The present retrospective cohort study was approved by the Ethics Committee of Hubei Cancer Hospital (approval no. 2020-KYLL-S08; Wuhan, China). The present study included a total of 883 patients diagnosed with breast cancer who underwent surgical procedures after NAC at the Hubei Cancer Hospital (Wuhan, China) from January 2021 to December 2022. Of these, 529 patients initially met the inclusion criteria. All data were derived from archived electronic medical records. The present study performed retrospective data extraction in early 2023, after all included patients had completed their full treatment courses and their electronic medical records were fully archived. No prospective patient enrollment or real-time data collection was conducted during the clinical treatment period. The decision to administer NAC was based on the guidelines set forth by the National Comprehensive Cancer Network guidelines (31). The inclusion criteria were as follows: i) Patients with a confirmed diagnosis for breast cancer via core needle biopsy at Hubei Cancer Hospital prior to NAC; ii) patients with complete pathological and immunohistochemical evaluations of both pre-NAC biopsies and post-operative samples; and iii) patients with US examinations and comprehensive reports before and after NAC, along with serum CA153 levels measured at both time points. A total of 206 patients were excluded from the analysis due to various reasons, including secondary breast cancer, inflammatory breast cancer or bilateral breast cancer, ambiguous clinical staging or molecular subtype, incomplete clinical and pathological data, subpar US images that hindered the extraction of ultrasonic characteristics and notable dysfunction of the heart, liver, kidneys or other organs (Fig. 1). It should be noted that the initial protocol preliminarily estimated a sample size of 200 cases based on historical admission data. However, the expansion of the Breast Cancer Center led to a marked rise in patient volume starting in 2021. The present study ultimately included all consecutive eligible patients within the predefined study period, yielding a total of 323 included cases.

US evaluations were performed by specialists with ≥5 years of professional experience. Each patient underwent standard brightness-mode US at the beginning (pre-NAC) and again within 1 week before surgery after completing NAC. Patients received 6–8 cycles of NAC. The timing of the last preoperative US was determined by the clinical team based on factors such as treatment response, side effects and surgical scheduling, occurring after the 6th, 7th or 8th cycle. All participants in the present study were examined using different US scanners and transducer probes, positioned supine with their arms raised to allow complete visibility of the breast. These assessments focused on various parameters, including morphological characteristics, margins, location (indicated by clock-point numbers and distance from the skin), maximum tumor diameter, posterior echo pattern, interaction with adjacent tissues, internal calcifications and blood flow. If elastography or contrast-enhanced US (CEUS) examinations were performed, those findings were documented concurrently. These examinations were conducted at the discretion of the senior ultrasound physician when additional assessment of tumor vascularity (CEUS) or stiffness (elastography) was considered necessary to aid in classification.

All pathological specimens obtained after surgery were analyzed by two experienced pathologists from Hubei Cancer Hospital, each with >5 years of experience in the field. The evaluation utilized the Miller and Payne (MP) method, comparing core needle biopsy samples taken before chemotherapy and those collected post-treatment. The assessment of residual tumor cells following NAC was categorized into five grades: i) G1, no tumor cell loss; ii) G2, loss <30%; iii) G3, 30–90% reduction; iv) G4, loss of >90%; and v) G5, absence of malignant cells. According to the MP grading system (32), a pathological incomplete response (non-pCR) is characterized by G1-G4, where tumor cells remain in the tumor bed. A pCR is classified as G5, indicating no invasive cancer cells are present at the tumor bed, although ductal carcinoma in situ may still be identified. Tumor cells were classified as ER⁺/PR⁺ if staining was ≥1%; otherwise, the tumor cells were considered negative. Immunohistochemistry results were interpreted as HER2 negative for scores of (−) and (+) (22,24), while (+++) indicated HER2 positivity. For scores of (++), additional fluorescence in situ hybridization (FISH) test was performed, with a FISH result of (+) considered positive and (−) negative. Tissue samples were retrieved from archived formalin-fixed paraffin-embedded (FFPE) blocks. Specimens were fixed in 10% neutral-buffered formalin at room temperature for 24 h. Sections (4 µm) were mounted on glass slides and subjected to routine hematoxylin-eosin (H&E) staining at room temperature (hematoxylin for 5 min, eosin for 1 min). FISH was performed on 4 µm-FFPE sections using the HER2/CEP17 dual-color FISH probe kit (Cat. No. F.01359-01; Guangzhou LBP Medicine Science & Technology Co., Ltd.), according to the manufacturer's instructions. Enzymatic permeabilization was performed with kit-supplied pepsin at 37°C for 6–10 min, with the incubation duration optimized according to tissue conditions, followed by post-fix using 10% neutral buffered formalin at room temperature for 10 min. Briefly, 10 µl of premixed hybridization buffer containing HER2/CEP17 probes was applied to tissue sections. Co-denaturation was conducted at 85°C for 5 min, followed by hybridization at 37°C for 10–18 h in a humidified chamber. Post-hybridization stringent washing was carried out using kit-provided buffers. Nuclei were counterstained with DAPI at room temperature for 5 min. Fluorescence signals were visualized and captured using an Olympus BX51 fluorescence microscope. Hormone receptor (HR) positivity was defined as positivity for ER and/or PR. A Ki-67 index indicating ≥20% positive cells was classified as high expression, while <20% was classified as low expression (33). According to the 2023 St. Gallen consensus (34), breast cancer was categorized into four subtypes: i) Luminal A (ER⁺ and/or PR⁺, HER2⁻ and Ki-67 <14%); ii) luminal B (ER⁺ and/or PR⁺, HER2⁻ and Ki-67 ≥14% or ER⁺ and/or PR⁺, HER2⁺ and any Ki-67); iii) HER2 upregulation (ER⁻/PR⁻, HER2⁺ and any Ki-67); and iv) TNBC (ER⁻/PR⁻, HER2⁻ and any Ki-67).

Patients underwent a core needle biopsy at Hubei Cancer Hospital, which pathology confirmed as primary invasive breast cancer. An immunohistochemical analysis was carried out to establish the molecular subtype. IHC was performed on 4 µm FFPE sections. Tissue specimens were fixed in 10% neutral buffered formalin at room temperature for 24 h and embedded in paraffin. Antigen retrieval was performed in pre-boiled 0.01 M citrate buffer (pH 6.0) with a pressure cooker at 121°C; slides were heated for 3 min, cooled for 20 min and rinsed under running tap water, followed by rehydration in a descending ethanol series. Endogenous peroxidase activity was blocked with 3% H₂O₂ for 10 min at room temperature. Sections were incubated overnight at 4°C with primary antibodies against ER (cat. No. 790-4324; 1:100), PR (cat. No. 790-2223; 1:100), HER2 (all Ventana, Cat. No. 790-4493; 1:200), and Ki-67 (Dako, Cat. No. M7240; 1:100). HRP-labelled secondary antibody from the EliVision™plus kit (KIT-9902; Fuzhou Maxim Biotech, China) was applied for 1 h at room temperature. Immunoreactivity was visualized with DAB, and sections were counterstained with Mayer's hematoxylin at room temperature for 1–3 min prior to examination under a light microscope. In cases where there was notable enlargement of the axillary lymph nodes, a lymph node aspiration was performed concurrently to assess if it was due to breast cancer metastasis. All patients included in the present study received 6–8 cycles of standard NAC, with the specific regimen tailored by the medical team based on the immunohistochemical molecular subtype results, following the Chinese Society of Clinical Oncology guidelines for breast cancer management (35). Comprehensive information on regimens, drugs, cycles and distribution of molecular subtype are provided in Table SI. The most frequently utilized regimens were doxorubicin + cyclophosphamide followed by docetaxel (AC-T; 8 cycles) and epirubicin + cyclophosphamide followed by docetaxel (EC-T; 8 cycles), accounting for 240 patients or 74.3%. For patients with HER2 positivity, regimens containing trastuzumab and/or pertuzumab (such as EC-TH, AC-THP, TCHP, TCbHP and TCH/TCbH) were administered in accordance with the Chinese Society of Clinical Oncology Breast Cancer guidelines [2021 edition (36) and 2022 edition (35)] available at the time of treatment.

The present study collected clinical, US and pathological information, which included age. While the frequency of monitoring was determined by the judgment of the physician, the present study documented the maximum tumor diameter and serum CA153 levels during both the initial (baseline) and final (preoperative) assessments for the present analysis. The present study also noted US characteristics before NAC, the surgical approach adopted post-NAC and the initial clinical T and N stages based on the 2017 American Joint Committee on Cancer 8th edition breast cancer TNM staging criteria (37). Furthermore, the present study evaluated biopsy pathological data before NAC, including ER and PR status, HER2 expression, Ki-67 index and the presence of axillary lymph node metastasis. Post-surgery, the present study recorded pathological data such as occurrence of vascular or nerve invasion and whether a pCR was achieved (Fig. 1).

For model development and internal validation, the dataset was divided randomly into a test set (70%) and a validation set (30%) using a straightforward random allocation technique [utilizing the SPSS random number generator (version 26.0, IBM Corp.)] (38). The percentage reduction in maximum tumor diameter was determined using the formula: (Initial diameter-preoperative diameter)/initial diameter ×100%. Receiver operating characteristic (ROC) curve analysis was utilized to ascertain the ideal cut-off point in predicting pCR, based on the Youden index (sensitivity + specificity-1). Measurement data are presented as mean ± standard deviation, while categorical data are indicated as frequencies (%). Normality assessments were conducted on the measurement data from both the test and validation set. The unpaired Student's t-test was used for comparisons of continuous variables between groups. Categorical data were compared between groups using either the χ² or Fisher's exact test. P<0.05 was considered to indicate a statistically significant difference.

In the test set, the data was categorized according to the achievement of pCR. A univariate logistic regression analysis was conducted to investigate the factors associated with pCR, and variables demonstrating P<0.05 were included into a multivariate logistic regression analysis. The present analysis used a forward stepwise selection approach, using the likelihood ratio method, using an entry threshold of α=0.05 and a removal threshold of α=0.10; IBM SPSS Statistics (version 26.0; IBM Corp.). The main variables, including US characteristics, CA153 levels, molecular subtype and pCR results, had no missing data. However, elastography and CEUS data, which were only available for certain subsets, were excluded from the primary model and thus, no imputation was performed.

The logistic regression model [using the ‘Enter’ method; IBM SPSS Statistics (version 26.0; IBM Corp.)] was used to examine two predetermined interactions by incorporating product terms: i) Molecular subtype in conjunction with US margin; and ii) the variation in CA153 alongside the variation in tumor diameter. An interaction of P<0.05 was considered to indicate a statistically significant difference.

An odds ratio (OR) >1 suggests an increased probability of achieving pCR. The assessment of the model was performed in both the test and validation sets. For the validation process, the calibration and discrimination were evaluated using the Hosmer-Lemeshow goodness-of-fit test and the area under the ROC curve (AUC) (39). To examine possible confounding effects associated with age and tumor size, a sensitivity analysis was conducted using forced entry (‘Enter’ method) logistic regression, incorporating these two factors alongside the initial predictors. In all evaluations, a two-sided P<0.05 was considered to indicate a statistically significant difference. Variance inflation factor (VIF) was calculated to assess multicollinearity between selected variables. All statistical analyses were carried out using the commercially available SPSS software (version 26.0; IBM Corp.).

A total of 323 patients were included in the present study (mean age ± standard deviation, 49.9±10.9 years; range, 21–74 years). Among the present study cohort, 226 patients were assigned to the test set and 97 patients to the validation set. Comprehensive clinical, US and pathological information were collected from all participants (Table I). The mean serum CA153 levels was recorded at 21.7±25.5 U/ml, while the maximum tumor diameter was 3.9±2.1 cm. US analysis revealed that a notable portion of the tumors exhibited irregular shapes (89.8%; 290/323) and the posterior echoes indicated no notable changes (59.1%; 191/323), with 29.1% (94/323) displaying shadowing. The majority of patients in the present study had axillary lymph node metastasis (81.1%; 262/323). The predominant molecular subtype identified was luminal B (49.2%; 159/323), with clinical stages T2 (71.5%; 231/323) and N1 (55.4%; 179/323) being the most prevalent. Of the 323 patients, 292 (90.4%) underwent total mastectomy as part of their radical treatment, while 31 (9.6%) opted for breast-conserving surgery after NAC, with 29.7% (96/323) achieving a pCR. The characteristics of test and validation sets are detailed in Table I, demonstrating no statistically significant differences in the distribution of all the variables between the two groups (all P>0.05).

In the test set (n=226), 30.1% (68/226) of patients achieved a pCR. Table II illustrates the comparisons between patients in the pCR and non-pCR groups. Tumor size reduction and the age of patients were categorized based on the cut-off values derived from the Youden index. The determined optimal cut-off value of tumor size reduction was 48.5% (with sensitivity at 0.892 and specificity at 0.882, yielding a Youden index of 0.774), which was rounded to 50% for practical application (data not shown). Among the pCR group (n=68), 22.1% (n=15) of them were aged <46 years of age and 91.2% (n=62) had tumors ≥2.1 cm. By contrast, in the non-pCR group, 33.5% (n=53/158) were aged <46 years, with a notable majority (88.6%; 140/158) having tumor diameters >2.1 cm. However, these differences were not statistically significant (all P>0.05). A reduction in maximum diameter by ≥50% was more frequently observed in the pCR group (73.5%; 50/68), while the non-pCR group predominantly exhibited diameter changes <50% (72.8%; 115/158), with these differences being statistically significant (P<0.001). In the pCR group, the most prevalent US features in the pCR group was a circumscribed margin (82.3%; 56/68; Fig. 2), whereas only 32.9% (52/158) of the non-PCR group displayed this characteristic (Fig. 3), indicating a statistically significant difference (P<0.001). Both in the pCR and non-PCR group, the absence of posterior features was common; however, tumors with post-echo attenuation were significantly more frequent in the non-pCR group (34.2%; 54/158) compared with the pCR group (17.6%; 12/68) (P=0.024. No significant differences were identified in morphological characteristics, twisted perforating vessels or calcification between the two groups (all P>0.05). Clinical N1 status (P=0.014 and a reduction in serum CA153 levels post-NAC (P=0.037) were significantly more prevalent in the pCR group (both 73.5%; 50/68) than in the non-pCR group (46.8%, 74/158 and 41.2%, 65/158). The pCR group also demonstrated elevated Ki-67 levels (83.8%, 57/68) and higher proportion of negative hormone receptors (ER⁻ and PR⁻; 64.8%, 44/68) (P=0.035 and P=0.006, respectively; Fig. S1) compared with the non-pCR group (70.2%, 111/158 and 36.1%, 57/158). Luminal B (55.7%, 88/158) and HER2 (21.5%, 34/158) were the most common subtypes in the non-pCR group (Fig. S2), while TNBC (39.7%, 27/68) and luminal B (30.8%, 21/68) subtypes were the most common subtypes in the pCR group. The overall distribution differed significantly (P=0.021). No significant differences were noted in clinical T stage or HER2 expression between the two groups (all P>0.05).

In the present study, variables with P<0.05 from the univariate analysis were incorporated into the multivariate logistic regression analysis. The results of the multivariate analysis (Table III) indicated that, in the US evaluation, a tumor diameter reduction ≥50% (adjusted OR, 5.197; 95%CI, 1.635–16.518) and a circumscribed margin (adjusted OR, 9.157; 95%CI, 4.242–19.765) are significantly associated with pCR (Figs. 2 and 3). Additional independent predictors of pCR included molecular subtype, with the TNBC subtype indicating a significant independent association with pCR when compared with luminal A (adjusted OR, 4.262; 95%CI, 1.647–11.028). A reduction in CA153 levels (adjusted OR, 2.227; 95%CI, 1.061–4.675) also demonstrated a significant relationship with pCR (all P<0.05).

Sensitivity analyses that adjusted for age and tumor size categories did not markedly alter the ORs (Table SII) for the AUC of the present study model (Fig. S3). To further determine whether a circumscribed margin serves only as a substitute for TNBC, stratified and interaction analyses were performed (Table SIII). In the non-TNBC subgroup (Luminal⁺/HER2⁺; n=176), the OR for a circumscribed margin was 1.827 [OR=1.827; 95% confidence interval (CI), 1.015–4.389; P=0.067]; for the TNBC subgroup (luminal⁻/HER2⁻; n=50), the OR was 1.875 (95% CI, 0.482–7.372; P=0.133). Interaction tests demonstrated no significant modification effects from molecular subtype × margin (P=0.667) or CA153 × tumor diameter reduction (P=0.42), suggesting that the influence of margin on pCR did not differ significantly between patients with TNBC and without TNBC. To evaluate the predictive value of conventional B-mode US, we found that among patients with a ≥50% size reduction, 27% had residual disease on pathology, corresponding to a false-positive rate of 27%. To explain the inconsistent significance of cN stage between univariate and multivariate analyses, collinearity testing was performed for cN stage and the ratio of tumor diameter reduction. Moderate multicollinearity was observed (VIF=3.713 for cN stage; VIF=3.554 for ≥50% diameter reduction), implying overlapping predictive information for pCR.

The integrated multivariate prediction model utilized the adjusted ORs derived from the multivariable analysis (Table III). The present study model demonstrated notable predictive capability (AUC, 0.823; 95% CI, 0.765–0.881) and effective calibration [Hosmer-Lemeshow goodness-of-fit test: χ²=2.779; degrees of freedom (df)=7; P=0.905] within the test set. Furthermore, it demonstrated robust predictive performance for pCR (AUC, 0.884; 95% CI, 0.816–0.953) and calibration ability (Hosmer-Lemeshow goodness-of-fit test: χ²=4.298; df=7; P=0.745) in the validation set (Fig. 4).