Prognostic Nomogram and Risk Stratification System for Breast Cancer Bone Metastasis: A SEER-Based Population Study

Hai Lu guang zhou zhong yi yao da xue: Guangzhou University of Chinese Medicine https://orcid.org/00000002-3795-6087 Jinqun Jiang Shantou University Medical College YuZhu Zhang Guangzhou University of Traditional Chinese Medicine: Guangzhou University of Chinese Medicine Rui Xu Guangzhou University of Traditional Chinese Medicine: Guangzhou University of Chinese Medicine Liping Ren Guangzhou University of Traditional Chinese Medicine: Guangzhou University of Chinese Medicine QianJun Chen (  cqj55@163.com ) Guangzhou University of Traditional Chinese Medicine: Guangzhou University of Chinese Medicine


Introduction
Breast cancer (BC) is the most frequently diagnosed female cancer and the second leading cause of cancer death [1,2].BC cells most frequently metastasize to bone, with up to 75% of stage IV BC patients developing bone metastases [3]. Multiple systemic organ metastases are common in BC, and in 17-37% of patients, the disease is limited to the skeleton. Patients with bone-only rst metastasis tend to experience a better prognosis than those with other-only rst metastasis. Patients in whom the disease remains con ned to the bone have longer survival than those with subsequent visceral involvement. Bone metastasis can result in poor survival, considerable morbidity, intractable pain, and decreased quality of life [4].
Additionally, surgery or radiotherapy for patients with bone metastasis can provide effective local control and improve quality of life, especially for patients with pathologic fractures.To our knowledge, the risk factors and their effects on prognosis of patients with BC and bone metastasis have rarely been explored.
Zhaoming Ye et al. reported that tumor grade, histologic type, primary tumor size, tumor subtype, surgery, chemotherapy, and number of metastatic organs except bone were signi cant independent variables of both overall survival (OS) and cancer-speci c survival (CSS) [7].
In recent years, nomograms have gained increasing attention as strong prognostic statistical models with intuitive graphs to quantify risk by incorporating important factors for oncology prognostics [8]. They are a useful and convenient tool to quantify and predict risk and prognosis in cancer patients. However, no systematic attempts have ever been made to develop prognostic nomograms for BCBM patients.
Therefore, in the present research, we intended to establish and validate a nomogram for those patients and assist clinicians to accurately predict patients' survival.

Study Design and Patients Selection
By using SEER*stat version 8.3.5, we selected eligible patients with breast cancer between 2010 and 2015. The inclusion criteria were histologically con rmed breast cancer, age at diagnosis (codes: ≥ 18), primary tumor site (breast), and initial bone metastasis. We excluded patients with unknown follow-up information and missing information on tumor stage, histologic grade, tumor size, therapy, or marital status. The selection ow chart is shown in Fig. 1. Finally, 3885 eligible patients were enrolled into the study and randomly divided into the training cohort (70%, n = 2721) and internal validation cohort (30%, n = 1164). For this type of retrospective study, there was no need for formal consent.

Statistical Methods
The chi-square test was used to compare the rates. Univariate comparisons of survival data were performed using the Kaplan-Meier method and Cox univariate analysis. Based on the results of the univariate analysis and combined with clinically important factors, further multivariate analysis using the Cox risk regression model with backward elimination was performed. Following the multivariate analysis, variables with a P < 0.05 were selected for developing the nomogram.
We used 1-, 3-, and 5-years OS for the analysis in the nomogram. One thousand bootstrap resamples were used to calculate C-indexes and generate calibration plots, which assessed the predictive accuracy of the nomogram. Furthermore, a risk strati cation model was developed on the basis of each patient's total scores in the nomogram to divide all of the cases into two prognostic groups according to its median value.
All of the analyses were performed using R (http://www.r-project.org) and Empower (R) (www.empowerstats.com, XY Solutions, inc. Boston MA). Statistical signi cance was reached when Pvalue was lower than 0.05 in a two-tailed test.

Patient characteristics
A total of 3885 eligible female patients with BCBM from 2010 to 2015 were retrieved from the SEER database (2721 patients for the training cohort and 1164 patients for the testing cohort). The sociodemographic and clinicopathologic characteristics of the three cohorts are summarized in Table 1. The major tumor subtype for the training and testing cohorts was luminal A. In the training cohort, 458 of 2721 patients (16.8%) were diagnosed with the luminal B subtype, whereas 158 patients (5.8%) were diagnosed with the HER2-enriched subtype. Most of the patients were White (76.0%) and the median age was 59 years. Survival analysis was conducted on the selected group of patients based on different clinical characteristics. The median OS for the training and testing cohorts was 24 months.

Survival Analysis
The Kaplan-Meier analysis determined the impact of variables on survival. The results generated by the log-rank test are listed in Table 2. We found that patients with bone metastasis had poorer survival than those without distant metastasis of breast cancer (P < 0.0001) ( Fig. 2A). In the training cohort, bone combined with other-site metastasis was associated with worse prognosis (P < 0.0001) (Fig. 2B). Among all of the tumor subtypes, the luminal B subtype had the most favorable survival, whereas triple negative breast cancer(TNBC) was associated with the worst prognosis (P < 0.0001) (Fig. 2C). Married patients had better prognosis than unmarried BCBM patients (P < 0.0001) (Fig. 2D). Black patients had a relatively worse prognosis than White and other races (P < 0.0001) (Fig. 2E). With respect to the factors associated with treatments, we found that patients who had undergone chemotherapy (yes vs. no/unknown) experienced prolonged survival (Fig. 2F).

Independent prognostic factors in the training cohort
The univariate Cox regression analysis demonstrated that age, race, marital status, tumor size, tumor subtype, grade, T classi cation, surgery, radiotherapy, chemotherapy, brain metastasis, liver metastasis, and lung metastasis were associated with OS. All of these factors were entered into the multivariate Cox regression analysis; age, race, marital status, tumor subtype, grade, T classi cation, surgery, chemotherapy, brain metastasis, liver metastasis, and lung metastasis were found to be independent prognostic factors after multivariate analysis ( Table 2).

Constructing and Validating of Nomogram
For predicting the overall survival of patients, these six signi cant independent factors were incorporated to construct a nomogram (Fig. 3). The score of each category was given on the point scale axis (Table 2). A total score was easily calculated by adding each single score; by projecting the total score to the bottom scale, we were able to estimate the probabilities of 1-, 3-, and 5-year OS for individual patients.
The C-index of the nomogram was larger than that of the seventh version AJCC-TNM staging system (0.704 vs 0.564, P < 0.001), suggesting that this model had an acceptable predictive accuracy. In addition, calibration plots of the nomogram were also developed (Fig. 4); they demonstrated that the predicted OS agreed well with the actual observations. In addition, decision curve analysis is a net bene t analysis that compares the true-positive to the weighted false-positive rates across different risk thresholds that a clinician/patient might want to accept; this analysis was performed evaluating 3-year OS of BCBM patients. As shown in Fig. 5, all of the models had a better net bene t compared with the "treat all" strategy. The net bene t of the nomogram was higher than that of the TNM stage model across most of the modeled decision threshold probabilities.

Risk strati cation system
These results proved the nomogram's e cacy in predicting survival. Thus, we calculated total points according to the nomogram-predicted score. Patients were classi ed into two risk groups according to the median points as follows: low risk (total score < 3094.04) and high risk (total score ≥ 3094.04). In addition, we strati ed the entire cohort according to the tumor subtype. Within each subtype, the survival rates predicted by the nomogram showed signi cant distinctions between the Kaplan-Meier curves ( Fig. 6).

Discussion
The AJCC-TNM staging classi cation is the most widely used system for predicting survival and selecting clinical strategies for patients with cancers [9,10]. However, this classical system cannot always accurately predict the difference in survival between different stages [11]. Furthermore, survival of patients with the same stage varies widely. An important reason may be that we ignore many of the factors that have been con rmed to be highly associated with survival. To solve this problem, we developed a nomogram, a more comprehensive, accurate, and useful prognostic model.
As few studies have established nomograms for predicting the survival of patients with BCBM; their sample size was small and the prognostic factors were limited. Thus, we developed a clinical nomogram to predict the survival based on the SEER database. The SEER registry is the largest population-based database of cancer patients in the United States, covering approximately 26% of patients diagnosed with cancer [12]. We reviewed patients' data from the latest version of the SEER as released in 2015 (covering 18 registries, 1973-2015) by using SEER*Stat version 8.3.5, and we also set a strict inclusion and exclusion criteria.
In this study, we identi ed 11 independent prognostic predictors of BC with bone metastasis. Tumor subtype was a signi cant factor of OS, which is in accordance with the previous studies [13]. It has also been reported that TNBC subtype, an aggressive form, shows the worst prognosis in BC patients with brain metastasis, which is consistent with our results [14]. Marital status was found to be an independent predictor of survival among BC patients with bone metastasis. Patients in the married group had better survival compared with those in the unmarried group. Tumor grade is usually recognized as an important factor of survival among BC patients [15,16]. Our multivariate analysis also revealed that well-differentiated tumor was signi cantly associated with increased survival. Moreover, patients with boneonly metastasis had better survival than those with additional distant metastases.
For metastatic BC, chemotherapy is recommended as it can prolong survival, decrease cancer-related complications, and improve quality of life [17,18]. Our research also revealed that BCBM patients who had received chemotherapy achieved survival bene ts. It is generally accepted that radiotherapy has the potential to alleviate pain and achieve good local control. Some studies reported that breast radiotherapy is associated with improved survival in patients with metastatic disease. 19 However, our multivariate analysis failed to identify radiotherapy as a signi cant predictor of OS. Nava et al. also supported this nding and showed no effect of breast radiotherapy on survival in patients with metastatic disease [19].
For validation of the nomogram, to guarantee that the model could be generally applied and to avoid over tting, it is necessary to evaluate the discrimination and calibration. Discrimination has usually been evaluated with the C-index, and calibration is assessed by comparing the agreement between the predicted and the actual survival of patients [20]. The results indicated that our nomogram had a better discriminating and predicting ability than the traditional staging system. Beyond that, we also performed the decision curves analysis to study the clinical net bene t of the nomogram for BCBM patients' prognosis [21,22]. The results showed that this model improved the clinical net bene t across all the threshold probabilities.
In addition, we constructed a system to classify patients to two risk subgroups based on predicting the total scores. When the risk strati cation system was applied in patients with the different tumor subtype, it discriminated OS well in each subtype (Fig. 6). Thus, the results con rmed that this risk strati cation system based on the nomogram was an accurate and reliable prognostic model. It could help clinicians to identify the patients with high risk and to perform individualized adjuvant treatment, which might also be helpful in this highly selective cohort.
Nevertheless, the present study had several limitations. First and foremost, this was a retrospective analysis, which is inherent for the SEER database. Although we performed multivariate analysis to minimize confounders associated with the heterogeneities, the retrospective nature of this study must be considered when interpreting the results. Second, the SEER database does not contain data about recurrence or speci c treatment, which may affect the clinical outcome. The third limitation was that other important factors, such as speci c site of bone metastasis or treatment for bone metastasis, were not included in the database.

Conclusion
The current study comprehensively analyzed the prognosis of patients with newly diagnosed BCBM based on the SEER population-level data; we developed a tool for assessing the individualized survival estimates in patients with BCBM. The developed nomogram can provide more accurate survival information for clinicians and help them to provide appropriate treatment measures for metastatic lesions. More external validations are recommended to further re ne our conclusions.   Abbreviations: 95% CI, 95% con dence interval; HR, hazard ratio; YOD, year of diagnosis. *p<0.1 was considered signi cant in univariate Cox-Regression analysis. ** p<0.05 was considered signi cant in multivariate Cox-Regression analysis. Figure 1 Flow chart of the selection of the study population.

Figure 4
The calibration curves predicting 1-, 3-, and 5-year overall survival in the training cohort (A, B, C) and validation cohort (D, E, F).