The data were published in the publicly available SEER database; therefore, all patient-identifying information was removed, and there was no need for approval from the Ethics Committee or informed consent from the patients. All data were approved by the National Institutes of Health before being obtained (approval number: 10934-Nov 2017). In this study, postoperative patient data on malignant PT-like breast tumors from 2004 to 2015 were obtained from the SEER database using the SEER-Stat 8.3.6 software.

The study included the patients according to the following inclusion and exclusion criteria. The inclusion criteria were 1) malignant breast PTs; 2) breast cancer as the first primary malignant tumor; and 3) a follow-up duration of >1 month. The exclusion criteria were 1) unknown tumor grade; 2) lack of follow-up data; 3) unknown information about tumor-node-metastasis (TNM) stage; 4) unknown tumor size; 5) unknown race; and 6) presence of multiple primary tumors. The following information was extracted from 508 eligible patients: pathologically proven mammary phyllodes malignancies (ICD-O-3 9020/3), single-sided breast cancer, American Joint Committee on Cancer (AJCC) stage, patients diagnosed from 2004 to 2015, size of primary tumors (T), tumor grade, race, age, primary lesions, regional lymph node metastasis (N), distant metastasis (M), radiation therapy, and marital status. Although all 508 patients who met the group criteria underwent surgery, the specific details of the interventions were not provided in the SEER database; therefore, surgery could not be included as an independent prognostic variable in the model study.

In the retrieved information, low-level tumors had high and moderate differentiation grades (ICD-O-3 1 and 2 levels), whereas high-level tumors had low or undifferentiated grades (ICD-O-3 3 and 4 levels). The critical values for diagnostic age and tumor size were determined using the X-tile software (Yale University, New Haven, CT, United States) (Camp et al., 2004), which has previously been shown to determine the critical point of the optimal tumor variable. The X-tile software was initially developed to determine the optimal threshold for breast cancer variables. The optimal threshold values for tumors were small (<70 mm), medium (70–160 mm), and large (>160 mm) (Figure 1). Because the optimal age limit was between 57 and 67 years (Figure 1), the patients were divided into three age groups (18–57, 57–67, or >67 years), as well as white, black, or other ethnic groups depending on race. Radiotherapy was divided into radiotherapy and non-radiotherapy. However, details such as radiotherapy type, site, and strength could not be obtained from the SEER database. Chemotherapy was divided into recipients and non-recipients.

Statistical analysis was conducted using the SPSS 25.0 software (IBM Corp., Armonk, NY, United States) and R language software (version 3.6.1, company, state, country). Patients were randomly divided into training (n = 356 cases) and validation (n = 152 cases) groups using R with a random sampling function. The univariate Cox regression analysis was performed to identify prognostic factors in the training group, and statistically significant variables (p < 0.05) were incorporated into a multivariate Cox risk regression model. Furthermore, the nomogram graph was constructed, with the C-index mapping the subject’s working receiver operating characteristic (ROC) curve and correction curve for decision curve analysis (DCA). In the correction curve, the closer the curve is to the ideal 45° guide, the closer the prediction is to the actual observation. The C-index and the area under the ROC curve (AUC) were used to evaluate the predicted value of the nomogram graph; the maximum C-index was 1 (representing 100% prediction accuracy), and the minimum value was 0.5 α. The closer the C-index was to 1, the better the accuracy of the prediction model. Using the DCA method, we could effectively determine the certainty and benefit of the prediction model. An α level of 0.05 was used.

According to the inclusion and exclusion criteria, 508 patients with malignant breast PTs between 2004 and 2015 were screened out of the SEER database. Table 1 shows the key characteristics of patients with PTs. The patients were randomly divided into training (n = 356) and validation (n = 152) groups. The training group was used to construct the nomogram model, and the validation group was used to validate the model.

For further stratified analysis of age and tumor size, the X-tile graph determined the critical values of age to be 57 and 67 years, whereas the critical value of tumor size was determined to be 70 and 160 mm. To explore the effect of age-size thresholds on overall survival, we first divided the patients into three risk groups using 57 and 67 truncation values: 18–57 years old(a), 57–67 years old(b), and >67 years old(c). Furthermore, we found that age was an important prognostic factor in patients with malignant breast PTs; the worst prognostic factor for patients with malignant PTs was an age of >67 years, whereas patients aged <57 years had better overall survival rates than other groups (Figure 1A). Additionally, patients with a tumor size of >160 mm(c) had the worst prognosis, whereas those with a tumor size of <70 mm(a) had a better prognosis than the other groups (Figure 1B).

To determine the independent prognostic factors for overall survival, we first performed univariate Cox regression analysis in the training group (n = 356). Factors such as older age, larger tumor size, race, tumor grade, chemotherapy, advanced AJCC stage, T stage, N stage, and distant metastasis (M) were significantly associated with poor overall survival (Table 2). These nine factors were further included and selected by the multivariate Cox proportional hazard regression analysis. These results showed that age at diagnosis, tumor grade, tumor size, N stage, and distant metastasis (M) were significant and independent prognostic factors for overall survival.

We first combined age, tumor size, N stage, distant metastasis (M), and tumor grade to construct a prognostic nomogram model for predicting the 1-, 3-, and 5-year overall survival in patients with malignant breast PTs (Figure 2). The model can score each prognostication factor on the score scale. By combining these scores with the total score of the lowest score, the 1-, 3-, and 5-year overall survival predictions for patients with malignant PTs of the breast can be calculated. The predictive accuracy of the final prognostication model was evaluated by the C-index. The C-index was 0.845 (95% confidence interval [CI], 0.802–0.888) and 0.784 (95% CI, 0.688–0.880) in the training and validation groups, respectively. The calibration plots revealed perfect consistency between the model-predicted and actual survival values (Figures 3A, C). Surgeons can easily use these prognostic models to determine the prognosis of patients with malignant PTs: age, tumor size, N stage, distant metastasis (M), and tumor grade.

Then, the predictive ability of the nomogram model was evaluated using the AUC (an AUC value of 0.5 indicates that the nomogram model has no predictive effect, and an AUC value of 1 indicates that the nomogram model has an excellent predictive effect). Patients with different survival rates were well distinguished (the higher the value between 0.5 and 1 is, the better the nomogram’s ability to distinguish), and other prognostic factors could be compared to further validate their superiority and determine which factor has the most superior predictive accuracy. In the training group, the advantages of the differentiation for overall survival shown in the nomogram were demonstrated compared to the other clinical information (1-year AUC: 0.908; 3-year AUC: 0.881; 5-year AUC: 0.879; Figures 4A–C). In addition, in the validation group, the AUC values for the 1-, 3-, and 5-year survival rates were 0.895, 0.843, and 0.807, respectively.

The nomogram offers a considerable positive net benefit from the risk of mortality in both the training and validation groups (Figures 3B, D). It can predict the 1-, 3-, and 5-year overall survival rates and has good value in clinical application. Using the nomogram model (Figure 2), we can predict a patient’s survival probability based on his/her personalized information. For example, in a middle-aged woman diagnosed with breast PTs with a medium-sized primary lesion and tumor grade III, the patient’s overall survival score was 10.5 according to the model. As a result, the patient’s 1-, 3-, and 5-year overall survival rates were estimated to be 90%, 82%, and 80%, respectively.