of CBCrisk of two independent datasets using 5185 and 6035 patients with 111 and 117 CBC assessed a discrimination between 0.61 and 0.65 [17]. The discrimination of our PredictCBC model at 5 and 10 years was similar; however, the geographic diversity of the studies gave a more complete overview of external validity [47]. Moreover, we showed the net benefit of our model using decision curve analysis since standard performance metrics of discrimination, calibration, sensitivity, and specificity alone are insufficient to assess the clinical utility [18, 53].

Some limitations of our study must be recognized. First, reporting of CBC was not entirely complete in all studies and information about CPM was limited in most datasets, which may have underestimated the cumulative incidence, although the overall 10-year cumulative incidence of 4.1% is in line with other data [5, 34]. Second, some women included in the Dutch studies (providing specific information on family history, BRCA mutation, or CPM) were also present in our selection of the Netherlands Cancer Registry population. Privacy and coding issues prevented linkage at the individual patient level, but based on the hospitals from which the studies recruited, and the age and period criteria used, we calculated a maximum potential overlap of 3.4%. Third, in the US and Australian datasets, the prediction performance was uncertain due to the limited sample size and missing values. Moreover, some important predictors such as family history and especially BRCA mutation status were only available in a subset of the women (from familial- and unselected hospital-based studies) and patients with data on BRCA mutation status might have been insufficiently represented for tested populations and further development and validation of PredictCBC-1A will be necessary. However, although BRCA1/2 mutation information was unavailable in 94% of our data, the approach of the imputation led to consistently good performing models [68‐70]. The remaining factors were quite complete: ~ 79% of patients had at most one missing factor, which provided good imputation diagnostic performances. Since most BC patients are not currently tested in the clinical practice for BRCA1/2 mutations, we assessed the clinical utility of PredictCBC version 1B to provide individualized CBC risk estimates for first BC patients not tested for BRCA1/2 germline mutations [60, 71]. Our PredictCBC version 1B model provides less precise estimates, but may be useful in providing general CBC risk estimates, which could steer women away from CPM or trigger BRCA testing.

Last but not the least, adequate presentation of the risk estimates from the PredictCBC-1A and PredictCBC-1B is crucial for effective communication about CBC risk during doctor-patient consultations [72, 73]. A nomogram is an important component to communicate the risk of modern medical decision-making, although it may be difficult to use and might potentially make it more difficult to interpret the risks for laymen [74] An online tool is being implemented, and a pilot study will be conducted among patients and clinicians to assess how the risk estimates from PredictCBC-1A and 1B can best be visualized to facilitate communication with patients. Other factors, which were not available in our study, predict breast cancer risk and their inclusion may further improve the discrimination and clinical utility of our CBC risk model: these factors include CHEK2 c.1100del mutation carriers, polygenic risk scores based on common genetic variants, breast density, and reproductive and lifestyle factors such as BMI and age at menarche [75]. Additional data with complete information of BRCA1/2 mutation should be also considered in the model upgrade to reduce uncertainty of CBC risk estimates. External validation in other studies, including patients of other ethnicities, will also be important. In the meantime, our model provides a reliable basis for CBC risk counseling.