This retrospective cohort study was conducted at Meizhou People’s Hospital, Meizhou, China. We reviewed medical records of consecutive patients with newly diagnosed HER2-positive breast cancer who received NAC followed by definitive surgery between January 2015 and December 2024. Inclusion criteria were: (1) histologically confirmed invasive breast carcinoma; (2) HER2-positive status defined as immunohistochemistry 3+ or FISH amplification ratio ≥2.0; (3) completion of preoperative DCE-MRI examination; (4) completion of planned NAC regimen; and (5) definitive surgical resection with pathological assessment. Exclusion criteria included bilateral breast cancer, distant metastasis at diagnosis, inflammatory breast cancer, incomplete imaging or clinical data, and prior breast cancer treatment.

Biomarker expression data were collected from pretreatment core biopsy specimens including HER2, ER, PR, and Ki-67 by immunohistochemistry. Differential expression analysis comparing pCR versus non-pCR groups was performed using log2 fold-change and P-value assessment. Gene set enrichment analysis (GSEA) was applied to identify biological pathways differentially activated between response groups using established pathway databases. GSEA was performed to identify biological pathways associated with pCR. Gene expression data from tumor samples were analyzed using a pre−ranked list based on the differential expression signal between pCR and non−pCR groups. Enrichment was assessed against a comprehensive collection of canonical pathways and gene ontology terms. A false discovery rate (FDR) < 0.25 was considered significant. Multi-omics feature integration combined expression biomarkers with imaging-derived features to comprehensively characterize the tumor microenvironment. Multi−omics integration was conducted on a subset of patients with available transcriptomic and proteomic data. Transcriptomic profiling was performed using RNA sequencing, and proteomic data were obtained by mass spectrometry from the same tumor specimens. An unsupervised integration algorithm (similarity network fusion) was applied to combine the two data types, followed by consensus clustering to identify multi−omics−defined subtypes. The association between these subtypes and pCR was then evaluated. Principal component analysis (PCA) was used to visualize the separation of pCR and non-pCR patient clusters in multidimensional feature space.

Random forest feature importance analysis was performed to rank candidate predictors of pCR. Pathological complete response (pCR) was defined as the absence of invasive residual disease in both the breast and axillary lymph nodes, allowing residual ductal carcinoma in situ (ypT0/is ypN0). LASSO (Least Absolute Shrinkage and Selection Operator) logistic regression with cross-validated regularization parameter selection identified the optimal feature subset. The correlation structure among final predictors was assessed to ensure model parsimony and avoid multicollinearity. Univariate and multivariate logistic regression analyses were performed to identify independent predictors of pCR. A nomogram was constructed incorporating the independently significant predictors identified through multivariate analysis.

All patients underwent standardized breast DCE-MRI examination on 1.5T or 3.0T MRI scanners prior to NAC initiation. The imaging protocol included T1-weighted, T2-weighted, and dynamic contrast-enhanced sequences with gadolinium-based contrast agent administration. Diffusion-weighted imaging (DWI) was performed with b-values of 0 and 800 s/mm². Quantitative parameters were extracted including maximum tumor diameter, tumor location, distance to skin, distance to chest wall, distance to nipple, axillary lymph node short-axis diameter, and minimum apparent diffusion coefficient (ADCmin) value. All measurements were performed by two experienced breast radiologists blinded to pathological outcomes.

All surgical specimens underwent standardized pathological examination. pCR was defined as absence of residual invasive carcinoma in the breast and sampled regional lymph nodes (ypT0/is ypN0). Residual disease was graded using the Miller-Payne system (grades 1-5), with grade 5 representing pCR.

Patients were randomly allocated to training (70%) and validation (30%) cohorts. Multivariate logistic regression with backward stepwise selection (P<0.05) determined independent predictors for nomogram construction. Model discrimination was assessed using area under the ROC curve (AUC). Calibration was evaluated by calibration plots and Hosmer-Lemeshow goodness-of-fit test. Clinical utility was assessed by decision curve analysis (DCA) (28). Statistical analyses were performed using R software version 4.2.0. Two-sided P<0.05 was considered statistically significant.

To comprehensively characterize the molecular landscape of the HER2-positive cohort and identify biologically relevant predictors of pCR, we conducted systematic biomarker expression analysis, pathway enrichment assessment, and multi-omics feature integration prior to clinical model development (Figure 1).

Differential biomarker expression analysis comparing pCR (n=74) and non-pCR (n=92) patients revealed significantly elevated HER2 expression (P<0.01) and Ki-67 proliferative index (P<0.01) in responders, consistent with the established association between HER2 pathway dependency and chemosensitivity. Conversely, estrogen receptor (ER) expression was significantly higher in non-pCR patients (P<0.001), reflecting the well-characterized inverse relationship between HR positivity and pCR achievement. Volcano plot analysis identified 23 upregulated and 18 downregulated biomarkers between response groups (fold-change threshold >1.5, FDR<0.05). Survival analysis demonstrated significantly improved disease-free survival in the HR-/HER2+ subtype (log-rank P = 0.031), supporting the prognostic relevance of molecular subtype classification.

To identify the optimal combination of predictors for nomogram development, we applied PCA-based dimensionality reduction, random forest feature importance analysis, and LASSO regularization to the integrated multi-omics feature set (Figure 2).

PCA of the combined molecular and imaging feature space demonstrated clear separation between pCR and non-pCR patient clusters along the first two principal components (PC1: 38.2%, PC2: 22.6% variance explained), confirming that the integrated feature set captures biologically meaningful response-associated variation. Random forest feature importance analysis ranked ADCmin value, Ki-67 index, HER2 expression level, and tumor size as the four highest-importance predictors (importance scores >0.13). LASSO regularization coefficient path analysis identified the optimal regularization parameter (lambda = 0.025) that retained five stable predictors: ADCmin value, Ki-67 index, tumor size, clinical N stage, and HR status. Correlation matrix analysis confirmed low inter-predictor correlations (all |r| < 0.60), supporting the independent contributions of selected features. Comparative ROC analysis demonstrated that the combined multi-omics nomogram (AUC = 0.823) outperformed models based on clinical features alone (AUC = 0.721) or DCE-MRI features alone (AUC = 0.748).

A total of 183 HER2-positive breast cancer patients met inclusion criteria. After excluding 17 patients with incomplete data, 166 patients were included in the final analysis. The median age was 52 years (range: 28–75 years), and median BMI was 23.4 kg/m² (range: 17.9-31.2 kg/m²). Approximately 55% of patients were postmenopausal. The median tumor size on MRI was 4.2 cm (range: 1.2-10.7 cm). Clinical stage distribution showed 5 (3.0%) stage I, 78 (47.0%) stage II, and 83 (50.0%) stage III patients. HR-positive (HR+/HER2+) and HR-negative (HR-/HER2+) subtypes accounted for 49.4% and 50.6% of cases, respectively (Figure 3).

Pretreatment DCE-MRI analysis revealed diverse tumor presentations. The median maximum tumor diameter was 4.2 cm. Among evaluable axillary lymph nodes, the median short−axis diameter was 1.2 cm. Supraclavicular and internal mammary lymph node involvement was observed in 13.0% and 23.7% of cases, respectively. The median ADCmin value was 0.84 × 10^-3 mm²/s, with pCR patients demonstrating significantly lower ADCmin values compared to non-pCR patients (0.72 × 10^-3 mm²/s vs 0.93 × 10^-3 mm²/s, P<0.001). Tumor location analysis showed predominant involvement of the upper outer quadrant (Figure 4).

All patients completed planned NAC regimens. The most common regimen was TCbHP (69.3%), followed by other taxane-based combinations. The median number of treatment cycles was 6. Overall breast tumor pCR rate was 44.6% (74/166), and regional lymph node pCR rate was 58.6% (95/162). Miller-Payne pathological grading showed grade 5 (pCR) in 71 patients (42.8%), grade 4 in 21 (12.7%), grade 3 in 18 (10.8%), and grade 2 in 20 (12.0%). pCR rates differed significantly between molecular subtypes: HR-/HER2+ patients achieved higher pCR rates (52.2%) compared to HR+/HER2+ patients (38.5%, P = 0.048), consistent with molecular profiling findings (Figure 5).

Univariate analysis in the training cohort (n=111) identified 12 variables with P<0.1 as candidate predictors, consistent with the features ranked by multi-omics analysis. Multivariate logistic regression with backward selection retained five independent predictors: tumor size (OR = 0.72, 95%CI: 0.58-0.89, P = 0.003), clinical N stage (OR = 0.65, 95%CI: 0.48-0.88, P = 0.006), HR status (OR = 0.52, 95%CI: 0.31-0.87, P = 0.012), Ki-67 index (OR = 1.03, 95%CI: 1.01-1.05, P = 0.008), and ADCmin value (OR = 2.15, 95%CI: 1.23-3.76, P = 0.007). These five variables, identified through both molecular feature importance analysis and multivariate clinical regression, were incorporated into a nomogram for individualized pCR probability prediction (Figure 6).

The nomogram demonstrated excellent discriminative ability in both cohorts. In the training set, AUC was 0.823 (95%CI: 0.754-0.892), with sensitivity of 78.4% and specificity of 72.6% at the optimal cutoff. External validation in the test cohort (n=55) confirmed robust performance with AUC of 0.795 (95%CI: 0.691-0.899). Calibration plots showed good agreement between predicted and observed pCR probabilities in both cohorts, with Hosmer-Lemeshow test P = 0.412 indicating no significant deviation from perfect calibration (Figure 7).

Decision curve analysis demonstrated that the nomogram provided superior net clinical benefit compared to treat-all or treat-none strategies across a wide range of threshold probabilities (0.15-0.75). This indicates that clinical decisions guided by the nomogram would result in better patient outcomes compared to indiscriminate treatment approaches. The model showed particular utility in the clinically relevant threshold range of 0.30-0.60, where treatment decisions are most uncertain (Figure 8).