Patients diagnosed with breast cancer by preoperative histopathologic biopsy in West China Hospital, Sichuan University, from 1 August, 2022 to 31 January, 2024 were prospectively included. The pathological diagnosis was based on the World Health Organization (WHO) Classification of Breast Tumors (35). Patients meeting the following criteria were excluded: 1) cancer primary focus and/or LN metastasis focus has been resected via breast surgery (including breast-conserving surgery, mastectomy, and even Mammotome) and/or axillary surgery (including ALND and SLNB), 2) inflammatory breast cancer, 3) a history of malignancy, and 4) necessary clinicopathological data were absent.

The study was conducted in accordance with the Declaration of Helsinki and approved by the Ethics Committee of West China Hospital, Sichuan University. Written informed consent was obtained from all patients.

Patients’ clinical information was mainly collected by face-to-face inquiry, and histopathologic information of tumors was obtained from the electronic medical records. Clinical features included age at diagnosis, body mass index (BMI), age at menarche, the number of pregnancies and deliveries, menopause status at diagnosis, and neoadjuvant therapy (NAT) or not. BMI was calculated by formula weight/height² (kg/m²), and the height and weight were measured in an outpatient setting. The histopathologic reports of preoperative US-guided CNB for breast primary tumors, including estrogen receptor (ER), progesterone receptor (PR), human epidermal growth factor receptor 2 (HER2), Ki-67, and histologic types, were recorded. The features of the largest tumor were recorded for patients with multiple tumors. The status of ER, PR, HER2, and Ki-67 was evaluated by immunohistochemical (IHC) staining, and HER2 needed extra detection by fluorescence in situ hybridization (FISH) when IHC showed 2+. ER or PR was identified as a positive result if the positivity rate was ≥1%, and HER2 was identified as a positive result if IHC staining presented 3+, or IHC showed 2+ but FISH showed positive. According to the St Gallen International Expert Consensus, 2013 (36), based on the status of ER, PR, HER2, and Ki-67, breast cancer was categorized into four molecular subtypes, as follows: luminal A subtype (ER+, PR ≥ 20%, HER2−, Ki-67 < 14%), luminal B subtype (ER+/PR+ but did not meet the condition of luminal A), HER2-enriched subtype (ER−, PR−, HER2+), and triple-negative subtype (ER−, PR−, HER2−).

The US scans were evaluated 3–5 days before surgery by two ultrasound experts (with 5–10 years of experience in breast US) blinded to the axillary surgery plan, and the images and features for breasts and ALN were collected. They reached a consensus through discussion when any disagreements in the independent analysis were encountered. US features of the tumor included tumor size (maximum diameter of breast lesion), location, distance to the nipple, blood flow signals, infiltration status of the subcutaneous layer, and retromammary space. In this study, the blood flow signals of tumors were divided into two categories based on Adler grades: poor (Adler grades 0–1) and abundant (Adler grades 2–3) blood flow signals (37). The characteristics of the largest tumor were recorded for patients with multiple tumors. In addition, US features of LNs contained long and short axes, blood flow signal types and grades, margin, corticomedullary demarcation, and cortical thickness and evenness. The cortical thickness was measured based on the thickest location. Then, the LN long/short (L/S) axis ratio by long and short axes was calculated. The characteristics of the most suspicious LN (the largest one usually) were recorded.

The surgery was performed by six specialists in breast surgery with more than 10 years of experience in breast cancer surgery. The ALN status reported by the postoperative histopathologic examination results of ALND or SLNB was recorded. Then, patients were divided into ALNM and non-ALNM groups by the ALN status. Macro-metastases (>2 mm) and micro-metastases (0.2–2 mm) were identified as ALNM, while isolated tumor cells (ITCs) (<0.2 mm) and negative SLN were considered non-ALNM (38).

All the patients were followed up. The end-point of follow-up was the acquirement of ALN status by postoperative histopathologic examination. During the follow-up, NAT status was checked via electronic medical records when the patients were admitted to the hospital for surgery. The patients who were not undergoing a surgical treatment or refused preoperative US in West China Hospital were lost to follow-up.

The patients’ data were registered in an Excel spreadsheet and were managed by a dedicated researcher. Clinicopathological feature collection was performed by two researchers in charge of acquisition and recording. The US features were recorded by one of the two ultrasound experts after a consensus was reached. Another two researchers were responsible for follow-up until postoperative pathological ALN status was recorded.

Statistical analysis was performed using SPSS 23.0 and the R software ver.4.3.2. Continuous variables were expressed as mean ± standard deviation (SD), while categorical variables were presented as numbers and percentages of the group they belong to and analyzed by χ² test (Yates’ correction if necessary) or Fisher’s exact test, which was used to compare the data distribution of the training and test cohorts.

The patients included were randomly divided into a training set and a test set according to the proportion of 7:3. In the training cohort, the variables associated with ALN status were preliminarily screened by univariate binary logistic regression analysis. Those significant variables in univariate analysis were included in multivariate binary logistic regression analysis, using the forward stepwise (likelihood ratio), to acquire the independent predictors of ALNM. The p-values, odds ratios (ORs), and 95% confidence intervals (CIs) were reported to present the analysis. The statistical analyses were two-sided, and p < 0.05 was considered statistically significant. The collinearity diagnosis of independent predictors was performed, and then the tolerance and variance inflation factor (VIF) were calculated.

The binary logistic regression model to predict ALNM was constructed using independent predictors and presented as a nomogram. The Hosmer–Lemeshow goodness-of-fit test was used to evaluate the fit level of the model. For the training and test cohorts, the discrimination of the model was evaluated by the receiver operating characteristic (ROC) curve to calculate the area under the ROC curve (AUC). Next, for both cohorts, the calibration of the model was evaluated by bootstrapping with 500 resamples and presented by the calibration curve. Last, to evaluate the clinical value of the nomogram, the net benefits of the model for both cohorts were measured using decision curve analysis (DCA) and shown by the DCA curve. Moreover, sensitivity, specificity, accuracy, PPV, negative predictive value (NPV), detection rate, and FNR of the model at different threshold values in the training and test sets were calculated to provide references for threshold value selection in situations with different requirements.

Overall, 854 patients were registered. A total of 176 patients were excluded because their cancer primary focus and/or LN metastasis focus had been resected, 40 patients were excluded due to inflammatory breast cancer, and 12 patients were excluded because their necessary clinicopathological information was absent. Ultimately, 228 patients were excluded, and 626 patients were included. Moreover, 14 patients were lost to follow-up because they did not undergo surgical treatment in West China Hospital, and two patients were lost to follow-up due to refusal to undergo preoperative US in West China Hospital. Finally, 16 patients were lost to follow-up, and the remaining 610 patients were included for analysis, which is illustrated in Figure 1 .

Of these 610 patients, 11 (1.8%) patients had unknown HER2 status, and one patient had unknown Ki-67 status. These missing data were not indispensable for model construction, so these patients were retained without handling their missing data. The mean age at diagnosis of the 610 patients was 50.9 ± 10.7 years (range, 25–86 years). As confirmed by postoperative histopathologic examination, 294 (48.2%) patients had ALNM, while the other 316 (51.8%) patients had negative ALN. In addition, 72 (11.8%) patients underwent NAT. More than half of the patients (55.6%) had tumors of the luminal B subtype, and most patients (94.8%) had ductal carcinoma. In addition, 84.4% of patients had LNs with a short axis <10 mm. The comparison of the clinicopathological and US features of patients between the training set (427 patients) and the test set (183 patients) is shown in Table 1 . The distribution of variables between the two sets was basically consistent, with a slight difference in the tumor location ( Table 1 ).

The clinicopathological and US features of patients with or without ALNM in the training and test cohorts are summarized in Table 2 . In the training cohort, 202 (47.3%) patients had ALNM, and the remaining 225 (52.7%) patients had negative ALN. In the test cohort, 92 (50.3%) patients had ALNM, and the remaining 91 (49.7%) patients had negative ALN.

Univariate binary logistic regression analysis was conducted in the training set ( Table 3 ). According to preoperative histopathologic biopsy, ER-positive, luminal A/B subtype, or ductal carcinoma was associated with ALNM. In addition, larger tumor size, abundant blood flow signals of the tumor, and tumor infiltration of the subcutaneous layer or retromammary space reported in preoperative US of breasts were also associated with ALNM. Moreover, according to preoperative US for ALN, longer LN long axis or short axis, smaller LN L/S axis ratio, peripheral or mixed blood flow signal type of LN, grade 3 blood flow signal of LN, obscure LN margin or corticomedullary demarcation, and thicker or more uneven LN cortex were also associated with ALNM.

As shown in Table 3 , multivariate binary regression analysis revealed that molecular subtypes, tumor infiltration of the subcutaneous layer, tumor infiltration of the retromammary space, LN short axis, LN L/S axis ratio, LN corticomedullary demarcation, and LN cortical thickness evenness were independent predictors of ALNM. Compared with triple-negative (TN) subtypes, luminal A (adjusted OR [95% CI], 3.611 [1.545–8.442], p = 0.003) and B (adjusted OR [95% CI], 2.288 [1.104–4.740], p = 0.026) subtypes were both more prone to ALNM. In addition, tumor infiltration of the subcutaneous layer (adjusted OR [95% CI], 2.755 [1.627–4.663], p < 0.001) and the retromammary space (adjusted OR [95% CI], 2.534 [1.443–4.452], p = 0.001) were both risk factors of ALNM. Moreover, the risk of metastasis greatly increased when the LN short axis was too long. Compared with short axis <10 mm, ≥10 and <15 mm indicated that ALNM risk was more than three times (adjusted OR [95% CI], 3.354 [1.329–8.461], p = 0.010), while short axis ≥15 mm indicated that ALNM risk was nearly 10 times (adjusted OR [95% CI], 9.696 [1.006–93.416], p = 0.049). LNs with L/S axis ratio ≤2 had a higher risk of metastasis than LNs with L/S axis ratio >2 (adjusted OR [95% CI], 2.003 [1.121–3.577], p = 0.019). In addition, LNs with obscure corticomedullary demarcation (adjusted OR [95% CI], 8.405 [4.186–16.876)], p < 0.001) had a great risk of metastasis, and uneven cortex (adjusted OR [95% CI], 1.842 [1.085–3.129], p = 0.024) also increased the metastasis risk of LN. According to the multicollinearity test, there was no collinearity among these independent predictors (tolerance > 0.2 and VIF < 5).

A binary logistic regression model to predict the probability of ALNM was developed based on the abovementioned independent predictors and presented as a nomogram ( Figure 2 ). The Hosmer–Lemeshow goodness-of-fit test suggested that the model fit well (p = 0.770).

According to further evaluation, the model had good discrimination ability with AUCs of 0.854 (95% CI, 0.818–0.890) and 0.822 (95% CI, 0.762–0.882) for the training set and the test set, respectively ( Figures 3A, B ). Moreover, the calibration curves indicated good agreement between predicted and observed probabilities, with the mean absolute error of 0.009 and 0.048 (500 repetitions) for the training set and the test set, respectively ( Figures 3C, D ). In the test set, when the observed probability was in the middle range (approximately 0.3–0.85), the model slightly overestimated the risk; however, when the observed probability was low or high (roughly <0.3 or >0.85), the model slightly underestimated the risk. DCA curves showed that the model could acquire net benefit with a threshold range of roughly 0.15–0.9 and 0.15–0.95 for the training group and the test group, respectively ( Figures 3E, F ). In particular, the model acquired greater net benefit (≥0.4) when the threshold range was approximately 0.15–0.65 and 0.15–0.8 for the training group and test group, respectively. Therefore, it was demonstrated that the model was of great benefit to guide clinical decisions for predicting ALNM.

As shown in Table 4 , with a risk threshold value of 0.5, the sensitivity, specificity, and accuracy of the model were 71.3%, 84.9%, and 78.5%, respectively, for the training set and 65.2%, 87.9%, and 76.5%, respectively, for the test set. Meanwhile, the detection rates were 41.7% and 38.8% in the training set and the test set, respectively, and these patients were identified to have ALNM and were likely to undergo further invasive examinations. In addition, the performance at different threshold values was presented to provide references for threshold value selection in situations with different requirements. For example, with a threshold value of 0.231, the sensitivity, specificity, and accuracy were 90.6%, 55,1%, and 71.9%, respectively, for the training set and 82.6%, 57.1%, and 69.9%, respectively, for the test set; the detection rates were 66.5% and 62.8% in the training set and the test set, respectively.