A nomogram for predicting breast cancer specific survival in elderly patients with breast cancer: a SEER population-based analysis

This study was based on SEER database data released in November 2020. Our target patients were extracted from SEER*Stat Version 8.3.9.2 (SEER ID: 26588-Nov2019), which included population-based data from 18 cancer registries covering approximately 28% of the United States (U.S.) cancer population between 1975 and 2018 and provided complete data regarding patient demographics, tumor characteristics, diagnosis, first course of treatment, and follow-up of vital status. Given that the data released by the SEER database are publicly available, this study did not require informed patient consent or ethical approval. We extracted data on patients with breast cancer, including chemotherapy records, who were treated between January 1, 2010, and December 31, 2016. A total of 252,472 breast cancer cases were identified in the database during this period (Supplementary 1). Among them, patients who met any of the following criteria were excluded:1) male sex; 2) aged <70 years at diagnosis; 3) breast cancer was not the first primary cancer diagnosed; 4) paired site; 5) without histologic confirmation; 6) missing stage or stage 0; 7) missing molecular type; 8) missing grade; 9) distant metastasis; or 10) death or loss to follow-up within six months of diagnosis. Ultimately, 33,118 eligible patients were included in the analysis. These patients were randomized at a 2:1 ratio into the training and testing groups.

We collected patient data, including age at diagnosis, race (white, Black, other, or unknown), marital status (married, divorced, separated, single, widowed, unmarried, domestic partner, or unknown), insurance status (insured, insured/no specifics, any medical, uninsured, or insurance unknown), grade (G1, G2, or G3), stage (I, II, III, or IV), tumor/node/metastasis (TNM) stage (T0–T4, N0–N3, or M0–M1), estrogen receptor (ER) status (negative, positive, or borderline), progesterone receptor (PR) status (negative, positive, or borderline), human epidermal growth factor receptor02 (HER2) status (negative, positive, or borderline), breast surgery procedure (partial mastectomy with or without axillary dissection, simple and subcutaneous mastectomy, modified radical mastectomy, radical and extended radical mastectomy with or without breast reconstruction, and other mastectomy or unknown), and chemotherapy and radiotherapy records. We defined the TNM stage according to the 7th edition guidelines of the American Joint Committee on Cancer (2010–2015). Detailed information about the variables can be found on the official SEER website (https://​seer.​cancer.​gov/​data-software/​documentation/​seerstat/​nov2020/​), and we strictly followed these definitions while conducting the analysis.

We defined BCSD as the time from diagnosis to death due to breast cancer. Death from other causes was defined as other cause-specific death (OCSD). We used the description from “SEER cause-specific death classification” to define the patient's cause of death.

We used a chi-square test to compare categorical variables. The cumulative occurrence of BCSD and OCSD was calculated using the cumulative incidence function (CIF). The difference between BCSD and OCSD CIFs in different subgroups, including those defined by age, race, insurance, marital status, grade, T stage, N stage, ER status, PR status, HER2 status, surgery method, and treatment with radiation or chemotherapy, were first compared using Gary’s test. Subsequently, the 1-, 3-, and 5-year CIFs for BCSD and OCSD in patients with breast cancer in the training cohort were calculated. In univariate analyses, risk factors were screened using the Fine-Gray model, and values with P < 0.05 were included in the subsequent multifactor analysis [16]. In the multivariate analysis for competing risks, the sub-distribution hazard ratio (sdHR) with a 95% confidence interval (CI) of each independent predictor associated with BCSD was calculated for the construction of nomograms. Based on the above analyses, a competing risk nomogram was constructed to predict the probability of BCSD in the 1st, 3rd, and 5th years after treatment [17]. The nomograph was verified using the 1000 resampling guidance method to assess its performance internally and externally. The concordance index (C-index) was chosen to quantify the predictive ability of the competing risk model [18]. The C-index ranges from 0.5 to 1, with values greater than 0.7 indicating better discrimination performance. And calibration curves were used to compare the predicted probability and observed frequencies, and the location of curve is closer to a 45° diagonal line, meaning a better-calibration. All analyses were performed using R software (version 4.1.3), and all tests were two-sided, and statistical significance was set at P < 0.05.

A total of 33,118 patients were included in this study, with 24,838 in the training group and 8,280 in the testing group. Clinicopathological and baseline characteristics are presented in Table 1.

Among all patients, 21,954 (66.3%) were aged between 70 and 80 years, 9,710 (29.3%) were aged between 80 and 90 years, and 1,454 (4.4%) were older than 90 years. Among all patients, 12.5% were ER-negative, 23.9% were PR-negative, and 90% were HER2-negative. In the entire population, a vast majority of patients underwent surgery, 82.8% of whom underwent partial mastectomy and 12.3% underwent mastectomy. Approximately half of the patients were treated with radiation; however, only a small proportion (15.9%) received chemotherapy.

In the training cohort, the number of patients (12.4%; 3,083/24,838) who died from non-breast cancer-related causes was higher than that of patients (5.7%; 1,412/24,838) who died from breast cancer. The 1-, 3-, and 5-year CIFs of BCSD and OCSD in patients with breast cancer in the training cohort are shown in Table 2 and Supplementary Table 2, respectively. The 1-, 3-, and 5-year CIFs of BCSD among the patients were 0.94%, 4.51%, and 7.23%, respectively, and those of OCSD were 1.46%, 8.11%, and 16.06%, respectively, which were almost twice the CIFs of BCSD. In the univariate analyses, most variables (P < 0.05) strongly correlated with the CIF of the BCSD, except for insurance (P = 0.110). Differences in the CIF of the BCSD are shown in Fig. 1. There was no relationship between CIF and ER status (P = 0.680), PR status (P = 0.069), HER2 status (P = 0.771), or N stage (P = 0.919). Multivariate analysis revealed the independent risk factors (age, race, marital status, grade, T stage, N stage, ER status, PR status, HER2 status, and treatment with surgery, radiation, or chemotherapy) associated with BCSD for the subsequent construction of a nomogram. The results of the multivariate analysis of competing risks in the training group are presented in Table 3. Among the identified factors, age, grade, T stage, and N stage positively correlated with the CIF of BCSD. Black and unmarried women have a higher risk of developing BCSD than those of other races and martial statuses. According to the sdHRs with 95% CI, the possibility of BCSD increased with grade, as observed in the T and N stages. Compared to patients who did not receive radiation or chemotherapy, those who underwent radiation (sdHR 0.760 [0.661–0.873]) or chemotherapy (sdHR 0.580 [0.514–0.653]) had a reduced probability of BCSD.

After model validation, all independent risk factors, including age, race, marital status, grade, T stage, N stage, ER status, PR status, HER2 status, and treatment with surgery, radiation, or chemotherapy were incorporated to construct a nomogram to predict the 1-, 3-, and 5-year CIFs of BCSD, as shown in Fig. 2. The probabilities of BCSD at 1-, 3-, and 5-years were predicted using the total score in the nomogram. As shown in Fig. 2, the N stage had the strongest effect on BCSD, followed by the T stage, age, and breast surgery method.

In the internal and external validation, the nomogram showed great predictive ability, with a C-index of 0.852 (0.842-0.862) in the training cohort and 0.876 (0.868-0.892) in the testing cohort, which indicated satisfactory discriminative ability of the nomogram. Calibration plots showed favorable consistency between the nomogram predictions and actual observations in both the training and validation cohorts. The calibration results are shown in Fig. 3.