Routine biomarkers map response probability in HER2-positive breast cancer treated with neoadjuvant chemotherapy and dual HER2 blockade

Abstract

Background

Pathologic complete response (pCR) after neoadjuvant chemotherapy and dual HER2 blockade is associated with improved outcomes in HER2-positive breast cancer, but treatment response remains heterogeneous. We aimed to develop and externally validate an interpretable model based on routine clinicopathological biomarkers to predict pCR and identify clinically meaningful response states.

Methods

In this multicenter retrospective study, 1,082 patients with HER2-positive breast cancer treated with neoadjuvant chemotherapy and dual HER2 blockade at two Chinese centers were included for model development and internal validation. An independent cohort of 307 patients from a third center was used for external validation. Machine-learning models were developed using routinely available pretreatment clinicopathological variables. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC). Probability-landscape analyses were performed to characterize biomarker-defined response states associated with pCR.

Results

The final gradient boosting machine model achieved AUCs of 0.78 in internal testing and 0.70 in external validation. Probability-landscape analyses identified distinct response states with substantially different pCR probabilities. Low hormone receptor expression combined with HER2 immunohistochemistry (IHC) 3 + status defined a high-probability pCR state, whereas high hormone receptor expression, low Ki-67 expression, and HER2 IHC 2+/fluorescence in situ hybridization (FISH)-positive disease defined persistently low-probability states. Among baseline node-positive patients who achieved breast pCR, residual nodal disease was observed in only 6.1% of cases.

Conclusions

Routine clinicopathological biomarkers can be integrated into an interpretable machine-learning model to predict pCR and enable scalable response probability mapping in HER2-positive breast cancer receiving neoadjuvant chemotherapy and dual HER2 blockade. This approach may facilitate individualized treatment selection and support future treatment-adaptation and surgical de-escalation strategies.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12885-026-16403-6.

Introduction

HER2-positive breast cancer represents a biologically and clinically distinct subtype of breast cancer, characterized by aggressive tumour behaviour but marked sensitivity to HER2-directed therapy. In early-stage disease, neoadjuvant systemic therapy provides an opportunity to eradicate micrometastatic disease, downstage the breast and axilla, and assess treatment sensitivity in vivo. Pathologic complete response (pCR) after neoadjuvant therapy is clinically meaningful because it is associated with favourable long-term outcomes at the individual patient level, particularly in aggressive breast cancer subtypes including HER2-positive disease [1].

The introduction of dual HER2 blockade with trastuzumab and pertuzumab has substantially improved neoadjuvant response rates and is now widely used for stage II–III HER2-positive breast cancer [2]. However, response remains heterogeneous. Some tumours achieve pCR with relatively limited chemotherapy exposure, whereas others show residual disease despite intensive cytotoxic treatment. This heterogeneity has become increasingly relevant as neoadjuvant treatment selection remains largely non-individualized, with most patients receiving standard guideline-recommended regimens rather than biologically tailored treatment strategies. Randomized evidence has shown that anthracycline-free regimens can achieve comparable outcomes to anthracycline-containing regimens in the presence of dual HER2 blockade, supporting efforts to refine chemotherapy intensity [3]. More recently, de-escalation studies have suggested that selected patients, especially hormone receptor-negative/HER2-positive tumours, may achieve high pCR rates with shorter or less intensive treatment backbones [4].

Biologically, HER2-positive breast cancer is not a single response entity [5–8]. Tumours enriched for HER2 signalling are more likely to achieve pCR, whereas hormone receptor-positive tumours often display lower sensitivity, partly reflecting interaction between endocrine and HER2 pathways [9–11]. A systematic review and meta-analysis showed that the HER2-enriched intrinsic subtype is strongly associated with pCR in HER2-positive breast cancer, underscoring the importance of underlying biology in determining treatment response [12]. Genomic and transcriptomic assays, such as HER2DX, can capture this biology and predict both pCR and survival outcomes, but their use in routine practice is constrained by cost, technical complexity, and limited availability in many healthcare settings [13, 14].

Most real-world treatment planning continues to rely on routinely available clinicopathological variables, including estrogen receptor (ER), progesterone receptor (PR), HER2 immunohistochemistry (IHC) status and Ki-67 [15–17]. These variables are individually associated with pCR response, but they are rarely integrated into an interpretable probability framework that can map pretreatment biomarker patterns to individualized response likelihood. Machine learning approaches may model nonlinear relationships among routine variables, yet many prior studies have focused mainly on discrimination metrics rather than clinically meaningful probability stratification, external validation, or biological interpretability [18]. Automated machine-learning frameworks may further standardize model development and lower barriers to clinical implementation, but their value depends on whether they generate transparent and clinically usable insights rather than black-box predictions [19].

Therefore, a key unmet need is to determine whether routinely collected pretreatment biomarkers can be integrated into an interpretable probability framework for mapping individualized pCR likelihood in HER2-positive breast cancer treated with neoadjuvant dual HER2 blockade. In this multicentre study, we developed and externally validated an interpretable routine-biomarker model to estimate individualized pCR probability in patients receiving trastuzumab- and pertuzumab-based neoadjuvant therapy. We further examined whether the resulting probability patterns identify biologically meaningful response states and explored regimen-conditional and nodal-response profiles to define candidate patient groups for prospective evaluation of treatment adaptation and surgical de-escalation.

Methods

Study design and participants

This multicentre retrospective cohort study developed and externally validated a prediction model to estimate the probability of pathologic complete response (pCR) after neoadjuvant therapy in HER2-positive breast cancer. The study was designed and reported in accordance with TRIPOD + AI guidance for prediction model development and validation.

The derivation cohort included consecutive patients treated at Tianjin Medical University Cancer Institute and Hospital, Tianjin, China, and the First Affiliated Hospital of Zhengzhou University, Zhengzhou, China, between 1 May 2019 and 30 June 2025. A total of 1,082 patients were included in the derivation cohort, comprising 966 patients from Tianjin and 116 patients from Zhengzhou. External validation was performed in an independent cohort of 307 consecutive patients treated at Sichuan Cancer Hospital and Institute, Chengdu, China, between 1 June 2019 and 31 August 2025. Data extraction procedures, variable definitions, and outcome definitions were harmonised across participating centres.

Eligible patients had histologically confirmed invasive ductal carcinoma, HER2-positive disease, trastuzumab- and pertuzumab-based neoadjuvant therapy, and definitive breast surgery with axillary evaluation. HER2-positive disease was defined as immunohistochemistry (IHC) 3 + or IHC 2 + with ERBB2 amplification confirmed by fluorescence in situ hybridisation.

Patients with prior or concurrent malignancy or bilateral breast cancer were excluded. For the derivation cohort, data collection was restricted to patients with complete information for all prespecified baseline predictors and surgical/pathological outcomes. Therefore, no additional patients were excluded from the assembled derivation cohort because of missing data. For the external validation cohort, the same eligibility criteria and complete-data requirement were applied. A total of 331 patients were initially identified from Sichuan Cancer Hospital. Twenty-two patients lacking histological grade information and two patients lacking Ki-67 data were excluded, leaving 307 patients for external validation.Consecutive inclusion was used to reduce selection bias.

Ethical approval was obtained from the institutional review boards of all participating centres: Tianjin Medical University Cancer Institute and Hospital, bc20260648; First Affiliated Hospital of Zhengzhou University, L2024-K006-002; and Sichuan Cancer Hospital and Institute, SCCHEC-02-2023-037. The study was conducted in accordance with the Declaration of Helsinki, and written informed consent was obtained according to local regulations.

Treatment regimens and outcome definition

Neoadjuvant treatment consisted of chemotherapy combined with dual HER2 blockade using trastuzumab and pertuzumab. Chemotherapy regimens (CR) included THP (taxane, trastuzumab, and pertuzumab), TCbHP (taxane, carboplatin, trastuzumab, and pertuzumab), and AC-THP (anthracycline/cyclophosphamide followed by taxane, trastuzumab, and pertuzumab), according to institutional protocols and physician discretion. THP was defined as taxane plus trastuzumab and pertuzumab; TCbHP as taxane, carboplatin, trastuzumab, and pertuzumab; and AC-THP as anthracycline and cyclophosphamide followed by taxane plus trastuzumab and pertuzumab.

The primary outcome was pCR, defined as absence of residual invasive carcinoma in both the breast and axillary lymph nodes after neoadjuvant therapy, corresponding to ypT0/is ypN0. Pathological assessments were performed by dedicated breast pathologists using standardised procedures.

For the nodal-response analysis, breast pCR was defined separately as absence of residual invasive carcinoma in the breast irrespective of nodal status, corresponding to ypT0/is. Residual nodal disease was assessed among patients who were clinically node-positive at baseline, achieved breast pCR, and underwent axillary evaluation at definitive surgery.

Machine learning model predictors and data processing

Model predictors were prespecified based on clinical availability, biological plausibility, and routine use before initiation of neoadjuvant therapy. Predictors included ER expression, PR expression, HER2 IHC status, Ki-67 index, tumour maximum diameter, baseline nodal status, menopausal status, age, histologic grade, and neoadjuvant regimen.

ER and PR were recorded as the percentage of positive tumour cells and retained as continuous variables during model development. HER2 IHC status was classified as IHC 2 + with positive amplification through fluorescence in situ hybridization (FISH) or IHC 3+. Ki-67 was recorded as a continuous percentage. Tumour maximum diameter was recorded in centimetres. Baseline nodal status was classified as clinically node-negative or node-positive. Neoadjuvant regimen was included as a treatment-context variable rather than a causal exposure.

Continuous variables were used without categorisation or transformation during model training to preserve information and allow flexible modelling of nonlinear relationships. Clinically relevant categories, such as ER or PR < 10%, 10–50%, and ≥ 90%, and Ki-67 thresholds used in subgroup summaries, were applied only for descriptive summaries and probability-landscape visualisation. These exploratory strata were not intended to replace established clinical cutoffs, including the ER/PR 1% threshold used to define hormone receptor positivity or commonly used Ki-67 cutoffs. Data were extracted using a standardised case report form and underwent internal consistency checks for impossible values, outliers, and inconsistent coding. Because complete baseline predictor and outcome data were required for study inclusion, all analyses were conducted using complete cases without imputation.

Model development

We developed a supervised machine learning model to predict the individual probability of pCR using routinely available clinicopathological predictors. Model development used H2O’s AutoML framework to train multiple candidate algorithms, tune hyperparameters, and rank models according to prespecified performance criteria [20]. The sample size was considered adequate for multivariable prediction modelling given the number of candidate predictors and outcome events, consistent with established methodological guidance.

The derivation dataset was randomly split into a training set (80%) and an internal test set (20%) using a fixed random seed. The internal test set was fully held out and not used for model selection, hyperparameter tuning, or threshold optimisation. Model development and selection were conducted exclusively within the training set using k-fold cross-validation implemented in the H2O.ai AutoML platform. To reduce the risk of overfitting, model development was restricted to the training set with k-fold cross-validation, and model performance was subsequently evaluated on a fully held-out internal test set and an independent external validation cohort.

A prespecified maximum of 50 candidate models spanning multiple algorithm families (including gradient boosting machines, extreme gradient boosting, extremely randomised trees, regularised generalised linear models, and deep learning) were trained and compared. Model selection was based on cross-validated performance (primary metric: AUC), with calibration considered as an important secondary criterion. The final model was selected within the training set and fixed (locked) prior to evaluation in the held-out internal test set and external validation cohort. Complete AutoML configuration details, leaderboard results, and hyperparameters are provided in Supplementary Tables S1–S3.

Model performance assessment and interpretability

Model discrimination was quantified using the area under the area under the receiver operating characteristic curve (AUC) with 95% confidence intervals estimated by non-parametric bootstrap resampling of the relevant evaluation cohort (internal test and external validation cohorts separately). ROC curves for all candidate base models in the internal test set are provided in Figure S1. For clinical interpretability, we also computed sensitivity, specificity, positive predictive value, and negative predictive value at a prespecified probability threshold of 0.50, and explored alternative thresholds using Youden’s index in sensitivity analyses.

Calibration was assessed by comparing predicted probabilities with observed pCR frequencies across risk deciles (calibration-in-the-large). Calibration curves for all candidate base models in the internal test set are shown in Figure S2. We visually inspected calibration plots and, where appropriate, estimated calibration intercept and slope by regressing the outcome on the logit of predicted risk. Because clinical decision-making depends on the reliability of predicted probabilities, calibration was considered alongside AUC when selecting the final model. The locked final model was applied to the external validation cohort from Sichuan Cancer Hospital without refitting. External validation was performed without model updating or recalibration to provide an unbiased assessment of model generalisability. All performance metrics (AUC, calibration plots, and threshold-based measures) were recomputed in the external cohort to quantify generalisability across institutions, time periods, and potential differences in case-mix. Comparative performance metrics across base models in the internal test set are summarized in Figure S3.

To ensure clinical interpretability, we used SHapley Additive exPlanations (SHAP) values to quantify each predictor’s contribution to model predictions. SHAP values were computed for individual patients and summarised as mean absolute SHAP values to rank overall variable importance. Directionality of associations was explored using dependence plots and partial dependence plots for key predictors (ER, PR, HER2 IHC status, Ki-67, tumour size, regimen). For categorical predictors (e.g., regimen), we reported average predicted pCR probabilities by category while holding other variables at their observed values. Interpretability analyses were performed after model selection and were not used to tune model performance. This separation helped avoid “double dipping” and preserved the validity of performance estimates. These analyses were intended to enhance model transparency and support clinical interpretability rather than to infer causal relationships.

Regimen-conditional probability mapping

To map pCR probability patterns across clinically relevant biomarker profiles, model-based simulations were conducted using biomarker grids defined by ER expression, PR expression, Ki-67 index, and HER2 IHC status. ER, PR, and Ki-67 were evaluated across their observed clinical ranges, and HER2 IHC status was evaluated as IHC 2 + with ERBB2 amplification or IHC 3+. For each biomarker profile, predicted pCR probabilities were generated under three regimen scenarios by setting the regimen variable to THP, TCbHP, or AC-THP. Other covariates were held at the prespecified values used for the simulations. Predicted probabilities were summarized as ER–PR and ER–Ki-67 probability landscapes, stratified by HER2 IHC status.

These regimen-conditional predictions were intended to characterize response-probability patterns under different real-world treatment contexts. Because treatment allocation was not randomized, they should not be interpreted as causal comparisons of regimen efficacy or as direct treatment recommendations.

Results

Multicentre cohort and baseline response patterns

A total of 1,082 patients with HER2-positive breast cancer were included in the derivation cohort, including 866 patients in the training set and 216 patients in the held-out internal test set for machine learning (Fig. 1). Overall, 605 patients (55.9%) achieved pCR. The mean age was 49.80 years, 563 patients (52.03%) were premenopausal, and most patients presented with node-positive disease (796 of 1,082, 73.57%). ER positivity was present in 555 tumours (51.29%), PR positivity in 363 tumours (33.55%), and HER2 IHC 3 + status in 877 tumours (81.05%). Most patients received TCbHP (738 of 1,082, 68.21%), whereas 152 (14.05%) received THP and 192 (17.74%) received AC-THP.

Fig. 1.

Multicentre study design and routine-biomarker modelling framework. This figure summarizes the study workflow, including cohort assembly, training and held-out internal testing, independent external validation, routine biomarker collection, model development, performance evaluation, and interpretation using SHAP and partial-dependence analyses. The framework illustrates how routinely available pretreatment clinicopathologic variables were used to generate individualized pCR probability estimates in HER2-positive breast cancer treated with neoadjuvant dual HER2 blockade

Patients achieving pCR were enriched for biomarker patterns associated with greater sensitivity to HER2-directed neoadjuvant therapy. Compared with patients without pCR, those achieving pCR more frequently had ER-negative tumours (61.98% vs. 31.87%), PR-negative tumours (80.66% vs. 48.43%), HER2 IHC 3 + disease (92.89% vs. 66.04%), and Ki-67 ≥ 30% (87.93% vs. 80.71%). Baseline characteristics according to pCR status are shown in Table 1.

Table 1.

Patient characteristics according to total pathological complete response

Variables	Total (n = 1082)	Non-pCR (n = 477)	pCR (n = 605)
Age (years), Mean ± SD	49.80 ± 10.09	49.60 ± 10.34	49.97 ± 9.90
Menopause, n (%)
No	563 (52.03)	250 (52.41)	313 (51.74)
Yes	519 (47.97)	227 (47.59)	292 (48.26)
Neoadjuvant chemotherapy Regimens, n (%)
THP*4	152 (14.05)	79 (16.56)	73 (12.07)
TCbHP*6	738 (68.21)	306 (64.15)	432 (71.40)
AC*4-THP*4	192 (17.74)	92 (19.29)	100 (16.53)
TumorMax (cm), M (Q1, Q3)	3.90 (2.90, 5.30)	4.00 (3.00, 5.50)	3.70 (2.80, 5.00)
Lymph node status, n (%)
Negative	286 (26.43)	143 (29.98)	143 (23.64)
Positive	796 (73.57)	334 (70.02)	462 (76.36)
ER Status, n (%)
Negative	527 (48.71)	152 (31.87)	375 (61.98)
Positive	555 (51.29)	325 (68.13)	230 (38.02)
PR Status, n (%)
Negative	719 (66.45)	231 (48.43)	488 (80.66)
Positive	363 (33.55)	246 (51.57)	117 (19.34)
HER2 Status, n (%)
2+	205 (18.95)	162 (33.96)	43 (7.11)
3+	877 (81.05)	315 (66.04)	562 (92.89)
Ki67, n (%)
<30%	165 (15.25)	92 (19.29)	73 (12.07)
≥30%	917 (84.75)	385 (80.71)	532 (87.93)
Miscarry, n (%)
No	528 (48.80)	232 (48.64)	296 (48.93)
Yes	554 (51.20)	245 (51.36)	309 (51.07)
Histological Grade, n (%)
I-II	549 (50.74)	261 (54.72)	288 (47.60)
III	533 (49.26)	216 (45.28)	317 (52.40)

SD Standard deviation, M Median, Q₁ 1st Quartile, Q₃ 3rd Quartile

Internal and external evaluation of the pCR probability model

Using routinely available clinicopathologic variables, the final gradient boosting machine (GBM) model from H2O AutoML achieved an area under the receiver operating characteristic curve (AUC) of 0.78 in the held-out internal test cohort (Fig. 2A). At a prespecified probability threshold of 0.50, the model showed a sensitivity of 83% and specificity of 61%, and false-negative rate of 17%. Calibration plots showed generally close agreement between predicted and observed pCR probabilities across risk strata, supporting the use of the model for probability estimation rather than binary classification alone (Supplementary Fig. S2). Performance summaries for candidate models are provided in Supplementary Table S4.

Fig. 2.

Internal discrimination and biomarker contributions of the routine-biomarker pCR model. A The routine-biomarker model showed good discrimination for predicting pCR in the held-out internal test cohort, with an AUC of 0.78. B SHAP-based variable importance identified chemotherapy regimen (CR), ER expression, tumour size, HER2 IHC status, nodal status, PR expression, and Ki-67 as key contributors to individualized response-probability estimates

The selected GBM model was externally validated in an independent cohort of 307 patients from Sichuan Cancer Hospital without refitting or recalibration. In this cohort, the model achieved an AUC of 0.70 (Fig. 3), showing moderate discrimination across an independent centre. The attenuation in performance compared with the internal test cohort highlights the importance of external validation and supports the need for prospective evaluation before routine clinical implementation. Baseline characteristics of the external validation cohort are provided in Table S5. Among the 307 patients in the external validation cohort, 51 (16.6%) had HER2 IHC 2+/ISH-positive tumors and 256 (83.4%) had HER2 IHC 3 + tumors.

Fig. 3.

Independent external validation of the routine-biomarker pCR probability model. The final selected model was evaluated in an independent external cohort from Sichuan Cancer Hospital without refitting or recalibration. The model retained moderate discrimination in external validation, with an AUC of 0.70, supporting the generalizability of routine-biomarker-based pCR probability estimation across centres

To facilitate clinical interpretation and transparent model application, the GBM model was implemented as a prototype web-based calculator for clinicians and clinical researchers, allowing users to input routine pretreatment variables and visualize individualized pCR probability estimates as an exploratory aid for probability stratification and future study design (Supplementary Fig. S4; see Data availability).

Routine clinicopathologic variables shape individualized pCR probability estimates

SHapley Additive exPlanations (SHAP)-based model interpretation identified neoadjuvant regimen and ER expression as major contributors to predicted pCR probability; among biomarkers, ER, PR, HER2 IHC status, Ki-67, tumour size and nodal status shaped the probability landscape (Fig. 2B). Because treatment allocation was not randomized, the contribution of regimen should be interpreted as a real-world conditional association rather than a causal treatment effect.

Among routine biomarker variables, lower ER and PR expression, HER2 IHC 3 + status, and higher Ki-67 were consistently associated with higher predicted pCR probabilities. Partial-dependence analyses further showed graded relationships between these pretreatment variables and modelled response probability, supporting their role in shaping the overall pCR probability landscape (Fig. 4).

Fig. 4.

Marginal effects of routine clinicopathologic predictors on pCR probability. Partial-dependence plots show the marginal relationship between each routine clinicopathologic predictor and predicted pCR probability while averaging over other variables. A Chemotherapy regimen (1 = AC-THP, 2 = TCbHP, 3 = THP). B ER expression (%). C Tumour size (maximum diameter, cm). D HER2 IHC status (2+, 3+). E Nodal status (0 = node-negative; 1 = node-positive). F Menopausal status (0 = premenopausal; 1 = postmenopausal). G PR expression (%). (H) Ki-67 expression (%). I Age (years). These plots illustrate how ER, PR, HER2 IHC status, Ki-67, tumour size, nodal status, age, menopausal status, and treatment regimen contribute to the modelled response-probability landscape.

ER–PR and ER–Ki-67 biomarker landscapes identify distinct pCR probability states

Regimen-conditional probability maps demonstrated distinct pCR probability patterns across ER, PR, Ki-67, and HER2 IHC strata (Figs. 5 and 6). These simulations were used to characterize biomarker-defined response states and should not be interpreted as causal comparisons between treatment regimens. Tumours with low ER and PR expression represented a high-probability response state, particularly when HER2 IHC status was 3+. In patients with ER < 10% and PR < 10%, predicted pCR probabilities exceeded 80% under the TCbHP scenario and exceeded 75% under the THP scenario, suggesting that this subgroup may be suitable for prospective evaluation in treatment de-escalation strategies.

Fig. 5.

ER–PR biomarker landscape of regimen-conditional pCR probability by HER2 IHC status. The figure maps predicted pCR probabilities across ER and PR expression levels, stratified by HER2 IHC status and neoadjuvant Chemotherapy Regimen (CR). Panels A–C show patients with HER2 IHC 2+/FISH-positive disease under regimen categories THP (CR = 3), TCbHP (CR = 2), and AC-THP (CR = 1), respectively; panels D–F show patients with HER2 IHC 3 + disease under regimen categories, respectively. The colour scale represents the predicted probability of pCR. The landscape highlights higher predicted pCR probabilities among tumours with low ER/PR expression and HER2 IHC 3 + status, and lower probabilities among tumours with high hormone receptor expression

Fig. 6.

ER–Ki-67 biomarker landscape of regimen-conditional pCR probability by HER2 IHC status. The figure shows predicted pCR probabilities across ER expression and Ki-67 levels, stratified by HER2 IHC status and neoadjuvant Chemotherapy Regimen (CR). Panels A–C show patients with HER2 IHC 2+/FISH-positive disease under regimen categories THP (CR = 3), TCbHP (CR = 2), and AC-THP (CR = 1), respectively; panels D–F show patients with HER2 IHC 3 + disease under regimen categories, respectively. The colour scale represents the predicted probability of pCR. The maps illustrate how Ki-67 modifies response probability across hormone receptor expression states, helping distinguish highly responsive, intermediate, and lower-probability biomarker profiles

An intermediate response-probability state was observed among tumours with ER expression between 10% and 50% and Ki-67 > 30%. In this subgroup, predicted pCR probabilities were approximately 10–20% points higher under the TCbHP scenario than under the THP scenario, whereas AC-THP did not show higher predicted probabilities. This pattern identifies a subgroup in which chemotherapy backbone optimization may be most relevant for future prospective studies. By contrast, tumours with high hormone receptor expression represented a low-probability response state. Patients with ER ≥ 90% or PR ≥ 90% had predicted pCR probabilities below 30% across evaluated regimen contexts, indicating that this hormone receptor-high subgroup may be less likely to achieve pCR with conventional chemotherapy plus dual HER2 blockade.

Breast pCR associates with low residual nodal burden

Among 501 patients with baseline node-positive disease who achieved breast pCR, defined as absence of residual invasive carcinoma in the breast irrespective of nodal status, 39 patients (7.8%) had residual nodal disease at surgery (Table S6). Of these, 37 had fewer than three positive nodes, resulting in a rate of three or more residual positive nodes below 0.5%. These findings suggest that patients achieving breast pCR after neoadjuvant dual HER2 blockade constitute a low residual nodal-burden subgroup, supporting their consideration as candidates for prospective evaluation of surgical de-escalation strategies.

Discussion

In this multicentre study, we developed and externally validated an interpretable model using routinely available clinicopathologic variables to estimate individualized pCR probability in HER2-positive breast cancer treated with neoadjuvant dual HER2 blockade. The model achieved an AUC of 0.78 in the held-out internal test cohort and 0.70 in the independent external validation cohort, which supports moderate and above model transportability. However, the slight decrease in AUC likely reflects the difficulty of applying a locked prediction model across different institutions, where differences in patient case mix, treatment patterns, pathology assessment, biomarker evaluation, and clinical workflows may reduce model performance. The external validation cohort was derived from an independent institution with distinct patient characteristics and treatment distributions, which may have altered the relationships between predictor variables and treatment response. Despite this attenuation, the model maintained reasonable discrimination without refitting or recalibration, supporting its transportability across institutions. Therefore, multicentre evaluation remains necessary, and local calibration may be required before the model can be implemented in routine clinical practice. Even though, the model mapped pretreatment biomarker patterns into clinically interpretable pCR probability states. Tumours with low HR expression and HER2 IHC 3 + status showed consistently high predicted pCR probabilities, whereas tumours with intermediate HR expression and high Ki-67 exhibited intermediate probabilities across regimen contexts. By contrast, tumours with high HR expression, low Ki-67, and particularly HER2 IHC 2+/in situ hybridization-positive disease showed persistently low predicted pCR probabilities. Among patients with baseline node-positive disease who achieved breast pCR, residual nodal disease was uncommon. These findings suggest that routine biomarkers can be used not only for response prediction, but also for mapping response heterogeneity in a way that may inform future treatment-adaptation and surgical-de-escalation studies.

A practical strength of probability stratification is that predicted pCR estimates can be interpreted as response-probability groups rather than as a single binary prediction. As an illustrative framework, patients with predicted pCR probabilities below 40%, 40–70%, and above 70% could be considered low-, intermediate-, and high-response-probability groups, respectively. These thresholds are not intended to guide treatment decisions directly, but may support multidisciplinary discussion, trial enrichment, and future response-adapted study design. For example, low-probability patients may be considered for closer monitoring or treatment-escalation studies, whereas high-probability patients may be candidates for prospective de-escalation studies. The final choice of clinically useful thresholds should depend on the intended decision, model calibration, decision-curve analysis, and prospective validation. To facilitate clinical interpretation, representative biomarker-defined probability states, their typical predicted pCR probabilities, and their potential clinical implications are summarized in Supplementary Table S7.

Another contribution of this study is the development of an accessible response-probability framework rather than a prediction model alone. pCR is already established as an individual-level marker of favourable prognosis in aggressive breast cancer subtypes, including HER2-positive disease [1]. However, the clinical challenge is not simply to predict whether pCR will occur, but to understand which pretreatment biomarker patterns define high-, intermediate-, and low-probability response states. Our findings are consistent with the known biological heterogeneity of HER2-positive breast cancer. HER2-enriched tumours are more likely to achieve pCR after anti-HER2-based neoadjuvant therapy, whereas HR-positive tumours often show lower response probability, reflecting less complete dependence on HER2 signalling [12]. Genomic assays such as HER2DX provide deeper biological resolution and have been associated with both pCR and survival outcomes, but they require molecular testing that may not be routinely available in many clinical settings [13]. In this context, our model shows that routinely collected markers can serve as practical proxies for underlying response biology, although they should be viewed as complementary to, rather than replacements for, genomic profiling. Unlike conventional approaches that rely on individual biomarkers or predefined categorical groupings, the proposed framework integrates multiple routinely available clinicopathologic variables and captures their interactions to generate individualized pCR probability estimates. This probabilistic approach allows patients with similar conventional biomarker classifications to be distinguished according to their predicted likelihood of response, thereby providing a more nuanced assessment of treatment sensitivity. Because all model inputs are routinely collected in standard clinical practice, the framework can be implemented without additional molecular testing or specialized assays, supporting its potential scalability and clinical applicability.

The high-probability response state has direct relevance to treatment de-escalation research. In our probability maps, tumours with low HR expression and HER2 IHC 3 + status showed high predicted pCR probabilities across common treatment contexts, including carboplatin-free scenarios. This pattern is consistent with clinical trial evidence showing that some patients with HER2-positive disease can achieve excellent outcomes with less intensive chemotherapy backbones. In TRAIN-2, anthracycline omission did not compromise survival outcomes in patients receiving dual HER2 blockade, while anthracycline-containing treatment was associated with greater toxicity [21]. In WSG-ADAPT HER2+/HR−, selected patients achieved high pCR rates with short-course paclitaxel plus trastuzumab and pertuzumab [4]. More recently, neoCARHP has further examined whether carboplatin can be omitted from taxane plus trastuzumab and pertuzumab in stage II–III HER2-positive disease [22]. Our findings do not establish that chemotherapy can be safely reduced in individual patients. Rather, they identify a biomarker-defined high-probability group that could be prioritized for prospective de-escalation trials.

The intermediate and low-probability states may be equally important. Tumours with intermediate HR expression and high Ki-67 retained partial chemosensitivity, and their predicted pCR probabilities differed across regimen scenarios. Because treatment allocation was not randomized, these differences should not be interpreted as causal treatment effects. Nevertheless, they identify a subgroup in which chemotherapy backbone optimization may be most clinically relevant. By contrast, tumours with high HR expression, low Ki-67, and particularly HER2 IHC 2+/in situ hybridization-positive disease showed persistently low predicted pCR probabilities. This pattern suggests that further cytotoxic escalation alone may have a limited ability to induce pCR in endocrine-dominant or lower-HER2-expression tumours. Preclinical work supports a biological basis for this observation, showing that interaction between ER and HER2 signalling can contribute to resistance to HER2-targeted therapy [23]. Early clinical studies have explored combined ER, HER2, and CDK4/6 blockade in ER-positive/HER2-positive disease, supporting the concept that endocrine-driven resistance may require pathway-specific treatment adaptation [24]. In parallel, antibody–drug conjugates (ADCs) are being evaluated in early HER2-positive breast cancer and may be particularly relevant where HER2 heterogeneity or incomplete HER2 dependence limits response to conventional HER2 blockade [25, 26]. These observations support future trials that use pretreatment probability states to enrich patients for treatment adaptation strategies rather than simply escalating standard chemotherapy.

The nodal-response findings provide an additional clinical perspective, but they should be interpreted conservatively. Among baseline node-positive patients who achieved breast pCR, residual nodal disease was uncommon, and extensive residual nodal burden was rare. This finding is consistent with the broader movement toward axillary surgery de-escalation after neoadjuvant therapy in patients who convert to clinically node-negative status [27]. However, pretreatment pCR probability estimates should not determine surgical extent. Surgical decisions still require post-treatment clinical assessment, imaging, targeted axillary approaches where appropriate, and pathological confirmation. The value of the model is therefore not to recommend omission or reduction of surgery, but to define candidate patient groups in whom surgical de-escalation strategies could be prospectively evaluated.

This study also has practical implications for accessible precision oncology. Precision stratification in HER2-positive breast cancer increasingly relies on molecular assays, transcriptomic subtyping, and emerging drug platforms. These tools are valuable, but they may not be available in all health-care systems. A framework based on ER, PR, HER2 IHC, Ki-67, tumour size, nodal status, and standard clinical variables is easier to implement, less expensive, and more scalable. The web-based calculator further illustrates how individualized probability estimates could be delivered at the point of care. The calculator is intended for clinicians and researchers managing patients with HER2-positive breast cancer treated with neoadjuvant chemotherapy and dual HER2 blockade. By integrating routinely available clinicopathologic variables, it may support patient counselling, multidisciplinary discussion, clinical trial design, and future treatment-adaptation or de-escalation research. Before clinical implementation, however, such tools should be evaluated for calibration, decision impact, and clinical utility in accordance with contemporary prediction-model reporting and evaluation standards [28].

Several strengths support the interpretation of this study. The cohort was multicentre and included a large number of patients treated with trastuzumab- and pertuzumab-based neoadjuvant therapy. The model was evaluated in a held-out internal test set and in an independent external validation cohort without refitting or recalibration. The use of SHAP and partial-dependence analyses improved interpretability and allowed the model outputs to be connected to known response biology. Finally, the model relied on routinely available variables, which increases the potential for broad implementation.

There are limitations that need to be acknowledged. First, this was a retrospective observational study. The external validation analysis was restricted to complete cases, which may have influenced cohort representativeness and introduced potential selection bias. Treatment allocation was not randomized, regimen-conditional predictions therefore represent associations within real-world treatment patterns, not causal estimates of regimen efficacy. Second, model discrimination was moderate in external validation, indicating that the tool is better suited for probability stratification than definitive individual-level treatment selection. Third, the study was conducted in Chinese cohorts, and performance should be evaluated in other populations and health-care systems. Fourth, some potentially informative variables, including tumour-infiltrating lymphocytes, quantitative HER2 measures, radiomic features, and genomic or transcriptomic data, were not available. Fifth, pCR was the primary endpoint and serves primarily as a surrogate endpoint. Future studies should evaluate whether the identified probability states are associated with longer-term outcomes, including event-free survival, invasive disease-free survival, overall survival, and locoregional control. Such analyses will be important for determining whether the probability-defined response states identified by the model are associated with clinically meaningful outcomes and for further establishing the clinical relevance of the proposed framework.

Future work should focus on prospective validation and clinical integration. The first priority is to test whether the model remains calibrated across institutions, populations, pathology workflows, and treatment protocols. The second priority is to evaluate whether biomarker-defined pCR probability states can improve trial design by enriching patients for de-escalation, treatment adaptation, or surgical de-escalation strategies. The third priority is to determine whether adding molecular, immune, imaging, or longitudinal response data meaningfully improves prediction beyond routine clinicopathologic variables. Ultimately, the clinical value of this framework will depend not only on discrimination, but on whether it improves patient selection, reduces unnecessary toxicity, and identifies patients who require alternative strategies.

Author contributions

X.C.C. and Z.J.Y. had full access to all study data and verified the integrity of the data and the accuracy of the analyses. X.C.C., X.C. and Z.J.Y. conceptualised and designed the study. Z.J.Y., Y.Q.D., L.X., Y.N.L., T.R.L., X.J.L. and J.J.L. contributed to data acquisition, analysis and interpretation. Y.Q.D., Y.N.L., X.C. and J.R.C. performed the statistical analyses. Z.J.Y., Y.Q.D. and L.X. drafted the manuscript. X.C., Y.Y. and X.C.C. critically revised the manuscript for important intellectual content. X.C. and X.C.C. obtained funding. All authors reviewed and approved the final manuscript.

Funding

This study was supported by the Noncommunicable Chronic Diseases–National Science and Technology Major Project (2024ZD0531600), the Natural Science Foundation of Tianjin (23JCZDJC00330), the Tianjin Key Medical Discipline Construction Project (TJYXZDXK-3-003A), and the Tianjin Municipal Education Commission Scientific Research Program (2022ZD056).

Data availability

Deidentified individual participant data supporting the findings of this study, together with the study protocol, are available from the corresponding author, Prof. Xu-Chen Cao, upon reasonable request and subject to relevant ethical and institutional approvals.

Code availability

Model development and prediction were performed using H2O.ai AutoML, and additional statistical analyses and visualisations were conducted in R version 4.5.2. Random seeds were fixed to support reproducibility of model training and evaluation. A web-based calculator for estimating individual predicted pCR probabilities is available at: https://aqpet.shinyapps.io/BC_pCR_app/ . Further details of the modelling workflow are provided in the Methods.

Declarations

Competing interests

The authors declare no competing interests.