Development and External Validation of a Preoperative Nomogram for Predicting Axillary Lymph Node Metastasis in Breast Cancer Using Routine Ultrasound and Core Needle Biopsy Variables

Abstract

Background

Accurate preoperative assessment of axillary lymph node metastasis (ALNM) is essential for individualized axillary management in breast cancer, yet imaging-based evaluation alone remains imperfect. We aimed to develop and externally validate a practical nomogram based on routinely available preoperative ultrasound and core needle biopsy (CNB) variables.

Methods

This retrospective, two-center cohort study included 295 women with pathologically confirmed primary breast cancer who underwent breast surgery with axillary nodal staging between January 2021 and January 2025. The training cohort included 182 patients from the Changji Branch, and the external validation cohort included 113 patients from the First Affiliated Hospital of Xinjiang Medical University. Candidate predictors were evaluated using univariable analysis, LASSO regression, and multivariable logistic regression. Model discrimination, calibration, and clinical utility were assessed using receiver operating characteristic analysis, calibration analysis, and decision curve analysis.

Results

 ALNM was confirmed in 88 of 182 patients (48.4%) in the training cohort and 50 of 113 patients (44.2%) in the external validation cohort. The final nomogram included ultrasound-reported calcifications, tumor size, histological grade, CNB-derived lymphovascular invasion, and HER2 status. The model showed good discrimination, with an AUC of 0.858 (95% CI 0.805–0.911) in the training cohort and 0.854 (95% CI 0.779–0.929) in the external validation cohort. Calibration and decision curve analyses supported acceptable model performance and potential clinical utility in both cohorts.

Conclusion: 

We developed and externally validated an interpretable preoperative nomogram using routinely obtainable ultrasound and CNB variables for individualized estimation of ALNM risk in breast cancer. The model may support preoperative risk stratification and axillary management planning, although prospective multicenter validation is warranted.

Introduction

Breast cancer remains the most commonly diagnosed malignancy in women worldwide and a leading cause of cancer-related mortality.¹ Axillary lymph node (ALN) status plays a central role in disease staging, prognosis, and treatment planning, making accurate preoperative assessment of axillary involvement clinically important.² Sentinel lymph node biopsy (SLNB) is widely used for axillary staging in clinically node-negative patients and substantially reduces morbidity compared with routine axillary lymph node dissection (ALND). Nevertheless, SLNB remains an invasive procedure with associated costs, perioperative morbidity, and a small but clinically relevant false-negative rate.^3,4 More accurate preoperative risk estimation may therefore complement conventional axillary assessment and support individualized surgical planning and patient counseling.

Preoperative axillary assessment relies primarily on imaging, particularly axillary ultrasonography, with mammography and magnetic resonance imaging used according to clinical indications. Although imaging can identify morphologically suspicious lymph nodes, its sensitivity for low-volume nodal metastasis is imperfect and may vary according to imaging technique, operator experience, and institutional practice.^5,6 At the same time, routine preoperative evaluation provides additional information on primary tumor characteristics and CNB-derived pathological features, including histological grade and biomarker status.^7,8 In clinical practice, however, these imaging and pathological findings are generally interpreted individually rather than integrated into a quantitative estimate of ALNM risk. An interpretable model that combines routinely available preoperative information may therefore provide complementary risk information without requiring additional specialized examinations.

Several models have been developed to predict ALNM using different combinations of clinicopathological and imaging features. Earlier clinicopathological nomograms demonstrated the feasibility of estimating nodal risk using readily available patient and tumor characteristics, although performance and calibration may vary when applied to external populations.^9,10 Ultrasound-based models subsequently incorporated primary tumor features together with axillary nodal characteristics such as cortical morphology and fatty-hilum abnormalities.^11–13 More recently, Wang et al developed a prospective nomogram combining preoperative histological grade and Ki-67 status with high-resolution computed tomography features of the axillary lymph nodes, including mean CT attenuation, short-axis diameter, border, and shape, achieving a C-index of 0.869.¹⁴ Radiomics and deep-learning approaches have further improved the ability to extract quantitative imaging information and have shown promising discrimination for ALNM prediction.^15–18 However, many of these approaches depend on detailed nodal imaging, additional imaging modalities, image segmentation, or computational processing, while external validation remains inconsistent across published models.

Accordingly, further efforts are needed to develop clinically applicable prediction tools that integrate routinely available preoperative information and demonstrate reproducible performance beyond the development cohort. In this study, we developed and externally validated a nomogram for individualized prediction of axillary lymph node metastasis in breast cancer using routine ultrasound findings and clinicopathological characteristics obtained from preoperative core needle biopsy. By integrating readily accessible imaging and biopsy-derived variables within a parsimonious and interpretable framework, the model was designed to support preoperative risk stratification and inform individualized axillary management.

Materials and Methods

Patients

This retrospective, two-center cohort study included female patients with pathologically confirmed primary breast cancer who underwent breast surgery with axillary nodal staging between January 2021 and January 2025. The training cohort was derived from the Changji Branch of the First Affiliated Hospital of Xinjiang Medical University, and the external validation cohort was derived from the First Affiliated Hospital of Xinjiang Medical University.

At the Changji Branch, 267 consecutive patients were screened during the study period. After application of the eligibility criteria, 85 patients were excluded and 182 were included in the training cohort. The reasons for exclusion were a history of previous malignancy (n = 12), severe cardiopulmonary comorbidities (n = 35), missing key variables (n = 27), and refusal of surgery by the patient or family (n = 11). The external validation cohort comprised 113 patients who met the same eligibility criteria at the First Affiliated Hospital of Xinjiang Medical University. Overall, 295 patients were included in the analysis (training cohort, n = 182; external validation cohort, n = 113; Figure 1).

Figure 1.

Study flowchart and model development workflow. A total of 267 patients were screened at the Changji Branch; 182 were included in the training cohort after exclusions. An independent external validation cohort included 113 patients from the First Affiliated Hospital. Model development used LASSO and multivariable logistic regression to construct the nomogram, which was evaluated by ROC, calibration (Hosmer–Lemeshow), and decision curve analysis.

Patients were eligible if they: (1) had pathologically confirmed primary breast cancer; (2) underwent breast surgery with axillary nodal staging by sentinel lymph node biopsy (SLNB) and/or axillary lymph node dissection (ALND); and (3) had the clinical, imaging, and pathological data required for analysis. Patients were excluded if they: (1) had a history of another malignancy; (2) had severe cardiopulmonary comorbidities precluding standard surgical management; (3) had received neoadjuvant chemotherapy, endocrine therapy, or radiotherapy before surgery; (4) had missing key predictor or outcome data; or (5) did not undergo surgery because of patient or family refusal.

Outcome Definition

The primary outcome was axillary lymph node metastasis (ALNM), defined as the presence of metastatic disease in at least one axillary lymph node on final postoperative histopathological examination. Axillary nodal status was determined from the final pathological assessment of SLNB specimens and, when performed, ALND specimens. Patients with a negative SLNB who did not undergo subsequent ALND were classified as ALNM-negative based on the final SLNB pathology. For patients undergoing ALND, the final axillary nodal status was determined from the postoperative pathological evaluation of the resected lymph nodes. Permanent formalin-fixed paraffin-embedded (FFPE) histopathology served as the reference standard for outcome determination.

Candidate Predictors and Data Collection

Predictors were obtained from preoperative clinical records, ultrasound reports, and core needle biopsy (CNB) pathology reports of the primary breast lesion, whereas the outcome (ALNM) was determined using final postoperative formalin-fixed paraffin-embedded (FFPE) histopathology of axillary specimens. Candidate predictors were selected based on clinical relevance and routine availability in standard preoperative assessment. Clinical variables included age, diabetes mellitus, and hypertension. Preoperative imaging variables were extracted from routine ultrasound reports and included tumor location, maximal tumor diameter, and the presence of calcifications as reported on ultrasound. Tumor size was recorded as the maximal diameter (cm) documented on preoperative ultrasound. All imaging variables were abstracted from the original radiology reports.

All clinicopathological candidate predictors were derived from preoperative CNB specimens of the primary breast lesion. Specifically, histological grade and lymphovascular invasion (LVI) were assessed on preoperative CNB specimens, while pathological type and biomarker status were determined on the same specimens. No postoperative histopathological characteristics of the primary breast tumor were used as predictors in model development. Biomarker variables included estrogen receptor (ER), progesterone receptor (PR), human epidermal growth factor receptor 2 (HER2), and Ki-67 proliferation index. Molecular subtype was derived from ER, PR, HER2, and Ki-67 findings. The complete candidate predictor set comprised 14 variables: age, tumor location, ultrasound-reported calcifications, diabetes mellitus, hypertension, tumor size, histological grade, LVI, pathological type, ER status, PR status, HER2 status, Ki-67 index, and molecular subtype.

Axillary Staging and Pathological Assessment

Axillary staging was performed by sentinel lymph node biopsy (SLNB) and/or axillary lymph node dissection (ALND) according to routine clinical practice. For SLNB, methylene blue dye was used for sentinel node mapping, and excised sentinel nodes were submitted for intraoperative frozen-section examination when applicable. Sentinel nodes were serially sectioned at approximately 2-mm intervals for frozen-section assessment. The final axillary pathological diagnosis used for outcome determination was based on postoperative histopathological examination of permanent formalin-fixed paraffin-embedded (FFPE) sections rather than intraoperative frozen-section findings. For patients undergoing ALND, resected axillary lymph nodes were submitted for routine histopathological examination on permanent FFPE sections. The total number of harvested lymph nodes and the number of metastatic lymph nodes were recorded for each patient.

Immunohistochemistry and Molecular Subtype

Immunohistochemistry (IHC) was performed on preoperative CNB specimens in the pathology laboratory of each participating center. Both centers followed the same pathological diagnostic criteria and interpretation standards according to the Chinese Anti-Cancer Association guidelines for breast cancer (2021 edition), and the pathology departments at the two centers were supervised by the same department director. Estrogen receptor (ER) and progesterone receptor (PR) status was assessed by manual visual interpretation, with positivity defined as nuclear staining in ≥1% of tumor cells; ER/PR expression was further categorized as low (1–10%) and high (>10%) for analysis. Human epidermal growth factor receptor 2 (HER2) status was assessed by IHC; cases scored as HER2 2+ underwent reflex in situ hybridization (ISH), and HER2 status was classified as negative or positive based on the combined IHC/ISH results. The Ki-67 proliferation index was assessed by manual visual evaluation and recorded as the percentage of positively stained tumor nuclei. At least 500 invasive tumor cells were evaluated using the hot-spot method. For prediction-model development, Ki-67 was analyzed as a continuous variable and was not dichotomized using a predefined cutoff.

ER, PR, HER2, and Ki-67 were interpreted by pathologists at the respective centers. Cases with uncertain interpretations were jointly reviewed by two pathologists and resolved by consensus. At the time of preoperative CNB interpretation, the final postoperative axillary pathological outcome was not yet available to the evaluating pathologists. Molecular subtype was assigned according to ER, PR, HER2, and Ki-67 findings as luminal A, luminal B, HER2-positive, or triple-negative breast cancer; luminal A and luminal B were combined as the luminal subtype for analysis.

Statistical Analysis

All analyses were performed using R software (version 4.2.1) and SPSS (version 26.0). Continuous variables were assessed for distributional assumptions. Normally distributed variables were presented as mean ± standard deviation and compared using the independent-samples t test, whereas non-normally distributed variables were presented as median (interquartile range, IQR) and compared using the Mann–Whitney U-test. Categorical variables were summarized as counts (percentages) and compared using the chi-square test or Fisher’s exact test, as appropriate. A two-sided P < 0.05 was considered statistically significant. As a retrospective cohort study, all eligible patients available during the predefined study period were included.

Model development was conducted in the training cohort (n = 182). The 14 candidate predictors described above were considered for model development. Variables with P < 0.20 in univariable analyses were entered into least absolute shrinkage and selection operator (LASSO) regression. The penalty parameter was selected by 10-fold cross-validation using the 1-standard-error (1-SE) criterion, and predictors with non-zero coefficients were subsequently entered into multivariable logistic regression. Regression coefficients, odds ratios (ORs) with 95% confidence intervals (CIs), and P values were reported. The LASSO-selected predictors were first entered into a full multivariable model (Model 1), and a reduced model excluding Ki-67 (Model 2) was also fitted. Model 1 and Model 2 were compared using the Akaike information criterion (AIC), likelihood-ratio testing, and bootstrap resampling with 1000 replicates. The final model was presented as a nomogram.

In a sensitivity analysis, LASSO regression was applied directly to the full 14-variable candidate predictor set without univariable P-value pre-screening, using the same 10-fold cross-validation procedure and 1-SE criterion, and the subsequent multivariable modeling procedure was repeated using the retained predictors. Model performance was evaluated in terms of discrimination, calibration, and clinical utility. Discrimination was assessed using receiver operating characteristic (ROC) curves and the area under the curve (AUC) with 95% CIs. Calibration was assessed using calibration plots, the Brier score, calibration intercept and slope, and the Hosmer–Lemeshow goodness-of-fit test. Internal validation was performed using 1,000 bootstrap resamples to estimate optimism-corrected discrimination and calibration. Clinical utility was evaluated using decision curve analysis (DCA).

A classification cutoff was determined in the training cohort using the Youden index. Sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy were calculated at this cutoff. For external validation, the final model developed in the training cohort was applied to the external validation cohort (n = 113) without model refitting, and the cutoff derived from the training cohort was applied without re-selection. Patients with missing key predictor or outcome data were excluded before analysis. No missing values remained among the analyzed variables in the final training or external validation cohorts, and no imputation was performed.

Results

Patient Characteristics

The study included 295 patients with breast cancer, comprising 182 patients in the training cohort and 113 patients in the external validation cohort. In the training cohort, the mean age was 55.06 ± 10.26 years, 117 patients (64.3%) had ultrasound-reported calcifications, and the median tumor size was 1.8 cm (IQR, 1.2–2.85). Histological grade was I in 49 patients (26.9%), II in 91 (50.0%), and III in 42 (23.1%); LVI was present in 46 patients (25.3%), and 60 (33.0%) were HER2-positive. In the external validation cohort, the mean age was 53.58 ± 9.59 years, 62 patients (54.9%) had ultrasound-reported calcifications, and the median tumor size was 1.8 cm (IQR, 1.1–2.6). Histological grade was I in 45 patients (39.8%), II in 40 (35.4%), and III in 28 (24.8%); LVI was present in 38 patients (33.6%), and 34 (30.1%) were HER2-positive. No statistically significant differences were observed in the baseline characteristics examined between the two cohorts (all P > 0.05; Table 1). No missing values remained among the analyzed variables in either cohort, and no imputation was performed.

Table 1.

Baseline Characteristics of the Training and External Validation Cohorts

Features		Training Cohort (n=182)	External Validation Cohort (n=113)	t/Z/χ2	P
Age		55.06±10.26	53.58±9.59	1.232	0.219
Location	Lateral	119(65.4)	66(58.4)	1.451	0.228
	Medial	63(34.6)	47(41.6)
Calcifications	No	65(35.7)	51(45.1)	2.592	0.107
	Yes	117(64.3)	62(54.9)
DM	No	142(78)	96(85)	2.150	0.143
	Yes	40(22)	17(15)
Hypertension	No	111(61)	80(70.8)	2.938	0.087
	Yes	71(39)	33(29.2)
Tumor size (cm)		1.8(1.2,2.85)	1.8(1.1,2.6)	−0.279	0.780
Histological grade	I	49(26.9)	45(39.8)	−1.346	0.178
	II	91(50)	40(35.4)
	III	42(23.1)	28(24.8)
Lymphovascular invasion (LVI)	No	136(74.7)	75(66.4)	2.389	0.122
	Yes	46(25.3)	38(33.6)
Pathological type	NST	163(89.6)	103(91.2)	0.199	0.656
	Special Type	19(10.4)	10(8.8)
PR	Negative	47(25.8)	20(17.7)	2.622	0.105
	Positive	135(74.2)	93(82.3)
ER	Negative	94(51.6)	47(41.6)	2.825	0.093
	Positive	88(48.4)	66(58.4)
HER2	Negative	122(67)	79(69.9)	0.266	0.606
	Positive	60(33)	34(30.1)
Ki-67		40(20,60)	30(20,50)	−1.721	0.085
Molecular subtype	Luminal	104(57.1)	65(57.5)	1.131	0.568
	HER2 positive	48(26.4)	34(30.1)
	Triple-negative	30(16.5)	14(12.4)

Notes: Data are presented as mean ± SD, median (IQR), or n (%), as appropriate. Between-cohort comparisons were performed using the independent-samples t test for normally distributed continuous variables, the Mann–Whitney U-test for non-normally distributed continuous or ordinal variables, and the Pearson chi-square test for categorical variables.

Abbreviations: DM, diabetes mellitus; ER, estrogen receptor; HER2, human epidermal growth factor receptor 2; IQR, interquartile range; LVI, lymphovascular invasion; NST, no special type; PR, progesterone receptor; SD, standard deviation.

ALNM was identified in 88 of 182 patients (48.4%) in the training cohort and in 50 of 113 patients (44.2%) in the external validation cohort. Among the 88 node-positive patients in the training cohort, 70 underwent SLNB as the initial axillary procedure and 18 underwent upfront ALND. Completion ALND was performed in 44 of the 70 patients initially undergoing SLNB; additional non-sentinel lymph node metastasis was identified in 30 of these 44 patients (68.2%), whereas 14 had no additional involved nodes.

Univariable Analysis

In univariable analyses (Table 2), patients with ALNM were more likely to have ultrasound-reported calcifications (78.4% vs 51.1%, P < 0.001), larger tumors (median, 2.5 vs 1.5 cm; P < 0.001), higher histological grade (grade III, 37.5% vs 9.6%; P < 0.001), and LVI (43.2% vs 8.5%, P < 0.001). HER2 positivity was also more frequent in the metastasis group (42.0% vs 24.5%, P = 0.012), and the Ki-67 index was higher (median, 42.5 vs 30.0; P = 0.001). Tumor location (P = 0.089), hypertension (P = 0.156), and molecular subtype (P = 0.149) also met the predefined screening criterion of P < 0.20. Accordingly, nine variables were taken forward for LASSO feature selection.

Table 2.

Univariable Analysis of Factors Associated with Axillary Lymph Node Metastasis in the Training Cohort

Features		Non-Metastasis Group (n=94)	Metastasis Group (n=88)	t/Z/χ2	P
Age		55.54±10.51	54.55±10.02	0.654	0.514
Tumor location	Lateral	56(59.6)	63(71.6)	2.900	0.089
	Medial	38(40.4)	25(28.4)
Calcifications	No	46(48.9)	19(21.6)	14.803	<0.001
	Yes	48(51.1)	69(78.4)
DM	No	70(74.5)	72(81.8)	1.432	0.231
	Yes	24(25.5)	16(18.2)
Hypertension	No	62(66)	49(55.7)	2.017	0.156
	Yes	32(34)	39(44.3)
Tumor size (cm)		1.5(1.1,2.0)	2.5(1.5,3.3)	−4.943	<0.001
Histological grade	I	37(39.4)	12(13.6)	−5.097	<0.001
	II	48(51.1)	43(48.9)
	III	9(9.6)	33(37.5)
Lymphovascular invasion (LVI)	No	86(91.5)	50(56.8)	28.928	<0.001
	Yes	8(8.5)	38(43.2)
Pathological type	NST	83(88.3)	80(90.9)	0.331	0.565
	Special Type	11(11.7)	8(9.1)
PR	Negative	23(24.5)	24(27.3)	0.187	0.666
	Positive	71(75.5)	64(72.7)
ER	Negative	51(54.3)	43(48.9)	0.529	0.467
	Positive	43(45.7)	45(51.1)
HER2	Negative	71(75.5)	51(58)	6.354	0.012
	Positive	23(24.5)	37(42)
Ki-67		30(18.8,50)	42.5(30,70)	−3.234	0.001
Molecular subtype	Luminal	60 (63.8)	44 (50.0)	3.801	0.149
	HER2 positive	22 (23.4)	26 (29.5)
	Triple-negative	12 (12.8)	18 (20.5)

Notes:Data are presented as mean ± SD, median (IQR), or n (%), as appropriate. Between-group comparisons were performed using the independent-samples t test for normally distributed continuous variables, the Mann–Whitney U-test for non-normally distributed continuous or ordinal variables, and the Pearson chi-square test for categorical variables.

Abbreviations: DM, diabetes mellitus; ER, estrogen receptor; HER2, human epidermal growth factor receptor 2; IQR, interquartile range; LVI, lymphovascular invasion; NST, no special type; PR, progesterone receptor; SD, standard deviation.

Feature Selection Using LASSO Regression

Nine candidate variables meeting the univariable screening criterion of P < 0.20 were entered into LASSO regression. Using 10-fold cross-validation, the penalty parameter was selected at λ = 0.072 according to the 1-standard-error (1-SE) criterion, resulting in six predictors with non-zero coefficients (Figure 2A and B): tumor size, ultrasound-reported calcifications, histological grade, LVI, HER2 status, and Ki-67. In the sensitivity analysis, LASSO regression was applied directly to the full set of 14 candidate predictors without univariable pre-screening. Using the same 10-fold cross-validation procedure and 1-SE criterion, the same six predictors were retained, and the subsequent multivariable modeling procedure resulted in the same final five-predictor model.

Figure 2.

Feature selection using LASSO regression. (A) Selection of the optimal penalty parameter (λ) by 10-fold cross-validation; the dotted lines indicate the minimum-criteria and 1-standard-error (1-SE) values (λ = 0.072 was chosen by the 1-SE criterion). (B) LASSO coefficient profiles of the nine candidate predictors plotted against log(λ); six predictors had non-zero coefficients at the selected λ.

Development of the Prediction Model

The six LASSO-selected predictors—tumor size, ultrasound-reported calcifications, histological grade, LVI, HER2 status, and Ki-67—were entered into multivariable logistic regression (Model 1). In Model 1, Ki-67 was not independently associated with ALNM (OR 1.012, 95% CI 0.996–1.028, P = 0.173). A reduced model excluding Ki-67 (Model 2) was also evaluated to determine whether a more parsimonious model could retain comparable predictive performance. Model 2 had a slightly lower AIC than Model 1 (182.322 vs 182.438), and the likelihood-ratio test showed no significant difference between the two models (P = 0.170). Bootstrap validation with 1,000 resamples yielded optimism-corrected AUCs of 0.838 and 0.840 and calibration slopes of 0.866 and 0.890 for Model 1 and Model 2, respectively. On the basis of these results and model parsimony, Model 2 was retained as the final prediction model.

In the final model, tumor size (OR 1.772 per 1-cm increase, 95% CI 1.245–2.521, P = 0.001), ultrasound-reported calcifications (OR 2.945, 95% CI 1.318–6.577, P = 0.008), histological grade II (OR 2.793, 95% CI 1.052–7.412, P = 0.039) and grade III (OR 12.231, 95% CI 3.722–40.194, P < 0.001) versus grade I, LVI (OR 7.622, 95% CI 2.783–20.873, P < 0.001), and HER2 positivity (OR 2.545, 95% CI 1.139–5.684, P = 0.023) were independently associated with ALNM (Table 3).

Table 3.

Multivariable Logistic Regression Models for Predicting Axillary Lymph Node Metastasis in the Training Cohort

Features		Model 1				Model 2
		β	SE	OR (CI 95%)	P	β	SE	OR (CI 95%)	P
Tumor size(cm)		0.565	0.179	1.759(1.239–2.499)	0.002	0.572	0.180	1.772(1.245–2.521)	0.001
Calcification	No			1.000				1.000
	Yes	1.072	0.414	2.921(1.298–6.576)	0.010	1.080	0.410	2.945(1.318–6.577)	0.008
Histological grade	I			1.000				1.000
	II	1.049	0.500	2.855(1.071–7.606)	0.036	1.027	0.498	2.793(1.052–7.412)	0.039
	III	2.289	0.623	9.865(2.909–33.451)	<0.001	2.504	0.607	12.231(3.722–40.194)	<0.001
Lymphovascular invasion (LVI)	No			1.000				1.000
	Yes	2.108	0.523	8.232(2.953–22.945)	<0.001	2.031	0.514	7.622(2.783–20.873)	<0.001
HER2	Negative			1.000				1.000
	Positive	0.884	0.414	2.421(1.075–5.449)	0.033	0.934	0.410	2.545(1.139–5.684)	0.023
Ki-67		0.012	0.008	1.012(0.996–1.028)	0.173	-	-	-	-
Intercept		−4.224	0.736	0.015(0.003–0.062)	<0.001	−3.805	0.646	0.022(0.006–0.079)	<0.001
AIC		182.438				182.322

Notes: Model 1 included all six predictors retained by LASSO regression, whereas Model 2 excluded Ki-67 and was retained as the final prediction model. Tumor size and Ki-67 were entered as continuous variables. The OR for tumor size represents the change in odds per 1-cm increase, and the OR for Ki-67 represents the change in odds per 1-percentage-point increase. Reference categories were absence of ultrasound-reported calcifications, histological grade I, absence of LVI, and HER2-negative status.

Abbreviations: AIC, Akaike information criterion; CI, confidence interval; HER2, human epidermal growth factor receptor 2; LASSO, least absolute shrinkage and selection operator; LVI, lymphovascular invasion; OR, odds ratio; SE, standard error.

The final logistic regression equation was:

logit(P)=−3.805+0.572×tumor size+1.080×calcifications+1.027×grade II+2.504×grade III+2.031×LVI+0.934×HER2 positivity.

where tumor size was entered as a continuous variable in centimeters; ultrasound-reported calcifications, LVI, and HER2 positivity were coded as 1 for presence/positivity and 0 for absence/negativity; and histological grades II and III were represented by indicator variables, with grade I as the reference category.

Construction of the Nomogram

A nomogram was constructed based on the final five-predictor logistic regression model to facilitate individualized estimation of ALNM risk (Figure 3). The nomogram incorporates tumor size, ultrasound-reported calcifications, histological grade, LVI, and HER2 status. Each predictor is assigned a weighted point value according to its regression coefficient, and the summed total score corresponds to the estimated probability of ALNM.

Figure 3.

Nomogram for predicting axillary lymph node metastasis in breast cancer. The nomogram integrates five predictors—ultrasound-reported calcifications, tumor size, lymphovascular invasion, histological grade, and HER2 status. Points assigned to each variable are summed to obtain a total score, which corresponds to the estimated probability of ALNM. The red markers illustrate an example patient and the resulting predicted risk.

Model Performance and Validation

The final model demonstrated good discrimination, with an AUC of 0.858 (95% CI 0.805–0.911) in the training cohort and 0.854 (95% CI 0.779–0.929) in the external validation cohort (Figure 4A and B). In the training cohort, the Brier score was 0.154, with an apparent calibration intercept of 0.000 and calibration slope of 1.000. The Hosmer–Lemeshow test showed no evidence of lack of fit (χ² = 8.524, P = 0.482). Bootstrap internal validation with 1,000 resamples yielded an optimism-corrected AUC of 0.840 and an optimism-corrected calibration slope of 0.890. In the external validation cohort, the Brier score was 0.144, the calibration intercept was −0.098, and the calibration slope was 0.778; the Hosmer–Lemeshow test was also non-significant (χ² = 12.793, P = 0.172) (Figure 4C and D).

Figure 4.

Discrimination, calibration, and clinical utility of the nomogram in the training and external validation cohorts. (A and B) Receiver operating characteristic (ROC) curves for the training cohort (A) and external validation cohort (B). (C and D) Calibration plots for the training cohort (C) and external validation cohort (D); the gray diagonal line indicates ideal calibration and the black line shows the nomogram’s performance. (E and F) Decision curve analysis (DCA) for the training cohort (E) and external validation cohort (F), showing net benefit across threshold probabilities.

Using the ROC-derived cutoff probability of 0.524 in the training cohort, sensitivity was 75.0% (95% CI 64.4–83.4%), specificity was 84.0% (95% CI 74.7–90.5%), PPV was 81.5% (95% CI 71.0–88.9%), NPV was 78.2% (95% CI 68.7–85.6%), and accuracy was 79.7% (95% CI 72.9–85.1%). When the same cutoff was applied to the external validation cohort without re-selection, sensitivity was 70.0% (95% CI 55.2–81.7%), specificity was 90.5% (95% CI 79.8–96.1%), PPV was 85.4% (95% CI 70.1–93.9%), NPV was 79.2% (95% CI 67.7–87.5%), and accuracy was 81.4% (95% CI 72.8–87.9%).

Decision curve analysis showed that the nomogram provided a higher net benefit than the treat-all and treat-none strategies over threshold probabilities of approximately 0.03–0.99 in the training cohort and 0.17–0.99 in the external validation cohort (Figure 4E and F).

Discussion

ALNM remains a pivotal determinant of staging, prognosis, and treatment planning in breast cancer, yet accurate preoperative identification of nodal involvement continues to be challenging.^19,20 While SLNB has reduced the morbidity associated with routine axillary lymph node dissection, it remains an invasive staging procedure and may represent overtreatment for some patients at very low risk of nodal disease.²¹ Current preoperative assessment relies largely on imaging, most commonly axillary ultrasound, but its diagnostic accuracy for small-volume metastasis varies across operators and institutions.²⁰ These limitations support the need for practical quantitative tools that integrate routinely available preoperative information. In this two-center retrospective study, we developed and externally validated an interpretable nomogram using routine preoperative ultrasound and core needle biopsy variables to estimate individualized ALNM risk. The final model incorporated five predictors—tumor size, ultrasound-reported calcifications, histological grade, lymphovascular invasion, and HER2 status—and demonstrated good discrimination in both the training and external validation cohorts, with AUCs of 0.858 and 0.854, respectively, together with acceptable calibration and potential clinical utility.

Among the predictors included in the final model, tumor size showed a robust association with nodal metastasis, consistent with previous evidence that the risk of ALNM increases with increasing primary tumor burden.^12,22 Several studies have reported higher rates of nodal involvement in patients with larger tumors.^22,23 Nevertheless, the relationship between tumor size and metastatic spread is biologically heterogeneous and may vary across molecular subtypes.²⁴ Small primary tumors may still exhibit substantial nodal or distant dissemination, indicating that tumor size alone is insufficient for individualized risk stratification.²⁵ Histological grade was also independently associated with ALNM, in agreement with previous studies identifying higher tumor grade as an important risk factor for nodal involvement.²⁵ These routinely available clinicopathological characteristics have also formed part of established ALNM prediction approaches, including conventional nomograms developed from clinical and pathological information.^9,10 Their continued contribution across different prediction frameworks supports their value as core components of multivariable risk estimation, while also highlighting the need to integrate them with complementary imaging and biomarker information.

Lymphovascular invasion is a biologically plausible and clinically informative marker of metastatic potential and has been recognized as a strong predictor of nodal involvement.^26,27 In the present study, LVI was strongly associated with ALNM, consistent with previous clinical and pathological evidence showing that tumor invasion of lymphatic or vascular channels is associated with a higher likelihood of nodal dissemination and adverse outcomes in breast cancer.^27,28 Mechanistically, interactions between tumor cells and the surrounding microenvironment may promote lymphangiogenesis and lymphatic invasion; for example, podoplanin-expressing tumor-associated macrophages have been implicated in facilitating these processes in experimental models.²⁹ The incorporation of LVI adds a tissue-based dimension to preoperative risk estimation and complements models that rely predominantly on clinical or imaging characteristics. In the preoperative setting, however, LVI is assessed from a limited core biopsy rather than the entire tumor specimen. Given its potentially focal distribution, CNB-based assessment may not capture all areas of lymphovascular involvement, which may influence the stability and reproducibility of this predictor across different pathological settings.

Ultrasound-reported calcifications emerged as an independent predictor of ALNM in our model. Previous studies have suggested that breast calcifications may be associated with tumor progression and unfavorable biological characteristics, while mechanistic evidence has implicated specific mineral components, such as hydroxyapatite, in tumor-related processes.³⁰ Clinical imaging studies, most commonly based on mammography, have also linked microcalcifications with adverse prognostic features in breast cancer.³¹ More recent ALNM prediction models have incorporated increasingly detailed imaging information, including breast and axillary sonographic characteristics, CT-derived nodal features, and radiomics signatures.^11–18 In contrast, the calcification variable in the present study was derived from routine preoperative ultrasound reports. Detailed characterization of breast microcalcifications is predominantly mammography-based, whereas ultrasonography detects only a subset of calcifications and is more dependent on lesion characteristics and reporting practice. Thus, ultrasound-reported calcifications should not be regarded as equivalent to mammographically characterized microcalcifications, but their independent association with ALNM suggests that this routinely available sonographic information may still contribute to multivariable preoperative risk estimation.

HER2 positivity remained independently associated with ALNM in our model, consistent with the recognized biological importance of HER2-positive breast cancer.³² HER2 status and Ki-67 have also been incorporated into previous biomarker-based ALNM prediction models.³³ In our cohort, Ki-67 was higher in patients with ALNM in univariable analysis, but its contribution became limited after adjustment for the other selected predictors, and exclusion of Ki-67 did not materially reduce model fit, discrimination, or calibration. Ki-67 reflects tumor proliferative activity and is widely used as a marker of cellular proliferation.³⁴ The relationship between molecular subtype and nodal metastasis has been less consistent across studies. Population-based and retrospective analyses have reported differences in nodal involvement across breast cancer molecular subtypes, although the direction and magnitude of these associations vary among populations.^35,36 These heterogeneous findings suggest that individual biomarkers or molecular subtype alone may not provide sufficiently stable estimation of nodal risk across populations, supporting their interpretation within a broader multivariable framework.

From a clinical perspective, the nomogram may serve as an adjunctive tool for preoperative risk stratification by integrating routinely available imaging and biopsy-derived information into an individualized estimate of ALNM probability. Decision curve analysis showed a higher net benefit than the treat-all and treat-none strategies across threshold probabilities of approximately 0.03–0.99 in the training cohort and 0.17–0.99 in the external validation cohort. Within clinically relevant portions of these ranges, a higher predicted probability may support more detailed axillary evaluation, including additional imaging or ultrasound-guided nodal biopsy when indicated, and may facilitate preoperative counseling and surgical planning. Conversely, a lower predicted probability may help contextualize the likelihood of nodal involvement during shared decision-making. The nomogram is intended to complement rather than replace established axillary staging procedures, and the clinical consequences of acting on specific probability thresholds require prospective evaluation before routine implementation.

Several aspects of this study may affect the generalizability of the findings. The retrospective design carries an inherent risk of selection and information bias, and the relatively modest sample was drawn from two clinical centers within the same regional healthcare system. The external validation cohort was retrospectively assembled using the same eligibility criteria as the training cohort; however, detailed screening and exclusion counts before cohort assembly were not available, which may introduce additional selection bias. Some predictors were derived from routine clinical assessments rather than standardized research protocols; in particular, ultrasound-reported calcifications may vary with image quality and reporting practice, while CNB-based assessment of LVI may be affected by limited tissue sampling. Detailed axillary ultrasound features, such as cortical thickness, hilar morphology, cortical asymmetry, and vascularity, were not systematically available and therefore could not be incorporated into the model. These factors may influence model performance when applied to populations with different imaging or pathological workflows. Further prospective validation in larger, geographically and institutionally diverse cohorts, with standardized imaging and pathological assessment, is needed to evaluate the model’s reproducibility and to define clinically actionable probability thresholds before routine implementation.

Conclusion

This study developed and externally validated a preoperative nomogram for predicting ALNM in breast cancer using five routinely available variables: tumor size, ultrasound-reported calcifications, histological grade, LVI, and HER2 status. The model demonstrated good discrimination and acceptable calibration in both the training and external validation cohorts, supporting its potential value as an adjunct to preoperative risk stratification and clinical decision-making. Further prospective validation in larger and more diverse multicenter populations is warranted before routine clinical implementation.

Funding Statement

This study was supported by Xinjiang Uygur Autonomous Region Regional Collaborative Innovation Special Program (Grant No.2022E02136).

Abbreviations

ALN, axillary lymph node; ALND, axillary lymph node dissection; ALNM, axillary lymph node metastasis; AUC, area under the curve; CI, confidence interval; cN0, clinically node-negative; CNB, core needle biopsy; DCA, decision curve analysis; ER, estrogen receptor; FFPE, formalin-fixed paraffin-embedded; HER2, human epidermal growth factor receptor 2; IHC, immunohistochemistry; IQR, interquartile range; ISH, in situ hybridization; Ki-67, Ki-67 proliferation index; LASSO, least absolute shrinkage and selection operator; LVI, lymphovascular invasion; MRI, magnetic resonance imaging; OR, odds ratio; PR, progesterone receptor; ROC, receiver operating characteristic; SLNB, sentinel lymph node biopsy.

Data Sharing Statement

The data used in this study are available from the corresponding author upon reasonable request.

Ethics Approval and Consent to Participate

This study was approved by the Ethics Committee of the First Affiliated Hospital of Xinjiang Medical University (Approval No. 230714-09), covering the First Affiliated Hospital of Xinjiang Medical University (Urumqi, Xinjiang, China) and its Changji Branch (Changji, Xinjiang, China). All procedures involving human participants were conducted in accordance with the Declaration of Helsinki. Written informed consent was obtained from all participants prior to enrollment.

Consent for Publication

This was deemed not applicable.

Author Contributions

All authors made a significant contribution to the work reported, whether that is in the conception, study design, execution, acquisition of data, analysis and interpretation, or in all these areas; took part in drafting, revising, or critically reviewing the article; gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work. Nijiati Taxifulati: Conceptualization; Methodology; Data curation; Validation; Formal analysis; Supervision; Writing – review & editing. Gulisumuhan Abulaiti: Methodology; Writing – original draft; Writing – review & editing. Yuan Zhang: Investigation; Data curation; Writing – review & editing. Chenming Guo: Investigation; Data curation; Writing – review & editing. Xin Yang: Data curation; Investigation; Validation; Supervision; Writing – review & editing.

Disclosure

The authors declare that they do not have a conflict of interest.