Towards a personalized organized screening in breast cancer : practical determination of thresholds of risk in women at average-risk

In France, more than 10 million women at ”average” risk of breast cancer (BC), are included in the organized BC screening. Existing predictive models of BC risk are not adapted to the French population. Thus, we set up a new score in the French Hérault region and looked for a graded level of risk in women at "average" risk. We recruited a retrospective cohort of women, aged 50 to 60, who underwent the organized BC screening, and included 2241 non-cancer women and 527 who developed a BC during a 12-year follow-up period (2006-2018). The risk factors identiﬁed were high breast density (ACR BI-RADS grading)(B vs A: HR 1.41, 95%CI [1.05; 1.9], p=0.023; C vs A: HR=1.65 [1.2; 2.27], p=0.02 ; D vs A: HR=2.11 [1.25;3.58],p=0.006), a history of maternal breast cancer (HR=1.61 [1.24; 2.09], p < 0.001), and socioeconomic difﬁculties (HR 1.23 [1.09; 1.55], p=0.003). While early menopause (HR=0.36 [0.13; 0.99], p=0.003) and an age at menarche after 12 years (HR=0.77 [0.63; 0.95], p=0.047) were protective factors. We identiﬁed 3 groups at risk: lower, average, and higher, respectively. A low threshold was characterized at 1.9% of risk and a high threshold at 4.5%. Mean risks in the 3 groups of risk were 1.37%, 2.68%, and 5.84%, respectively. Thus, 12% of women presented a level of risk different from the average risk group, corresponding to 600,000 women involved in the French organized BC screening, enabling to propose a new strategy for performing an organized and personalized national BC screening.


Introduction
Breast cancer is the most common cancer in women and the leading cause of death from cancer in women in France. The earlier breast cancer is detected, the greater the chances of recovery. For a breast cancer detected at an early stage, 5-year survival is 99%, and 27% for a metastatic cancer 1 . The difference in the detection stage also influences the cost of medical care 2 . While high risk women 3 receive an individualized monitoring, the majority of women are included in the nationwide organized breast cancer screening. In France, this screening consists of a mammography (2 views) every two years with a double reading of mammographic images.
In April 2017, the Ministry of Health published an action plan to renovate the organized breast cancer screening 4 based on the recommendations of the National Cancer Institute. This reorganization aimed to enable a more personalized organized screening, and encouraged research projects on tools and methods for assessing the level of risk, including scoring. We wondered whether, among the 10 million women at "average-risk" targeted for the French organized screening 5 , some women would be at lower-risk than average risk and others at higher risk.
Several predictive models already exist, such as Gail 6 , BRCAPRO 7 , BODICEA 8 , or Tyrer-Cuzick 9 models. But they are not adapted to the European population 10 and most of them include BRCA1 and BRCA2 mutations. However, in France, women presenting with these mutations are not included in the organized nationwide screening. Moreover, it is difficult to estimate an individual risk precisely 11,12 . We, therefore, intended to set up a new risk model, including new variables, adapted to the French population, and enabling to explore whether in this average risk population it was possible to identify women at lower risk of breast cancer (or higher risk, respectively) in order to adapt the organized screening accordingly. Database We used the database of the Center that coordinates breast cancer screening (CRCDC-OC) in the Hérault department, France. This database registers women with precise, verified, and consolidated information on the development or not of breast cancer during the follow-up of their screening. We enrolled women, aged 50 to 60 at inclusion, who experienced a 12 year breast cancer screening from 2006. We extracted a database of 33,858 women, including 805 women who developed a breast cancer in 3-12 years following inclusion. We excluded women who have been diagnosed with breast cancer during the first two year of follow-up to avoid including women who might have already developed a cancer but not yet diagnosed at inclusion. Data were recovered by a telephone survey. It was not possible to interview 33,000 women. Therefore, on one hand we selected all the women who developed a breast cancer during the follow-up period, and on the other hand, we randomly selected 5000 women among those who did not develop breast cancer.

Power
We hypothesized a risk ratio ≥ 2, with a ratio of 4 to 1 between women who did not develop breast cancer and those who did during the follow-up period, together with a power of 80%. It gave a required sample size of 543 women with breast cancer during the follow-up period and 2174 women without breast cancer.

Study questionnaire
We looked for the risk factors for breast cancer in average risk women reported in the literature prior to creating the study questionnaire. We elaborated a text mining algorithm-assisted search (TEMAS) (available at shiny.temas-bonnet.site) to optimize the bibliographic search on very large volumes of scientific articles. We also included the variables used in Gail's model relative to women at average risk. The study questionnaire is provided in Supplementary material 1.

Data collection
Data were collected by telephone by two trained investigators. An explanatory leaflet (Supplementary material 3) was sent to all participants prior to the phone call and informed consent was obtained from all individual participants prior to the telephone interview. Responses to the study questionnaire were collected between November 2018 and May 2019. The inclusion flowchart is shown on Figure 1. This survey enabled collecting 3077 questionnaires, including 597 questionnaires from women who developed a breast cancer. The Case Report Form was developed with the ACCESS software. The analysis database was provided by the CRCDC-OC while having anonymized the data beforehand. This study complies the MR003 CNIL's (National Commission for Informatics and Liberties) methodology (Declaration: 2200402v0). The study was also agreed by the ethics committee of the CPP (Personal Protection Committee) (A0246039). All methods were performed in accordance with the relevant guidelines and regulations.

Statistical analysis
Data management was performed on the fly. Statistical analysis was carried out using R software 13 .

Descriptive analysis
We discretized the variables according to the literature. For duration of oral contraceptive 14 we retained the classes 0, [1,5[, [5,10[, and ≥ 10. For the number of biopsies, we complied with Gail's classes 6 0, 1, ≥ 2. Breast density was categorized in four classes according to the ACR BI-RADS (Breast Imaging-Reporting and Data System) Atlas 15 : A: the breasts are almost entirely fatty, B: there are scattered areas of fibroglandular density, C: the breasts are heterogeneously dense, and D: the breasts are extremely dense. For the number of hours per week of physical activity we followed WHO recommendations 16 : 0, [1-5[, and ≥ 5. Alcohol consumption (in glasses per week) was discretized as follows : 0, [1,10[, ≥ 10 according to Santé Publique France (French public health service) recommendations 17 . Age at menarche was divided as follow: < 12 and ≥ 12 18 . Age at first live birth was discretized in 3 classes 19 : nulliparous, ≤ 24, and > 24. For number of children, we used the average number of children in the generation of interest 20 : 0, [1,3[, and ≥ 3. For the total duration of breastfeeding we followed WHO and UNICEF recommendations 21 : 0, [1,7[, and ≥ 7. For age at menopause, we used the definition of the national college of French gynecologists of early menopause 22 : < 41, ≥ 41. We created a socio-economic hardship binary variable that was "yes" if the interviewee presented financial hardship and / or was not going on vacation, "no" otherwise. We also studied overweight (i.e. BMI≥25). We also studied detailed family history of breast cancer : mother, sister, daughter, and aunt. All the quantitative variables were also tested as continuous variables.

Cox model
All variables with a P-value <0.2 in the univariate models were included in a global Cox model, then a selection with the stepwise Akaiké information criterion (AIC) was achieved (age was forced in the model) in order to obtainβ , the estimated coefficients of the Cox proportional hazards regression model. Schoenfeld residuals were computed to check the proportional hazards assumption.

2/10
Our goal was to define an individual score. Therefore, to estimateΛ 0 , the cumulative base-line hazard function, we used Breslow's estimator 23 defined below. with:

the estimated coefficients of the Cox proportional hazards regression model
It enabled to estimate a risk function F for each individual as a function of its covariates : As we performed a random sampling among women without breast cancer, in our database, the distribution between women who developed cancer and the others was no longer representative of the starting population. To remedy this, we performed bootstrap samples. We computed 100,000 bootstrap databases using stratified random sampling with replacement among women for whom we obtained answers to all the covariates of the final Cox model, in order to obtain a database close to the initial CRCDC database (i.e. 33,858 women, including 896 who developed a breast cancer within the 12 years of follow-up) In these databases, mean risk was identical to the one of the CRCDC database. Then, we estimated, for each woman, Λ 0 (t), and calculated F(12|Z) to obtain a risk score of developing a breast cancer during the follow-up period.

Thresholds and likelihood ratios
In France, the target population of the organized screening is greater than 10 million women. Current participation is about 50% 5 with a target of 65%. The number of women actually screened is therefore between 5 and 6.5 million. Among women at average risk, we hypothesized that 10% of them might have a risk different from the "average" risk, either lower or higher. This average risk was 2.6% in this population. From a clinical standpoint, we looked for women that might have a risk two times lower than this "average" risk. This group was considered as at lower-risk. We also looked for women that might have a risk two times higher than this "average" risk. This group was considered as at higher-risk. We then searched for the corresponding thresholds. We looked for s 1 , the higher risk threshold, and s 2 , the lower risk threshold, such as: where: • M represents the event "developing a breast cancer during the 12 years follow-up"; • R represents the risk as determined by the score.
This method then enabled to delineate 3 groups according to the level of risk: lower risk, average risk, and higher risk. Given that, for each threshold s, the likelihood ratios (LR) either positive (+) or negative (-) where as follows: We plotted the curves LR+(s) and LR-(s) curves, and looked for s 1 and s 2 according to the calculations available in Supplementary material 2.
Women with a risk score below the "low threshold" will be considered at lower risk of breast cancer, women with a risk level between "low threshold" and "high threshold" will be considered at average risk and finally, women with a risk score above the "high threshold" will be considered at higher risk of breast cancer in this population of women at average risk.

Results
Personalized organized breast cancer screening retrospective cohort: descriptive analysis Descriptive analysis results are presented in Table 1. The significant risk factors in the univariate analysis were the number of biopsies (p = 0.042), breast density (p = 0.012), total duration of breastfeeding (p = 0.026), financial difficulties and/or not going on vacation (p = 0.002), mother history of breast cancer (p = 0.001), and aunt history of breast cancer (p = 0.029).
We explored whether a selection bias might have occurred. We analyzed the distribution of the two variables (age and breast density) we had for women who could not be reached or who refused to answer our questions. The results are shown in Tables 2 and 3. The only significant difference was observed for breast density of women who did not develop cancer during follow-up. The women who responded to the survey had significantly higher breast density on average. So the women selected, in the group without breast cancer, were more at risk than those who did not respond. Under these conservative conditions, we considered that the selection bias would not penalize the analysis. Moreover, we did not retained the fact that breast density as such was a criterion of choice for women to answer the questionnaire or not.

Cox model
The final model, presented in Table 4, has been built with 2768 women for whom we had answers to all the covariates of the model, including 2241 non-cancer patients and 527 women who developed cancer during the follow-up period.

Bootstrap
We computed 100,000 bootstrap databases using random sampling with replacement among the 2768 women for whom we had answers to all the covariates of the final Cox model. The risk function F(t) was calculated for all women in the 100,000 databases.

Threshold determination
Pre-test prevalence of the disease was the percentage of breast cancer in the CRCDC-OC database i.e. 2.6%.
So, as described in Supplementary material 2, we looked for : • s 1 such as ∀s ≥ s 1 : LR + (s) ≥ 2.054 • s 2 such as ∀s ≤ s 2 : We then plot LR+(s) and LR-(s) in Figure 2. This method gave a low threshold at 1.9% and a high threshold at 4.5%.

Figure 2.
Determination of the three risk groups with optimal thresholds in women at average risk of breast cancer: "Lower risk", "Average risk" and "Higher risk".
On one hand, 9.39% of women presented a risk lower than 1.9%. Among this group, the mean risk was 1.37%. On the other hand, 2.85% of women presented a risk higher than 4.5% with a mean risk of 5.84%.

Discussion
In a retrospective cohort of French women, aged 50 to 60, at average risk of breast cancer and followed-up for 12 years, from 2006 to 2018, who underwent the organized screening, we identified two additional levels of risk, either twice lower than the average-risk, or twice higher.
Our partition into three groups at risk of breast cancer in these women showed that the level of risk in the so-called "average-risk" population is not homogeneous. Depending on the presence or level of clinical, genetic or biological markers, treatment strategies are personalized to be as effective as possible. This approach can be extended to public health and in particular for primary and secondary prevention. Indeed, offering personalized prevention strategies could increase adherence to the program, and effectiveness as well. Thus, identifying new groups at risk among women at average risk offers new perspectives.
Women's wishes for a more personalized organized screening 24 therefore appears attainable. Indeed, women interviewed during the French citizen consultation on breast cancer screening in 2017, expected the organized screening to be more targeted, thanks to advances in research, and to become more personalized according to individual risk factors. Physicians, also consulted, mentioned the value of determining the level of individual risk 25,26 . They also ask for a score in French women in order to substantiate the levels of risk. This risk assessment would be used by the screening management structures to offer a personalized approach of the organized screening, adapted to the level of risk of each woman. We developed a user friendly computer application in this perspective.
The value of the organized screening is undeniable. In France, it enables a regular follow-up of invited women, aged 50 to 74, to carry out a free mammogram in the radiology office of their choice every two years. Moreover, if the initial assessment does not present any anomaly, a second reading of the mammography is systematically performed by another radiologist. The second independent reading is a guarantee of reliability, with almost 8% of the breast cancers detected by the independent second reading 27,28 . The principle of an organized breast cancer screening has been challenged given the claims of over-diagnosis or over-treatment. A reassessment of the organized screening that would allow a personalization according to the individual level of risk, would enable a standardized and quality monitoring of women while producing evaluative data.
We identified three levels of risk with a risk ratio of at least 2 between the group at average risk and the group at lower risk on one hand, and on the other hand with the group at higher risk. It lead to fairly large group sizes. The choice of a risk divided or multiplied by a factor two remains arbitrary though clinically relevant. Moreover, sensitivity analysis could be carried out by varying this ratio according to the complementary hypotheses of epidemiologists or physicians specialized in this field. Furthermore, the choice could be supported by a cost function taking into account monetary cost and years of life adjusted for quality of life.
According to the two thresholds of lower and higher risk, 12% of women were thus identified, corresponding to more than 600,000 women in the French population. In this population at "average" risk of breast cancer, a new screening strategy could therefore be proposed for the groups at lower or higher risk. For the group at average-risk of breast cancer, corresponding to 88% of the presently screened population, the current screening rate, i.e. a mammogram every two years, would be maintained as such.
For the group at lower risk that includes 9.4% of the screened population with a level of risk of 1.37% of developing breast cancer within the 12 years of follow-up, we propose to reduce the frequency of mammograms with a mammogram every three or more years instead of the two years currently in force. This mammogram spacing could reduce the risk of over-diagnosis while maintaining a regular follow-up. Moreover, it would reduce, the risk of radiation-induced cancer, even if it is already very low. However, the risk of interval cancer can't be excluded. This proposal will have thus to be tested with an appropriate design.
For the 2.9% of women at higher risk, increasing the frequency of mammograms is possible. However, high breast density is an important risk factor for this group. In this group, the distribution of high or very high breast density was 60% compared to 28% in the group at average risk and 3% in the group at lower risk. Furthermore, the sensitivity of the mammogram falls to 63% 29 for very dense breasts. The sensitivity of cancer detection via a mammogram appeared inversely proportional to the breast density 30 . Thus, it did not seem that increasing the frequency of mammograms would bring a significant benefit. Based on recent results, an alternative would be to offer these women an abbreviated breast MRI, which appeared associated with a higher rate of breast cancer detection among women with dense breasts 31 , while maintaining a two year frequency for surveillance. This would bring the care strategy for these women at higher-risk closer to the strategy applied to women already identified at high risk (i.e. different from women here identified in the "average" risk population), and benefiting from an individual follow-up 3 .
Such a personalization of the organized screening might make women feeling more involved. For women at lower-risk, a mammogram every three years may be less binding than every two years. For women at higher risk, offering a more specific examination may increase adhesion to the organized screening. These strategies could then increase the participation of French women to the organized screening which is currently around 50% 5 , though the target is 65%.
As previously stated, the cost of the case finding (including diagnosis and treatment of patients diagnosed) should be economically balanced in relation to possible expenditure on medical care as a whole 32 . An analysis of the medico-economic gain must be carried out taking into account, in particular, interval cancers, over-diagnosis and the stage of the cancers detected. This study proposes a new approach to personalize the organized screening by stratifying the risk within the population at "average" risk and currently considered as homogeneous. However, the results of this study remain to be confirmed in other contexts to validate whether the proposed strategy could be implemented.
Six variables were included our the score of breast cancer risk, two of which (age and breast density) are already included in the databases of the different French CRCDCs. Only four more questions are required to be asked in order to establish the score. This information is easy and simple to collect at the time of mammogram. Moreover, an internal validation on another database, for example of women aged 40 to 50 in the Hérault department, could be performed. A double external geographic validation could be carried out on the databases from other CDCRCs and a database from another country 33 .
We collected information of interest for the creation of our score at t 0 , the inclusion time. But, over time, the value of variables retained at t 0 may change and the risk at t 0 + x years may also change. It will therefore be necessary to regularly recalculate women risk of breast cancer in order to assign them to the appropriate group of risk.
A re-evaluation of the score will also be necessary on regular time intervals to reappraise the coefficients estimated by the model in order to evaluate whether they would need to be adjusted.
Our proposal for a new personalized organized screening must be validated by other studies.

Availability of data and material
The data used and analyzed during the current study are available from the corresponding author (EB).

Code availability
The code used during the current study is available from the corresponding author (EB).

Authors' contributions
All authors contributed to the study conception and design. Material preparation and analysis were performed by E.B. The first draft of the manuscript was written by E.B. and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.