Nomograms for the prognosis prediction model of early-stage triple-negative breast cancer - based on SEER database

Abstract

Aims

Triple-negative breast cancer (TNBC) is aggressive and unresponsive to hormone therapy, drawing widespread attention to its treatment and prognosis. This study aimed to summarize the prognosis of early-stage TNBC and construct nomograms to accurately predict survival and improve treatment strategies.

Patients & Methods

The modeling group comprised early-stage TNBC patients from the SEER database. Prognostic factors were identified via Cox regression, and nomograms were developed using R software. A validating group with identical screening criteria was assembled from Sun Yat-sen University Cancer Center. Validation involved the Concordance index, Receiver Operating Characteristic curves, and calibration curves.

Results

The modeling group included 14,389 patients and the validating group 1,146. Age and seven other factors were independent predictors (P < 0.05) and were incorporated into the nomograms. The C-indices for disease-free survival were 0.884 (95% CI: 0.876–0.892) in the modeling group and 0.686 (95% CI: 0.653–0.719) in the validating group. For overall survival, C-indices were 0.873 (95% CI: 0.863–0.883) and 0.741 (95% CI: 0.702–0.780), respectively.

Conclusions

The constructed nomograms demonstrated good predictive performance and may serve as practical tools for risk stratification in early-stage TNBC. However, given the observational nature of the data, the associations between treatment-related factors and survival should be interpreted with caution and warrant confirmation in prospective studies.

1. Background

Breast cancer is the most prevalent female malignant tumor worldwide [1]. Although the mortality rate of breast cancer has decreased due to current advances in oncological treatment strategies, the disease remains one of the leading causes of death in female cancer patients [2]. Breast cancer is categorized into several different molecular types. Triple-negative breast cancer (TNBC), which is defined as the absence of estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2), accounts for 15% to 20% of all invasive breast cancers [3]. Despite its relatively low incidence, TNBC lacks effective drug targets compared with other breast cancer subtypes, exhibits unique biological behavior and clinicopathological features, and consequently has a higher recurrence rate and a poorer prognosis [4]. Due to the special pathologic properties of TNBC, how to accurately predict the prognosis of TNBC patients so as to formulate a better treatment plan has become an important issue.

In the past, patients with TNBC were prone to relapse and distant metastasis within 2 years [5]. In the event of recurrent or metastatic disease, the overall survival (OS) is only from 13 to 18 months. Although several clinical studies in recent years have shown that programmed cell death protein 1 antibody and anti-programmed death ligand 1 antibody are effective at treating TNBC, they are mainly used to treat late-stage recurrent or metastatic patients, and their clinical benefits are limited [6,7]. Thus, early identification of high-risk patients and the adoption of aggressive, individualized treatment and follow-up strategies are essential for patients’ prognosis and clinical practice.

Survival of patients with TNBC is influenced by age at diagnosis, tumor stage, proportion of positive lymph nodes, lymphvascular invasion and Ki-67 level [8]. The traditional treatment strategy for early-stage TNBC is based on surgery and postoperative adjuvant chemotherapy as a means of preventing disease recurrence and metastasis [9]. TNBC is also one of the indications for neoadjuvant chemotherapy (NAC), which decreases the stage and increases the likelihood of complete surgical resection [10]. Patients with a pathologic complete response (pCR) have a survival benefit compared to those who do not achieve a pCR after NAC [11], especially in TNBC and HER2-positive breast cancer [12]. However, due to its genomic heterogeneity and lack of targeted molecular alterations, TNBC has shown only partial improvement in survival outcomes [13].

As one of the most authoritative global oncology databases, the Surveillance, Epidemiology, and End Results (SEER) program of the National Cancer Institute (https://seer.cancer.gov/) provides diverse and reliable data on early‑stage TNBC patients, encompassing a wealth of clinical information on pathology, treatment, and prognosis [14]. The SEER database was established in 1973, and in recent years the number of registries in the database has increased to 18, covering approximately 34.6% of the entire US population [15]. SEER is an open-access database available to the public for research and data downloaded from this database do not contain personal patient information.

At present, clinical studies on TNBC drug therapy have yielded encouraging results, and patient survival has been prolonged after comprehensive treatment [16]. Nevertheless, breast cancer mortality remains high, and treatment strategies for TNBC patients still require further investigation. Nomograms can integrate multiple key predictors and estimate an endpoint intuitively and simply, thereby enabling prognostic assessment of patients, particularly those with cancer. By integrating multiple prognostically relevant clinicopathological factors and constructing a prognostic prediction model through visualization and quantitative methods, an objective description of a patient’s prognosis can be obtained, leading to more accurate predictions.

Although several prognostic models have been proposed to predict the outcome of breast cancer patients [17–19], most are only partially clinically applicable, primarily to hormone receptor‑positive breast cancer. While several prognostic prediction models exist for TNBC patients [20–22], their complexity and high technical requirements have limited their widespread clinical use. Moreover, studies on nomogram prediction models specifically for early-stage TNBC are still rare. Therefore, in this study we constructed nomograms for disease-free survival (DFS) and overall survival (OS) using data from early-stage TNBC patients in the SEER database. An external validation cohort was assembled using data from Sun Yat-sen University Cancer Center (SYSUCC), with the aim of better predicting the survival of early-stage TNBC patients and ultimately supporting more accurate clinical decision-making.

2. Methods

2.1. Inclusion and exclusion criteria

The modeling group comprised early-stage TNBC patients identified from the SEER database using SEER*Stat software, which was used to extract clinicopathologic data. The validating group was obtained from the clinical history of patients diagnosed with early-stage TNBC from the SYSUCC pathology system. The inclusion and exclusion criteria were as follows:

Inclusion criteria

1. Age between 18–80 years.

2. Surgery time between January 1, 2013, and December 31, 2020.

3. Pathologically confirmed TNBC, i.e., negative for ER, PR, and HER2.

4. Pathologically confirmed early breast cancer, i.e., AJCC 8th edition TNM staging of stage I-II.

Exclusion criteria

1. Male breast cancer patients.

2. Patients diagnosed with bilateral breast cancer.

3. Patients with combination of other malignant tumors.

4. Patients diagnosed with breast cancer at autopsy.

5. Patients with ambiguous clinicopathologic data and lost to visit.

Candidate predictor variables were selected based on clinical relevance, established literature on TNBC prognosis, and availability within the SEER database. The initial set of variables included: age at diagnosis, race, tumor size (T stage), nodal status (N stage), TNM stage, histologic grade, surgery type (including plastic surgery), radiotherapy, and chemotherapy. Univariate Cox proportional hazards regression was first performed to screen for potential prognostic factors. Variables with a p-value < 0.10 in the univariate analysis were subsequently entered into a multivariate Cox regression model to identify independent predictors of OS and DFS. To ensure model stability and avoid overfitting, we maintained an events-per-variable ratio of approximately 10:1, in accordance with recommended guidelines for clinical prediction models.

2.2. Definition of study endpoints

OS in this study was defined as the interval from pathologic diagnosis to death, last follow-up, or database cutoff. DFS was defined as the interval from pathologic diagnosis to the first occurrence of local recurrence, regional lymph node metastasis, distant metastasis, death, or last follow-up/database cutoff.

2.3. Statistical methods

Categorical variables were presented as counts and percentages, and continuous variables were expressed as medians. Survival probabilities were estimated using the Kaplan-Meier method, and differences were compared using the log-rank test. To identify independent prognostic factors for DFS and OS, univariate and multivariate Cox proportional hazards regression analyses were performed. Variables with P < 0.05 in univariate analysis were entered into the multivariate model. Hazard ratios (HRs) and their 95% confidence intervals (CIs) were calculated. Categorical variables were compared using the chi-square test. All statistical analyses were performed using SPSS version 27.0 (IBM Corp., Armonk, NY, USA).

Variables with P < 0.05 in the univariate analysis were included in the multivariate Cox regression analysis. Factors that remained significant (P < 0.05) in the multivariate model were selected for nomogram construction. Nomograms predicting DFS and OS were generated using the “rms” package in R software (version 4.2.3; R Foundation for Statistical Computing, Vienna, Austria).

The predictive performance of the nomograms was evaluated by calculating the concordance index (C-index), plotting receiver operating characteristic (ROC) curves and calibration curves, and calculating the area under the ROC curve (AUC) using the “rms” package in R. Internal validation was performed using bootstrapping with 1,000 resamples. External validation was conducted using the independent SYSUCC cohort. A two-sided P < 0.05 was considered statistically significant.

The C-index was used to evaluate the discrimination of the nomogram. A C-index between 0.5 and 0.7 indicates low discrimination, between 0.7 and 0.9 indicates moderate to high discrimination, and a value >0.9 suggests excellent discrimination. The ROC curve was used to assess the predictive accuracy of the nomogram. The closer the ROC curve is to the top-left corner, the higher its practical value. The AUC was calculated to quantify the accuracy, with values ranging from 0.5 to 1.0. Calibration curves were used to evaluate the goodness of fit of the nomogram; the closer the calibration curve is to the 45° diagonal, the better the fit.

For external validation, the originally developed nomogram was applied directly to the independent cohort without any recalibration or adjustment. This approach was chosen to rigorously evaluate the model’s transportability and performance in a real-world setting where patient populations may differ from the derivation cohort.

LASSO regression with 10-fold cross-validation was applied to the same set of candidate predictors in the modeling cohort. The optimal penalty parameter (λ) was chosen based on the minimum partial likelihood deviance. Variables with non-zero coefficients at the optimal λ were considered selected. The predictive performance of the LASSO-derived model was then assessed in both the modeling and external validation cohorts using the concordance index (C-index), with the reverse = TRUE option in the concordance function of the survival package to account for the direction of the linear predictor. This analysis was conducted using the glmnet package in R.

2.4. Fellow-up methods

Patients in the modeling group were followed up from the time of the latest update of the SEER database, which was the time of initial diagnosis to April 3, 2023. The follow-up was the patients’ survival status, recurrence or metastasis. The follow-up time of the validating group was more accurate, which was the date of patients’ surgical diagnosis until January 1, 2023, the cutoff date. Recorded through the SYSUCC follow-up system, individual patients with missing follow-up data were followed up by telephone.

2.5. Data sources and ethical considerations

This retrospective study utilized data from the SEER database. The SEER program is a publicly available, population-based cancer registry supported by the National Cancer Institute. Use of SEER data does not require individual patient consent, as the database contains de-identified information. Access to and use of the SEER research data for this study use complies with the SEER Data Use Agreement, and the study protocol was approved by the Ethics Committee of SYSUCC. Data for the validating group were derived from the medical records of SYSUCC. The protocol for this retrospective analysis of de-identified data was approved by the Institutional Ethics Committee and granted a waiver of informed consent.

3. Results

3.1. Clinicopathological characteristics

The modeling cohort included 14,389 patients with early-stage TNBC from the SEER database, with a median follow-up of 72 months. During this period, 1,860 (12.9%) patients experienced a DFS event, and 1,224 (8.5%) died (OS events). The external validation cohort comprised 1,146 patients from SYSUCC, with a median follow-up of 66 months. In this cohort, 245 (21.4%) DFS events and 181 (15.8%) OS events were observed (Supplementary Figures 3_1 and 3_2).

The modeling group comprised 14,389 patients from the SEER database, and the validation group comprised 1,146 patients from SYSUCC. The patients in the modeling group had no specific recorded age; therefore, they were classified based on a median age cutoff of 50 years. Patients aged >50 years in the modeling group were 10,639 (73.9%), whereas younger patients were more common in the validating group, with a total of 738 (64.4%) patients aged ≤50 years, and the chi-square test showed that there was a statistically significant difference in the age distribution of patients in the two groups (P < 0.001). Histologic grades II and III were more common in both groups and grades I, and other grads were less common. Therefore, for statistical purposes, we categorized cases other than grades II and III as other types. Patients with histologic grade III were 7,429 patients (51.6%) and 788 patients (68.1%) in the modeling and validating groups, respectively, and there were more patients with grade III in the validating group (P = 0.001). However, the total number of stage II patients in the modeling group amounted to 8,512 (59.2%), and the number of stage II patients in the validating group was more frequent(P = 0.037). The rest of the details are shown in Table 1.

Table 1.

Clinicopathologic characteristics of patients in the modeling and validating group.

Characteristics		Modeling group(n = 14389)	Proportion	Validation group(n = 1146)	Proportion	P
Age	≤50	3750	26.10%	738	64.40%	<0.001*
	>50	10639	73.90%	408	35.60%
Race	White	9920	68.90%	–	–	–
	Black	2980	20.70%	–	–	–
	Other	1489	10.40%	1146	100.00%	–
Location	Left	7435	51.70%	593	51.70%	1
	Right	6954	48.30%	553	48.30%
Histological grading	II	6836	47.50%	234	20.40%	0.001*
	III	7429	51.60%	780	68.10%
	Other	124	8.60%	132	11.50%
T stage	T1	6884	47.80%	431	37.60%	0.021*
	T2	6445	44.80%	690	66.20%
	T3	1060	7.40%	25	2.20%
N stage	N0	10899	75.70%	751	65.50%	0.097
	N1	3490	24.30%	395	34.50%
TNM stage	I	5877	40.80%	310	27.10%	0.037*
	II	8512	59.20%	836	72.90%

T, N, TNM staging refer to AJCC 8th edition staging.

3.2. Treatment

Table 2 shows the treatment situation of the patients in the two groups. A total of 8,709 patients (60.5%) underwent sentinel lymph node biopsy (SLNB) in the modeling group, whereas the proportion of axillary lymph node dissection (ALND) was higher in the validating group, with 861 patients accounting for 75.1%. There was a difference between the two groups in terms of axillary surgical approaches (P < 0.001). Regarding breast surgery, a total of 8,301 patients (57.7%) underwent breast-conserving surgery (BCS) and a total of 5,157 patients (35.8%) underwent cosmetic plastic surgery in the modeling group, whereas only 273 patients (23.8%) underwent BCS, 76.2% underwent total excision, and only 8.1% underwent cosmetic plastic surgery in the modeling group. There were differences between the two groups in terms of breast surgery methods (P < 0.001). There was no difference in postoperative radiotherapy between the two groups, with the proportion of radiotherapy in the modeling group and the validating group being 52.7% and 41.1%, respectively. Detailed radiotherapy records are not available in the SEER database, so the technique, fractionation, dose, and side associations of radiotherapy in the validating group could not be counted in this study.

Table 2.

Treatment of patients in the modeling and validating group.

Treatment		Modeling group(n = 14389)	Proportion	Validation group(n = 1146)	Proportion	P
Axillary surgery	SLNB	8709	60.50%	285	24.90%	<0.001*
	ALND	5680	39.50%	861	75.10%
Breast surgery	BCS	8301	57.70%	273	23.80%	<0.001*
	Mastectomy	6088	42.30%	873	76.20%
Plastic surgery	Yes	5157	35.80%	93	8.10%	<0.001*
	No	9232	64.20%	1053	91.90%
Construction type	Tissue	512	9.90%	26	28.00%	0.005*
	Implant	4517	87.60%	64	68.80%
	Combined	128	2.50%	3	3.20%
Chemotherapy	Yes	11707	81.40%	1004	87.60%	0.171
	N0	2682	18.60%	142	12.40%
Radiotherapy	Yes	7581	52.70%	471	41.10%	0.089
	No	6808	47.30%	675	58.90%

SLNB, Sentinel lymph node biopsy; ALND, Axillary lymph node dissection; BCS, Breast-conserving surgery.

3.3. Analysis of survival outcomes and prognostic factors

The median follow-up time for the two groups was 72 and 66 months, respectively. Univariate analyses were performed using the Kaplan-Meier method and Cox proportional hazards regression model for all variables collected in Tables 1 and 2. The results of the univariate analysis showed that DFS in the modeling group was associated with age, race, T stage, N stage, histologic grading, TNM staging, plastic surgery, breast surgery approach, axillary surgery approach, and radiotherapy. OS was associated with age, race, T-stage, N-stage, histologic grading, TNM staging, plastic surgery, breast surgery approach, and radiotherapy. Variables with P < 0.05 in the univariate analysis were included in the multivariate analysis.

The results of multivariate analysis showed that in the modeling group, DFS was worse in patients with early-stage TNBC who were >50 years of age, had T3, N1stage, histologic grade III or other types, stage II, underwent reconstructive surgery, and did not undergo radiotherapy. The age at diagnosis was one of the factors influencing DFS, which was worse in patients >50 compared with those ≤50 years of age (HR = 1.30; 95% CI, 1.16–1.47; P < 0.001). DFS was worse in T3 patients compared with those with T1 stage (HR = 1.60; 95% CI, 1.32–1.94; P < 0.001). In terms of lymph nodes, DFS was worse in patients with N1 staging compared with N0 (HR = 2.05; 95% CI, 1.85–2.28; P < 0.001). Histologic grading was also a factor influencing recurrence and metastasis. In the modeling group, grade III and other rare histologic grades were poor prognostic factors compared with grade II patients. Patients with TNM staging of stage II had poorer DFS than stage I (HR = 11.06; 95% CI, 8.65–14.13; P < 0.001). No survival differences were seen in patients who underwent BCS or total mastectomy, SLNB, or ALND. However, patients who underwent plastic surgery had a higher risk of recurrence than those who did not (HR = 0.75; 95% CI, 0.62–0.90; P = 0.002). Patients without chemotherapy had a higher risk of metastasis compared with those who received chemotherapy (HR = 6.23; 95% CI: 5.45–7.13; P < 0.001). Radiotherapy was likewise an independent predictor of patients’ DFS, which was worse in patients without radiotherapy compared to those who received it (HR = 10.07; 95% CI: 9.06–11.18, P < 0.001). The rest of the results are detailed in Table 3.

Table 3.

Univariate and multivariate analysis of DFS in the modeling group.

DFS		Univariate		Multivariate
characteristic		HR(95%CI)	P	HR(95%CI)	P
Age	>50 vs ≤50	1.12(1.00-1.24)	0.045*	1.30(1.16–1.47)	<0.001*
Race	Black vs White	1.24(1.11-1.38)	<0.001*	1.07(0.95-1.19)	0.265
	Other vs White	1.14(0.98-1.32)	0.082	0.97(0.84-1.13)	0.687
Location	Right vs Left	0.98(0.90-1.08)	0.724	–	–
T stage	T2 vs T1	4.34(3.84-4.90)	<0.001*	1.67(0.99-1.38)	0.072
	T3 vs T1	6.83(5.84-7.99)	<0.001*	1.60(1.32–1.94)	<0.001*
N stage	N1 vs N0	2.67(2.44-2.93)	<0.001*	2.05(1.85–2.28)	<0.001*
Histological grading	III vs II	2.41(2.18-2.67)	<0.001*	2.70(2.42-3.02)	<0.001*
	Other vs II	2.15(1.36-3.40)	0.001*	1.68(1.06-2.66)	0.027*
TNM stage	II vs I	9.42(7.92-11.20)	<0.001*	11.06(8.65–14.13)	<0.001*
Plastic surgery	No vs Yes	0.62(0.57-0.68)	<0.001*	0.75(0.62–0.90)	0.002*
Breast surgery	Mastectomy vs BCS	1.53(1.40-1.68)	<0.001*	0.81(0.58-1.13)	0.211
Axillary surgery	ALND vs SLNB	1.54(1.41-1.69)	<0.001*	0.97(0.73-1.29)	0.839
Radiotherapy	No vs Yes	6.93(6.10-7.87)	<0.001*	6.23(5.45–7.13)	<0.001*
Chemotherapy	No vs Yes	4.54(4.14-4.97)	<0.001*	10.07(9.06–11.18)	<0.001*

SLNB, Sentinel lymph node biopsy; ALND, Axillary lymph node dissection; BCS, Breast-conserving surgery.

In the modeling group, OS was worse in patients with early-stage TNBC who were >50 years old, had T2 or T3 stage, N1 stage, histologic grade III or other types, stage II, without plastic surgery, and without radiotherapy. Survival was shorter in patients older than 50 years than in patients 50 years old and younger (HR = 2.23; 95% CI, 1.91–2.61; P < 0.001). The mortality rate increased with higher T stage, with patients in the T2 and T3 groups having worse outcomes compared with those in the T1 group compared with patients in the T1 group (P = 0.021and P < 0.001, respectively). The prognosis of N1 patients was worse than that of N0 patients (HR = 2.64; 95% CI, 2.31–3.00; P < 0.001). Histologic grading similarly affected OS, III and other rare histologic grades were poor prognostic factors compared with grade II patients (P < 0.001). Patients with TNM stage II had poorer OS than stage I (HR = 3.40; 95% CI, 2.59–4.46; P < 0.001). In contrast to the association with DFS, having plastic surgery was a protective factor for OS, and patients without plastic surgery had a worse prognosis (HR = 1.41; 95% CI, 1.13–1.77; P = 0.003). In addition, both chemotherapy and radiotherapy were protective factors (P < 0.001). The remaining results are detailed in Table 4.

Table 4.

Univariate and multivariate analysis of OS in the modeling group.

OS		Univariate		Multivariate
Characteristic		HR(95%CI)	P	HR(95%CI)	P
Age	>50 vs ≤50	1.85(1.59-2.15)	<0.001*	2.23(1.91–2.61)	<0.001*
Race	Black vs White	1.36(1.20-1.55)	<0.001*	1.08(0.95-1.23)	0.258
	Other vs White	1.01(0.83-1.22)	0.933	0.94(0.77-1.14)	0.497
Location	Right vs Left	1.04(0.93-1.16)	0.533	–	–
T stage	T2 vs T1	2.93(2.55-3.36)	<0.001*	1.26(1.04-1.53)	0.021*
	T3 vs T1	5.38(4.50-6.43)	<0.001*	1.93(1.55-2.41)	<0.001*
N stage	N1 vs N0	3.54(3.17-3.96)	<0.001*	2.64(2.31–3.00)	<0.001*
Histological grading	III vs II	4.71(4.05-5.48)	<0.001*	4.83(4.11-5.67)	<0.001*
	Other vs II	4.50(2.67-7.61)	<0.001*	2.90(1.71-4.91)	<0.001*
TNM stage	II vs I	5.68(4.76-6.78)	<0.001*	3.40(2.59–4.46)	<0.001*
Plastic surgery	No vs Yes	1.25(1.11-1.41)	<0.001*	1.41(1.13–1.77)	0.003*
Breast surgery	Mastectomy vs BCS	0.84(0.75-0.94)	0.003*	0.87(0.70-1.08)	0.196
Axillary surgery	ALND vs SLNB	0.90(0.81-1.02)	0.087	–	–
Radiotherapy	No vs Yes	4.63(4.02-5.33)	<0.001*	4.65(4.01-5.40)	<0.001*
Chemotherapy	No vs Yes	3.53(3.15-3.95)	<0.001*	5.86(5.16-6.66)	<0.001*

SLNB, Sentinel lymph node biopsy; ALND, Axillary lymph node dissection; BCS, Breast-conserving surgery.

In contrast, in the validating group, the results of univariate and multivariate Cox regression analyses showed that patients’ DFS and OS were related to N stage, TNM stage, plastic surgery, and radiotherapy. The results of multivariate analysis showed that lymph node-positive patients had poorer DFS than lymph node-negative patients (HR = 2.09; 95% CI, 1.90–2.30; P < 0.001). Stage II patients had a higher risk of recurrence compared with stage I (HR = 7.86; 95% CI, 6.57–9.41, P < 0.001). The validating group differed from the modeling group in that patients who underwent reconstructive surgery had better DFS (HR = 1.23; 95% CI, 1.13–1.35; P < 0.001). Adjuvant chemotherapy did not reflect a survival difference in the validating group, but patients who did not undergo radiotherapy had a higher risk of recurrence or metastasis (HR = 9.33; 95% CI, 8.19–10.63; P < 0.001). The remaining results are detailed in Table 5.

Table 5.

Univariate and multivariate analysis of DFS in the validating group.

DFS		Univariate		Multivariate
Characteristic		HR(95%CI)	P	HR(95%CI)	P
Age	>50 vs ≤50	0.90(0.69-1.17)	0.424
Location	Right vs Left	1.02(0.79-1.31)	0.892
T stage	T2 vs T1	1.28(0.98-1.68)	0.070
	T3 vs T1	1.72(0.79-3.72)	0.170
N stage	N1 vs N0	1.76(1.37-2.26)	<0.001*	2.09(1.90–2.30)	<0.001*
Histological grading	III vs II	0.82(0.60-1.11)	0.193
	Other vs II	1.07(0.70-1.65)	0.757
TNM stage	II vs I	1.45(1.06-1.97)	0.019*	7.86(6.57–9.41)	<0.001*
Plastic surgery	No vs Yes	5.79(2.16-15.56)	<0.001*	1.23(1.13–1.35)	<0.001*
Breast surgery	Mastectomy vs BCS	0.94(0.70-1.25)	0.661
Axillary surgery	ALND vs SLNB	1.05(0.78-1.41)	0.747
Radiotherapy	No vs Yes	1.44(1.10-1.88)	0.007*	9.33(8.19–10.63)	<0.001*
Chemotherapy	No vs Yes	1.16(0.81-1.68)	0.420

SLNB, Sentinel lymph node biopsy; ALND, Axillary lymph node dissection; BCS, Breast-conserving surgery.

Regarding OS, in the validating group, N1 patients had poorer OS than N0 patients (HR = 2.81; 95% CI, 2.49–3.18; P < 0.001)。 Stage II patients had a higher risk of death than stage I (HR = 3.81; 95% CI, 3.15–4.62; P < 0.001). Similarly, OS was better in patients who underwent plastic surgery (HR = 2.28; 95% CI, 2.02–2.58; P < 0.001). Patients without radiotherapy had poorer OS than those with radiotherapy (HR = 7.03; 95% CI, 6.10–8.11; P < 0.001). The rest are detailed in Supplementary Table 1.

3.4. Nomograms and validating of predictive ability

Based on the results of multivariate Cox analysis, statistically significant (P < 0.05) predictors were selected, and the prognostic model nomograms were constructed using R software (version 4.3.2). The nomograms of 3-year and 5-year DFS and OS in early-stage TNBC patients are shown in Figure 1. The scores were read out on the Points axis based on patient characteristics, and the patient’s survival were assigned through Total Points by summing up the scores.

Figure 1.

1.1. Nomogram for predicting DFS in patients with early-stage TNBC. 1.2. Nomogram predicting OS in patients with early-stage TNBC.

The C-index for DFS in the modeling group was 0.884 (95% CI, 0.876–0.892), and that for OS in the modeling group was 0.873 (95% CI, 0.863–0.883), which were above 0.7, suggesting that the prognostic model had a high degree of good discrimination. The “survival” function method was used to calculate the risk score for each independent risk factor, and ROC curves were plotted. As shown in Figure 2, the AUC values of the ROC curves for DFS and OS were 0.889 and 0.866, respectively, suggesting that the predictive model had good fit for assessing the prognosis of early-stage TNBC patients.

Figure 2.

ROC curves for DFS and OS in the modeling group. 2.1. ROC curve for DFS. 2.2. ROC curve for OS.

The Bootstrap method was used with 1,000 resamples to perform internal validation of the nomograms constructed in this study. The calibration curves were plotted, and the results showed that the calibration curves for 3- and 5-year DFS and OS were close to the 45° reference line (Figure 3), indicating that the model had good accuracy for prognostic prediction in early-stage TNBC patients.

Figure 3.

Calibration curves of DFS and OS for 3 and 5 year for the modeling group. 3.1. Calibration curves of DFS 3 year; 3.2. Calibration curves of DFS 5 year. 3.3. Calibration curves of OS 3 year; 3.4. Calibration curves of DFS 5 year.

To further validate whether the nomograms constructed from the modeling group could distinguish the prognosis of early-stage TNBC patients with different clinicopathological features and treatment modalities, we used the survival data of 1,146 early-stage TNBC patients in the validation group. The validation results showed that the C-index for DFS in the validation group was lower than that in the modeling group [0.686 (95% CI, 0.653–0.719)], while the C-index for OS was 0.741 (95% CI, 0.702–0.780), suggesting differences in risk factors for DFS between the two groups. ROC curves were plotted using the same method (Figure 4), and the AUC values for DFS and OS in the validation group were 0.688 and 0.717, respectively, indicating that the DFS nomogram had moderate fit and the OS nomogram had good fit.

Figure 4.

ROC curves of DFS and OS in the validating group. 4.1. ROC curve for DFS. 4.2. ROC curve for OS.

The same Bootstrap method was used to validate the nomogram in the validating group by plotting the calibration curve. The results showed that the 3-year and 5-year DFS and OS calibration curves were partially offset from the 45° reference line, but the alignments matched, and the predicted and actual values of the DFS and OS nomograms had good accuracy. The calibration curves for the validating group are shown in Supplementary Figure 1.

The LASSO regression identified similar sets of predictors for both overall survival (OS) and disease-free survival (DFS) compared to our original multivariable Cox model. For OS, the LASSO selected 11 variables (including age, T stage, N stage, TNM stage, histology, plastic surgery, breast surgery, axillary surgery, radiotherapy, and chemotherapy), with coefficients detailed in Supplementary Table 2. For DFS, LASSO selected 13 variables (Supplementary Table 3). Notably, the C-indices of the LASSO-based models were nearly identical to those of our original nomograms in the modeling cohort (OS: 0.874 vs. 0.873; DFS: 0.884 vs. 0.884). In the external validation cohort, the LASSO models yielded C-indices of 0.648 for OS and 0.590 for DFS, compared to 0.741 and 0.686 from the original models(Supplementary Figures 2_1 and 2_2). These results confirm that our original variable selection and model development approach did not lead to substantial overfitting, as the predictive accuracy remained consistent with that of a penalized regression model.

4. Discussion

Breast cancer is one of the most common malignant tumors among female cancer patients worldwide, with high morbidity and mortality rates [23,24]. Compared with other types of breast cancer, TNBC is characterized by poor differentiation and high rates of recurrence and metastasis [25]. Understanding the clinical and epidemiological characteristics of TNBC can help clinicians choose treatment options and make prognostic assessments for patients. In this study, we utilized patient data from the SEER public database and from SYSUCC to construct nomograms for predicting DFS and OS, which provides a scientific basis for disease management by clinicians and patients. The results showed that N stage, TNM stage, plastic surgery, and radiotherapy are important factors in determining the prognosis of early-stage TNBC patients. The current research status of TNBC and the results of this study are discussed further.

Chemotherapy is one of the means of systemic treatment for TNBC patients, while surgery and radiotherapy are the main local treatments. Adjuvant chemotherapy for breast cancer has shown sustained survival benefit and improved DFS [26], especially in TNBC, where it demonstrates a greater survival benefit than in hormone receptor-positive breast cancer. Adjuvant chemotherapy improved the prognosis of patients with T1c TNBC to a greater extent than that of patients with T1a and T1b [27]. Similarly, it has been shown that patients with T1a/bN0M0 TNBC have a favorable prognosis with or without adjuvant chemotherapy [28]. The role of immune checkpoint inhibitors has now been extended to NAC with the results of the KEYNOTE-355 and IMpassion130 clinical studies, which showed that pembrolizumab and atezolizumab demonstrated PFS benefit in patients with advanced TNBC in phase III trials [29–30]. A study indicated the potential for guiding individualized therapeutic approaches, such as immune checkpoint gene therapy and chemotherapy strategies, by utilizing the risk scores generated from the SRlncRNA signature [31]. Although current studies have shown that it does not show significant survival benefits in early-stage TNBC [32], pembrolizumab has better PFS and OS benefits in advanced TNBC [33].

While our model includes patients receiving NAC within the broader ‘‘chemotherapy “variable, it does not incorporate treatment response data, such as pCR. This is a recognized limitation of population-based registries like SEER. Consequently, our nomogram is best interpreted as a baseline risk assessment tool prior to the initiation of therapy. It is effective for stratifying patients who may benefit most from aggressive systemic treatment at diagnosis. However, for patients who undergo NAC, their ultimate prognosis is dynamically refined by their response to therapy. Our model does not aim to replace post-NAC response assessment but rather to complement the clinical workflow by providing initial risk stratification. Future models integrating real-world or clinical trial data with treatment response information are needed to provide post-NAC prognostic refinement.

The effect of radiotherapy after total mastectomy in patients with T1–2N1M0 TNBC is controversial. Currently, the absolute indications for post-mastectomy radiotherapy (PMRT) are a primary tumor diameter of ≥ 5 cm or axillary lymph node metastases of ≥ 4. The necessity of adjuvant radiotherapy for patients with T1–2 breast cancer with one to three axillary lymph node metastases remains controversial across different guidelines [34]. Studies have found no significant improvement in DFS and OS in T1–2N1M0 TNBC patients who received PMRT [35]. Conversely, other studies have confirmed that patients with T1–2N1M0 TNBC could benefit from radiotherapy [36]. The value of PMRT for patients with T1–2N1M0 TNBC still needs to be further clarified. Therefore, some researchers conducted a more in-depth study based on the number of lymph nodes [37], which showed that PMRT significantly improved breast cancer-specific survival (BCSS) in T1–2N1M0 TNBC patients with three positive lymph nodes, but no significant advantage was observed in patients with one or two positive lymph nodes. Chen et al. also reported that T1–2 breast cancer patients with 1 or 2 positive lymph nodes do not benefit from PMRT [38]. It may be that these patients have low tumor burden and recurrence risk, and radiotherapy has limited significance in improving their survival, but instead brings side effects such as arm edema and cardiopulmonary radiation damage. For these patients, radiotherapy should be considered carefully to avoid over-treatment.

The primary modalities of plastic surgery currently include prosthesis implantation and autologous tissue reconstruction. Specifically, prosthesis implantation involves the use of silicone or saline-filled implants to restore the shape of the breast. Autologous tissue reconstruction, on the other hand, utilizes the patient’s own tissue, such as the transverse rectus abdominis myocutaneous flap, deep inferior epigastric perforator flap, or superficial gluteal artery perforator flap, to reconstruct the breast. Additionally, some patients may undergo a combination of autologous tissue and prosthesis, which integrates the benefits of both methods to achieve optimal aesthetic and functional outcomes [39]. Prosthetic implant surgery has the advantages of short surgical time, fast postoperative recovery, and no scarring in the donor area. However, there are some drawbacks such as unnatural shape of the reconstructed breast, hard texture, and poor tolerance to radiotherapy. Autologous tissue reconstruction provides a breast with natural contour, soft texture, and better tolerance to radiotherapy than prosthesis plastic surgery, but at the same time, at the cost of donor site scarring. A study has shown that women of European descent are more likely to undergo breast reconstructive surgery, while Asian and Pacific Islander women are underrepresented [40]. A higher proportion of breast cancer patients in the United States undergo reconstructive surgery (54.4%), whereas fewer Chinese women undergo plastic surgery (12.3%) [41]. This may be attributed to the fact that women in Europe and the United States have larger breast volumes and more space for cosmetic purposes, whereas Asian women have smaller breast volumes and more traditional and conservative treatment concepts. Despite the current lack of level 1 evidence, available studies suggest that plastic surgery has superior or at least equivalent outcomes compared with traditional BCS and can result in better aesthetic outcomes for patients [42]. The results of this study suggest that plastic surgery may be a poor prognostic factor for DFS but beneficial for OS in patients with early-stage TNBC. There are few studies on early-stage TNBC and plastic surgery, and further prospective exploration is needed in the future.

An interesting finding from our prognostic model was an observed association between procedures involving plastic surgery techniques and an increased risk of local recurrence and metastasis. This warrants careful interpretation. It is important to emphasize that this association likely does not implicate the surgical techniques of oncoplastic breast conservation or immediate reconstruction as direct contributors to worse oncological outcomes. A substantial body of evidence confirms the safety of these procedures in appropriately selected patients. Rather, we hypothesize that in the real-world cohort captured by our database, the variable “plastic surgery” may act as a surrogate for a more complex clinical scenario. Patients undergoing these procedures may present with disease characteristics that inherently pose a higher biological risk or may experience a treatment pathway where surgical complexity interacts with other factors, such as the timing or tolerance of adjuvant therapies. This perspective aligns with discussions in the literature that call for careful patient selection.

Therefore, our finding primarily highlights a need for refined risk stratification. It suggests that the decision to pursue complex breast-conserving or reconstructive surgery should be integrated with a comprehensive assessment of tumor biology and treatment plans. This observation underscores the importance of future prospective, multicenter studies designed to collect granular data on surgical margins, complication profiles, and systemic therapy sequencing to clarify the independent prognostic role of surgical complexity in early-stage TNBC.

Currently, numerous researchers have developed prognostic models for TNBC. For instance, a study that utilized data from the SEER database and multicentric cohorts constructed the TNMpBC-NeoBCSS model [43], which integrates clinical and pathological features as well as the response to neoadjuvant therapy to predict 3-year and 5-year BCSS in TNBC patients receiving neoadjuvant treatment. The model demonstrated a C-index of 0.82. In the training cohort, the AUC values for 3-year and 5-year BCSS were 0.838 and 0.866, respectively, while in the validation cohort, the corresponding AUC values were 0.922 and 0.868. The consistency between the model’s predictions and actual observations indicates a high level of accuracy. However, the predictive capacity of single-modality models is relatively limited. To address this limitation, other researchers have developed multiple models. For example, Guo R et al. integrated radiomics from ultrasound imaging and clinical features and demonstrated favorable performance in identifying germline BRCA1/2 mutations in patients with invasive breast cancer through the development of a machine-learning-based nomogram model, which may help avoid unnecessary genetic testing in clinical practice [44]. These findings suggest that multimodal prediction models will likely become one of the hot research topics in breast cancer in the future.

So far, several researchers have developed a predictive model for the prognosis of TNBC. Most studies, however, employ a conventional survival analysis approach that fails to consider the impact of competing events on the occurrence of the event of interest. Specifically, deaths due to causes other than breast cancer are not considered. Patients’ tumor-specific mortality can be overstated [45]. To analyze the probability of a specific outcome in survival data that included a variety of possible outcome measures, Fine and Gray roposed a statistical method called the Competing Risk Model. Since then, the competing risks model has been used in medical statistics to analyze survival data involving competing events[46]. However, because the cause of death of some of the patients in the SEER database and in our center is not known, this study also does not apply a competitive risk model to investigate the prognosis.

Several limitations of this study should be acknowledged. First, due to the lack of cause-specific mortality data in the SEER database, we were unable to perform a competing risks analysis for DFS. As DFS includes death from any cause as an event, non-breast cancer deaths may theoretically compete with recurrence events. However, given that our study population consists of early-stage TNBC patients with a relatively young age, the incidence of non-breast cancer deaths is expected to be low, and the potential bias from ignoring competing events is likely minimal. Nonetheless, future studies with access to detailed cause-of-death information should validate our findings using competing risks models such as the Fine-Gray subdistribution hazard model.

In recent years, there has been a growing number of studies utilizing the SEER database to develop prognostic prediction models for TNBC, reflecting the clinical demand for such tools. However, these studies vary in their target populations, clinical focus, and methodological approaches, and our research aims to fill some of these gaps and strengthen certain aspects.

First, regarding the target population, our study complements the work of Feng et al. [47] and Zhang et al. [48]. Their research focuses specifically on special subpopulations, such as older women and patients with stage I disease, providing valuable subgroup-specific insights. In contrast, our model is designed to serve as a universal baseline risk assessment tool for a broader cohort of early-stage (I–II) TNBC patients, applicable to the non-elderly majority at initial diagnosis as well as across the entire age spectrum.

Second, in terms of research objectives, while Zhang et al.’s study primarily addresses the efficacy of adjuvant chemotherapy in a specific subgroup [48], and Zhou and Chen’s work offers a broad analysis of prognostic factors [20], our study advances further by integrating these key prognostic and treatment variables—including surgery, radiotherapy, and systemic chemotherapy—to construct a visual nomogram that can be directly applied to predict individualized survival probabilities, placing greater emphasis on direct clinical translation and utility.

Finally, concerning model validation, we acknowledge and note the rigorous step of external validation adopted in Qiu Y’s study [49]. Although our study conducted comprehensive internal validation—including Bootstrap validation and time-dependent ROC analysis based on the SEER database to ensure internal robustness—we also recognize that future collaboration to obtain independent external cohort data will be an essential direction for validating the generalizability of our model.

In summary, the core contributions of this study are: 1) developing a comprehensive, individualized prognostic tool for early-stage TNBC patients of all ages; 2) specifically evaluating and integrating “systemic chemotherapy” (whether neoadjuvant or adjuvant) as a key treatment variable to capture its overall real-world impact; and 3) supplementing purely statistical validation with an in-depth assessment of the model’s clinical utility through ROC analysis, quantifying the potential net benefit of applying the model across different risk thresholds for clinical decision-making. Therefore, our work is not merely a repetition but a refinement and extension of existing research, offering a new tool and perspective for the precise prognostic management of early-stage TNBC.

Our nomogram demonstrated robust predictive performance in both internal and external validations. However, we acknowledge that this study did not directly compare its incremental value against standard staging systems (e.g., the AJCC TNM staging) or existing TNBC-specific prognostic models. Such comparisons are essential to establish the clinical utility of a new model. Nevertheless, by incorporating variables not routinely included in traditional staging—such as plastic surgery and radiotherapy status—our nomogram may capture additional prognostic dimensions. The strong discrimination and calibration observed in our validation cohorts suggest that this tool could complement existing methods. Future research should focus on head-to-head comparisons using formal metrics like net reclassification improvement (NRI) or integrated discrimination improvement (IDI) to quantify the added benefit of our model. Additionally, prospective studies are needed to assess whether use of this nomogram leads to improved clinical outcomes or more personalized treatment decisions.

The nomograms developed in this study have potential implications for clinical management. First, they enable individualized risk stratification, allowing physicians to identify patients at high risk of recurrence or death, who may benefit from more intensive counseling and closer monitoring. Second, the nomograms could support treatment decisions by providing a quantitative risk estimate that complements guideline-based recommendations when considering adjuvant therapies. Third, they may guide surveillance strategies, with high-risk patients receiving more frequent follow-up and low-risk patients potentially avoiding overtreatment. However, these applications remain exploratory and require prospective validation in clinical implementation studies.

This study also has some limitations. First, some data extracted from the SEER database were missing, such as specific Ki-67 values, comparisons between preoperative and postoperative pathology, and genetic testing results. The lack of this information may have reduced the predictive ability of the model. Second, other relevant treatment information—such as chemotherapeutic agents, treatment duration, radiotherapy fractionation, and techniques—was also not included in this study. The SEER database contains limited data on Asian populations, and the models may therefore have biased predictive ability for the SYSUCC cohort, in which all patients are Asian. Finally, although the modeling group from SEER is large and population-based, the external validation group from SYSUCC was derived from a single institution, which may introduce selection bias. Therefore, our findings require further validation in large-scale, multicenter, and geographically diverse prospective studies.

5. Conclusion

The prognosis of patients with early-stage TNBC is independently associated with age, N stage, TNM stage, histologic grade, and the receipt of plastic surgery and radiotherapy. Notably, the observed association between plastic surgery and an increased risk of local recurrence and metastasis warrants further investigation in future multicenter prospective studies to clarify causality. After rigorous internal and external validation, the nomograms constructed in this study demonstrated good prognostic predictive ability. These models may serve as practical tools to more accurately assess DFS and OS in patients with early-stage TNBC, thereby providing an objective reference for clinical decision-making. However, due to the observational nature of the SEER database, the associations between treatment-related factors (e.g., radiotherapy) and survival should be interpreted with caution and require confirmation in prospective studies.

Ethics approval and consent to participate

The need for Informed consent was waived off by the ethical committee of SYSUCC because of that patients in retrospective clinical studies did not need to sign informed consent. The Sun Yat-sen University Cancer Center Ethics Committee approved this project (Ethical Consent Letter No.: B2022-638-01).

Consent for publication

The figures and tables included in this article are original and do not involve any conflict of interest on the part of others. The authors of this article agreed to publish it.

Authors’ contributions

Yaxue Wang: Conceptualization, Methodology, Formal Analysis, Writing – Original Draft, Writing – Review & Editing.

Yaxue Wang and Fengyan Li: Data Curation, Formal Analysis, Investigation, Visualization.

Yaxue Wang: Methodology, Resources, Software, Validation.

Fengyan Li: Project Administration, Supervision, Funding Acquisition, Writing – Review & Editing.

Fengyan Li: Investigation, Data Collection, Writing – Review & Editing.

Preprint statement

A preprint has previous published on Research Square [50] (DOI: https://doi.org/10.21203/rs.3.rs-4792314/v1). This version includes significant updates, additional data, and revisions in response to peer review.

Competing interests

The authors have no relevant affiliations or financial involvement with any organization or entity with a financial interest in or financial conflict with the subject matter or materials discussed in the manuscript. This includes employment, consultancies, honoraria, stock ownership or options, expert testimony, grants or patents received or pending, or royalties.

Availability of data and materials

The raw data used in this study were uploaded to the Research Data Deposit platform (https://www.researchdata.org.cn/), with the approval RDD number RDDA2024812699. All data and tables can be viewed through the website.