Prediction model incorporating dynamic changes in serum tumor markers for evaluating the efficacy of neoadjuvant chemotherapy in patients with breast cancer: a retrospective cohort study Authors: Xiaoqian Li, Junjie Liu, Zirui Wang, Xiaoduo Li, Kexuan Feng, Rui Zhang, Jianjun He, Huimin Zhang Published: 2025 ABSTRACT BackgroundCurrent prediction models for the efficacy of neoadjuvant chemotherapy (NAC) in patients with breast cancer (BC) include only static measurements of serum tumor markers, while the dynamic measurement data of these markers have not been fully utilized. This study aimed to develop and validate a prediction model for evaluating the efficacy of NAC in BC patients on the basis of dynamic changes in CEA, CA125, and CA15-3 levels.MethodsWe retrospectively screened 565 patients with BC who received NAC at our department from December 2016 to November 2021. A total of 446 patients were included and randomly divided into a training cohort (n = 312) and a test cohort (n = 134) at a 7:3 ratio. Least absolute shrinkage and selection operator (LASSO) regression with tenfold cross-validation was applied to select the most relevant features, and multivariate logistic regression was used to construct the predictive model on the basis of the selected features. The performance of the model was evaluated by the area under the receiver operating characteristic (ROC) curve, calibration curves, and decision curve analysis (DCA).ResultsA nomogram integrating age, human epidermal growth factor receptor 2 (HER2) status, pre-chemotherapy levels of CEA, principal component 1 (PC1) for CEA, PC2 for CA125, and PC2 for CA15-3 was developed for the dynamic CEA&CA125&CA15-3 model. Another nomogram integrating estrogen receptor (ER) status, HER2 status, and pre-chemotherapy CEA levels was developed for the pre-chemotherapy CEA&CA125&CA15-3 model. The area under the receiver operating characteristic curve (AUC) of the dynamic model was 0.739 (95% CI 0.680–0.797) in the training cohort and 0.712 (95% CI 0.613–0.811) in the test cohort. The AUC was 0.658 (95% CI 0.593–0.722) in the training cohort and 0.715 (95% CI 0.624–0.807) in the test cohort for the pre-chemotherapy model. Compared with the pre-chemotherapy model, the dynamic model demonstrated significantly improved predictive accuracy. Our dynamic model also exhibited good predictive performance in subgroup analyses.ConclusionsThis study developed and validated nomogram models using clinicopathological features and serum tumor marker characteristics to predict NAC efficacy in BC patients. Although the dynamic model demonstrated comparable discriminative ability (AUC) to the pre-chemotherapy model in the test cohort, it showed significantly improved performance in net reclassification. While promising, this model requires further validation in multi-center prospective studies before clinical application.Supplementary InformationThe online version contains supplementary material available at 10.1186/s40001-025-03468-5. Introduction Breast cancer (BC) represents the most prevalent malignancy among women, accounting for approximately 30% of all newly diagnosed cancer cases, with its global incidence rising steadily at an annual rate of approximately 0.5% [1]. Neoadjuvant chemotherapy (NAC) has become a cornerstone in the management of locally advanced BC, offering tumor downsizing to facilitate surgical resection and providing deeper insights into individual patient sensitivity to chemotherapy [2]. Despite its widespread application, individual responses to NAC vary significantly, with pathological complete response (pCR) rates ranging from 20 to 60% depending on the tumor subtype [3]. Clinical data indicate that approximately 5% of patients receiving neoadjuvant therapy experience progressive disease (PD) [4]. These patients not only face the risk of losing the opportunity for curative surgery but are also more likely to experience irreversible clinical outcomes such as tumor metastasis. For patients with potential benefits from pCR, the current evaluation model, which overly relies on postoperative pathological verification, results in redundant treatment intensity. These findings highlight the critical need for reliable biomarkers to facilitate early evaluation of therapeutic efficacy and guide personalized treatment strategies. Current assessment modalities, such as imaging and histopathological evaluation, have inherent limitations. Imaging-based approaches often lack sensitivity for early response detection, whereas pathological assessment remains confined to post-surgical analysis, delaying timely therapeutic adjustments [5, 6]. Serum tumor markers, such as carcinoembryonic antigen (CEA), cancer antigen 125 (CA125) and carbohydrate antigen 15–3 (CA15-3), have emerged as minimally invasive tools for monitoring disease progression in various cancers. CEA is a cell adhesion molecule whose elevation in the bloodstream is associated with tumor metastasis [7]. CA125, the product of the MUC16 gene, serves as a key regulator of multiple cell survival pathways in ovarian cancer and BC cells [8]. CA15-3 is a member of the mucin-1 glycoprotein family, which is overexpressed in cancer and has been identified as a useful serum tumor marker driven by altered glycosylation [9]. These biomarkers reflect tumor burden and biological activity, offering potential for real-time treatment monitoring [10, 11]. Previous retrospective studies have demonstrated that elevated preoperative CA15-3 and CEA levels are associated with tumor burden and show independent prognostic significance [12]. A recent cohort study involving 10,836 Chinese BC patients revealed that preoperative CA15-3 and CEA expression exhibited a negative correlation with survival rates and tumor progression [13]. Currently, the predictive models for the NAC efficacy of BC include only static measurements of serum tumor markers [14], while the dynamic measurement data of these markers have not been fully utilized. Therefore, we propose that incorporating dynamic measurements of serum tumor markers may increase the predictive accuracy of the models. In this study, we developed and validated a predictive model for evaluating the efficacy of NAC in BC patients on the basis of dynamic changes in serum tumor marker levels. Our retrospective cohort study aimed to characterize the temporal patterns of multiple serum tumor markers during NAC and establish a clinically actionable predictive model for early efficacy assessment. The construction of this novel model is expected to provide critical insights for guiding individualized precision treatment in clinical practice. Methods Patients In this study, we retrospectively screened 565 patients with BC who received NAC at the First Affiliated Hospital of Xi’an Jiaotong University from December 2016 to November 2021. The inclusion criteria were as follows: (1) patients with BC aged ≥ 18 years who received NAC. (2) The diagnosis of BC was confirmed via ultrasound-guided core needle biopsy. (3) Comprehensive clinical imaging evaluations confirmed the diagnosis of stage I–III BC. (4) Serum tumor marker levels were measured at the baseline before NAC and at least twice before the fourth cycle of NAC prior to surgery. The exclusion criteria were as follows: (1) Patients with bilateral or occult BC. (2) Patients combined with secondary primary cancer. (3) Patients who completed the NAC but were uncertain whether they had undergone surgery. (4) Incomplete laboratory or clinical data. Following the application of inclusion and exclusion criteria, a total of 446 patients were included in the final analysis. A comprehensive check for missing data was performed on all clinicopathological variables and serum tumor marker measurements for these 446 patients. No missing data were found in any variables for the included patients. Therefore, a complete-case analysis was inherently applied, and no data imputation was necessary. The study flowchart is presented in Fig. 1.Fig. 1Flowchart of this study Flowchart of this study Serum tumor marker determination Blood samples for serum tumor marker (CEA, CA125, and CA15-3) measurement were collected as part of routine clinical practice. The pre-chemotherapy value was defined as the measurement closest to the initiation of NAC, with a median interval of 6 days (interquartile range, IQR: 4–9 days). Subsequent measurements were typically obtained within 3–5 days before each subsequent chemotherapy cycle. The median interval between consecutive measurements during NAC was 21 days (IQR: 19–23 days). This reflects real-world clinical scheduling while maintaining close adherence to the standard 21-day cycle. All measurements were performed using the ROCHECCM8K_C instrument via electrochemiluminescence immunoassay (ECLIA). Data collection and classification We collected the following information from the electronic medical records of all patients, including clinicopathological data, treatment characteristics, and serum tumor marker levels. Clinicopathological data included age, menopausal status, hormonal status, human epidermal growth factor receptor 2 (HER2) status, Ki-67 status, clinical tumor stage (cT stage), clinical lymph node stage (cN stage), and pCR. The treatment characteristics included the regimen of NAC and the number of NAC cycles. Immunohistochemical analysis (IHC) was used to determine the estrogen receptor (ER) and the progesterone receptor (PR) status, and hormone receptor (HR) status was classified as positive when more than 1% of cells showed positive staining for either ER or PR. HER2 expression was considered positive if the IHC result was 3 + or if the HER2 was confirmed amplified by in situ hybridization when the IHC staining was 2 +. Tumor staging was determined according to the American Joint Committee on Cancer (AJCC) 8th edition criteria. The absence of invasive cancer in both regional lymph nodes and breast tissue can be identified as pCR, regardless of the presence of residual ductal carcinoma in situ (DCIS) within the breast tissue [3]. pCR was assessed by reviewing the final pathological reports of all patients. This assessment was conducted by two independent, experienced pathologists who examined all surgical specimens, with any discrepancies resolved through a consensus review. Serum tumor marker levels include dynamic measurements of CEA, CA125, and CA15-3. Extraction of dynamic trajectories Given the high intrinsic variability of serum tumor markers, we defined the maximum value as ten times the upper reference limits for CEA, CA125, and CA15-3. Logarithmic transformation of the measured values was applied to extract dynamic trajectories [15]. To capture the dynamic changes in serum tumor markers over time, we applied functional principal component analysis (FPCA) [16]. FPCA is an extension of traditional Principal Component Analysis for functional data, which projects high-dimensional time-series data onto a low-dimensional orthogonal basis (principal components, PCs) to retain the primary variation patterns. By decomposing the covariance operator of the functional data, FPCA extracts principal component functions and their corresponding eigenvalues, which represent the dominant modes of variation in the data [17]. The calculated individual-specific functional principal component (FPC) scores can reflect different trajectory patterns. The optimal number of FPCs is determined on the basis of the Akaike information criterion (AIC) [15]. Construction and evaluation of prediction models A total of 446 patients were included and randomly divided into a training cohort (n = 312) and a test cohort (n = 134) at a 7:3 ratio. The 7:3 ratio for splitting the data set into training and test cohorts is a well-established convention in predictive modeling that provides sufficient data for robust model development while retaining an adequate sample size for reliable performance evaluation in the test set. No significant differences were observed in demographic or clinical characteristics between the training and test cohorts. To mitigate the risk of overfitting and enhance model generalizability, we implemented two key strategies in the training cohort: (a) employing Least Absolute Shrinkage and Selection Operator (LASSO) regression with tenfold cross-validation for robust feature selection, which penalizes model complexity by shrinking the coefficients of less relevant variables; (b) strictly evaluating the final model’s performance on the independent test cohort that was not involved in any step of the model building process. For the pre-chemotherapy CEA&CA125&CA15-3 model, pre-chemotherapy levels of CEA, CA125, and CA15-3, along with other clinical variables, were included as predictors. LASSO regression with tenfold cross-validation was applied to select the most relevant features and to determine the optimal value of the penalty hyperparameter (λ), which was chosen based on the minimum cross-validation error criterion. Multivariate logistic regression was used to construct the predictive model on the basis of the selected features. A nomogram was developed to visualize the predictive capacity of pre-chemotherapy biomarkers for treatment efficacy. To further improve the model, we constructed a dynamic CEA&CA125&CA15-3 model incorporating dynamic serum tumor marker changes by integrating dynamic variations in CEA, CA125, and CA15-3 during NAC. For early prediction of treatment response, dynamic trajectories of serum tumor markers were derived from measurements taken at baseline (pre-chemotherapy) and throughout the first three NAC cycles. To capture these dynamic patterns, FPCA was applied separately to the longitudinal measurements of each serum marker. The first two PCs for each marker, which represent the major modes of variation in their dynamic trajectories, were extracted as features. The resulting FPC scores, along with clinical variables, were combined as candidate predictors. Similarly, LASSO regression was employed to select the most predictive features, and a new logistic regression model was developed on the basis of these selected features. A corresponding nomogram was created to illustrate the predictive ability of dynamic biomarkers to predict treatment efficacy. The performance of both models was tested in the test cohort. The discriminative ability was evaluated via receiver operating characteristic (ROC) curves, and the area under the curve (AUC) was calculated [18]. Calibration curves were plotted to assess the agreement between the predicted and observed outcomes, and decision curve analysis (DCA) was performed to evaluate the clinical utility of both models [19]. To quantify the incremental improvement of the dynamic CEA&CA125&CA15-3 model over the pre-chemotherapy CEA&CA125&CA15-3 model, net reclassification improvement (NRI) and integrated discrimination improvement (IDI) were calculated [20, 21], providing insights into the enhanced predictive performance of the new model. Subgroup analysis Owing to the significant biological and treatment response distinctions across subtypes, we stratified the subgroups based on HER2 status and chemotherapy regimens. We performed stratified analyses within clinically meaningful subgroups. For each subgroup, individual models were constructed using an identical methodology with cohort-stratified test, guaranteeing the robustness and comparability of the outcomes. Statistical analysis Continuous variables were described as median (interquartile ranges, IQRs), and categorical variables were summarized as numbers (percentages). To compare characteristics between the pCR and non-pCR groups, as well as between the training and test cohorts, statistical tests were selected on the basis of data distribution: the Mann–Whitney U test was used for non-normally distributed data, whereas the Student’s t test was applied for normally distributed data. Categorical variables were compared via the chi-square (χ2) test, with Fisher’s exact test substituted for small expected frequencies (< 5). All tests were two-sided, and a p value < 0.05 was considered statistically significant. As an exploratory predictive modeling study, a formal sample size calculation was not performed a priori. However, our sample size of 446 patients with 145 pCR events exceeds the commonly recommended minimum of 10 events per variable for logistic regression modeling, providing reasonable statistical power for model development. To control for Type I errors (false positives) in variable selection, we employed LASSO regression with tenfold cross-validation, which inherently penalizes model complexity and reduces overfitting. While these measures mitigate statistical error risks, the possibility of Type II errors (false negatives) remains, particularly for variables with modest effects. FPCA was implemented with the package “fdapace”. Logistic regression analysis and nomogram construction were performed via the “rms” package. All analyses were conducted in the R programming environment (version 4.4.2; http://www.r-project.org/). Results Patient characteristics As shown in Table 1, a total of 446 eligible patients were included in this study, among who 145 achieved pCR. Patients achieving pCR had a significantly higher proportion of ER-negative, PR-negative, and HER2-positive status compared to non-pCR patients. No significant differences were observed in age, menopausal status, Ki-67 status, clinical stage (cT stage and cN stage), or pre-chemotherapy serum tumor marker levels between the two groups. The entire cohort was randomly allocated into a training cohort (n = 312) and a test cohort (n = 134), as presented in Table 2. Statistical analysis revealed no significant differences in clinical characteristics between the training and test cohorts (P > 0.05). Table 1Comparison of patient characteristics between the pCR and non-pCR groupsVariablesAll patientsN = 446pCR groupN = 145non-pCR groupN = 301p valueAge (years)48.00 (40.00–55.00)48.00 (41.00–57.00)48.00 (40.00–55.00)0.234Menopause, n (%) No269 (60.31%)83 (57.24%)186 (61.79%)0.357 Yes177 (39.69%)62 (42.76%)115 (38.21%)ER status, n (%) Positive254 (56.95%)66 (45.52%)188 (62.46%) < 0.001 Negative192 (43.05%)79 (54.48%)113 (37.54%)PR status, n (%) Positive200 (44.84%)54 (37.24%)146 (48.50%)0.025 Negative246 (55.16%)91 (62.76%)155 (51.50%)HER2 IHC, n (%) 0101 (22.65%)22 (15.17%)79 (26.25%) < 0.001 1 + 91 (20.40%)21 (14.48%)70 (23.26%) 2 + 78 (17.49%)19 (13.10%)59 (19.60%) 3 + 176 (39.46%)83 (57.24%)93 (30.90%)Ki-67, n (%) < 15%13 (2.91%)3 (2.07%)10 (3.32%)0.461 ≥ 15%433 (97.09%)142 (97.93%)291 (96.68%)cT stage, n (%) T198 (21.97%)37 (25.52%)61 (20.27%)0.566 T2278 (62.33%)87 (60.00%)191 (63.46%) T346 (10.31%)15 (10.34%)31 (10.30%) T424 (5.38%)6 (4.14%)18 (5.98%)cN stage, n (%) N0112 (25.11%)43 (29.66%)69 (22.92%)0.213 N1295 (66.14%)94 (64.83%)201 (66.78%) N222 (4.93%)5 (3.45%)17 (5.65%) N317 (3.81%)3 (2.07%)14 (4.65%) Pre-chemotherapy CEA1.89 (1.21–3.10)1.80 (1.20–2.72)2.00 (1.24–3.26)0.134 Pre-chemotherapy CA12514.25 (10.30–21.17)13.70 (9.60–21.20)15.00 (10.40–21.10)0.361 Pre-chemotherapy CA15-312.16 (8.65–17.80)11.30 (8.35–16.60)12.50 (8.80–17.96)0.145pCR pathological complete response, ER estrogen receptor, PR progesterone receptor, HER2 human epidermal growth factor receptor 2, IHC immunohistochemical analysis, cT stage clinical tumor stage, cN stage clinical lymph node stageTable 2Comparison of patient characteristics between the training and test cohortsVariablesAll patientsN = 446Training cohortN = 312Test cohortN = 134p valueAge (years)48.00 (40.00–55.00)48.00 (40.00–55.00)49.00 (41.00–56.00)0.439Menopause, n (%) No269 (60.31%)193 (61.86%)76 (56.72%)0.309 Yes177 (39.69%)119 (38.14%)58 (43.28%)ER status, n (%) Positive254 (56.95%)177 (56.73%)77 (57.46%)0.886 Negative192 (43.05%)135 (43.27%)57 (42.54%)PR status, n (%) Positive200 (44.84%)139 (44.55%)61 (45.52%)0.850 Negative246 (55.16%)173 (55.45%)73 (54.48%)HER2 IHC, n (%) 0101 (22.65%)69 (22.12%)32 (23.88%)0.882 1 + 91 (20.40%)66 (21.15%)25 (18.66%) 2 + 78 (17.49%)56 (17.95%)22 (16.42%) 3 + 176 (39.46%)121 (38.78%)55 (41.04%)Ki-67, n (%) < 15%13 (2.91%)8 (2.56%)5 (3.73%)0.502 ≥ 15%433 (97.09%)304 (97.44%)129 (96.27%)cT stage, n (%) T198 (21.97%)68 (21.79%)30 (22.39%)0.992 T2278 (62.33%)194 (62.18%)84 (62.69%) T346 (10.31%)33 (10.58%)13 (9.70%) T424 (5.38%)17 (5.45%)7 (5.22%)cN stage, n (%) N0112 (25.11%)81 (25.96%)31 (23.13%)0.760 N1295 (66.14%)202 (64.74%)93 (69.40%) N222 (4.93%)17 (5.45%)5 (3.73%) N317 (3.81%)12 (3.85%)5 (3.73%)pCR, n (%) No301 (67.49%)209 (66.99%)92 (68.66%)0.730 Yes145 (32.51%)103 (33.01%)42 (31.34%) Pre-chemotherapy CEA1.89 (1.21–3.10)1.85 (1.22–2.89)2.22 (1.20–3.20)0.489 Pre-chemotherapy CA12514.25 (10.30–21.17)15.00 (10.40–22.60)13.75 (10.10–19.55)0.175 Pre-chemotherapy CA15-312.16 (8.65–17.80)12.00 (8.53–17.52)12.60 (8.78–19.32)0.387ER estrogen receptor, PR progesterone receptor, HER2 human epidermal growth factor receptor 2, IHC immunohistochemical analysis, cT stage clinical tumor stage, cN stage clinical lymph node stage, pCR pathological complete response Table 1Comparison of patient characteristics between the pCR and non-pCR groupsVariablesAll patientsN = 446pCR groupN = 145non-pCR groupN = 301p valueAge (years)48.00 (40.00–55.00)48.00 (41.00–57.00)48.00 (40.00–55.00)0.234Menopause, n (%) No269 (60.31%)83 (57.24%)186 (61.79%)0.357 Yes177 (39.69%)62 (42.76%)115 (38.21%)ER status, n (%) Positive254 (56.95%)66 (45.52%)188 (62.46%) < 0.001 Negative192 (43.05%)79 (54.48%)113 (37.54%)PR status, n (%) Positive200 (44.84%)54 (37.24%)146 (48.50%)0.025 Negative246 (55.16%)91 (62.76%)155 (51.50%)HER2 IHC, n (%) 0101 (22.65%)22 (15.17%)79 (26.25%) < 0.001 1 + 91 (20.40%)21 (14.48%)70 (23.26%) 2 + 78 (17.49%)19 (13.10%)59 (19.60%) 3 + 176 (39.46%)83 (57.24%)93 (30.90%)Ki-67, n (%) < 15%13 (2.91%)3 (2.07%)10 (3.32%)0.461 ≥ 15%433 (97.09%)142 (97.93%)291 (96.68%)cT stage, n (%) T198 (21.97%)37 (25.52%)61 (20.27%)0.566 T2278 (62.33%)87 (60.00%)191 (63.46%) T346 (10.31%)15 (10.34%)31 (10.30%) T424 (5.38%)6 (4.14%)18 (5.98%)cN stage, n (%) N0112 (25.11%)43 (29.66%)69 (22.92%)0.213 N1295 (66.14%)94 (64.83%)201 (66.78%) N222 (4.93%)5 (3.45%)17 (5.65%) N317 (3.81%)3 (2.07%)14 (4.65%) Pre-chemotherapy CEA1.89 (1.21–3.10)1.80 (1.20–2.72)2.00 (1.24–3.26)0.134 Pre-chemotherapy CA12514.25 (10.30–21.17)13.70 (9.60–21.20)15.00 (10.40–21.10)0.361 Pre-chemotherapy CA15-312.16 (8.65–17.80)11.30 (8.35–16.60)12.50 (8.80–17.96)0.145pCR pathological complete response, ER estrogen receptor, PR progesterone receptor, HER2 human epidermal growth factor receptor 2, IHC immunohistochemical analysis, cT stage clinical tumor stage, cN stage clinical lymph node stage Comparison of patient characteristics between the pCR and non-pCR groups pCR pathological complete response, ER estrogen receptor, PR progesterone receptor, HER2 human epidermal growth factor receptor 2, IHC immunohistochemical analysis, cT stage clinical tumor stage, cN stage clinical lymph node stage Table 2Comparison of patient characteristics between the training and test cohortsVariablesAll patientsN = 446Training cohortN = 312Test cohortN = 134p valueAge (years)48.00 (40.00–55.00)48.00 (40.00–55.00)49.00 (41.00–56.00)0.439Menopause, n (%) No269 (60.31%)193 (61.86%)76 (56.72%)0.309 Yes177 (39.69%)119 (38.14%)58 (43.28%)ER status, n (%) Positive254 (56.95%)177 (56.73%)77 (57.46%)0.886 Negative192 (43.05%)135 (43.27%)57 (42.54%)PR status, n (%) Positive200 (44.84%)139 (44.55%)61 (45.52%)0.850 Negative246 (55.16%)173 (55.45%)73 (54.48%)HER2 IHC, n (%) 0101 (22.65%)69 (22.12%)32 (23.88%)0.882 1 + 91 (20.40%)66 (21.15%)25 (18.66%) 2 + 78 (17.49%)56 (17.95%)22 (16.42%) 3 + 176 (39.46%)121 (38.78%)55 (41.04%)Ki-67, n (%) < 15%13 (2.91%)8 (2.56%)5 (3.73%)0.502 ≥ 15%433 (97.09%)304 (97.44%)129 (96.27%)cT stage, n (%) T198 (21.97%)68 (21.79%)30 (22.39%)0.992 T2278 (62.33%)194 (62.18%)84 (62.69%) T346 (10.31%)33 (10.58%)13 (9.70%) T424 (5.38%)17 (5.45%)7 (5.22%)cN stage, n (%) N0112 (25.11%)81 (25.96%)31 (23.13%)0.760 N1295 (66.14%)202 (64.74%)93 (69.40%) N222 (4.93%)17 (5.45%)5 (3.73%) N317 (3.81%)12 (3.85%)5 (3.73%)pCR, n (%) No301 (67.49%)209 (66.99%)92 (68.66%)0.730 Yes145 (32.51%)103 (33.01%)42 (31.34%) Pre-chemotherapy CEA1.89 (1.21–3.10)1.85 (1.22–2.89)2.22 (1.20–3.20)0.489 Pre-chemotherapy CA12514.25 (10.30–21.17)15.00 (10.40–22.60)13.75 (10.10–19.55)0.175 Pre-chemotherapy CA15-312.16 (8.65–17.80)12.00 (8.53–17.52)12.60 (8.78–19.32)0.387ER estrogen receptor, PR progesterone receptor, HER2 human epidermal growth factor receptor 2, IHC immunohistochemical analysis, cT stage clinical tumor stage, cN stage clinical lymph node stage, pCR pathological complete response Comparison of patient characteristics between the training and test cohorts ER estrogen receptor, PR progesterone receptor, HER2 human epidermal growth factor receptor 2, IHC immunohistochemical analysis, cT stage clinical tumor stage, cN stage clinical lymph node stage, pCR pathological complete response Extraction of dynamic features The dynamic trajectories of CEA, CA125, and CA15-3 levels in pCR and non-pCR patients are shown in Figure S1. In non-pCR patients, the serum levels of CEA, CA125, and CA15-3 during NAC tended to be higher and exhibited an upward trend. The first two PCs for the dynamic trajectories of CEA, CA125, and CA15-3 were selected on the basis of the AIC. The corresponding eigenfunctions are shown in Figure S2. Predictor identification and prediction model development To mitigate multicollinearity in the training cohort regression analysis, we implemented LASSO regression with tenfold cross-validation for preliminary predictor screening. For the dynamic CEA&CA125&CA15-3 model, at the optimal lambda value with minimal bias, we identified eleven predictors with non-zero coefficients associated with pCR (Fig. 2a, b). These predictors included age, ER status, HER2 IHC, Ki-67 status, cT stage, cN stage, pre-chemotherapy levels of CEA, PC1 for CEA, PC1 for CA125, PC2 for CA125, and PC2 for CA15-3. Subsequent multivariate logistic regression analysis revealed that age (OR = 1.03, 95% CI 1.00–1.06, P = 0.022), HER2 IHC 3 + (OR = 4.62, 95% CI 2.21–10.24, P < 0.001), pre-chemotherapy levels of CEA (OR = 1.20, 95% CI 1.10–1.34, P < 0.001), PC1 for CEA (OR = 0.05, 95% CI 0.01–0.19, P < 0.001), PC2 for CA125 (OR = 0.06, 95% CI 0.01–0.49, P = 0.009), and PC2 for CA15-3 (OR = 57.47, 95% CI 2.23–1685.52, P = 0.016) emerged as independent predictors of pCR (It should be noted that the wide confidence interval for PC2 for CA15-3 indicates considerable uncertainty in this estimate, likely due to the relatively small number of patients with extreme trajectory patterns. This finding requires validation in larger cohorts.). The forest plot of the multivariate logistic regression analysis is presented in Fig. 2c. A nomogram integrating these six independent predictors was developed, visually quantifying their relative contributions to pCR through scaled point allocations (Fig. 3). For the pre-chemotherapy CEA&CA125&CA15-3 model, predictors associated with pCR were identified and a nomogram was constructed via the same methodology, as demonstrated in Figure S3.Fig. 2Predictor identification for the dynamic CEA&CA125&CA15-3 model via LASSO and multivariate logistic regression. a Coefficient profile plot of 17 factors associated with pCR in the training cohort. The bottom horizontal axis represents the log lambda (λ) values of the independent variable, the top horizontal axis indicates the number of variables with non-zero coefficients, and the vertical axis represents the coefficients of the independent variable. Different-colored curves represent different variables, with each curve depicting the trajectory of the coefficient variation for each independent variable. b Eleven non-zero coefficient variables selected through tenfold cross-validation via optimal lambda. Dotted vertical lines were placed at the optimal log(λ) values determined by both the minimum criteria and one standard error. λ = 0.0085, corresponding to a logarithm of λ equal to − 4.7670, was selected as the optimal choice. c Forest plot of the multivariate logistic regression analysis. HER2 human epidermal growth factor receptor 2, IHC immunohistochemical analysis, PC principal componentFig. 3Nomogram for the dynamic CEA&CA125&CA15-3 model for predicting the probability of pCR. HER2 human epidermal growth factor receptor 2, IHC immunohistochemical analysis, PC principal components, pCR pathological complete response Predictor identification for the dynamic CEA&CA125&CA15-3 model via LASSO and multivariate logistic regression. a Coefficient profile plot of 17 factors associated with pCR in the training cohort. The bottom horizontal axis represents the log lambda (λ) values of the independent variable, the top horizontal axis indicates the number of variables with non-zero coefficients, and the vertical axis represents the coefficients of the independent variable. Different-colored curves represent different variables, with each curve depicting the trajectory of the coefficient variation for each independent variable. b Eleven non-zero coefficient variables selected through tenfold cross-validation via optimal lambda. Dotted vertical lines were placed at the optimal log(λ) values determined by both the minimum criteria and one standard error. λ = 0.0085, corresponding to a logarithm of λ equal to − 4.7670, was selected as the optimal choice. c Forest plot of the multivariate logistic regression analysis. HER2 human epidermal growth factor receptor 2, IHC immunohistochemical analysis, PC principal component Nomogram for the dynamic CEA&CA125&CA15-3 model for predicting the probability of pCR. HER2 human epidermal growth factor receptor 2, IHC immunohistochemical analysis, PC principal components, pCR pathological complete response Prediction model test The discriminative capacities of nomograms were quantified using AUC analysis (Fig. 4). For the dynamic CEA&CA125&CA15-3 model, the nomogram demonstrated robust predictive performance with an AUC of 0.739 (95% CI 0.680–0.797) in the training cohort and maintained generalizability in the test cohort (AUC = 0.712, 95% CI 0.613–0.811). For the pre-chemotherapy CEA&CA125&CA15-3 model, the AUC was 0.658 (95% CI 0.593–0.722) in the training cohort and 0.715 (95% CI 0.624–0.807) in the test cohort. Compared with the pre-chemotherapy CEA&CA125&CA15-3 model, the prediction performance of the dynamic CEA&CA125&CA15-3 model incorporating the dynamic features of CEA, CA125, and CA15-3 was enhanced. The calibration plot revealed good prediction accuracy between the actual probability and the predicted probability for both prediction models (Fig. 5). DCA validated the clinical utility of both prediction models, demonstrating superior net benefit for pCR prediction following NAC across clinically actionable threshold probabilities (Fig. 6). The dynamic CEA&CA125&CA15-3 model achieved significant net benefit in threshold ranges of 10–53% (training cohort) and 15–65% (test cohort). Similarly, the pre-chemotherapy CEA&CA125&CA15-3 model showed clinical utility across thresholds of 17–48% (training cohort) and 10–50% (test cohort). Both models consistently outperformed the default ‘treat-all’ and ‘treat-none’ strategies, with the dynamic model exhibiting broader clinically useful ranges particularly in the test cohort. Compared with the pre-chemotherapy model, the dynamic CEA&CA125&CA15-3 model demonstrated significantly improved predictive accuracy, as evidenced by both the NRI (0.203, 95% CI 0.068–0.338, P = 0.003) and IDI (0.053, 95% CI 0.026–0.080, P < 0.001) metrics (Table 3). These results indicate that the dynamic model correctly reclassified 20.3% more cases and improved the average predicted probability difference between pCR and non-pCR cases by 5.3 percentage points.Fig. 4ROC curves for quantifying the discriminative capacity of the models. a ROC curves of the dynamic CEA&CA125&CA15-3 model. b ROC curves of the pre-chemotherapy CEA&CA125&CA15-3 model. ROC receiver operating characteristic, AUC area under the receiver operating characteristic curve, CI confidence intervalFig. 5Calibration curves for revealing the prediction accuracy between the actual probability and the predicted probability. a Training cohort of the dynamic CEA&CA125&CA15-3 model. b Test cohort of the dynamic CEA&CA125&CA15-3 model. c Training cohort of the pre-chemotherapy CEA&CA125&CA15-3 model. d Test cohort of the pre-chemotherapy CEA&CA125&CA15-3 modelFig. 6Decision curve analysis (DCA) for validating the clinical utility of the models. a Dynamic CEA&CA125&CA15-3 model showed net benefit in threshold ranges of 10–53% (training cohort) and 15–65% (test cohort). b Pre-chemotherapy CEA&CA125&CA15-3 model showed net benefit in threshold ranges of 17–48% (training cohort) and 10–50% (test cohort)Table 3NRI and IDI of the dynamic CEA&CA125&CA15-3 modelIndexEstimate95% CIP valueNRI (vs. the pre-chemotherapy CEA&CA125&CA15-3 model)0.2030.068–0.3380.003IDI (vs. the pre-chemotherapy CEA&CA125&CA15-3 model)0.0530.026–0.080 < 0.001NRI net reclassification improvement, IDI integrated discrimination improvement, CI confidence interval ROC curves for quantifying the discriminative capacity of the models. a ROC curves of the dynamic CEA&CA125&CA15-3 model. b ROC curves of the pre-chemotherapy CEA&CA125&CA15-3 model. ROC receiver operating characteristic, AUC area under the receiver operating characteristic curve, CI confidence interval Calibration curves for revealing the prediction accuracy between the actual probability and the predicted probability. a Training cohort of the dynamic CEA&CA125&CA15-3 model. b Test cohort of the dynamic CEA&CA125&CA15-3 model. c Training cohort of the pre-chemotherapy CEA&CA125&CA15-3 model. d Test cohort of the pre-chemotherapy CEA&CA125&CA15-3 model Decision curve analysis (DCA) for validating the clinical utility of the models. a Dynamic CEA&CA125&CA15-3 model showed net benefit in threshold ranges of 10–53% (training cohort) and 15–65% (test cohort). b Pre-chemotherapy CEA&CA125&CA15-3 model showed net benefit in threshold ranges of 17–48% (training cohort) and 10–50% (test cohort) Table 3NRI and IDI of the dynamic CEA&CA125&CA15-3 modelIndexEstimate95% CIP valueNRI (vs. the pre-chemotherapy CEA&CA125&CA15-3 model)0.2030.068–0.3380.003IDI (vs. the pre-chemotherapy CEA&CA125&CA15-3 model)0.0530.026–0.080 < 0.001NRI net reclassification improvement, IDI integrated discrimination improvement, CI confidence interval NRI and IDI of the dynamic CEA&CA125&CA15-3 model NRI net reclassification improvement, IDI integrated discrimination improvement, CI confidence interval Subgroup analysis The 446 included patients were distributed as follows: 209 patients were HER2-positive with anti-HER2 therapy, 226 patients were HER2-negative, and 11 patients were HER2-positive without anti-HER2 therapy. Given the limited sample size (n = 11) and reduced clinical relevance of the HER2 + non-targeted subgroup (representing non-standard therapy scenarios), subgroup analyses focused on the HER2-positive patients receiving guideline-recommended anti-HER2 therapy (n = 209) and the HER2-negative patients (n = 226). This approach is consistent with clinical decision-making priorities [22]. The HER2-positive with anti-HER2 therapy cohort was divided into a training cohort of 146 patients and a test cohort of 63 patients, with a pCR rate of 42.58%. The HER2-negative cohort was divided into a training cohort of 158 patients and a test cohort of 68 patients, with a pCR rate of 22.57%. Baseline characteristics differed significantly between subgroups (Table 4), particularly in pCR rate and pre-chemotherapy CEA levels (P < 0.001). Table 4Comparison of patient characteristics between the HER2-positive with anti-HER2 therapy and HER2-negative subgroupsVariablesHER2-positive with anti-HER2 therapyN = 209HER2-negativeN = 226p valueAge (years)50.00 (42.00–56.00)47.00 (39.00–54.00)0.071Menopause, n (%) No115 (55.02%)148 (65.49%)0.026 Yes94 (44.98%)78 (34.51%)ER status, n (%) Positive116 (55.50%)133 (58.85%)0.481 Negative93 (44.50%)93 (41.15%)PR status, n (%) Positive89 (42.58%)108 (47.79%)0.276 Negative120 (57.42%)118 (52.21%)HER2 IHC, n (%) 00 (0.00%)101 (44.69%) < 0.001 1 + 0 (0.00%)91 (40.27%) 2 + 40 (19.14%)34 (15.04%) 3 + 169 (80.86%)0 (0.00%)Ki-67, n (%) < 15%3 (1.44%)10 (4.42%)0.067 ≥ 15%206 (98.56%)216 (95.58%)cT stage, n (%) T145 (21.53%)51 (22.57%)0.995 T2131 (62.68%)140 (61.95%) T322 (10.53%)23 (10.18%) T411 (5.26%)12 (5.31%)cN stage, n (%) N048 (22.97%)59 (26.11%)0.725 N1142 (67.94%)149 (65.93%) N212 (5.74%)9 (3.98%) N37 (3.35%)9 (3.98%)pCR, n (%) No120 (57.42%)175 (77.43%) < 0.001 Yes89 (42.58%)51 (22.57%) Pre-chemotherapy CEA2.12 (1.44–3.90)1.79 (1.12–2.72) < 0.001 Pre-chemotherapy CA12514.00 (10.40–19.80)14.80 (10.20–24.93)0.277 Pre-chemotherapy CA15-312.00 (9.17–17.36)12.12 (8.36–17.75)0.918ER estrogen receptor, PR progesterone receptor, HER2 human epidermal growth factor receptor 2, IHC immunohistochemical analysis, cT stage clinical tumor stage, cN stage clinical lymph node stage, pCR pathological complete response Table 4Comparison of patient characteristics between the HER2-positive with anti-HER2 therapy and HER2-negative subgroupsVariablesHER2-positive with anti-HER2 therapyN = 209HER2-negativeN = 226p valueAge (years)50.00 (42.00–56.00)47.00 (39.00–54.00)0.071Menopause, n (%) No115 (55.02%)148 (65.49%)0.026 Yes94 (44.98%)78 (34.51%)ER status, n (%) Positive116 (55.50%)133 (58.85%)0.481 Negative93 (44.50%)93 (41.15%)PR status, n (%) Positive89 (42.58%)108 (47.79%)0.276 Negative120 (57.42%)118 (52.21%)HER2 IHC, n (%) 00 (0.00%)101 (44.69%) < 0.001 1 + 0 (0.00%)91 (40.27%) 2 + 40 (19.14%)34 (15.04%) 3 + 169 (80.86%)0 (0.00%)Ki-67, n (%) < 15%3 (1.44%)10 (4.42%)0.067 ≥ 15%206 (98.56%)216 (95.58%)cT stage, n (%) T145 (21.53%)51 (22.57%)0.995 T2131 (62.68%)140 (61.95%) T322 (10.53%)23 (10.18%) T411 (5.26%)12 (5.31%)cN stage, n (%) N048 (22.97%)59 (26.11%)0.725 N1142 (67.94%)149 (65.93%) N212 (5.74%)9 (3.98%) N37 (3.35%)9 (3.98%)pCR, n (%) No120 (57.42%)175 (77.43%) < 0.001 Yes89 (42.58%)51 (22.57%) Pre-chemotherapy CEA2.12 (1.44–3.90)1.79 (1.12–2.72) < 0.001 Pre-chemotherapy CA12514.00 (10.40–19.80)14.80 (10.20–24.93)0.277 Pre-chemotherapy CA15-312.00 (9.17–17.36)12.12 (8.36–17.75)0.918ER estrogen receptor, PR progesterone receptor, HER2 human epidermal growth factor receptor 2, IHC immunohistochemical analysis, cT stage clinical tumor stage, cN stage clinical lymph node stage, pCR pathological complete response Comparison of patient characteristics between the HER2-positive with anti-HER2 therapy and HER2-negative subgroups ER estrogen receptor, PR progesterone receptor, HER2 human epidermal growth factor receptor 2, IHC immunohistochemical analysis, cT stage clinical tumor stage, cN stage clinical lymph node stage, pCR pathological complete response In terms of model performance, for the HER2-positive with anti-HER2 therapy cohort, the dynamic model had an AUC of 0.717 (95%CI 0.634–0.800) in the training cohort and 0.723 (95%CI 0.595–0.852) in the test cohort, while the pre-chemotherapy model had an AUC of 0.659 (95%CI 0.570–0.747) in the training cohort and 0.600 (95%CI 0.454–0.746) in the test cohort. For the HER2-negative cohort, the dynamic model had an AUC of 0.705 (95%CI 0.610–0.800) in the training cohort and 0.693 (95%CI 0.494–0.893) in the test cohort, while the pre-chemotherapy model had an AUC of 0.688 (95%CI 0.591–0.785) in the training cohort and 0.663 (95%CI 0.482–0.844) in the test cohort (Fig. 7). Subgroup-specific nomograms are presented in Figure S4. These subgroup analyses consistently demonstrated the superior predictive performance of the dynamic model over the pre-chemotherapy baseline, aligning with the overall cohort trends.Fig. 7ROC curves for quantifying the discriminative capacity of subgroups. a ROC curves of the dynamic CEA&CA125&CA15-3 model for the HER2-positive with anti-HER2 therapy subgroup. b ROC curves of the pre-chemotherapy CEA&CA125&CA15-3 model for the HER2-positive with anti-HER2 therapy subgroup. c ROC curves of the dynamic CEA&CA125&CA15-3 model for the HER2-negative subgroup. d ROC curves of the pre-chemotherapy CEA&CA125&CA15-3 model for the HER2-negative subgroup. ROC receiver operating characteristic, AUC area under the receiver operating characteristic curve, CI confidence interval ROC curves for quantifying the discriminative capacity of subgroups. a ROC curves of the dynamic CEA&CA125&CA15-3 model for the HER2-positive with anti-HER2 therapy subgroup. b ROC curves of the pre-chemotherapy CEA&CA125&CA15-3 model for the HER2-positive with anti-HER2 therapy subgroup. c ROC curves of the dynamic CEA&CA125&CA15-3 model for the HER2-negative subgroup. d ROC curves of the pre-chemotherapy CEA&CA125&CA15-3 model for the HER2-negative subgroup. ROC receiver operating characteristic, AUC area under the receiver operating characteristic curve, CI confidence interval Discussion The present study established a novel predictive model incorporating the dynamic trajectories of CEA, CA125, and CA15-3 levels, which achieved greater predictive accuracy than did static biomarker assessment. The identification of PCs as key predictors underscores the critical value of temporal biomarker patterns in monitoring chemotherapeutic response dynamics. Our findings highlight the critical role of longitudinal biomarker monitoring in capturing tumor biological evolution under chemotherapy pressure, thereby enabling early stratification of responders and non-responders. Many researchers have investigated whether standard clinicopathologic characteristics at baseline can predict the response to NAC, albeit with contradictory results. Some articles reported an association between the smaller, lower grade, non-metastasized tumors and the pCR rate, whereas other articles reported no association, indicating that these characteristics cannot be solely used for the prediction of a pCR [23]. The superiority of our dynamic biomarker model over conventional static assessment aligns with emerging trends in precision oncology. Unlike static assessments, monitoring changes in CEA, CA125, and CA15-3 levels over time offers a window into the tumor’s response to treatment, reflecting both the efficacy of the chemotherapy and the biological evolution of the cancer. This dynamic approach allows clinicians to identify patients who are unlikely to benefit from continued treatment earlier, potentially sparing them unnecessary side effects and costs. Furthermore, it facilitates timely adjustments to treatment regimens, optimizing therapeutic outcomes. Our findings are in line with recent studies that emphasize the importance of temporal patterns in biomarker levels. For instance, research has shown that the kinetics of serum tumor marker changes during therapy can predict treatment response and prognosis in various cancers [15, 24–26]. In BC, dynamic monitoring has been used to effectively predict the tumor response to NAC by developing a comprehensive model that combines ultrasound radiomics and deep learning features of pre-NAC, post-treatment 2 cycle, and Delta ultrasound images [27]. A previous study showed that personalized monitoring of circulating tumor DNA (ctDNA) during NAC of high-risk early BC may aid in real-time assessment of treatment response and help fine-tune pCR as a surrogate endpoint of survival [28]. However, the detection of ctDNA demands advanced technical expertise, specialized equipment, and trained personnel. Currently, the lack of standardized protocols and unified detection criteria across platforms results in poor comparability of outcomes between different methodologies, which may compromise the accuracy and reliability of the test results. In contrast, the measurement of serum tumor biomarkers has been highly standardized and is routinely performed in clinical practice during NAC for all BC patients. Our improvement in the predictive performance of serum tumor biomarkers for NAC was achieved through algorithmic refinement, which preserved their inherent advantage of convenient accessibility through minimally invasive blood sampling. The developed dynamic prediction model holds significant clinical and health economic implications. Clinically, it enables early identification of patients unlikely to respond to NAC, allowing timely consideration of alternative treatment strategies, such as switching chemotherapy regimens or proceeding directly to surgery. This could prevent unnecessary toxicity from ineffective treatment and potentially improve surgical outcomes. From a health economic perspective, the model utilizes routinely available serum marker data, making it a cost-effective tool compared to more expensive monitoring techniques, such as advanced imaging or ctDNA analysis. By potentially reducing the number of ineffective chemotherapy cycles, it could generate substantial savings in drug costs, supportive care medications, and management of treatment-related adverse events. The strong association between HER2 IHC 3 + and pCR (OR = 4.62, p < 0.001) corroborates known chemotherapy sensitivity in HER2-positive subtypes [29]. The HER2 protein is crucial for tumor cell proliferation, survival, and invasion. Its overexpression increases chemotherapeutic agent sensitivity. When HER2 is amplified or overexpressed, intracellular signaling pathways are abnormally activated, accelerating tumor cell growth and altering their drug response [30]. In clinical practice, combining anti-HER2 targeted drugs such as trastuzumab with chemotherapy has significantly improved treatment outcomes and survival rates for HER2-positive BC patients [31]. This study underscores the importance of accurate HER2 status detection in clinical practice, enabling more precise and effective treatment strategies. Interestingly, the finding that PC2 for CA125 demonstrated greater predictive power than its baseline level potentially indicates treatment-induced remodeling of the tumor microenvironment. This is corroborated by recent research showing that fluctuations in CA125 levels during NAC correlate with dynamic changes in the tumor immune microenvironment, particularly in macrophage infiltration [32]. The tumor microenvironment is a complex ecosystem comprising various cells, including immune cells, fibroblasts, and endothelial cells, as well as extracellular matrix components [33], and its dynamic state significantly influences therapy response. The predictive capacity of CA125 trajectories, as captured by PC2, may thus reflect underlying alterations in immune cell infiltration, cytokine secretion, and other immunological processes. As key players in this environment, macrophages exhibit a spectrum of functional states [34], and their infiltration dynamics are known to impact treatment sensitivity and recurrence risk [35, 36]. Consequently, monitoring dynamic CA125 patterns during NAC could serve as a valuable surrogate for assessing the tumor’s immune status and predicting outcomes. Furthermore, this principle extends beyond CA125. The dynamic patterns captured by FPCA for other markers likely represent the complex interplay between chemotherapy-induced cell death, tumor repopulation between cycles, and the emergence of resistance mechanisms. Collectively, these dynamic trajectories provide a more holistic and biologically relevant view of tumor behavior under therapeutic pressure than static, single-timepoint measurements. In the era of precision medicine, the assessment of cancer treatment responses is evolving toward multidimensional and dynamic monitoring. Future integration with ctDNA kinetics and radiomic features may further increase the predictive precision. The integration of these multi-dimensional data will enable a more comprehensive understanding of tumor biology and treatment response heterogeneity. By capturing the dynamic evolution of tumors at the molecular level, ctDNA kinetics can provide real-time insights into the genetic changes and clonal selection occurring during neoadjuvant chemotherapy [37]. Radiomics features extracted from medical imaging not only reflect the spatial and temporal heterogeneity of tumor morphology and microenvironment but also offer non-invasive means to monitor tumor burden and response [38]. Combining these factors with serum tumor marker dynamics creates a synergistic effect, allowing for more accurate prediction of pCR and long-term outcomes, such as disease-free survival and overall survival. Despite the novelty and potential of this study, several limitations warrant consideration. First, the retrospective single-center design may introduce selection bias and limit generalizability of our findings. We attempted to minimize this by applying strict inclusion criteria and consecutive enrollment, but multi-center validation remains essential. Second, the potential for overfitting was addressed through LASSO regression with tenfold cross-validation for feature selection, yet external validation is needed to confirm model robustness. Third, the heterogeneity of our study population, particularly the merging of triple-negative and HR-positive/HER2-negative subtypes in subgroup analyses due to sample size constraints, may obscure subtype-specific therapeutic effects. Fourth, the retrospective nature precludes causal inferences and is subject to inherent biases. Fifth, some model estimates, particularly for the dynamic component of CA15-3 (PC2), were unstable, as evidenced by a wide confidence interval. This imprecision likely stems from the relatively small number of patients exhibiting the extreme trajectory patterns that this component captures, highlighting the need for validation in larger cohorts to obtain more reliable effect estimates. Finally, while the variability in sampling intervals improves real-world applicability, it may influence the precision of the dynamic trajectory estimates. Future research directions will focus on several key areas to advance this work. First, multi-center prospective studies are essential for external validation of our model and to confirm its generalizability. Building upon this, we plan to enhance predictive accuracy by incorporating additional biomarkers, such as ctDNA and radiomic features, to create a more comprehensive multi-modal prediction tool. Concurrently, the clinical utility of the model must be rigorously assessed through impact studies that evaluate its effect on real-world treatment decisions and patient outcomes. Finally, the imprecise effect estimates, reflected in the wide confidence intervals of some coefficients, underscore the necessity for larger validation studies to obtain more definitive results and to enable the development of subtype-specific models that truly account for the biological heterogeneity of BC. While this study focused on predicting pCR as a primary endpoint, future work will incorporate long-term survival outcomes such as disease-free survival and overall survival to establish the model’s prognostic value. Conclusions This study developed and validated nomogram models using clinicopathological features and serum marker characteristics to predict NAC efficacy in BC patients. The dynamic model demonstrated comparable discriminative performance to the pre-chemotherapy model in the test set (AUC: 0.712 vs. 0.715) but showed significant improvement in net reclassification (NRI = 0.203, P = 0.003), indicating its potential to refine risk stratification. However, external validation and prospective evaluation are necessary before clinical implementation. Supplementary Information Supplementary material 1. Supplementary material 1.