Risk prediction for breast Cancer in Han Chinese women based on a cause-specific Hazard model

Breast cancer is one of the most prevalent malignancies among women worldwide [1]. Although the incidence of breast cancer is low compared with Western countries, China is currently experiencing increasing trends in both breast cancer incidence and mortality [2]. However, the mammography screening participation rate is only 21.7% in China, far lower than in Western countries [3]. Considering the limited medical resources, especially in rural areas of China, a risk prediction model that is suitable for general population screening is urgently needed.

Competing risks are said to be present when an individual is at risk for more than one mutually exclusive event, such as death from a different cause, and the occurrence of one of these competing events will prevent the event of interest from ever happening. A cause-specific hazard model considers competing risks and therefore has better performance than the Cox model when used for disease risk prediction. Gail proposed a method to estimate individual probabilities of developing breast cancer, based on a cause-specific hazard model [4]. Several risk prediction models have been developed in Western countries; the two most widely used are the Gail cause-specific hazard model with traditional risk factors as predictors [4], and the International Breast Cancer Intervention Study (IBIS)model, which includes genetic markers [5]. The Gail II model was developed based on the Gail model and provides a feasible web-based instrument [6]; however, that model was first developed in a Caucasian ethnic population so the effects maybe uncertain when directly applied to Han Chinese women [4, 6‐11]. Moreover, because biopsies are not widely used in China (especially in rural areas), information on the number of previous breast biopsies is not readily available for most Chinese women. Although other prediction models include genetic risk prediction models such as the IBIS model [12], their higher cost may lead to limited application considering the large population in China. In addition, the most important genetic factor of breast cancer, the BRCA gene mutation rate, is lower in the Chinese population than those in Western countries [13]. The aim of our study was to develop a breast cancer risk prediction model using non-laboratory indicators in a population of Han Chinese women, and to convert the model into a simple tool that can be conveniently used in clinical and public health practice.

We developed a risk prediction model for breast cancer for use in Han Chinese women. The results of validation showed good calibration and discriminative ability, with E/O ratio 1.03, and C-statistic 0.64 (95% CI: 0.55, 0.72). Using this model, women with a high risk of developing breast cancer can be identified using simple data collection. With the model, women identified as having high risk can be selected for further breast cancer-related examination, such as mammographic screening. Consequently, there is greater likelihood of identifying women with early-stage breast cancer. Women with higher risk might be motivated to maintain their current health status and take measures to prevent or delay the onset of breast cancer.

Competing risk is commonly existing in disease risk prediction. Traditional survival analysis models (like the Cox model) treat deaths from competing causes as independent censoring events. Hence, bias would be induced into risk prediction [19]. The proportional cause-specific hazard model is commonly used to analyze competing risks. The model computes absolute risk without assuming that these deaths act as “independent censoring” [19]. Therefore, we proposed a cause-specific hazard model to estimate absolute risk of an individual for the development of breast cancer.

Selection of the risk factors included in the current model was based on a systematic review of epidemiological studies as well as statistical analyses. In this study, we assessed a variety of factors that are considered to be associated with breast cancer including: number of abortions, age at first live birth, benign breast disease history, BMI, diabetes, breast cancer family history, and life satisfaction score; most risk factors have been confirmed in previous studies.

The relationship between diabetes and breast cancer is controversial. In our study, diabetes was a risk factor in univariate logistic regression (Additional file 2: Table S1) but it was not significant after multivariate-adjustment. This might be owing to the correlation between BMI and diabetes. Obesity is a major risk factor of diabetes. Weight-loss programs can lead to successful long-term weight loss and a decrease in the onset of diabetes [20, 21]. BMI has been found to be positively associated with increased risk of diabetes onset [22, 23]. Among women with gestational diabetes mellitus, BMI is significantly and positively associated with risk of progression from gestational diabetes mellitus to type 2 diabetes [24, 25]. However, in a meta-analysis, type 2 diabetes was found to increase the risk of breast cancer by 16% after adjustment for BMI [26]. One hypothesis is that hyperinsulinemia, as a potential risk factor of breast cancer and a marker of insulin resistance in obesity and type 2 diabetes, may account for the association among BMI, diabetes, and breast cancer [27‐29]. Some studies have reported that increased BMI is associated with increased insulin [30]; therefore, BMI may be a confounding factor of diabetes and breast cancer. Compared with the model including diabetes, the model without diabetes showed a change in the E/O ratio from 0.99 to 1.03 as well as in the C-statistic, from 0.642 to 0.637; the 95% CI of the C-statistic also varied from 0.56–0.73 to 0.55–0.72), as did the standard error (from 0.043 to 0.044). Although the inclusion of diabetes in the final risk prediction model increased the C-statistic by 0.005, it was not significant in multivariate regression and the increase in discrimination was not distinct; thus, we did not include diabetes in the final model.

Life satisfaction scores have not been used in previous breast cancer prediction models. Many articles have reported that negative life events, depression, anxiety, and other harmful psychological and mental factors are related to breast cancer [31‐33]. In our results, life satisfaction score showed a significance difference in both univariate and multivariate logistic regression. Life satisfaction status is a modifiable risk factor for which specific prevention measures can be implemented, which is important in community prevention of breast cancer.

Although additional nutrient variables including intakes of calcium, soy products, and iron have been related to breast cancer risk in some studies [34‐36], a detailed dietary assessment and supporting nutritional database would be needed to accurately capture nutritional intake, making such assessment unfeasible in most clinical settings. Therefore, we did not include dietary variables that require a detailed assessment.

In China, some studies have added genetic markers to prediction models, to improve discriminative accuracy. Several prediction models have been developed that include some genetic markers and limited environmental predictors, with discriminative accuracy of around 0.6 [18, 37]. Together with the high cost of testing for genetic markers and the need to obtain blood samples, it is very difficult to implement such risk prediction tools among the general population of China.

In our developed model, all included risk factors are simple and feasible to measure, which improves its convenience and lowers the cost of implementation in a large population. In Western countries, the Gail model is used widely in clinic decision-making [38]. However, application of the Gail model in China is limited because biopsy tests are uncommon. Moreover, several validation studies have been conducted in Asian populations and have reported poor performance of the Gail model [39]. In a comparison with the Gail model in which race was defined as Chinese-American, the predicted incidence rate may still be higher than the actual rate in Taixing. Considering the lack of risk predictors and lower incidence rate in Taixing, Gail model’s prediction accuracy may be biased. In China, Yuan et al. developed a risk assessment tool for breast cancer prediction among Chinese women. Although that model’s C-statistic was 0.64 (95% CI: 0.50, 0.78), the authors included patients from a breast cancer database with data obtained in first-round screening only, with no further follow-up; these data may not be appropriate to test the reliability of the model [18]. In our study, the results of comparison showed that without including competing risks in the prediction model, the HRA model seemed to overestimate the probability of developing breast cancer in the Taixing cohort study. In our validation cohort, the E/O ratio was near 1 and showed good projection.

The strength of this study is we developed a simple and efficient prediction model with good calibration and discrimination. The E/O ratio was 1.03, which is very close to 1. The C-statistic was 0.64, which showed that the HCBCP model performed well in the validation study. Modifiable risk factors were selected, such that intervention measures can be carried out in high-risk groups. Using this model, populations at risk can easily be screened, to identify those with increased risk who would benefit from health management advice, to avoid or delay the onset of breast cancer.

In conclusion, we developed a risk prediction model including fertility status and relevant disease history as well as other modifiable risk factors. The developed model demonstrated good discriminative accuracy. We expect that the newly developed model can be used to screen populations with a high risk of breast cancer and will contribute to primary and secondary prevention of breast cancer in China.