Correlation Study between Neoadjuvant Chemotherapy Response and Long-term Prognosis in Breast Cancer Based on Deep Learning Models

Correlation Study between Neoadjuvant Chemotherapy Response and Long-term Prognosis in Breast Cancer Based on Deep Learning Models | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Correlation Study between Neoadjuvant Chemotherapy Response and Long-term Prognosis in Breast Cancer Based on Deep Learning Models Wang Ke, Luo Yikai, Zhang Peng, Yang Bing, Tao Yubo This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6407870/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background: Neoadjuvant chemotherapy (NAC) is a critical component of breast cancer treatment; however, patient responses and long-term prognoses vary significantly. Accurately predicting post-NAC prognosis is essential for guiding individualized treatment plans. This study aims to develop a deep learning-based prediction model to analyze the correlation between NAC efficacy and long-term outcomes in breast cancer patients, providing a new approach to identifying high-risk populations. Objective: To construct a deep learning model that integrates multi-dimensional clinical and pathological parameters to predict recurrence and metastasis risk in breast cancer patients following NAC, thereby facilitating personalized treatment strategies. Methods: A retrospective analysis was conducted on 832 breast cancer patients who received NAC at our hospital from 2013 to 2022. Comprehensive clinical, pathological, and molecular subtype data including：pre- and post-NAC tumor characteristics, Ki-67 index, lymph node status, lymphovascular invasion, and Miller-Payne grading were collected. A Multi-layer Perceptron (MLP) based deep learning model was developed, incorporating ensemble learning strategies to integrate multi-modal prediction results. The model’s performance in assessing recurrence and metastasis risk was evaluated across different breast cancer subtypes. Results: The analysis identified key prognostic factors, including tumor size reduction, post-NAC lymph node status, Ki-67 index, lymphovascular invasion, and Miller-Payne grading. The MLP model achieved AUC values of 0.86 (95% CI: 0.82-0.93) for HER2+,0.82 (95% CI: 0.70-0.92)for triple-negative breast cancer, and 0.76 (95% CI: 0.66-0.82) for HR+/HER2−. The model successfully stratified high-risk subgroups with significant differences in prognosis, providing valuable insights for clinical decision-making. Conclusions: The deep learning-based prediction model developed in this study effectively assesses the prognostic risk of breast cancer patients after NAC. Its clinical application holds potential for optimizing individualized treatment and follow-up strategies, ultimately improving patient outcomes. Breast cancer Neoadjuvant chemotherapy Artificial intelligence Recurrence and metastasis Prognosis prediction Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Background Breast cancer is the most common malignant tumor among women worldwide, with its incidence and mortality rates showing a continuous upward trend. Neoadjuvant chemotherapy (NAC), as an important component of the standard treatment regimen for locally advanced breast cancer, plays a key role in downstaging surgery, improving breast conservation rates, early control of micrometastases, and evaluating chemotherapy sensitivity [ 1 , 2 ]. However, clinical practice indicates that patients' responses to NAC show significant heterogeneity, ranging from pathological complete response (pCR) to disease progression. This variation in response directly affects patients' long-term prognosis [ 3 – 4 ]. Traditional prognostic assessment methods primarily rely on clinical pathological indicators such as TNM staging and molecular subtypes. Meta-analyses show that even among patients achieving pCR, significant differences exist in long-term survival prognosis across different molecular subtypes [ 6 – 8 ]. For instance, Cortazar et al.'s study found that triple-negative breast cancer patients, even when achieving pCR, might still have inferior long-term prognosis compared to hormone receptor-positive subtypes [ 5]. The variations in patient longevity outcomes highlight the intricate nature of this condition and indicate that traditional single-source information approaches might not adequately capture the wide spectrum of breast cancer manifestations, calling for an integrated assessment that incorporates multiple types of complementary data sources. However, these single-dimensional evaluation approaches struggle to comprehensively reflect disease biological characteristics and prognostic risks. Comprehensive evaluation of temporal morphological alterations in conjunction with treatment-related pathological changes yields clinically actionable knowledge of tumor behavior patterns. Accordingly, synergistic multimodal analysis is fundamentally required to exhaustively map disease heterogeneity and engineer individualized therapeutic approaches. In recent years, Artificial Intelligence (AI) technology, especially deep learning algorithms, has achieved significant breakthroughs in medical applications[ 9 – 11 ]. AI technology possesses unique advantages in processing multi-dimensional, heterogeneous data, enabling the extraction of deep features from multi-source data including clinical, radiological, and pathological information, and identifying complex biological patterns. In the field of breast cancer, existing research has demonstrated that AI technology shows promising application prospects in radiological diagnosis and prognosis prediction [ 12 , 13 ]. Nevertheless, the application of machine learning in predicting therapeutic responses remains nascent, particularly for intricate pathologies such as breast cancer, a substantially more complex undertaking compared to diagnostic classification. [ 14 – 16 ]. Assessment tools for disease identification generally depend on extensive collections of both healthy and pathological examples. In contrast, forecasting therapeutic outcomes demands more specialized information sets tailored to specific progression phases and capable of capturing the nuanced impacts of various treatment protocols across time periods. In clinical practice, incomplete imaging records represent a common yet understudied limitation that may substantially compromise model performance and subsequent clinical applicability. This study focuses on several key scientific questions: First, the non-linear relationship between tumor size changes and prognosis requires more precise quantitative assessment. Second, while pCR is generally considered a favorable prognostic indicator, there remains a need to identify pCR patients with high recurrence risk. Additionally, the treatment response grading information provided by the Miller-Payne grading system may contain important prognostic value requiring depth analysis [ 17 , 18 ]. Based on these considerations, we propose constructing a comprehensive prediction model based on deep learning, integrating clinical features, treatment response indicators, and molecular subtype information before and after NAC to more accurately assess patient prognostic risk. This model not only focuses on studying the correlation between NAC efficacy and prognosis, more importantly, identifies high-risk populations that traditional assessment systems might overlook, providing more precise evidence for clinical decision-making. This research not only helps optimize existing prognostic assessment systems but also provides new insights for developing individualized treatment plans and adjusting follow-up strategies for breast cancer patients, holding significant clinical translational value. Materials and Methods Study Population and Design This retrospective cohort study included breast cancer patients who received NAC treatment at the Second Affiliated Hospital of Zhejiang University School of Medicine from January 1, 2013, to December 31, 2022. Inclusion criteria: (1) age ≥ 18 years; (2) invasive breast cancer confirmed by core needle biopsy; (3) completed standard NAC regimen; (4) complete clinical pathological data and follow-up information. Exclusion criteria: (1) distant metastasis at initial diagnosis; (2) history of malignancy; (3) incomplete planned NAC regimen; (4) pregnancy-associated breast cancer; (5) male breast cancer patients. This study was approved by the Ethics Committee of the Second Affiliated Hospital of Zhejiang University School of Medicine (approval number: IRB-2021-930). Informed consent was waived due to the retrospective nature of the study. Treatment Protocol Treatment Protocol All enrolled patients received standard NAC regimens based on taxanes combined with anthracyclines or platinum agents. HER2-positive patients received concurrent trastuzumab or trastuzumab plus pertuzumab targeted therapy. Hormone receptor-positive patients began endocrine therapy after completing chemotherapy. Following NAC, surgical treatment included breast-conserving surgery or unilateral mastectomy, with Sentinel Lymph Node Biopsy (SLNB) and/or Axillary Lymph Node Dissection (ALND) based on clinical staging. Clinicopathological Assessment Tissue Specimens were obtained through ultrasound-guided core needle biopsy, and molecular markers were detected using Immunohistochemistry (IHC). Hormone Receptor (HR) positivity was defined as ≥ 1% nuclear staining for Estrogen Receptor (ER) or Progesterone Receptor (PR). HER2 status assessment followed ASCO/CAP guidelines: IHC 0/1 + as negative, 3 + as positive, and 2 + requiring Fluorescence In Situ Hybridization (FISH) confirmation. Ki-67 proliferation index used 20% as the threshold for high/low expression. Patients were categorized into HR+/HER2−, HER2+, and triple-negative subtypes based on receptor status. Response Evaluation and Follow-up Treatment response assessment included both radiological and pathological dimensions. Radiological assessment used RECIST 1.1 criteria, recording tumor maximum diameter changes before and after NAC. Pathological assessment employed the Miller-Payne grading system (Grade 1–5) to evaluate treatment response. Pathological Complete Response (pCR) was defined as no residual invasive cancer in breast and axillary nodes (ypT0/is ypN0). Follow-up endpoints included Disease-free Survival (DFS), defined as the interval from surgery date to first recurrence/metastasis or last follow-up. Deep Learning Model Construction The Multi-layer Perceptron (MLP) model was constructed using Python 3.8 and PyTorch 1.9 framework. Data preprocessing included feature standardization and missing value imputation. The model architecture adopted a three-layer structure (input-hidden-output layers), with ReLU activation function in hidden layers and Softmax function in the output layer. Model training used Adam optimizer with cross-entropy loss function and early stopping strategy to prevent overfitting. Hyperparameters were optimized through grid search, including learning rate, hidden layer neurons, and batch size [ 19 , 20 ]. Statistical Analysis Model performance was evaluated using the Area Under the Receiver Operating Characteristic (ROC) curve (AUC) with 95% confidence intervals (CI), sensitivity, specificity, positive predictive values, negative predictive values, and F1 score. Prognostic factors were identified through the Multi-layer Perceptron (MLP) risk model. The optimal cut-off values for radiomics scores were determined by maximizing the Youden index in the primary cohort, which were then applied to the validation cohort. The DeLong test was used to compare AUCs among different radiomics scores. All statistical tests were two-sided, with P < 0.05 considered statistically significant. Results Patient Characteristics Initially, 904 breast cancer patients who received NAC were included in this study. After excluding 72 cases with incomplete clinical data or lost to follow-up, 832 patients were ultimately included for analysis. Among them, 208 patients (25.0%) achieved pCR, while 624 patients (75.0%) had residual invasive lesions and/or axillary lymph node metastases. All enrolled patients received standard NAC regimens based on anthracyclines and/or taxanes. Regarding surgical approaches, 193 patients (23.2%) underwent breast-conserving surgery, while 639 patients (76.8%) underwent modified radical mastectomy. The median follow-up time was 60.2 months, during which 135 cases (16.2%) developed recurrence or metastasis, with a 5-year DFS rate of 83.8%. Detailed baseline clinicopathological characteristics of patients are shown in Table 1 . Table 1 Baseline Clinicopathological Characteristics [n(%)] Characteristics Non-pCR PCR p -value Age (year) NS Median 51 52 Cancer subtype p < 0.05 HR+/HER2- 307 20 HER2+ 106 32 TN 211 156 Clinical T stage, n (%) NS cT1-2 5 3 cT3-4 616 208 Clinical stage p < 0.05 Stage I 10 2 Stage II 369 161 Stage III 245 45 Ki-67 status (≥ 20%) p < 0.05 Positive 497 188 Negative 126 21 Regarding molecular subtypes, the non-pCR group comprised 307 HR+/HER2-, 106 HER2+, and 211 triple-negative cases, while the pCR group included 20, 32, and 156 cases, respectively ( P < 0.05). High Ki-67 expression was observed in 79.8% of the non-pCR group and 90.4% of the pCR group ( P < 0.05). While age distribution showed no statistical difference between groups (median age: 51 vs. 52 years), significant differences were observed in clinical staging, with the non-pCR group having 10, 369, and 245 patients in stages I, II, and III, respectively, compared to 2, 161, and 45 patients in the pCR group ( P < 0.05). Model Performance Evaluation Key predictive variables were initially selected through random forest feature importance analysis combined with clinical expert consensus. These primarily included molecular subtypes, histological grade, tumor size changes, node status, Miller-Payne grade, lymph pathological features, and biological markers. Post-NAC residual tumor size, Ki-67 level, and Miller-Payne grade emerged as the three most significant prognostic factors (Fig. 1). The deep learning model, evaluated through five-fold cross-validation, achieved satisfactory prediction performance in the overall cohort, with an accuracy of 0.86 (95% CI: 0.81–0.92), sensitivity of 0.85 (95% CI: 0.81–0.89). Figure 1. Model performance visualization showing: (A) SHAP value analysis of predictive features, where post-chemotherapy tumor size, Ki-67, and Miller-Payne grade emerged as the top contributing factors (mean SHAP values: 0.16, 0.17, and 0.09, respectively), and (B) ROC demonstrating robust model discrimination with AUC of 0.86 (95% CI: 0.81–0.92; sensitivity: 0.85, 95% CI: 0.81–0.89). AUC = area under the receiver operating. In molecular subtype analysis, the model demonstrated strong predictive capability across different subtypes, albeit with some variations. HER2-positive subtype achieved the best predictive performance with an AUC of 0.86 (95% CI: 0.82–0.93), followed by triple-negative subtype (AUC 0.82, 95% CI: 0.70–0.92) and HR+/HER2 − subtype (AUC 0.76, 95% CI: 0.66–0.82) (Fig. 2). These differences may reflect the distinct biological behaviors of tumors with different molecular subtypes. Figure 2. Receiver operating characteristic (ROC) curves showing the model's predictive performance across different molecular subtypes: (A) HR+/HER2- subtype with AUC of 0.76 (95% CI: 0.66–0.82), (B) HER2-positive subtype demonstrating the best performance with AUC of 0.86 (95% CI: 0.82–0.93), and (C), triple-negative subtype with AUC of 0.82 (95% CI: 0.70–0.93), reflecting the distinct predictive capabilities of the model across different breast cancer molecular subtypes. In surgical approach stratification analysis, the breast-conserving surgery group showed superior prediction accuracy (0.88, 95% CI: 0.79–0.95) compared to the mastectomy group (0.85, 95% CI: 0.70–0.89) (Fig. 3). The model demonstrated high sensitivity (0.88, 95% CI: 0.80–0.96) and good specificity (0.86, 95% CI: 0.78–0.94) in the breast-conserving group, while showing relatively balanced sensitivity (0.78, 95% CI: 0.72–0.83) and specificity (0.75, 95% CI: 0.69–0.81) in the mastectomy group. Figure 3. ROC curves illustrating the model's predictive performance in different surgical groups: (A) mastectomy group with AUC of 0.85 (95% CI: 0.70–0.89), and (B) breast-conserving surgery group with slightly better performance (AUC of 0.88, 95% CI: 0.79–0.95), demonstrating robust predictive capability across both surgical approaches. In pCR status stratification analysis, prediction accuracy was comparable between non-pCR and pCR groups (0.75, 95% CI: 0.62–0.88 vs 0.79, 95% CI: 0.66–0.82) (Fig. 4). Notably, both groups demonstrated high sensitivity, with the pCR group at 0.84 (95% CI: 0.73–0.95) and the non-pCR group at 0.89 (95% CI: 0.83–0.94), though specificity was relatively lower (pCR group 0.63, 95% CI: 0.49–0.77; non-pCR group 0.62, 95% CI: 0.53–0.71). This finding suggests that the model has similar predictive capability in identifying high-risk individuals within both patient categories. Figure 4. ROC curves comparing model performance between different pathological response groups: (A) non-pCR group showing moderate predictive ability with AUC of 0.75 (95% CI: 0.62–0.88), while (B) pCR group demonstrating better discrimination with AUC of 0.79 (95% CI: 0.66–0.82), indicating the model's enhanced capability in risk stratification among patients achieving pathological complete response. Predictive Factors and High-Risk Identification The deep learning model attempted to identify high-risk patient populations by integrating the aforementioned predictive factors. Even among patients who achieved pCR, 45 out of 208 patients (21.6%) experienced recurrence. Based on SHAP value analysis, the contribution of each predictive factor to recurrence and metastasis risk was systematically evaluated. Figure 5. Model performance visualization showing: (A) SHAP value analysis revealing Ki-67 as the strongest predictor (SHAP value: 0.129), followed by molecular typing (0.064) and surgical approach (0.025), and (B) ROC curves demonstrats robust model discrimination with AUC of 0.84 (95% CI: 0.81–0.89), indicating strong predictive capability for patient outcomes. AUC = area under the receiver operating. The results showed that Ki-67 proliferation index was the most valuable predictive factor, followed by molecular subtype, mastectomy or breast conservation, pre-NAC tumor size, age, and stage (Fig. 5). These findings highly correlate with clinical experience while also revealing some new predictive features. Discussion This study demonstrated the accuracy and utility of multi-modal data analysis in predicting response to NAC and prognosis in breast cancer patients. Our comprehensive dataset encompasses diverse medical modalities, integrating pre-treatment and post-neoadjuvant chemotherapy parameters, histopathological markers, molecular classification signatures, and systematically organized clinical documentation. Our deep learning framework exhibits exceptional efficacy in prognosticating both neoadjuvant chemotherapy response and longitudinal clinical outcomes across the spectrum of breast cancer molecular phenotypes. Notably, the robust verification of our model's predictive performance substantiates its prospective utility across heterogeneous clinical environments. In terms of practical implementation, our system's component-based structure and uniform framework enable straightforward expansion to incorporate additional abnormalities, data types, and temporal measurements. It utilizes a comprehensive and rich patient profile to improve the accuracy of treatment response and prognostic predictions. Our model demonstrated good discriminative ability across different molecular subtypes, with its predictive performance being best in HER2-positive subtypes, consistent with previous research reports on the reliability of prognostic prediction in HER2-positive breast cancer. Our model’s superiority stems from its dynamic fusion of multi-modal data, capturing complex interactions between pre-/post-NAC features and molecular subtypes, which a capability unmatched by conventional approaches [ 21 , 22 ]. Our study identified residual tumor size after NAC, Ki-67 expression level, and Miller-Payne grade as the most predictive factors. Notably, we quantified for the first time the importance of tumor size changes in prognostic prediction, a finding that transcends traditional binary pCR assessment methods and provides new reference indicators for prognostic evaluation in non-pCR patients. This echoes the findings of von Minckwitz et al., who also found that the degree of tumor burden reduction correlates positively with survival benefit [ 23 ]. Most existing predictive strategies for breast cancer are based on elementary data fusion techniques—such as concatenation, sequential alignment followed by averaging, or traditional multivariate machine learning. These techniques inadequately capture the distinct and overlapping characteristics inherent in various medical modalities. In contrast, our model dynamically processes varied data types, facilitating a more rational process of multi-modal fusion. Additionally, our model uniquely employs a learnable feature approach to address the issue of missing modalities in real world scenarios, enabling accurate treatment response predictions for patients with partially incomplete data. From a clinical perspective, the model's modular structure and unified architecture enable flexible expansion to support a broader range of lesions, diverse data modalities, and multiple temporal stages. The model dynamic structure can adapt to increasingly complex treatment response prediction tasks, such as incorporating patient information from multiple time points. This standardized framework allows for a novel continual learning paradigm that can be incorporated into the entire medical imaging and reporting workflow. Once new information is generated, our model can be trained without waiting for the complete collection of patient information. This is particularly important for providing personalized treatment, as the model can immediately consider the latest diagnostic information, treatment response, and changes in patient status to enable more accurate treatment predictions and improve patient management. An additional important advantage of our model lies in its future capacity to incorporate multiple modalities, a crucial step toward enabling personalized treatment strategies. Our model consolidated diverse data sources, including pre- and post-NAC features, pathological indicators, and molecular subtypes, and can incorporate even more modalities in the future, such as cancer biomarkers, gene expression, and lifestyle and health history information. Adding such data may significantly improve the model's accuracy in predicting treatment response while also helping physicians formulate more targeted treatment plans. The flexibility of our model makes it an ideal platform for interdisciplinary collaboration, facilitating knowledge fusion between clinicians, and data scientists. Through such collaborations, the model can continuously assimilate the latest research discoveries and clinical feedback to iteratively update and refine its algorithms. Additionally, our framework is highly extensible and can be rapidly expanded to other types of breast cancer and diagnostic markers. This adaptability not only enhances the model's utility across various oncological applications but also supports a more comprehensive approach to personalized medicine. The main advantages of this study include: (1) complete follow-up data enhancing result reliability; (2) advanced deep learning algorithms improving prediction accuracy; (3) SHAP value analysis enhancing model interpretability; (4) successful identification of special high-risk populations providing new insights for clinical practice. Despite the encouraging nature of our preliminary findings, it is essential to critically acknowledge the inherent limitations of this study. Firstly, although the dataset employed was relatively comprehensive, its scope could be further broadened to enhance generalizability. Second, this is a single-center retrospective study design potentially limiting result generalizability. Third, there is a lack of external validation cohorts. Finally, despite performing accurate treatment response predictions, the model still relies on human input for certain sub-tasks. Future research directions should include: (1) conducting multicenter prospective validation studies; (2) integrating multi-omics data to enhance predictive performance; (3) exploring model application in clinical decision support systems; (4) evaluating the model's guidance value for treatment optimization; and (5) further reducing the need for human input and moving closer to fully autonomous end-to-end treatment response prediction systems. In conclusion, the deep learning predictive model developed in this study provides a new tool for prognostic evaluation of breast cancer patients receiving NAC, with its excellent predictive performance and unique ability to identify high-risk populations providing strong support for achieving precision medicine and individualized treatment decisions. The results of this study not only advance the current understanding of breast cancer prognosis prediction but also serve as a meaningful reference for evidence-based clinical decision-making. Declarations Author contributions WK and YL designed the study, developed the deep learning model, and performed data analysis. PZ and BY collected and processed clinical data. YT supervised the project, interpreted the results, and critically revised the manuscript. All authors reviewed and approved the final manuscript. Acknowledgements We would like to express our sincere gratitude to Dr. Jiang Jingxin for his valuable assistance with this paper. Funding This study was conducted without financial support from any external funding agencies. Data availability All data generated or analyzed during this study are included in this published article and its supplementary files. Ethics approval and consent to participate This study was conducted in accordance with the ethical principles of the Declaration of Helsinki and approved by the Institutional Review Board (IRB) of the Second Affiliated Hospital, Zhejiang University School of Medicine (Approval No. IRB-2021-930). Informed consent was waived by the ethics committee due to the retrospective design of this study, which exclusively utilized anonymized clinical data. All procedures involving human participants complied with institutional and national ethical standards for the protection of privacy and confidentiality. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Clinical trial number Not applicable. References Shepherd J H, Ballman K, Polley M C, et al. CALGB 40603 (Alliance): Long-Term Outcomes and Genomic Correlates of Response and Survival After Neoadjuvant Chemotherapy With or Without Carboplatin and Bevacizumab in Triple-Negative Breast Cancer. J Clin Oncol, 2022,40(12):1323-1334. Gu J, Tong T, Xu D, et al. Deep learning radiomics of ultrasonography for comprehensively predicting tumor and axillary lymph node status after neoadjuvant chemotherapy in breast cancer patients: A multicenter study. Cancer, 2023,129(3):356-366. Ogston K N, Miller I D, Payne S, et al. A new histological grading system to assess response of breast cancers to primary chemotherapy: prognostic significance and survival. Breast, 2003,12(5):320-327. Vasseur A, Carton M, Guiu S, et al. Efficacy of taxanes rechallenge in first-line treatment of early metastatic relapse of patients with HER2-negative breast cancer previously treated with a (neo)adjuvant taxanes regimen: A multicentre retrospective observational study. Breast, 2022,65:136-144. Cortazar P, Zhang L, Untch M, et al. Pathological complete response and long-term clinical benefit in breast cancer: the CTNeoBC pooled analysis[J]. Lancet, 2014,384(9938):164-172. von Minckwitz G, Untch M, Blohmer J U, et al. Definition and impact of pathologic complete response on prognosis after neoadjuvant chemotherapy in various intrinsic breast cancer subtypes. J Clin Oncol, 2012,30(15):1796-1804. Mukhtar R A, Chau H, Woriax H, et al. Breast Conservation Surgery and Mastectomy Have Similar Locoregional Recurrence After Neoadjuvant Chemotherapy: Results From 1462 Patients on the Prospective, Randomized I-SPY2 Trial. Ann Surg, 2023,278(3):320-327. Yau C, Osdoit M, van der Noordaa M, et al. Residual cancer burden after neoadjuvant chemotherapy and long-term survival outcomes in breast cancer: a multicentre pooled analysis of 5161 patients. Lancet Oncol, 2022,23(1):149-160. Esteva A, Kuprel B, Novoa R A, et al. Dermatologist-level classification of skin cancer with deep neural networks. Nature, 2017,542(7639):115-118. Skarping I, Larsson M, Fornvik D. Analysis of mammograms using artificial intelligence to predict response to neoadjuvant chemotherapy in breast cancer patients: proof of concept. Eur Radiol, 2022,32(5):3131-3141. Zhou P, Qian H, Zhu P, et al. Machine learning for predicting neoadjuvant chemotherapy effectiveness using ultrasound radiomics features and routine clinical data of patients with breast cancer. Front Oncol, 2024,14:1485681. Padhani A R, Papanikolaou N. AI and human interactions in prostate cancer diagnosis using MRI. Eur Radiol, 2025. Saha A, Bosma J S, Twilt J J, et al. Artificial intelligence and radiologists in prostate cancer detection on MRI (PI-CAI): an international, paired, non-inferiority, confirmatory study. Lancet Oncol, 2024,25(7):879-887. Potnis K C, Ross J S, Aneja S, et al. Artificial Intelligence in Breast Cancer Screening: Evaluation of FDA Device Regulation and Future Recommendations. JAMA Intern Med, 2022,182(12):1306-1312. McKinney S M, Sieniek M, Godbole V, et al. International evaluation of an AI system for breast cancer screening. Nature, 2020,577(7788):89-94. Lowry K P, Zuiderveld C C. Artificial Intelligence for Breast Cancer Risk Assessment. Radiol Clin North Am, 2024,62(4):619-625. Ribelles N, Perez-Villa L, Jerez J M, et al. Pattern of recurrence of early breast cancer is different according to intrinsic subtype and proliferation index. Breast Cancer Res, 2013,15(5):R98. Hammond M E, Hayes D F, Dowsett M, et al. American Society of Clinical Oncology/College Of American Pathologists guideline recommendations for immunohistochemical testing of estrogen and progesterone receptors in breast cancer. J Clin Oncol, 2010,28(16):2784-2795. Quinn K N, Wilber H, Townsend A, et al. Chebyshev Approximation and the Global Geometry of Model Predictions. Phys Rev Lett, 2019,122(15):158302. Lorencin I, Andelic N, Spanjol J, et al. Using multi-layer perceptron with Laplacian edge detector for bladder cancer diagnosis. Artif Intell Med, 2020,102:101746. Prat A, Fan C, Fernandez A, et al. Response and survival of breast cancer intrinsic subtypes following multi-agent neoadjuvant chemotherapy. BMC Med, 2015,13:303. Gan S, Macalinao D G, Shahoei S H, et al. Distinct tumor architectures and microenvironments for the initiation of breast cancer metastasis in the brain. Cancer Cell, 2024,42(10):1693-1712. Cortazar P, Zhang L, Untch M, et al. Pathological complete response and long-term clinical benefit in breast cancer: the CTNeoBC pooled analysis. Lancet, 2014,384(9938):164-172. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6407870","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":462855189,"identity":"38996014-3de3-4b5a-87af-934403fea5fa","order_by":0,"name":"Wang Ke","email":"","orcid":"","institution":"Zhejiang University School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Wang","middleName":"","lastName":"Ke","suffix":""},{"id":462855190,"identity":"b0b6411b-ec09-4a38-bf28-b3088014e3f7","order_by":1,"name":"Luo Yikai","email":"","orcid":"","institution":"Zhejiang University","correspondingAuthor":false,"prefix":"","firstName":"Luo","middleName":"","lastName":"Yikai","suffix":""},{"id":462855191,"identity":"7d171ff7-01d4-4557-80d1-76ddda8b0441","order_by":2,"name":"Zhang Peng","email":"","orcid":"","institution":"Zhejiang University","correspondingAuthor":false,"prefix":"","firstName":"Zhang","middleName":"","lastName":"Peng","suffix":""},{"id":462855192,"identity":"cd98db29-8baf-4e5a-a0bb-b10d48a09158","order_by":3,"name":"Yang Bing","email":"","orcid":"","institution":"Zhejiang University School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Yang","middleName":"","lastName":"Bing","suffix":""},{"id":462855193,"identity":"264cd646-230f-4ce3-8712-098369aa20ab","order_by":4,"name":"Tao Yubo","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA70lEQVRIiWNgGAWjYLACxgYGOSBlwMDAA+Im4FfNA9ViTLqWxAawFgYitNizHz4mzbvDJr2//fAGxh8yhxn42XMMGH7uwGMLT1qaNO+ZtNwZZ9IKGCR4DjNI9rwxYOw9g89hOWbSvG2HczcwAA03AGoxuJFjwMzYhkcL/xuwlnQD/jcGDAlALfYEtUhAbEkwkADacgBkiwQhLTeeJVvOPZNmOOPGs4KDDTzpPBJngIxePFrY+5MP3ni7w0aevz9548OfPdZy/O3JGx/8xKMFCFgkYKwDjD2QiDqAVwMDA/MHBPsHAbWjYBSMglEwIgEAA9pLEiexWd0AAAAASUVORK5CYII=","orcid":"","institution":"Zhejiang University","correspondingAuthor":true,"prefix":"","firstName":"Tao","middleName":"","lastName":"Yubo","suffix":""}],"badges":[],"createdAt":"2025-04-09 04:23:20","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6407870/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6407870/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":83648335,"identity":"6af35b02-78db-46ba-b453-084b60f60477","added_by":"auto","created_at":"2025-05-30 06:21:31","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":418664,"visible":true,"origin":"","legend":"\u003cp\u003eModel performance visualization showing: (A) SHAP value analysis of predictive features, where post-chemotherapy tumor size, Ki-67, and Miller-Payne grade emerged as the top contributing factors (mean SHAP values: 0.16, 0.17, and 0.09, respectively), and (B) ROC demonstrating robust model discrimination with AUC of 0.86 (95% CI: 0.81-0.92; sensitivity: 0.85, 95% CI: 0.81-0.89). AUC=area under the receiver operating.\u003c/p\u003e","description":"","filename":"Fig.1.png","url":"https://assets-eu.researchsquare.com/files/rs-6407870/v1/a793ea73f9eb86c5cb0f3087.png"},{"id":83648339,"identity":"890d08f3-9065-47e5-9369-e16e4d061bd3","added_by":"auto","created_at":"2025-05-30 06:21:31","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":454432,"visible":true,"origin":"","legend":"\u003cp\u003eReceiver operating characteristic (ROC) curves showing the model's predictive performance across different molecular subtypes: (A) HR+/HER2- subtype with AUC of 0.76 (95% CI: 0.66-0.82), (B) HER2-positive subtype demonstrating the best performance with AUC of 0.86 (95% CI: 0.82-0.93), and (C), triple-negative subtype with AUC of 0.82 (95% CI: 0.70-0.93), reflecting the distinct predictive capabilities of the model across different breast cancer molecular subtypes.\u003c/p\u003e","description":"","filename":"Fig.2.png","url":"https://assets-eu.researchsquare.com/files/rs-6407870/v1/f09da8c52ce9e4b2be8c4589.png"},{"id":83649026,"identity":"1515709b-380a-402a-ac0d-0259b7a828e4","added_by":"auto","created_at":"2025-05-30 06:29:31","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":298793,"visible":true,"origin":"","legend":"\u003cp\u003eROCcurves illustrating the model's predictive performance in different surgical groups: (A) mastectomy group with AUC of 0.85 (95% CI: 0.70-0.89), and (B) breast-conserving surgery group with slightly better performance (AUC of 0.88, 95% CI: 0.79-0.95), demonstrating robust predictive capability across both surgical approaches.\u003c/p\u003e","description":"","filename":"Fig.3.png","url":"https://assets-eu.researchsquare.com/files/rs-6407870/v1/4c4fd9430e77ff98d38b4df5.png"},{"id":83648341,"identity":"f5512bb7-cec6-4f08-afe6-ca534faf7696","added_by":"auto","created_at":"2025-05-30 06:21:31","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":320069,"visible":true,"origin":"","legend":"\u003cp\u003eROC curves comparing model performance between different pathological response groups: (A) non-pCR group showing moderate predictive ability with AUC of 0.75 (95% CI: 0.62-0.88), while (B) pCR group demonstrating better discrimination with AUC of 0.79 (95% CI: 0.66-0.82), indicating the model's enhanced capability in risk stratification among patients achieving pathological complete response.\u003c/p\u003e","description":"","filename":"Fig.4.png","url":"https://assets-eu.researchsquare.com/files/rs-6407870/v1/1ad9bfefa1b810efb625ebe8.png"},{"id":83648343,"identity":"a2352dd0-77c8-4cf4-bc34-6132f720ba3c","added_by":"auto","created_at":"2025-05-30 06:21:31","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":334163,"visible":true,"origin":"","legend":"\u003cp\u003eModel performance visualization showing: (A) SHAP value analysis revealing Ki-67 as the strongest predictor (SHAP value: 0.129), followed by molecular typing (0.064) and surgical approach (0.025), and (B) ROC curves demonstrats robust model discrimination with AUC of 0.84 (95% CI: 0.81-0.89), indicating strong predictive capability for patient outcomes. AUC=area under the receiver operating.\u003c/p\u003e","description":"","filename":"Fig.5.png","url":"https://assets-eu.researchsquare.com/files/rs-6407870/v1/ee71ecbbb67bf7afbd7243ec.png"},{"id":86309518,"identity":"77a1221c-6276-4ede-bcca-28a2edc04c11","added_by":"auto","created_at":"2025-07-09 07:55:13","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2276004,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6407870/v1/a3dc0d96-4ce6-426d-b6fe-5e57ab40309d.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Correlation Study between Neoadjuvant Chemotherapy Response and Long-term Prognosis in Breast Cancer Based on Deep Learning Models","fulltext":[{"header":"Background","content":"\u003cp\u003eBreast cancer is the most common malignant tumor among women worldwide, with its incidence and mortality rates showing a continuous upward trend. Neoadjuvant chemotherapy (NAC), as an important component of the standard treatment regimen for locally advanced breast cancer, plays a key role in downstaging surgery, improving breast conservation rates, early control of micrometastases, and evaluating chemotherapy sensitivity [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. However, clinical practice indicates that patients' responses to NAC show significant heterogeneity, ranging from pathological complete response (pCR) to disease progression. This variation in response directly affects patients' long-term prognosis [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eTraditional prognostic assessment methods primarily rely on clinical pathological indicators such as TNM staging and molecular subtypes. Meta-analyses show that even among patients achieving pCR, significant differences exist in long-term survival prognosis across different molecular subtypes [\u003cspan additionalcitationids=\"CR7\" citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. For instance, Cortazar et al.'s study found that triple-negative breast cancer patients, even when achieving pCR, might still have inferior long-term prognosis compared to hormone receptor-positive subtypes [ 5]. The variations in patient longevity outcomes highlight the intricate nature of this condition and indicate that traditional single-source information approaches might not adequately capture the wide spectrum of breast cancer manifestations, calling for an integrated assessment that incorporates multiple types of complementary data sources.\u003c/p\u003e \u003cp\u003eHowever, these single-dimensional evaluation approaches struggle to comprehensively reflect disease biological characteristics and prognostic risks. Comprehensive evaluation of temporal morphological alterations in conjunction with treatment-related pathological changes yields clinically actionable knowledge of tumor behavior patterns. Accordingly, synergistic multimodal analysis is fundamentally required to exhaustively map disease heterogeneity and engineer individualized therapeutic approaches.\u003c/p\u003e \u003cp\u003eIn recent years, Artificial Intelligence (AI) technology, especially deep learning algorithms, has achieved significant breakthroughs in medical applications[\u003cspan additionalcitationids=\"CR10\" citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. AI technology possesses unique advantages in processing multi-dimensional, heterogeneous data, enabling the extraction of deep features from multi-source data including clinical, radiological, and pathological information, and identifying complex biological patterns. In the field of breast cancer, existing research has demonstrated that AI technology shows promising application prospects in radiological diagnosis and prognosis prediction [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. Nevertheless, the application of machine learning in predicting therapeutic responses remains nascent, particularly for intricate pathologies such as breast cancer, a substantially more complex undertaking compared to diagnostic classification. [\u003cspan additionalcitationids=\"CR15\" citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eAssessment tools for disease identification generally depend on extensive collections of both healthy and pathological examples. In contrast, forecasting therapeutic outcomes demands more specialized information sets tailored to specific progression phases and capable of capturing the nuanced impacts of various treatment protocols across time periods. In clinical practice, incomplete imaging records represent a common yet understudied limitation that may substantially compromise model performance and subsequent clinical applicability.\u003c/p\u003e \u003cp\u003eThis study focuses on several key scientific questions: First, the non-linear relationship between tumor size changes and prognosis requires more precise quantitative assessment. Second, while pCR is generally considered a favorable prognostic indicator, there remains a need to identify pCR patients with high recurrence risk. Additionally, the treatment response grading information provided by the Miller-Payne grading system may contain important prognostic value requiring depth analysis [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eBased on these considerations, we propose constructing a comprehensive prediction model based on deep learning, integrating clinical features, treatment response indicators, and molecular subtype information before and after NAC to more accurately assess patient prognostic risk. This model not only focuses on studying the correlation between NAC efficacy and prognosis, more importantly, identifies high-risk populations that traditional assessment systems might overlook, providing more precise evidence for clinical decision-making. This research not only helps optimize existing prognostic assessment systems but also provides new insights for developing individualized treatment plans and adjusting follow-up strategies for breast cancer patients, holding significant clinical translational value.\u003c/p\u003e"},{"header":"Materials and Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eStudy Population and Design\u003c/h2\u003e \u003cp\u003eThis retrospective cohort study included breast cancer patients who received NAC treatment at the Second Affiliated Hospital of Zhejiang University School of Medicine from January 1, 2013, to December 31, 2022. Inclusion criteria: (1) age\u0026thinsp;\u0026ge;\u0026thinsp;18 years; (2) invasive breast cancer confirmed by core needle biopsy; (3) completed standard NAC regimen; (4) complete clinical pathological data and follow-up information. Exclusion criteria: (1) distant metastasis at initial diagnosis; (2) history of malignancy; (3) incomplete planned NAC regimen; (4) pregnancy-associated breast cancer; (5) male breast cancer patients. This study was approved by the Ethics Committee of the Second Affiliated Hospital of Zhejiang University School of Medicine (approval number: IRB-2021-930). Informed consent was waived due to the retrospective nature of the study.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eTreatment Protocol\u003c/h3\u003e\n\u003cdiv class=\"Heading\"\u003eTreatment Protocol\u003c/div\u003e \u003cp\u003eAll enrolled patients received standard NAC regimens based on taxanes combined with anthracyclines or platinum agents. HER2-positive patients received concurrent trastuzumab or trastuzumab plus pertuzumab targeted therapy. Hormone receptor-positive patients began endocrine therapy after completing chemotherapy. Following NAC, surgical treatment included breast-conserving surgery or unilateral mastectomy, with Sentinel Lymph Node Biopsy (SLNB) and/or Axillary Lymph Node Dissection (ALND) based on clinical staging.\u003c/p\u003e\n\u003ch3\u003eClinicopathological Assessment Tissue\u003c/h3\u003e\n\u003cp\u003eSpecimens were obtained through ultrasound-guided core needle biopsy, and molecular markers were detected using Immunohistochemistry (IHC). Hormone Receptor (HR) positivity was defined as \u0026ge;\u0026thinsp;1% nuclear staining for Estrogen Receptor (ER) or Progesterone Receptor (PR). HER2 status assessment followed ASCO/CAP guidelines: IHC 0/1\u0026thinsp;+\u0026thinsp;as negative, 3\u0026thinsp;+\u0026thinsp;as positive, and 2\u0026thinsp;+\u0026thinsp;requiring Fluorescence In Situ Hybridization (FISH) confirmation. Ki-67 proliferation index used 20% as the threshold for high/low expression. Patients were categorized into HR+/HER2\u0026minus;, HER2+, and triple-negative subtypes based on receptor status.\u003c/p\u003e\n\u003ch3\u003eResponse Evaluation and Follow-up\u003c/h3\u003e\n\u003cp\u003eTreatment response assessment included both radiological and pathological dimensions. Radiological assessment used RECIST 1.1 criteria, recording tumor maximum diameter changes before and after NAC. Pathological assessment employed the Miller-Payne grading system (Grade 1\u0026ndash;5) to evaluate treatment response. Pathological Complete Response (pCR) was defined as no residual invasive cancer in breast and axillary nodes (ypT0/is ypN0). Follow-up endpoints included Disease-free Survival (DFS), defined as the interval from surgery date to first recurrence/metastasis or last follow-up.\u003c/p\u003e\n\u003ch3\u003eDeep Learning Model Construction\u003c/h3\u003e\n\u003cp\u003eThe Multi-layer Perceptron (MLP) model was constructed using Python 3.8 and PyTorch 1.9 framework. Data preprocessing included feature standardization and missing value imputation. The model architecture adopted a three-layer structure (input-hidden-output layers), with ReLU activation function in hidden layers and Softmax function in the output layer. Model training used Adam optimizer with cross-entropy loss function and early stopping strategy to prevent overfitting. Hyperparameters were optimized through grid search, including learning rate, hidden layer neurons, and batch size [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e].\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eStatistical Analysis\u003c/h2\u003e \u003cp\u003eModel performance was evaluated using the Area Under the Receiver Operating Characteristic (ROC) curve (AUC) with 95% confidence intervals (CI), sensitivity, specificity, positive predictive values, negative predictive values, and F1 score. Prognostic factors were identified through the Multi-layer Perceptron (MLP) risk model. The optimal cut-off values for radiomics scores were determined by maximizing the Youden index in the primary cohort, which were then applied to the validation cohort. The DeLong test was used to compare AUCs among different radiomics scores. All statistical tests were two-sided, with \u003cem\u003eP\u0026thinsp;\u0026lt;\u003c/em\u003e\u0026thinsp;0.05 considered statistically significant.\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003ePatient Characteristics\u003c/h2\u003e \u003cp\u003eInitially, 904 breast cancer patients who received NAC were included in this study. After excluding 72 cases with incomplete clinical data or lost to follow-up, 832 patients were ultimately included for analysis. Among them, 208 patients (25.0%) achieved pCR, while 624 patients (75.0%) had residual invasive lesions and/or axillary lymph node metastases. All enrolled patients received standard NAC regimens based on anthracyclines and/or taxanes. Regarding surgical approaches, 193 patients (23.2%) underwent breast-conserving surgery, while 639 patients (76.8%) underwent modified radical mastectomy. The median follow-up time was 60.2 months, during which 135 cases (16.2%) developed recurrence or metastasis, with a 5-year DFS rate of 83.8%. Detailed baseline clinicopathological characteristics of patients are shown in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eBaseline Clinicopathological Characteristics [n(%)]\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colspan=\"4\" nameend=\"c4\" namest=\"c1\"\u003e \u003cp\u003eCharacteristics Non-pCR PCR \u003cem\u003ep\u003c/em\u003e-value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAge (year)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eNS\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMedian\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e51\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCancer subtype\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.05\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHR+/HER2-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e307\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHER2+\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e106\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e32\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e211\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e156\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eClinical T stage, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eNS\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ecT1-2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ecT3-4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e616\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e208\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eClinical stage\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.05\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStage I\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStage II\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e369\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e161\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStage III\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e245\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e45\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eKi-67 status (\u0026ge;\u0026thinsp;20%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.05\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e497\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e188\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNegative\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e126\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e21\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eRegarding molecular subtypes, the non-pCR group comprised 307 HR+/HER2-, 106 HER2+, and 211 triple-negative cases, while the pCR group included 20, 32, and 156 cases, respectively (\u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.05). High Ki-67 expression was observed in 79.8% of the non-pCR group and 90.4% of the pCR group (\u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.05). While age distribution showed no statistical difference between groups (median age: 51 vs. 52 years), significant differences were observed in clinical staging, with the non-pCR group having 10, 369, and 245 patients in stages I, II, and III, respectively, compared to 2, 161, and 45 patients in the pCR group (\u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.05).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eModel Performance Evaluation\u003c/h2\u003e \u003cp\u003eKey predictive variables were initially selected through random forest feature importance analysis combined with clinical expert consensus. These primarily included molecular subtypes, histological grade, tumor size changes, node status, Miller-Payne grade, lymph pathological features, and biological markers. Post-NAC residual tumor size, Ki-67 level, and Miller-Payne grade emerged as the three most significant prognostic factors (Fig.\u0026nbsp;1). The deep learning model, evaluated through five-fold cross-validation, achieved satisfactory prediction performance in the overall cohort, with an accuracy of 0.86 (95% CI: 0.81\u0026ndash;0.92), sensitivity of 0.85 (95% CI: 0.81\u0026ndash;0.89).\u003c/p\u003e \u003cp\u003eFigure 1. Model performance visualization showing: (A) SHAP value analysis of predictive features, where post-chemotherapy tumor size, Ki-67, and Miller-Payne grade emerged as the top contributing factors (mean SHAP values: 0.16, 0.17, and 0.09, respectively), and (B) ROC demonstrating robust model discrimination with AUC of 0.86 (95% CI: 0.81\u0026ndash;0.92; sensitivity: 0.85, 95% CI: 0.81\u0026ndash;0.89). AUC\u0026thinsp;=\u0026thinsp;area under the receiver operating.\u003c/p\u003e \u003cp\u003eIn molecular subtype analysis, the model demonstrated strong predictive capability across different subtypes, albeit with some variations. HER2-positive subtype achieved the best predictive performance with an AUC of 0.86 (95% CI: 0.82\u0026ndash;0.93), followed by triple-negative subtype (AUC 0.82, 95% CI: 0.70\u0026ndash;0.92) and HR+/HER2\u0026thinsp;\u0026minus;\u0026thinsp;subtype (AUC 0.76, 95% CI: 0.66\u0026ndash;0.82) (Fig.\u0026nbsp;2). These differences may reflect the distinct biological behaviors of tumors with different molecular subtypes.\u003c/p\u003e \u003cp\u003eFigure 2. Receiver operating characteristic (ROC) curves showing the model's predictive performance across different molecular subtypes: (A) HR+/HER2- subtype with AUC of 0.76 (95% CI: 0.66\u0026ndash;0.82), (B) HER2-positive subtype demonstrating the best performance with AUC of 0.86 (95% CI: 0.82\u0026ndash;0.93), and (C), triple-negative subtype with AUC of 0.82 (95% CI: 0.70\u0026ndash;0.93), reflecting the distinct predictive capabilities of the model across different breast cancer molecular subtypes.\u003c/p\u003e \u003cp\u003eIn surgical approach stratification analysis, the breast-conserving surgery group showed superior prediction accuracy (0.88, 95% CI: 0.79\u0026ndash;0.95) compared to the mastectomy group (0.85, 95% CI: 0.70\u0026ndash;0.89) (Fig.\u0026nbsp;3). The model demonstrated high sensitivity (0.88, 95% CI: 0.80\u0026ndash;0.96) and good specificity (0.86, 95% CI: 0.78\u0026ndash;0.94) in the breast-conserving group, while showing relatively balanced sensitivity (0.78, 95% CI: 0.72\u0026ndash;0.83) and specificity (0.75, 95% CI: 0.69\u0026ndash;0.81) in the mastectomy group.\u003c/p\u003e \u003cp\u003eFigure 3. ROC curves illustrating the model's predictive performance in different surgical groups: (A) mastectomy group with AUC of 0.85 (95% CI: 0.70\u0026ndash;0.89), and (B) breast-conserving surgery group with slightly better performance (AUC of 0.88, 95% CI: 0.79\u0026ndash;0.95), demonstrating robust predictive capability across both surgical approaches.\u003c/p\u003e \u003cp\u003eIn pCR status stratification analysis, prediction accuracy was comparable between non-pCR and pCR groups (0.75, 95% CI: 0.62\u0026ndash;0.88 vs 0.79, 95% CI: 0.66\u0026ndash;0.82) (Fig.\u0026nbsp;4). Notably, both groups demonstrated high sensitivity, with the pCR group at 0.84 (95% CI: 0.73\u0026ndash;0.95) and the non-pCR group at 0.89 (95% CI: 0.83\u0026ndash;0.94), though specificity was relatively lower (pCR group 0.63, 95% CI: 0.49\u0026ndash;0.77; non-pCR group 0.62, 95% CI: 0.53\u0026ndash;0.71). This finding suggests that the model has similar predictive capability in identifying high-risk individuals within both patient categories.\u003c/p\u003e \u003cp\u003eFigure 4. ROC curves comparing model performance between different pathological response groups: (A) non-pCR group showing moderate predictive ability with AUC of 0.75 (95% CI: 0.62\u0026ndash;0.88), while (B) pCR group demonstrating better discrimination with AUC of 0.79 (95% CI: 0.66\u0026ndash;0.82), indicating the model's enhanced capability in risk stratification among patients achieving pathological complete response.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003ePredictive Factors and High-Risk Identification\u003c/h2\u003e \u003cp\u003eThe deep learning model attempted to identify high-risk patient populations by integrating the aforementioned predictive factors. Even among patients who achieved pCR, 45 out of 208 patients (21.6%) experienced recurrence. Based on SHAP value analysis, the contribution of each predictive factor to recurrence and metastasis risk was systematically evaluated.\u003c/p\u003e \u003cp\u003eFigure 5. Model performance visualization showing: (A) SHAP value analysis revealing Ki-67 as the strongest predictor (SHAP value: 0.129), followed by molecular typing (0.064) and surgical approach (0.025), and (B) ROC curves demonstrats robust model discrimination with AUC of 0.84 (95% CI: 0.81\u0026ndash;0.89), indicating strong predictive capability for patient outcomes. AUC\u0026thinsp;=\u0026thinsp;area under the receiver operating.\u003c/p\u003e \u003cp\u003eThe results showed that Ki-67 proliferation index was the most valuable predictive factor, followed by molecular subtype, mastectomy or breast conservation, pre-NAC tumor size, age, and stage (Fig.\u0026nbsp;5). These findings highly correlate with clinical experience while also revealing some new predictive features.\u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis study demonstrated the accuracy and utility of multi-modal data analysis in predicting response to NAC and prognosis in breast cancer patients. Our comprehensive dataset encompasses diverse medical modalities, integrating pre-treatment and post-neoadjuvant chemotherapy parameters, histopathological markers, molecular classification signatures, and systematically organized clinical documentation. Our deep learning framework exhibits exceptional efficacy in prognosticating both neoadjuvant chemotherapy response and longitudinal clinical outcomes across the spectrum of breast cancer molecular phenotypes. Notably, the robust verification of our model's predictive performance substantiates its prospective utility across heterogeneous clinical environments.\u003c/p\u003e \u003cp\u003eIn terms of practical implementation, our system's component-based structure and uniform framework enable straightforward expansion to incorporate additional abnormalities, data types, and temporal measurements. It utilizes a comprehensive and rich patient profile to improve the accuracy of treatment response and prognostic predictions. Our model demonstrated good discriminative ability across different molecular subtypes, with its predictive performance being best in HER2-positive subtypes, consistent with previous research reports on the reliability of prognostic prediction in HER2-positive breast cancer. Our model\u0026rsquo;s superiority stems from its dynamic fusion of multi-modal data, capturing complex interactions between pre-/post-NAC features and molecular subtypes, which a capability unmatched by conventional approaches [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eOur study identified residual tumor size after NAC, Ki-67 expression level, and Miller-Payne grade as the most predictive factors. Notably, we quantified for the first time the importance of tumor size changes in prognostic prediction, a finding that transcends traditional binary pCR assessment methods and provides new reference indicators for prognostic evaluation in non-pCR patients. This echoes the findings of von Minckwitz et al., who also found that the degree of tumor burden reduction correlates positively with survival benefit [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eMost existing predictive strategies for breast cancer are based on elementary data fusion techniques\u0026mdash;such as concatenation, sequential alignment followed by averaging, or traditional multivariate machine learning. These techniques inadequately capture the distinct and overlapping characteristics inherent in various medical modalities. In contrast, our model dynamically processes varied data types, facilitating a more rational process of multi-modal fusion. Additionally, our model uniquely employs a learnable feature approach to address the issue of missing modalities in real world scenarios, enabling accurate treatment response predictions for patients with partially incomplete data.\u003c/p\u003e \u003cp\u003eFrom a clinical perspective, the model's modular structure and unified architecture enable flexible expansion to support a broader range of lesions, diverse data modalities, and multiple temporal stages. The model dynamic structure can adapt to increasingly complex treatment response prediction tasks, such as incorporating patient information from multiple time points. This standardized framework allows for a novel continual learning paradigm that can be incorporated into the entire medical imaging and reporting workflow. Once new information is generated, our model can be trained without waiting for the complete collection of patient information. This is particularly important for providing personalized treatment, as the model can immediately consider the latest diagnostic information, treatment response, and changes in patient status to enable more accurate treatment predictions and improve patient management.\u003c/p\u003e \u003cp\u003eAn additional important advantage of our model lies in its future capacity to incorporate multiple modalities, a crucial step toward enabling personalized treatment strategies. Our model consolidated diverse data sources, including pre- and post-NAC features, pathological indicators, and molecular subtypes, and can incorporate even more modalities in the future, such as cancer biomarkers, gene expression, and lifestyle and health history information. Adding such data may significantly improve the model's accuracy in predicting treatment response while also helping physicians formulate more targeted treatment plans.\u003c/p\u003e \u003cp\u003eThe flexibility of our model makes it an ideal platform for interdisciplinary collaboration, facilitating knowledge fusion between clinicians, and data scientists. Through such collaborations, the model can continuously assimilate the latest research discoveries and clinical feedback to iteratively update and refine its algorithms. Additionally, our framework is highly extensible and can be rapidly expanded to other types of breast cancer and diagnostic markers. This adaptability not only enhances the model's utility across various oncological applications but also supports a more comprehensive approach to personalized medicine.\u003c/p\u003e \u003cp\u003eThe main advantages of this study include: (1) complete follow-up data enhancing result reliability; (2) advanced deep learning algorithms improving prediction accuracy; (3) SHAP value analysis enhancing model interpretability; (4) successful identification of special high-risk populations providing new insights for clinical practice.\u003c/p\u003e \u003cp\u003eDespite the encouraging nature of our preliminary findings, it is essential to critically acknowledge the inherent limitations of this study. Firstly, although the dataset employed was relatively comprehensive, its scope could be further broadened to enhance generalizability. Second, this is a single-center retrospective study design potentially limiting result generalizability. Third, there is a lack of external validation cohorts. Finally, despite performing accurate treatment response predictions, the model still relies on human input for certain sub-tasks.\u003c/p\u003e \u003cp\u003eFuture research directions should include: (1) conducting multicenter prospective validation studies; (2) integrating multi-omics data to enhance predictive performance; (3) exploring model application in clinical decision support systems; (4) evaluating the model's guidance value for treatment optimization; and (5) further reducing the need for human input and moving closer to fully autonomous end-to-end treatment response prediction systems.\u003c/p\u003e \u003cp\u003eIn conclusion, the deep learning predictive model developed in this study provides a new tool for prognostic evaluation of breast cancer patients receiving NAC, with its excellent predictive performance and unique ability to identify high-risk populations providing strong support for achieving precision medicine and individualized treatment decisions. The results of this study not only advance the current understanding of breast cancer prognosis prediction but also serve as a meaningful reference for evidence-based clinical decision-making.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAuthor contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWK and YL designed the study, developed the deep learning model, and performed data analysis. PZ and BY collected and processed clinical data. YT supervised the project, interpreted the results, and critically revised the manuscript. All authors reviewed and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe would like to express our sincere gratitude to Dr. Jiang Jingxin for his valuable assistance with this paper.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was conducted without financial support from any external funding agencies.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll data generated or analyzed during this study are included in this published article and its supplementary files.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was conducted in accordance with the ethical principles of the Declaration of Helsinki and approved by the Institutional Review Board (IRB) of the Second Affiliated Hospital, Zhejiang University School of Medicine (Approval No. IRB-2021-930). Informed consent was waived by the ethics committee due to the retrospective design of this study, which exclusively utilized anonymized clinical data. All procedures involving human participants complied with institutional and national ethical standards for the protection of privacy and confidentiality.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eClinical trial number\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eShepherd J H, Ballman K, Polley M C, et al. CALGB 40603 (Alliance): Long-Term Outcomes and Genomic Correlates of Response and Survival After Neoadjuvant Chemotherapy With or Without Carboplatin and Bevacizumab in Triple-Negative Breast Cancer. J Clin Oncol, 2022,40(12):1323-1334.\u003c/li\u003e\n\u003cli\u003eGu J, Tong T, Xu D, et al. Deep learning radiomics of ultrasonography for comprehensively predicting tumor and axillary lymph node status after neoadjuvant chemotherapy in breast cancer patients: A multicenter study. Cancer, 2023,129(3):356-366.\u003c/li\u003e\n\u003cli\u003eOgston K N, Miller I D, Payne S, et al. A new histological grading system to assess response of breast cancers to primary chemotherapy: prognostic significance and survival. Breast, 2003,12(5):320-327.\u003c/li\u003e\n\u003cli\u003eVasseur A, Carton M, Guiu S, et al. Efficacy of taxanes rechallenge in first-line treatment of early metastatic relapse of patients with HER2-negative breast cancer previously treated with a (neo)adjuvant taxanes regimen: A multicentre retrospective observational study. Breast, 2022,65:136-144.\u003c/li\u003e\n\u003cli\u003eCortazar P, Zhang L, Untch M, et al. Pathological complete response and long-term clinical benefit in breast cancer: the CTNeoBC pooled analysis[J]. Lancet, 2014,384(9938):164-172.\u003c/li\u003e\n\u003cli\u003evon Minckwitz G, Untch M, Blohmer J U, et al. Definition and impact of pathologic complete response on prognosis after neoadjuvant chemotherapy in various intrinsic breast cancer subtypes. J Clin Oncol, 2012,30(15):1796-1804.\u003c/li\u003e\n\u003cli\u003eMukhtar R A, Chau H, Woriax H, et al. Breast Conservation Surgery and Mastectomy Have Similar Locoregional Recurrence After Neoadjuvant Chemotherapy: Results From 1462 Patients on the Prospective, Randomized I-SPY2 Trial. Ann Surg, 2023,278(3):320-327.\u003c/li\u003e\n\u003cli\u003eYau C, Osdoit M, van der Noordaa M, et al. Residual cancer burden after neoadjuvant chemotherapy and long-term survival outcomes in breast cancer: a multicentre pooled analysis of 5161 patients. Lancet Oncol, 2022,23(1):149-160.\u003c/li\u003e\n\u003cli\u003eEsteva A, Kuprel B, Novoa R A, et al. Dermatologist-level classification of skin cancer with deep neural networks. Nature, 2017,542(7639):115-118.\u003c/li\u003e\n\u003cli\u003eSkarping I, Larsson M, Fornvik D. Analysis of mammograms using artificial intelligence to predict response to neoadjuvant chemotherapy in breast cancer patients: proof of concept. Eur Radiol, 2022,32(5):3131-3141.\u003c/li\u003e\n\u003cli\u003eZhou P, Qian H, Zhu P, et al. Machine learning for predicting neoadjuvant chemotherapy effectiveness using ultrasound radiomics features and routine clinical data of patients with breast cancer. Front Oncol, 2024,14:1485681.\u003c/li\u003e\n\u003cli\u003ePadhani A R, Papanikolaou N. AI and human interactions in prostate cancer diagnosis using MRI. Eur Radiol, 2025.\u003c/li\u003e\n\u003cli\u003eSaha A, Bosma J S, Twilt J J, et al. Artificial intelligence and radiologists in prostate cancer detection on MRI (PI-CAI): an international, paired, non-inferiority, confirmatory study. Lancet Oncol, 2024,25(7):879-887.\u003c/li\u003e\n\u003cli\u003ePotnis K C, Ross J S, Aneja S, et al. Artificial Intelligence in Breast Cancer Screening: Evaluation of FDA Device Regulation and Future Recommendations. JAMA Intern Med, 2022,182(12):1306-1312.\u003c/li\u003e\n\u003cli\u003eMcKinney S M, Sieniek M, Godbole V, et al. International evaluation of an AI system for breast cancer screening. Nature, 2020,577(7788):89-94.\u003c/li\u003e\n\u003cli\u003eLowry K P, Zuiderveld C C. Artificial Intelligence for Breast Cancer Risk Assessment. Radiol Clin North Am, 2024,62(4):619-625.\u003c/li\u003e\n\u003cli\u003eRibelles N, Perez-Villa L, Jerez J M, et al. Pattern of recurrence of early breast cancer is different according to intrinsic subtype and proliferation index. Breast Cancer Res, 2013,15(5):R98.\u003c/li\u003e\n\u003cli\u003eHammond M E, Hayes D F, Dowsett M, et al. American Society of Clinical Oncology/College Of American Pathologists guideline recommendations for immunohistochemical testing of estrogen and progesterone receptors in breast cancer. J Clin Oncol, 2010,28(16):2784-2795.\u003c/li\u003e\n\u003cli\u003eQuinn K N, Wilber H, Townsend A, et al. Chebyshev Approximation and the Global Geometry of Model Predictions. Phys Rev Lett, 2019,122(15):158302.\u003c/li\u003e\n\u003cli\u003eLorencin I, Andelic N, Spanjol J, et al. Using multi-layer perceptron with Laplacian edge detector for bladder cancer diagnosis. Artif Intell Med, 2020,102:101746.\u003c/li\u003e\n\u003cli\u003ePrat A, Fan C, Fernandez A, et al. Response and survival of breast cancer intrinsic subtypes following multi-agent neoadjuvant chemotherapy. BMC Med, 2015,13:303.\u003c/li\u003e\n\u003cli\u003eGan S, Macalinao D G, Shahoei S H, et al. Distinct tumor architectures and microenvironments for the initiation of breast cancer metastasis in the brain. Cancer Cell, 2024,42(10):1693-1712.\u003c/li\u003e\n\u003cli\u003eCortazar P, Zhang L, Untch M, et al. Pathological complete response and long-term clinical benefit in breast cancer: the CTNeoBC pooled analysis. Lancet, 2014,384(9938):164-172.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Breast cancer, Neoadjuvant chemotherapy, Artificial intelligence, Recurrence and metastasis, Prognosis prediction","lastPublishedDoi":"10.21203/rs.3.rs-6407870/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6407870/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground: \u003c/strong\u003eNeoadjuvant chemotherapy (NAC) is a critical component of breast cancer treatment; however, patient responses and long-term prognoses vary significantly. Accurately predicting post-NAC prognosis is essential for guiding individualized treatment plans. This study aims to develop a deep learning-based prediction model to analyze the correlation between NAC efficacy and long-term outcomes in breast cancer patients, providing a new approach to identifying high-risk populations.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eObjective:\u003c/strong\u003e To construct a deep learning model that integrates multi-dimensional clinical and pathological parameters to predict recurrence and metastasis risk in breast cancer patients following NAC, thereby facilitating personalized treatment strategies.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods: \u003c/strong\u003eA retrospective analysis was conducted on 832 breast cancer patients who received NAC at our hospital from 2013 to 2022. Comprehensive clinical, pathological, and molecular subtype data including：pre- and post-NAC tumor characteristics, Ki-67 index, lymph node status, lymphovascular invasion, and Miller-Payne grading were collected. A Multi-layer Perceptron (MLP) based deep learning model was developed, incorporating ensemble learning strategies to integrate multi-modal prediction results. The model’s performance in assessing recurrence and metastasis risk was evaluated across different breast cancer subtypes.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults: \u003c/strong\u003eThe analysis identified key prognostic factors, including tumor size reduction, post-NAC lymph node status, Ki-67 index, lymphovascular invasion, and Miller-Payne grading. The MLP model achieved AUC values of 0.86 (95% CI: 0.82-0.93) for HER2+,0.82 (95% CI: 0.70-0.92)for triple-negative breast cancer, and 0.76 (95% CI: 0.66-0.82) for HR+/HER2−. The model successfully stratified high-risk subgroups with significant differences in prognosis, providing valuable insights for clinical decision-making.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusions: \u003c/strong\u003eThe deep learning-based prediction model developed in this study effectively assesses the prognostic risk of breast cancer patients after NAC. Its clinical application holds potential for optimizing individualized treatment and follow-up strategies, ultimately improving patient outcomes.\u003c/p\u003e","manuscriptTitle":"Correlation Study between Neoadjuvant Chemotherapy Response and Long-term Prognosis in Breast Cancer Based on Deep Learning Models","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-05-30 06:21:26","doi":"10.21203/rs.3.rs-6407870/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"5ba7ce2c-73e3-402b-a56c-89ab41277719","owner":[],"postedDate":"May 30th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-07-09T07:54:50+00:00","versionOfRecord":[],"versionCreatedAt":"2025-05-30 06:21:26","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6407870","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6407870","identity":"rs-6407870","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}