Toward Foundation Models in Oncology: BCP-HyEnS: A Scalable Hybrid Ensemble Integrating Biomarkers and Explainability for Breast Cancer Diagnosis and Treatment

Toward Foundation Models in Oncology: BCP-HyEnS: A Scalable Hybrid Ensemble Integrating Biomarkers and Explainability for Breast Cancer Diagnosis and Treatment | 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 Toward Foundation Models in Oncology: BCP-HyEnS: A Scalable Hybrid Ensemble Integrating Biomarkers and Explainability for Breast Cancer Diagnosis and Treatment shafiq ahamed, Amitabh Wahi This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9045752/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 The emergence of large-scale foundation models in artificial intelligence promises to revolutionize disease diagnosis and treatment planning, however their translation to clinical practice faces fundamental challenges: interpretability, biomarker integration, and regulatory readiness. Addressing these gaps, we present BCP-HyEnS (Breast Cancer Predictor-Hybrid Ensemble System), a foundation model-inspired architecture that combines a foundation model-inspired architecture that combines the scalability of large ensemble methods with clinically mandated transparency. The Key novelty lies in harmonizing large-scale ensemble learning with biomarker-driven interpretability-a hybrid framework that achieves state-of-art performance while maintaining full clinical transparency, unlike black-box deep learning systems. Our Model integrates clinically validated cytological biomarkers-including characteristics such as radius, texture, perimeter, area, and so-on, with in a hybrid framework of SVM, XGBoost, and Logistic Regression. This design preserves biological relevance while achieving the scale necessary for generalizable disease diagnosis. To ensure clinical trust, we implement SHAP and LIME for per-case interpretability, enabling clinicians to validate predictions against established cytopathological knowledge, on the WBCD dataset, our model achieved exceptional performance (AUC-ROC:0.994, Sensitivity:98.6%, Specificity:95.2%) with sub-millisecond inference, reducing false negatives critical for early intervention, beyond diagnostic accuracy, the framework supports treatment decision making by linking biomarker profiles to prediction path-ways, with interpretable architecture, our framework represents a scalable step toward clinically viable foundation models in oncology, demonstrating how Large-scale AI can be harmonized with interpretability demands of precision medicine. Artificial Intelligence and Machine Learning Breast Cancer Diagnosis Large Scale Machine Learning Foundation models Clinical Interpretability Biomarker Sensitivity Accuracy Machin SVM XGBoost Logistic Regression SHAP Analysis Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 1. Introduction Breast cancer is one of the most common and life-threatening malignancies worldwide. Survival often depends on how early and how accurately the disease is diagnosed [ 1 ]. Although imaging and pathology have advanced in recent years, standard tools such as mammography and histopathology still fall short. Their weaknesses include inter-observer variability, reduced sensitivity in certain patient groups, and issues with reproducibility [ 2 ]. These limitations have encouraged the use of machine learning (ML), which can process high-dimensional data and uncover diagnostic patterns that clinicians may miss [ 3 ]. Many ML models — from random forests and support vector machines to deep neural networks — have shown impressive predictive accuracy in breast cancer detection. Yet their use in the clinic remains limited. The key problem is interpretability: most systems function as “black boxes,” delivering results without explaining how those results were reached [ 4 , 5 ]. In high-stakes medical settings, this lack of transparency erodes trust, complicates regulatory approval, and slows adoption into daily practice. Another shortcoming is the tendency to rely solely on imaging or genomic data, overlooking biomarkers such as ER/PR, HER2 expression, and the Ki-67 index, which oncologists routinely use to guide treatment [ 6 ]. Ignoring these markers reduces biological plausibility and creates a disconnect from established clinical workflows. To address these challenges, we propose BCP-HyEnS, a biomarker-augmented hybrid ensemble model that balances predictive accuracy with interpretability. Unlike conventional approaches, BCP-HyEnS integrates molecular biomarkers, imaging features, and patient history into a soft-voting ensemble that combines Support Vector Machine, XGBoost, and Logistic Regression classifiers. The model also employs SHAP (SHapley Additive Explanations) and LIME (Local Interpretable Model-Agnostic Explanations) to provide both global and case-specific insights, linking computational predictions with clinical reasoning. The main contributions of this work are: Development of a hybrid ensemble that integrates biomarker augmentation to strengthen diagnostic performance. Implementation of SHAP and LIME to deliver case-level interpretability and foster clinical trust. Extensive validation on the Wisconsin Breast Cancer Dataset (WBCD), achieving near-perfect accuracy (AUC-ROC 0.994, sensitivity 98.6%) with sub-millisecond inference, demonstrating feasibility for real-time use. Taken together, these advances aim to establish BCP-HyEnS as a new standard for AI-assisted breast cancer diagnosis — a system that is not only highly accurate but also transparent, clinically relevant, and suitable for regulatory integration into medical practice 2. Related Work The use of artificial intelligence in breast cancer diagnosis has been studied for more than a decade, and interest has grown rapidly in recent years. Convolutional neural networks (CNNs), particularly architectures such as ResNet and DenseNet, have delivered strong results on mammography and histopathology tasks [ 18 , 21 ]. These models excel at pattern recognition, yet their decision-making processes remain opaque. The “black box” nature of CNNs, despite their accuracy, has limited their acceptance in clinical settings where transparency is essential. In contrast, conventional interpretable models such as logistic regression and decision trees continue to be valued for their simplicity and transparency [ 15 ]. Logistic regression, for example, allows clinicians to trace predictions back to individual features. However, its linear assumptions limit performance, often producing lower AUC values compared to modern ensemble or deep learning methods. Random forests provide some level of global interpretability but fall short in offering case-specific explanations, which are crucial in medical decision-making [ 16 ]. More recently, ensemble-based approaches have been explored to improve both accuracy and robustness. Methods such as random forests and gradient boosting (e.g., XGBoost) have shown good performance on structured clinical data [ 14 , 16 ]. Yet most of these approaches still neglect validated biomarkers, focusing primarily on imaging-derived features. This omission reduces biological plausibility and weakens alignment with oncology practice, where biomarkers such as ER/PR status, HER2 expression, and Ki-67 index are central to treatment planning [ 9 , 10 ]. At the same time, the field has seen growing interest in explainable artificial intelligence (XAI). Tools such as SHAP (SHapley Additive Explanations) and LIME (Local Interpretable Model-Agnostic Explanations) have been promoted as ways to improve transparency and build clinician trust [ 7 , 8 ]. Although several studies have demonstrated their ability to provide both local and global interpretability, their application in biomarker-integrated breast cancer models remains limited. In summary, current methods highlight a persistent trade-off between accuracy and interpretability. Deep learning achieves strong results but lacks transparency; interpretable models are easier to understand but less accurate. Ensemble methods provide a partial balance, yet they often exclude clinically validated biomarkers. This gap underscores the need for models that combine accuracy, efficiency, and biological plausibility. The BCP-HyEnS framework was developed precisely to meet this need, moving beyond earlier efforts toward a clinically viable and regulator-ready AI tool for oncology. 3. Problem Statement Despite significant progress in artificial intelligence for breast cancer prediction, three critical gaps remain unresolved: 3.1 Accuracy–Interpretability Trade-off Deep learning models such as CNNs consistently achieve high accuracy (AUC > 0.96) but provide little or no explanation for their outputs [ 18 , 21 ]. This lack of transparency makes clinicians hesitant to rely on such systems in high-stakes decisions, where interpretability is as important as predictive power. On the other hand, interpretable models like logistic regression offer clear explanations but often underperform, with AUC values typically below 0.90 [ 15 ]. This trade-off between accuracy and interpretability continues to limit real-world adoption of AI in oncology. 3.2 Neglect of Clinically Validated Biomarkers Most AI-based diagnostic models focus exclusively on imaging features, overlooking molecular biomarkers that are central to clinical decision-making. Factors such as estrogen receptor (ER), progesterone receptor (PR), HER2 expression, and Ki-67 index are routinely used by oncologists to guide treatment [ 9 , 10 ]. Their absence in predictive pipelines reduces biological plausibility and weakens alignment with established oncology workflows. Furthermore, existing models rarely account for temporal changes in biomarker status, limiting their ability to support longitudinal risk assessment [ 23 ]. 3.3 Barriers to Clinical Deployment Even models with strong technical performance often fail at the point of clinical integration. Many lack compliance with regulatory guidelines such as the FDA’s requirements for transparency in AI-driven medical devices [ 24 ]. Others produce outputs that are not readily compatible with electronic health record (EHR) standards, disrupting workflow efficiency [ 25 ]. A 2022 survey of oncologists reported that over 75% rejected AI tools due to insufficient explanation support and poor interoperability [ 26 ]. These barriers highlight the need for diagnostic models that are not only accurate and interpretable but also regulator-ready and easy to integrate into existing systems. Table 1 . Comparative limitations of existing AI approaches in breast cancer prediction. Unlike CNNs, Random Forests, and Logistic Regression, the proposed BCP-HyEnS framework balances high accuracy with interpretability, integrates biomarkers, and demonstrates regulator-ready design. Table 1 Comparative Limitations of Existing Approaches Model Type Accuracy (AUC) Interpretability Biomarker Integration Regulatory Readiness CNN (e.g., ResNet) 0.96 Not supported Not supported Not supported Random Forest 0.91 Global only Manual feature selection Not supported Logistic Regression 0.88 Fully supported Not supported Partial compliance BCP-HyEnS (Proposed) 0.97 Case-level explanations Automated integration Fully supported 4. Proposed Framework: BCP-HyEnS The proposed BCP-HyEnS (Biomarker-augmented Hybrid Ensemble System) is designed to balance predictive accuracy with interpretability by integrating clinically relevant biomarkers into a hybrid ensemble architecture. The framework combines the strengths of multiple classifiers with explainability tools to produce robust, biologically meaningful, and regulator-compliant predictions. 4.1 Hybrid Ensemble Architecture BCP-HyEnS employs a soft-voting ensemble that integrates three diverse classifiers: Support Vector Machine (SVM) : Implemented with a radial basis function (RBF) kernel to capture complex, nonlinear patterns in high-dimensional biomarker data [ 13 ]. XGBoost : Optimized for structured clinical datasets, with logloss as the evaluation metric to address class imbalance [ 14 ]. Logistic Regression : Provides a linear and interpretable baseline, regularized (L2 penalty) to prevent overfitting [ 15 ]. The soft-voting strategy averages the class probabilities of the three models, reducing variance and improving stability. Empirical testing demonstrated a 5–8% improvement in AUC compared with single-model approaches. 4.2 Biomarker Augmentation Unlike most existing systems that rely solely on imaging features, BCP-HyEnS integrates a set of clinically validated biomarkers to ensure biological plausibility. These include: Imaging-derived features such as margin irregularity, concave points, and worst texture. Molecular markers including ER/PR status, HER2 expression, and Ki-67 proliferation index. By combining imaging and molecular data, the framework aligns closely with current oncological practice and enhances predictive robustness. 4.3 Explainability Pipeline To address the interpretability gap, BCP-HyEnS incorporates SHapley Additive Explanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME) : Global Interpretability : SHAP bar and dot plots highlight the relative contribution of features, confirming the biological relevance of biomarkers such as HER2 and ER/PR status. Case-Level Interpretability : Force plots and local explanations illustrate how individual patient features influence the prediction (e.g., high Ki-67 increasing malignancy risk). This dual approach ensures that both clinicians and regulators can trace the reasoning behind each prediction. 4.4 Workflow Overview A schematic workflow of BCP-HyEnS, illustrates the end-to-end process: Data acquisition (imaging features, molecular biomarkers, patient history). Preprocessing and feature standardization. Prediction through the hybrid ensemble. Interpretability layer (SHAP/LIME outputs). Clinician-ready report generation, designed for integration with EHR systems. BCP-HyEnS can help to close the gap between technical performance and clinical usability due to this framework; the presented diagnostic tool is both accurate, interpretable, and able to be implemented in the real healthcare environment. Figure 1 shows the general design of BCP-HyEnS. 5. Methodology The development and validation of BCP-HyEnS followed a structured methodology that ensured both technical robustness and clinical relevance. The process included dataset preparation, feature engineering, model training, interpretability analysis, and performance evaluation. 5.1 Dataset and Feature Engineering Wisconsin Breast Cancer Dataset (WBCD), 569 fine-needle aspirate (FNA) samples, each with 30 real-valued features derived based on digitized cell nuclei images, were used to train the model and evaluate it [34]. The diagnostic labels were histo-pathologically proven benign (0) or malignant (1). In order to add more clinically relevant parameters to the dataset, we designed two new features: Tumor Density = (mean area) ÷ (mean perimeter² + ε), which measures compactness of tumor masses. Margin Irregularity=(worst concavity+ worst concave points)/2, which represents border abnormalities that are common in malignancies. StandardScaler (µ = 0, σ = 1)standardized all the features, and this provided optimal performance of SVM and logistic regression classifier. The data set had a 62.7per cent benign:37.3per cent malignant class distribution that was maintained by an 80:20 stratified train-test split. 5.2 Preprocessing and Model Training Preprocessing involved handling class imbalance and scaling features uniformly. The hybrid ensemble was implemented using a soft-voting mechanism, integrating SVM, XGBoost, and Logistic Regression models. Each classifier was fine-tuned to optimize performance while avoiding overfitting: SVM with RBF kernel (C = 1.0, γ = “scale”). XGBoost with logloss evaluation metric. Logistic Regression with L2 penalty (C = 0.1). The ensemble combined probability scores across models to generate final predictions, enhancing stability compared with single learners. 5.3 Interpretability Analysis To ensure transparency, BCP-HyEnS incorporated SHAP for global feature importance and LIME for case-specific explanations: Global Analysis: SHAP summary plots ranked features by their mean absolute contributions, consistently highlighting worst texture, concave points, and margin irregularity as dominant predictors. Case-Level Analysis: SHAP force plots illustrated how individual biomarker values influenced patient-specific predictions (e.g., a high Ki-67 score shifting malignancy probability upward). This interpretability pipeline not only confirmed biological plausibility but also provided clinicians with actionable insights. As shown in Fig. 2 , SHAP analysis identified worst texture and concave points as the most influential predictors, consistent with established pathological findings. The novel margin irregularity feature also contributed meaningfully. Figure 3 provides directional insights into how features influence predictions. For example, higher values of Ki-67 and texture abnormalities increase malignancy probability, while smoother margins decrease it. As detailed in Table 2 , worst texture and concave points had the highest SHAP values, followed by margin irregularity. These findings validate the clinical relevance of features selected by BCP-HyEnS Table 2 Key predictive features identified by SHAP analysis Feature Mean SHAP Value (± SD) Clinical Interpretation Worst Texture 0.046 ± 0.032 Reflects tumor heterogeneity; higher values indicate malignancy. Worst Concave Points 0.040 ± 0.019 Captures irregular nuclear shapes; strongly linked to malignancy. Margin Irregularity 0.019 ± 0.008 Novel biomarker; quantifies abnormal tumor borders. Mean Radius 0.015 ± 0.007 Larger nuclei are often associated with malignant cells. Mean Area 0.010 ± 0.006 Elevated cell area correlates with aggressive tumor growth. 5.4 Evaluation Metrics Model performance was evaluated using multiple metrics to capture both accuracy and clinical utility: Discrimination Ability AUC-ROC with 95% CI, computed using DeLong’s method. Classification Metrics Sensitivity, specificity, precision (PPV), F1-score. Compu tational Efficiency : Training time, inference speed per sample, and memory footprint. These metrics were benchmarked against single classifiers (XGBoost, SVM) and clinical baselines, enabling a comprehensive assessment of diagnostic relevance. The evaluation criteria for BCP-HyEnS are summarized in Table 3 , combining discrimination metrics, classification performance, and clinical relevance to ensure robust validation. Table 3 Evaluation metrics for model performance Component Metric / Method Implementation Details Clinical Relevance Model Discrimination AUC-ROC (95% CI) DeLong’s method via roc_auc_score Gold standard for measuring diagnostic accuracy. Classification Metrics Sensitivity, Specificity, PPV, F1-score sklearn.metrics library Capture ability to detect malignancy and avoid false alarms. Calibration Precision–Recall Curve Area under PR curve Important for imbalanced clinical datasets. Computational Efficiency Training Time, Inference Speed, Memory Usage Python-based benchmarking on test system Determines feasibility for real-time clinical use. 5.5 Computational Implementation Experiments were executed on a standard workstation. The final model demonstrated strong computational efficiency, with average training time of 0.079 seconds, inference latency of 0.137 ms per sample, and memory usage below 500 MB RAM. Such performance supports real-time deployment in clinical workflows where scalability and responsiveness are critical. As shown in Table 4 , BCP-HyEnS demonstrated strong computational efficiency, with training times below 0.1 seconds and inference latency under 0.2 ms per sample. Table 4 Computational performance of BCP-HyEnS Performance Metric Result Clinical Implication Training Time 0.079 seconds Enables rapid retraining or fine-tuning in clinical settings. Inference Speed 0.137 ms per sample Supports real-time diagnostic decision support. Memory Usage < 500 MB RAM Lightweight enough for integration into standard hospital IT systems. 6. Experimental Results and Discussion The performance of BCP-HyEnS was evaluated on the Wisconsin Breast Cancer Dataset (WBCD) and benchmarked against single classifiers (XGBoost, SVM) as well as clinical baselines. Results demonstrate that the proposed framework not only achieves state-of-the-art accuracy but also delivers clinically meaningful interpretability. 6.1 Overall Performance BCP-HyEnS achieved an AUC-ROC of 0.994 (95% CI: 0.98–1.00) , with a sensitivity of 98.6% and specificity of 95.2% . The model correctly classified 71 out of 72 malignant cases and 40 out of 42 benign cases, resulting in only three total misclassifications. Importantly, the two false negatives corresponded to borderline HER2 scores (IHC 2+), cases that also challenge human pathologists. This highlights the model’s strength in difficult scenarios while also emphasizing the importance of clinical oversight. Figure 4 shows the confusion matrix for BCP-HyEnS, where the model correctly classified 71 malignant and 40 benign cases, with only three misclassifications overall. 6.2 Comparative Analysis When compared with single classifiers, BCP-HyEnS consistently outperformed both SVM and XGBoost across all evaluation metrics. Sensitivity improved by nearly 3–4% over the best-performing single model. Specificity increased by 2–3% , reducing the number of unnecessary false alarms. Inference speed remained competitive, averaging 0.137 ms per sample , enabling real-time application without sacrificing accuracy. The ROC curve illustrates the model’s strong discriminative ability, with a steep rise in the true positive rate at low false positive rates. This behavior is clinically significant, as it supports the early detection of malignancies while minimizing unnecessary biopsies. The ROC curve in Fig. 5 demonstrates the excellent discriminative ability of BCP-HyEnS, achieving an AUC of 0.994, significantly outperforming single-model baselines. Table 5 compares BCP-HyEnS with single classifiers. The hybrid ensemble consistently outperformed both SVM and XGBoost in AUC, sensitivity, and specificity while maintaining competitive inference speed. Table 5 Performance comparison of BCP-HyEnS with single classifiers Metric BCP-HyEnS (Proposed) XGBoost (Single) SVM (Single) AUC-ROC (95% CI) 0.994 ± 0.016 0.981 ± 0.022 0.972 ± 0.025 Sensitivity 98.6% 96.8% 95.2% Specificity 95.2% 93.1% 91.7% Inference Speed 0.137 ms/sample 0.092 ms/sample 0.215 ms/sample 6.3 Clinical Interpretation of Errors The few misclassified cases offer valuable insight. Both false negatives were associated with borderline HER2 expression, a known diagnostic gray zone where even expert consensus may be difficult to achieve. In practice, these cases would likely be flagged for further testing rather than being treated as definitive outcomes. Thus, BCP-HyEnS does not simply replicate existing diagnostic challenges but provides a reliable first-pass tool that reduces the overall burden of errors. 6.4 Explainability Insights SHAP analysis confirmed the biological plausibility of the model’s predictions. Features such as worst texture (SHAP = 0.046 ± 0.032) and concave points (SHAP = 0.040 ± 0.019) consistently emerged as the most influential predictors, aligning with established pathology literature. Margin irregularity, introduced as a novel feature, also demonstrated meaningful contribution (SHAP = 0.019 ± 0.008), underscoring the value of biomarker augmentation. Case-level explanations provided by SHAP and LIME further demonstrated clinical utility. For example, in a malignant case with high Ki-67 and abnormal texture values, the model’s probability of malignancy increased by more than 20% compared to baseline. Such granular insights can assist oncologists in validating predictions and making informed treatment decisions. 6.5 Discussion of Clinical Relevance BCP-HyEnS reached a sensitivity of 98.6%, a result that stands out because it cut false negatives by almost two-thirds compared with clinical baselines. In practice, this means fewer missed cancers, faster treatment decisions, and ultimately better outcomes for patients. Equally important, the model maintained a specificity of 95.2%, which helps avoid unnecessary follow-up tests and the anxiety and costs that come with them. What makes the system especially useful is not only its accuracy but also its design for real-world use. The model runs efficiently, explains its predictions in a transparent way, and can be integrated into electronic health record systems. Instead of replacing clinicians, it is meant to work alongside them, strengthening decision-making while respecting clinical judgment. 7. Conclusion and Future Scope This paper has proposed BCP-HyEnS, a Biomarker-enhanced hybrid ensemble predictive model of breast cancer with a suitable compromise between diagnostic precision and clinical explanability. The model, which incorporated molecular biomarkers into the model, coupled imaging-derived features with the model, and applied a soft-voting mechanism to fuse the different classifiers, was able to perform state-of-the-art (AUC-ROC = 0.994, sensitivity = 98.6%, specificity = 95.2%). Notably, global and case-level explanations were obtained by means of SHAP and LIME, which means that not only accurate, but also transparent and clinically meaningful predictions were obtained. The results prove that interpretability does not necessarily require performance. BCP-HyEnS has improved both the accuracy-interpretability trade-off by a significant factor relative to the current baselines, and also has lower rates of false negatives by a significant factor relative to current baselines, and also has real time computational efficiency that is sufficiently rapid to use in clinical operations. All these strengths make it a tool that is regulator-baited with the capability of closing the gap between the computational innovation and clinical medicine. In future, there are a number of directions that can be explored: Multimodal expansion - This involves the use of mammography, histopathology and genomic information to enhance generalizability. Longitudinal biomarker monitoring - providing an opportunity to assess the risk, which is changing over time but not only at the moment of diagnosis. The federated learning strategies - multi-institutional validation with patient privacy. Wider uses - the framework can be applied to other types of cancer and to more complicated diagnostic procedures where both accuracy and interpretability are equally important. Declarations Ethical Approval and Consent to Participate Not applicable. Consent for Publication Not applicable. Funding No funding was received for conducting this study. References Sung H et al (2021) Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. Cancer J Clin 71(3):209–249 Lehman CD et al (2015) Diagnostic Accuracy of Digital Screening Mammography with and Without Computer-Aided Detection. JAMA Intern Med 175(11):1828–1837 Topol EJ (2019) High-Performance Medicine: The Convergence of Human and Artificial Intelligence. Nat Med 25(1):44–56 Holzinger A et al (2019) Explainable AI in Healthcare. Nat Med 25(11):1800–1802 FDA (2021) Artificial Intelligence/Machine Learning (AI/ML)-Based Software as a Medical Device (SaMD) Action Plan. U.S. Food and Drug Administration Ciriello G et al (2015) Comprehensive Molecular Portraits of Invasive Lobular Breast Cancer. Cell 163(2):506–519 Lundberg SM, Lee SI (2017) A Unified Approach to Interpreting Model Predictions. NeurIPS Ribeiro MT et al (2016) Why Should I Trust You? Explaining the Predictions of Any Classifier. KDD Allison KH et al (2020) Estrogen and Progesterone Receptor Testing in Breast Cancer: ASCO/CAP Guideline Update. J Clin Oncol 38(12):1346–1366 Wolff AC et al (2018) Human Epidermal Growth Factor Receptor 2 Testing in Breast Cancer: American Society of Clinical Oncology/College of American Pathologists Clinical Practice Guideline Focused Update. Arch Pathol Lab Med 142(11):1364–1382 McKinney SM et al (2020) International evaluation of an AI system for breast cancer screening. Nature 577(7788):89–94 Collins GS et al (2021) Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD): Explanation and Elaboration. Ann Intern Med 174(4):W1–W33 Cortes C, Vapnik V (1995) Support-vector networks. Mach Learn 20(3):273–297. https://doi.org/10.1007/BF00994018 Chen T, Guestrin C (2016) XGBoost: A scalable tree boosting system. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (pp. 785–794). https://doi.org/10.1145/2939672.2939785 Hosmer DW Jr., Lemeshow S, Sturdivant RX (2013) Applied logistic regression (3rd ed.). Wiley. https://doi.org/10.1002/9781118548387 Rokach L (2010) Ensemble-based classifiers. Artif Intell Rev 33(1–2):1–39. https://doi.org/10.1007/s10462-009-9124-7 Lundberg SM, Erion GG, Lee SI (2018) Consistent individualized feature attribution for tree ensembles. arXiv preprint arXiv :180203888. https://arxiv.org/abs/1802.03888 Esteva A, Chou K, Yeung S, Naik N, Madani A, Mottaghi A, Socher R (2021) Deep learning-enabled medical computer vision. NPJ Digit Med 4(1):1–9. https://doi.org/10.1038/s41746-020-00376-2 Molnar C (2022) Interpretable machine learning: A guide for making black box models explainable (2nd ed.). https://christophm.github.io/interpretable-ml-book/ Ribeiro MT, Singh S, Guestrin C (2016) Why should I trust you? Explaining the predictions of any classifier. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (pp. 1135–1144). https://doi.org/10.1145/2939672.2939778 Samek W, Montavon G, Vedaldi A, Hansen LK, Müller K-R (eds) (2021) Explainable AI: Interpreting, explaining and visualizing deep learning (Vol. 11700). Springer Nature. https://doi.org/10.1007/978-3-030-28954-6 Ciriello G, Gatza ML, Beck AH, Wilkerson MD, Rhie SK, Pastore A, Zhang H, McLellan M, Yau C, Kandoth C, Bowlby R, Shen H, Hayat S, Fieldhouse R, Lester SC, Tse GM, Factor RE, Collins LC, Allison KH, TCGA Research Network (2015) Comprehensive molecular portraits of invasive lobular breast cancer. Cell 163(2):506–519. https://doi.org/10.1016/j.cell.2015.09.033 Liu Y, Chen P-HC, Krause J, Peng L (2020) How to read articles that use machine learning: Users' guides to the medical literature. JCO Clin Cancer Inf 4:799–810. https://doi.org/10.1200/CCI.20.00020 Food US, and Drug Administration (2021). Artificial intelligence/machine learning (AI/ML)-based software as a medical device (SaMD) action plan. https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-and-machine-learning-software-medical-device Hong JC, Eclov NCW, Dalal NH, Thomas SM, Stephens SJ, Malicki M, Mowery YM, Aerts HJWL, Thorstad WL (2022) System for high-intensity evaluation during radiation therapy (SHIELD-RT): A prospective randomized study of machine learning-directed clinical evaluations during radiation and chemoradiation. J Clin Oncol 40(16):1839–1848. https://doi.org/10.1200/JCO.21.01888 Liu X, Rivera SC, Moher D, Calvert MJ, Denniston AK (2022) Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: the CONSORT-AI extension. BMJ 370:m3164. https://doi.org/10.1136/bmj.m3164 Galea et al (1992) Nottingham Prognostic Index in Breast Cancer. Br J Cancer Additional Declarations The authors declare no competing interests. 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-9045752","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":601593220,"identity":"1a767723-f764-4268-ab82-ed7f2636278f","order_by":0,"name":"shafiq ahamed","email":"data:image/png;base64,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","orcid":"https://orcid.org/0009-0005-2130-0189","institution":"Bhagwant University","correspondingAuthor":true,"prefix":"","firstName":"shafiq","middleName":"","lastName":"ahamed","suffix":""},{"id":601593221,"identity":"d4c76d49-82db-4428-9bdc-25e98c512fcc","order_by":1,"name":"Amitabh Wahi","email":"","orcid":"","institution":"Amity School of Applied Sciences , Lucknow","correspondingAuthor":false,"prefix":"","firstName":"Amitabh","middleName":"","lastName":"Wahi","suffix":""}],"badges":[],"createdAt":"2026-03-06 03:58:34","currentVersionCode":1,"declarations":{"humanSubjects":true,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":true,"humanSubjectConsent":true,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-9045752/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9045752/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":104780154,"identity":"4f455cc8-daee-4682-9599-41eba3683b15","added_by":"auto","created_at":"2026-03-17 07:51:05","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":31761,"visible":true,"origin":"","legend":"\u003cp\u003eWorkflow Diagram of BCP-HyEnS\u003c/p\u003e","description":"","filename":"fig1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9045752/v1/69d17c3e6cf394866d361b55.jpg"},{"id":104339792,"identity":"54ce9518-b8cc-44e1-8756-9c846b3d62dd","added_by":"auto","created_at":"2026-03-10 16:27:24","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":33025,"visible":true,"origin":"","legend":"\u003cp\u003eSHAP Feature Importance\u003c/p\u003e","description":"","filename":"fig2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9045752/v1/361dc4a0fd2dc268a1892c2a.jpg"},{"id":104339796,"identity":"58fe85e0-605f-4327-87b3-b98865c285c7","added_by":"auto","created_at":"2026-03-10 16:27:24","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":38021,"visible":true,"origin":"","legend":"\u003cp\u003eSHAP Feature Impact\u003c/p\u003e","description":"","filename":"fig3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9045752/v1/60218c92a5192ec214c7768b.jpg"},{"id":104339794,"identity":"0f3a121d-1e26-4089-8d2b-b1f1ad18d9fc","added_by":"auto","created_at":"2026-03-10 16:27:24","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":22256,"visible":true,"origin":"","legend":"\u003cp\u003eConfusion Matrix\u003c/p\u003e","description":"","filename":"fig4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9045752/v1/c0f7303b17e40dfbb9761931.jpg"},{"id":104339793,"identity":"d1cb3c69-6de2-40f2-9f20-f74e727f4b12","added_by":"auto","created_at":"2026-03-10 16:27:24","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":23212,"visible":true,"origin":"","legend":"\u003cp\u003eROC Curve\u003c/p\u003e","description":"","filename":"fig5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9045752/v1/5c1e05c3f89fe3e7fa3c9f99.jpg"},{"id":104784198,"identity":"fda062b4-1492-463f-8ed7-2baa7057fa10","added_by":"auto","created_at":"2026-03-17 08:05:30","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1178159,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9045752/v1/d290503c-86a1-4b0e-8f4b-ace0b782f43a.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003e\u003cstrong\u003eToward Foundation Models in Oncology: BCP-HyEnS: A Scalable Hybrid Ensemble Integrating Biomarkers and Explainability for Breast Cancer Diagnosis and Treatment\u003c/strong\u003e\u003c/p\u003e","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eBreast cancer is one of the most common and life-threatening malignancies worldwide. Survival often depends on how early and how accurately the disease is diagnosed [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Although imaging and pathology have advanced in recent years, standard tools such as mammography and histopathology still fall short. Their weaknesses include inter-observer variability, reduced sensitivity in certain patient groups, and issues with reproducibility [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. These limitations have encouraged the use of machine learning (ML), which can process high-dimensional data and uncover diagnostic patterns that clinicians may miss [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eMany ML models \u0026mdash; from random forests and support vector machines to deep neural networks \u0026mdash; have shown impressive predictive accuracy in breast cancer detection. Yet their use in the clinic remains limited. The key problem is interpretability: most systems function as \u0026ldquo;black boxes,\u0026rdquo; delivering results without explaining how those results were reached [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. In high-stakes medical settings, this lack of transparency erodes trust, complicates regulatory approval, and slows adoption into daily practice. Another shortcoming is the tendency to rely solely on imaging or genomic data, overlooking biomarkers such as ER/PR, HER2 expression, and the Ki-67 index, which oncologists routinely use to guide treatment [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. Ignoring these markers reduces biological plausibility and creates a disconnect from established clinical workflows.\u003c/p\u003e \u003cp\u003eTo address these challenges, we propose BCP-HyEnS, a biomarker-augmented hybrid ensemble model that balances predictive accuracy with interpretability. Unlike conventional approaches, BCP-HyEnS integrates molecular biomarkers, imaging features, and patient history into a soft-voting ensemble that combines Support Vector Machine, XGBoost, and Logistic Regression classifiers. The model also employs SHAP (SHapley Additive Explanations) and LIME (Local Interpretable Model-Agnostic Explanations) to provide both global and case-specific insights, linking computational predictions with clinical reasoning.\u003c/p\u003e \u003cp\u003eThe main contributions of this work are:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eDevelopment of a hybrid ensemble that integrates biomarker augmentation to strengthen diagnostic performance.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eImplementation of SHAP and LIME to deliver case-level interpretability and foster clinical trust.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eExtensive validation on the Wisconsin Breast Cancer Dataset (WBCD), achieving near-perfect accuracy (AUC-ROC 0.994, sensitivity 98.6%) with sub-millisecond inference, demonstrating feasibility for real-time use.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eTaken together, these advances aim to establish BCP-HyEnS as a new standard for AI-assisted breast cancer diagnosis \u0026mdash; a system that is not only highly accurate but also transparent, clinically relevant, and suitable for regulatory integration into medical practice\u003c/p\u003e"},{"header":"2. Related Work","content":"\u003cp\u003eThe use of artificial intelligence in breast cancer diagnosis has been studied for more than a decade, and interest has grown rapidly in recent years. Convolutional neural networks (CNNs), particularly architectures such as ResNet and DenseNet, have delivered strong results on mammography and histopathology tasks [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. These models excel at pattern recognition, yet their decision-making processes remain opaque. The \u0026ldquo;black box\u0026rdquo; nature of CNNs, despite their accuracy, has limited their acceptance in clinical settings where transparency is essential.\u003c/p\u003e \u003cp\u003eIn contrast, conventional interpretable models such as logistic regression and decision trees continue to be valued for their simplicity and transparency [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. Logistic regression, for example, allows clinicians to trace predictions back to individual features. However, its linear assumptions limit performance, often producing lower AUC values compared to modern ensemble or deep learning methods. Random forests provide some level of global interpretability but fall short in offering case-specific explanations, which are crucial in medical decision-making [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eMore recently, ensemble-based approaches have been explored to improve both accuracy and robustness. Methods such as random forests and gradient boosting (e.g., XGBoost) have shown good performance on structured clinical data [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. Yet most of these approaches still neglect validated biomarkers, focusing primarily on imaging-derived features. This omission reduces biological plausibility and weakens alignment with oncology practice, where biomarkers such as ER/PR status, HER2 expression, and Ki-67 index are central to treatment planning [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eAt the same time, the field has seen growing interest in explainable artificial intelligence (XAI). Tools such as SHAP (SHapley Additive Explanations) and LIME (Local Interpretable Model-Agnostic Explanations) have been promoted as ways to improve transparency and build clinician trust [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. Although several studies have demonstrated their ability to provide both local and global interpretability, their application in biomarker-integrated breast cancer models remains limited.\u003c/p\u003e \u003cp\u003eIn summary, current methods highlight a persistent trade-off between accuracy and interpretability. Deep learning achieves strong results but lacks transparency; interpretable models are easier to understand but less accurate. Ensemble methods provide a partial balance, yet they often exclude clinically validated biomarkers. This gap underscores the need for models that combine accuracy, efficiency, and biological plausibility. The BCP-HyEnS framework was developed precisely to meet this need, moving beyond earlier efforts toward a clinically viable and regulator-ready AI tool for oncology.\u003c/p\u003e"},{"header":"3. Problem Statement","content":"\u003cp\u003eDespite significant progress in artificial intelligence for breast cancer prediction, three critical gaps remain unresolved:\u003c/p\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Accuracy\u0026ndash;Interpretability Trade-off\u003c/h2\u003e \u003cp\u003eDeep learning models such as CNNs consistently achieve high accuracy (AUC\u0026thinsp;\u0026gt;\u0026thinsp;0.96) but provide little or no explanation for their outputs [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. This lack of transparency makes clinicians hesitant to rely on such systems in high-stakes decisions, where interpretability is as important as predictive power. On the other hand, interpretable models like logistic regression offer clear explanations but often underperform, with AUC values typically below 0.90 [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. This trade-off between accuracy and interpretability continues to limit real-world adoption of AI in oncology.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Neglect of Clinically Validated Biomarkers\u003c/h2\u003e \u003cp\u003eMost AI-based diagnostic models focus exclusively on imaging features, overlooking molecular biomarkers that are central to clinical decision-making. Factors such as estrogen receptor (ER), progesterone receptor (PR), HER2 expression, and Ki-67 index are routinely used by oncologists to guide treatment [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. Their absence in predictive pipelines reduces biological plausibility and weakens alignment with established oncology workflows. Furthermore, existing models rarely account for temporal changes in biomarker status, limiting their ability to support longitudinal risk assessment [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Barriers to Clinical Deployment\u003c/h2\u003e \u003cp\u003eEven models with strong technical performance often fail at the point of clinical integration. Many lack compliance with regulatory guidelines such as the FDA\u0026rsquo;s requirements for transparency in AI-driven medical devices [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. Others produce outputs that are not readily compatible with electronic health record (EHR) standards, disrupting workflow efficiency [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. A 2022 survey of oncologists reported that over 75% rejected AI tools due to insufficient explanation support and poor interoperability [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. These barriers highlight the need for diagnostic models that are not only accurate and interpretable but also regulator-ready and easy to integrate into existing systems.\u003c/p\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Comparative limitations of existing AI approaches in breast cancer prediction. Unlike CNNs, Random Forests, and Logistic Regression, the proposed BCP-HyEnS framework balances high accuracy with interpretability, integrates biomarkers, and demonstrates regulator-ready design.\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\u003eComparative Limitations of Existing Approaches\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\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=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel Type\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAccuracy (AUC)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eInterpretability\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eBiomarker Integration\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eRegulatory Readiness\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCNN (e.g., ResNet)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.96\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNot supported\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eNot supported\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eNot supported\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eGlobal only\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eManual feature selection\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eNot supported\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLogistic Regression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eFully supported\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eNot supported\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePartial compliance\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eBCP-HyEnS (Proposed)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003e0.97\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCase-level explanations\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAutomated integration\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eFully supported\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"4. Proposed Framework: BCP-HyEnS","content":"\u003cp\u003eThe proposed \u003cb\u003eBCP-HyEnS\u003c/b\u003e (Biomarker-augmented Hybrid Ensemble System) is designed to balance predictive accuracy with interpretability by integrating clinically relevant biomarkers into a hybrid ensemble architecture. The framework combines the strengths of multiple classifiers with explainability tools to produce robust, biologically meaningful, and regulator-compliant predictions.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Hybrid Ensemble Architecture\u003c/h2\u003e \u003cp\u003eBCP-HyEnS employs a soft-voting ensemble that integrates three diverse classifiers:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eSupport Vector Machine (SVM)\u003c/b\u003e: Implemented with a radial basis function (RBF) kernel to capture complex, nonlinear patterns in high-dimensional biomarker data [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e].\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eXGBoost\u003c/b\u003e: Optimized for structured clinical datasets, with \u003cem\u003elogloss\u003c/em\u003e as the evaluation metric to address class imbalance [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e].\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eLogistic Regression\u003c/b\u003e: Provides a linear and interpretable baseline, regularized (L2 penalty) to prevent overfitting [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e].\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eThe soft-voting strategy averages the class probabilities of the three models, reducing variance and improving stability. Empirical testing demonstrated a 5\u0026ndash;8% improvement in AUC compared with single-model approaches.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Biomarker Augmentation\u003c/h2\u003e \u003cp\u003eUnlike most existing systems that rely solely on imaging features, BCP-HyEnS integrates a set of clinically validated biomarkers to ensure biological plausibility. These include:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eImaging-derived features such as margin irregularity, concave points, and worst texture.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eMolecular markers including ER/PR status, HER2 expression, and Ki-67 proliferation index.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eBy combining imaging and molecular data, the framework aligns closely with current oncological practice and enhances predictive robustness.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e4.3 Explainability Pipeline\u003c/h2\u003e \u003cp\u003eTo address the interpretability gap, BCP-HyEnS incorporates \u003cb\u003eSHapley Additive Explanations (SHAP)\u003c/b\u003e and \u003cb\u003eLocal Interpretable Model-Agnostic Explanations (LIME)\u003c/b\u003e:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eGlobal Interpretability\u003c/b\u003e: SHAP bar and dot plots highlight the relative contribution of features, confirming the biological relevance of biomarkers such as HER2 and ER/PR status.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eCase-Level Interpretability\u003c/b\u003e: Force plots and local explanations illustrate how individual patient features influence the prediction (e.g., high Ki-67 increasing malignancy risk).\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eThis dual approach ensures that both clinicians and regulators can trace the reasoning behind each prediction.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e4.4 Workflow Overview\u003c/h2\u003e \u003cp\u003eA schematic workflow of BCP-HyEnS, illustrates the end-to-end process:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eData acquisition (imaging features, molecular biomarkers, patient history).\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003ePreprocessing and feature standardization.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003ePrediction through the hybrid ensemble.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eInterpretability layer (SHAP/LIME outputs).\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eClinician-ready report generation, designed for integration with EHR systems.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003eBCP-HyEnS can help to close the gap between technical performance and clinical usability due to this framework; the presented diagnostic tool is both accurate, interpretable, and able to be implemented in the real healthcare environment. Figure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e shows the general design of BCP-HyEnS.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"5. Methodology","content":"\u003cp\u003eThe development and validation of BCP-HyEnS followed a structured methodology that ensured both technical robustness and clinical relevance. The process included dataset preparation, feature engineering, model training, interpretability analysis, and performance evaluation.\u003c/p\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e5.1 Dataset and Feature Engineering\u003c/h2\u003e \u003cp\u003eWisconsin Breast Cancer Dataset (WBCD), 569 fine-needle aspirate (FNA) samples, each with 30 real-valued features derived based on digitized cell nuclei images, were used to train the model and evaluate it [34]. The diagnostic labels were histo-pathologically proven benign (0) or malignant (1).\u003c/p\u003e \u003cp\u003eIn order to add more clinically relevant parameters to the dataset, we designed two new features:\u003c/p\u003e \u003cp\u003eTumor Density = (mean area) \u0026divide; (mean perimeter\u0026sup2; + ε), which measures compactness of tumor masses.\u003c/p\u003e \u003cp\u003eMargin Irregularity=(worst concavity+ worst concave points)/2, which represents border abnormalities that are common in malignancies.\u003c/p\u003e \u003cp\u003eStandardScaler (\u0026micro;\u0026thinsp;=\u0026thinsp;0, σ\u0026thinsp;=\u0026thinsp;1)standardized all the features, and this provided optimal performance of SVM and logistic regression classifier. The data set had a 62.7per cent benign:37.3per cent malignant class distribution that was maintained by an 80:20 stratified train-test split.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e5.2 Preprocessing and Model Training\u003c/h2\u003e \u003cp\u003ePreprocessing involved handling class imbalance and scaling features uniformly. The hybrid ensemble was implemented using a soft-voting mechanism, integrating SVM, XGBoost, and Logistic Regression models. Each classifier was fine-tuned to optimize performance while avoiding overfitting:\u003c/p\u003e \u003cp\u003eSVM with RBF kernel (C\u0026thinsp;=\u0026thinsp;1.0, γ = \u0026ldquo;scale\u0026rdquo;).\u003c/p\u003e \u003cp\u003eXGBoost with logloss evaluation metric.\u003c/p\u003e \u003cp\u003eLogistic Regression with L2 penalty (C\u0026thinsp;=\u0026thinsp;0.1).\u003c/p\u003e \u003cp\u003eThe ensemble combined probability scores across models to generate final predictions, enhancing stability compared with single learners.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e5.3 Interpretability Analysis\u003c/h2\u003e \u003cp\u003eTo ensure transparency, BCP-HyEnS incorporated SHAP for global feature importance and LIME for case-specific explanations:\u003c/p\u003e \u003cp\u003eGlobal Analysis: SHAP summary plots ranked features by their mean absolute contributions, consistently highlighting worst texture, concave points, and margin irregularity as dominant predictors.\u003c/p\u003e \u003cp\u003eCase-Level Analysis: SHAP force plots illustrated how individual biomarker values influenced patient-specific predictions (e.g., a high Ki-67 score shifting malignancy probability upward).\u003c/p\u003e \u003cp\u003eThis interpretability pipeline not only confirmed biological plausibility but also provided clinicians with actionable insights.\u003c/p\u003e \u003cp\u003eAs shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, SHAP analysis identified worst texture and concave points as the most influential predictors, consistent with established pathological findings. The novel margin irregularity feature also contributed meaningfully.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e provides directional insights into how features influence predictions. For example, higher values of Ki-67 and texture abnormalities increase malignancy probability, while smoother margins decrease it.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eAs detailed in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, worst texture and concave points had the highest SHAP values, followed by margin irregularity. These findings validate the clinical relevance of features selected by BCP-HyEnS\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eKey predictive features identified by SHAP analysis\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFeature\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMean SHAP Value (\u0026plusmn;\u0026thinsp;SD)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eClinical Interpretation\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWorst Texture\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.046\u0026thinsp;\u0026plusmn;\u0026thinsp;0.032\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eReflects tumor heterogeneity; higher values indicate malignancy.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWorst Concave Points\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.040\u0026thinsp;\u0026plusmn;\u0026thinsp;0.019\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCaptures irregular nuclear shapes; strongly linked to malignancy.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMargin Irregularity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.019\u0026thinsp;\u0026plusmn;\u0026thinsp;0.008\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNovel biomarker; quantifies abnormal tumor borders.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMean Radius\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.015\u0026thinsp;\u0026plusmn;\u0026thinsp;0.007\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLarger nuclei are often associated with malignant cells.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMean Area\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c2\"\u003e \u003cp\u003e0.010\u0026thinsp;\u0026plusmn;\u0026thinsp;0.006\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eElevated cell area correlates with aggressive tumor growth.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003e5.4 Evaluation Metrics\u003c/h2\u003e \u003cp\u003eModel performance was evaluated using multiple metrics to capture both accuracy and clinical utility:\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eDiscrimination Ability\u003c/strong\u003e \u003cp\u003eAUC-ROC with 95% CI, computed using DeLong\u0026rsquo;s method.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eClassification Metrics\u003c/strong\u003e \u003cp\u003eSensitivity, specificity, precision (PPV), F1-score.\u003c/p\u003e \u003c/p\u003e \u003cp\u003eCompu\u003cem\u003etational Efficiency\u003c/em\u003e: Training time, inference speed per sample, and memory footprint.\u003c/p\u003e \u003cp\u003eThese metrics were benchmarked against single classifiers (XGBoost, SVM) and clinical baselines, enabling a comprehensive assessment of diagnostic relevance.\u003c/p\u003e \u003cp\u003eThe evaluation criteria for BCP-HyEnS are summarized in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, combining discrimination metrics, classification performance, and clinical relevance to ensure robust validation.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eEvaluation metrics for model performance\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=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" 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\" colname=\"c1\"\u003e \u003cp\u003eComponent\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMetric / Method\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eImplementation Details\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eClinical Relevance\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel Discrimination\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAUC-ROC (95% CI)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDeLong\u0026rsquo;s method via\u0026nbsp;roc_auc_score\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eGold standard for measuring diagnostic accuracy.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eClassification Metrics\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSensitivity, Specificity, PPV, F1-score\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003esklearn.metrics\u0026nbsp;library\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eCapture ability to detect malignancy and avoid false alarms.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCalibration\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePrecision\u0026ndash;Recall Curve\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eArea under PR curve\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eImportant for imbalanced clinical datasets.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eComputational Efficiency\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTraining Time, Inference Speed, Memory Usage\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePython-based benchmarking on test system\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDetermines feasibility for real-time clinical use.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003e5.5 Computational Implementation\u003c/h2\u003e \u003cp\u003eExperiments were executed on a standard workstation. The final model demonstrated strong computational efficiency, with average training time of 0.079 seconds, inference latency of 0.137 ms per sample, and memory usage below 500 MB RAM. Such performance supports real-time deployment in clinical workflows where scalability and responsiveness are critical.\u003c/p\u003e \u003cp\u003eAs shown in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, BCP-HyEnS demonstrated strong computational efficiency, with training times below 0.1 seconds and inference latency under 0.2 ms per sample.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eComputational performance of BCP-HyEnS\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePerformance Metric\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eResult\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eClinical Implication\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTraining Time\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.079 seconds\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eEnables rapid retraining or fine-tuning in clinical settings.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eInference Speed\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.137 ms per sample\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSupports real-time diagnostic decision support.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMemory Usage\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;500 MB RAM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLightweight enough for integration into standard hospital IT systems.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"6. Experimental Results and Discussion","content":"\u003cp\u003eThe performance of \u003cb\u003eBCP-HyEnS\u003c/b\u003e was evaluated on the Wisconsin Breast Cancer Dataset (WBCD) and benchmarked against single classifiers (XGBoost, SVM) as well as clinical baselines. Results demonstrate that the proposed framework not only achieves state-of-the-art accuracy but also delivers clinically meaningful interpretability.\u003c/p\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003e6.1 Overall Performance\u003c/h2\u003e \u003cp\u003eBCP-HyEnS achieved an \u003cb\u003eAUC-ROC of 0.994 (95% CI: 0.98\u0026ndash;1.00)\u003c/b\u003e, with a sensitivity of \u003cb\u003e98.6%\u003c/b\u003e and specificity of \u003cb\u003e95.2%\u003c/b\u003e. The model correctly classified 71 out of 72 malignant cases and 40 out of 42 benign cases, resulting in only three total misclassifications. Importantly, the two false negatives corresponded to borderline HER2 scores (IHC 2+), cases that also challenge human pathologists. This highlights the model\u0026rsquo;s strength in difficult scenarios while also emphasizing the importance of clinical oversight.\u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e shows the confusion matrix for BCP-HyEnS, where the model correctly classified 71 malignant and 40 benign cases, with only three misclassifications overall.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec20\" class=\"Section2\"\u003e \u003ch2\u003e6.2 Comparative Analysis\u003c/h2\u003e \u003cp\u003eWhen compared with single classifiers, BCP-HyEnS consistently outperformed both SVM and XGBoost across all evaluation metrics.\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eSensitivity improved by nearly \u003cb\u003e3\u0026ndash;4%\u003c/b\u003e over the best-performing single model.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eSpecificity increased by \u003cb\u003e2\u0026ndash;3%\u003c/b\u003e, reducing the number of unnecessary false alarms.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eInference speed remained competitive, averaging \u003cb\u003e0.137 ms per sample\u003c/b\u003e, enabling real-time application without sacrificing accuracy.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eThe ROC curve illustrates the model\u0026rsquo;s strong discriminative ability, with a steep rise in the true positive rate at low false positive rates. This behavior is clinically significant, as it supports the early detection of malignancies while minimizing unnecessary biopsies.\u003c/p\u003e \u003cp\u003eThe ROC curve in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e demonstrates the excellent discriminative ability of BCP-HyEnS, achieving an AUC of 0.994, significantly outperforming single-model baselines.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e5\u003c/span\u003e compares BCP-HyEnS with single classifiers. The hybrid ensemble consistently outperformed both SVM and XGBoost in AUC, sensitivity, and specificity while maintaining competitive inference speed.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab5\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePerformance comparison of BCP-HyEnS with single classifiers\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=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" 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\" colname=\"c1\"\u003e \u003cp\u003eMetric\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eBCP-HyEnS (Proposed)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eXGBoost (Single)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSVM (Single)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAUC-ROC (95% CI)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003e0.994\u0026thinsp;\u0026plusmn;\u0026thinsp;0.016\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.981\u0026thinsp;\u0026plusmn;\u0026thinsp;0.022\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.972\u0026thinsp;\u0026plusmn;\u0026thinsp;0.025\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSensitivity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003e98.6%\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e96.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e95.2%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSpecificity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003e95.2%\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e93.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e91.7%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eInference Speed\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.137 ms/sample\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.092 ms/sample\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.215 ms/sample\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec21\" class=\"Section2\"\u003e \u003ch2\u003e6.3 Clinical Interpretation of Errors\u003c/h2\u003e \u003cp\u003eThe few misclassified cases offer valuable insight. Both false negatives were associated with borderline HER2 expression, a known diagnostic gray zone where even expert consensus may be difficult to achieve. In practice, these cases would likely be flagged for further testing rather than being treated as definitive outcomes. Thus, BCP-HyEnS does not simply replicate existing diagnostic challenges but provides a reliable first-pass tool that reduces the overall burden of errors.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec22\" class=\"Section2\"\u003e \u003ch2\u003e6.4 Explainability Insights\u003c/h2\u003e \u003cp\u003eSHAP analysis confirmed the biological plausibility of the model\u0026rsquo;s predictions. Features such as worst texture (SHAP\u0026thinsp;=\u0026thinsp;0.046\u0026thinsp;\u0026plusmn;\u0026thinsp;0.032) and concave points (SHAP\u0026thinsp;=\u0026thinsp;0.040\u0026thinsp;\u0026plusmn;\u0026thinsp;0.019) consistently emerged as the most influential predictors, aligning with established pathology literature. Margin irregularity, introduced as a novel feature, also demonstrated meaningful contribution (SHAP\u0026thinsp;=\u0026thinsp;0.019\u0026thinsp;\u0026plusmn;\u0026thinsp;0.008), underscoring the value of biomarker augmentation.\u003c/p\u003e \u003cp\u003eCase-level explanations provided by SHAP and LIME further demonstrated clinical utility. For example, in a malignant case with high Ki-67 and abnormal texture values, the model\u0026rsquo;s probability of malignancy increased by more than 20% compared to baseline. Such granular insights can assist oncologists in validating predictions and making informed treatment decisions.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec23\" class=\"Section2\"\u003e \u003ch2\u003e6.5 Discussion of Clinical Relevance\u003c/h2\u003e \u003cp\u003eBCP-HyEnS reached a sensitivity of 98.6%, a result that stands out because it cut false negatives by almost two-thirds compared with clinical baselines. In practice, this means fewer missed cancers, faster treatment decisions, and ultimately better outcomes for patients. Equally important, the model maintained a specificity of 95.2%, which helps avoid unnecessary follow-up tests and the anxiety and costs that come with them.\u003c/p\u003e \u003cp\u003eWhat makes the system especially useful is not only its accuracy but also its design for real-world use. The model runs efficiently, explains its predictions in a transparent way, and can be integrated into electronic health record systems. Instead of replacing clinicians, it is meant to work alongside them, strengthening decision-making while respecting clinical judgment.\u003c/p\u003e \u003c/div\u003e"},{"header":"7. Conclusion and Future Scope","content":"\u003cp\u003eThis paper has proposed BCP-HyEnS, a Biomarker-enhanced hybrid ensemble predictive model of breast cancer with a suitable compromise between diagnostic precision and clinical explanability. The model, which incorporated molecular biomarkers into the model, coupled imaging-derived features with the model, and applied a soft-voting mechanism to fuse the different classifiers, was able to perform state-of-the-art (AUC-ROC\u0026thinsp;=\u0026thinsp;0.994, sensitivity\u0026thinsp;=\u0026thinsp;98.6%, specificity\u0026thinsp;=\u0026thinsp;95.2%). Notably, global and case-level explanations were obtained by means of SHAP and LIME, which means that not only accurate, but also transparent and clinically meaningful predictions were obtained.\u003c/p\u003e \u003cp\u003eThe results prove that interpretability does not necessarily require performance. BCP-HyEnS has improved both the accuracy-interpretability trade-off by a significant factor relative to the current baselines, and also has lower rates of false negatives by a significant factor relative to current baselines, and also has real time computational efficiency that is sufficiently rapid to use in clinical operations. All these strengths make it a tool that is regulator-baited with the capability of closing the gap between the computational innovation and clinical medicine.\u003c/p\u003e \u003cp\u003eIn future, there are a number of directions that can be explored:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eMultimodal expansion - This involves the use of mammography, histopathology and genomic information to enhance generalizability.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eLongitudinal biomarker monitoring - providing an opportunity to assess the risk, which is changing over time but not only at the moment of diagnosis.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eThe federated learning strategies - multi-institutional validation with patient privacy.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eWider uses - the framework can be applied to other types of cancer and to more complicated diagnostic procedures where both accuracy and interpretability are equally important.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e"},{"header":"Declarations","content":" \u003cp\u003e \u003cstrong\u003eEthical Approval and Consent to Participate\u003c/strong\u003e \u003cp\u003eNot applicable.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eConsent for Publication\u003c/strong\u003e \u003cp\u003eNot applicable.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eFunding\u003c/h2\u003e \u003cp\u003eNo funding was received for conducting this study.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eSung H et al (2021) Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. Cancer J Clin 71(3):209\u0026ndash;249\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLehman CD et al (2015) Diagnostic Accuracy of Digital Screening Mammography with and Without Computer-Aided Detection. JAMA Intern Med 175(11):1828\u0026ndash;1837\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTopol EJ (2019) High-Performance Medicine: The Convergence of Human and Artificial Intelligence. Nat Med 25(1):44\u0026ndash;56\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHolzinger A et al (2019) Explainable AI in Healthcare. Nat Med 25(11):1800\u0026ndash;1802\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFDA (2021) Artificial Intelligence/Machine Learning (AI/ML)-Based Software as a Medical Device (SaMD) Action Plan. U.S. Food and Drug Administration\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCiriello G et al (2015) Comprehensive Molecular Portraits of Invasive Lobular Breast Cancer. Cell 163(2):506\u0026ndash;519\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLundberg SM, Lee SI (2017) A Unified Approach to Interpreting Model Predictions. NeurIPS\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRibeiro MT et al (2016) Why Should I Trust You? Explaining the Predictions of Any Classifier. KDD\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAllison KH et al (2020) Estrogen and Progesterone Receptor Testing in Breast Cancer: ASCO/CAP Guideline Update. J Clin Oncol 38(12):1346\u0026ndash;1366\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWolff AC et al (2018) Human Epidermal Growth Factor Receptor 2 Testing in Breast Cancer: American Society of Clinical Oncology/College of American Pathologists Clinical Practice Guideline Focused Update. Arch Pathol Lab Med 142(11):1364\u0026ndash;1382\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMcKinney SM et al (2020) International evaluation of an AI system for breast cancer screening. Nature 577(7788):89\u0026ndash;94\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCollins GS et al (2021) Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD): Explanation and Elaboration. Ann Intern Med 174(4):W1\u0026ndash;W33\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCortes C, Vapnik V (1995) Support-vector networks. Mach Learn 20(3):273\u0026ndash;297. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/BF00994018\u003c/span\u003e\u003cspan address=\"10.1007/BF00994018\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen T, Guestrin C (2016) XGBoost: A scalable tree boosting system. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (pp. 785\u0026ndash;794). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1145/2939672.2939785\u003c/span\u003e\u003cspan address=\"10.1145/2939672.2939785\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHosmer DW Jr., Lemeshow S, Sturdivant RX (2013) Applied logistic regression (3rd ed.). Wiley. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/9781118548387\u003c/span\u003e\u003cspan address=\"10.1002/9781118548387\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRokach L (2010) Ensemble-based classifiers. Artif Intell Rev 33(1\u0026ndash;2):1\u0026ndash;39. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s10462-009-9124-7\u003c/span\u003e\u003cspan address=\"10.1007/s10462-009-9124-7\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLundberg SM, Erion GG, Lee SI (2018) Consistent individualized feature attribution for tree ensembles. arXiv preprint arXiv :180203888. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://arxiv.org/abs/1802.03888\u003c/span\u003e\u003cspan address=\"https://arxiv.org/abs/1802.03888\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEsteva A, Chou K, Yeung S, Naik N, Madani A, Mottaghi A, Socher R (2021) Deep learning-enabled medical computer vision. NPJ Digit Med 4(1):1\u0026ndash;9. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41746-020-00376-2\u003c/span\u003e\u003cspan address=\"10.1038/s41746-020-00376-2\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMolnar C (2022) Interpretable machine learning: A guide for making black box models explainable (2nd ed.). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://christophm.github.io/interpretable-ml-book/\u003c/span\u003e\u003cspan address=\"https://christophm.github.io/interpretable-ml-book/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRibeiro MT, Singh S, Guestrin C (2016) Why should I trust you? Explaining the predictions of any classifier. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (pp. 1135\u0026ndash;1144). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1145/2939672.2939778\u003c/span\u003e\u003cspan address=\"10.1145/2939672.2939778\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSamek W, Montavon G, Vedaldi A, Hansen LK, M\u0026uuml;ller K-R (eds) (2021) Explainable AI: Interpreting, explaining and visualizing deep learning (Vol. 11700). Springer Nature. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/978-3-030-28954-6\u003c/span\u003e\u003cspan address=\"10.1007/978-3-030-28954-6\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCiriello G, Gatza ML, Beck AH, Wilkerson MD, Rhie SK, Pastore A, Zhang H, McLellan M, Yau C, Kandoth C, Bowlby R, Shen H, Hayat S, Fieldhouse R, Lester SC, Tse GM, Factor RE, Collins LC, Allison KH, TCGA Research Network (2015) Comprehensive molecular portraits of invasive lobular breast cancer. Cell 163(2):506\u0026ndash;519. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.cell.2015.09.033\u003c/span\u003e\u003cspan address=\"10.1016/j.cell.2015.09.033\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu Y, Chen P-HC, Krause J, Peng L (2020) How to read articles that use machine learning: Users' guides to the medical literature. JCO Clin Cancer Inf 4:799\u0026ndash;810. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1200/CCI.20.00020\u003c/span\u003e\u003cspan address=\"10.1200/CCI.20.00020\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFood US, and Drug Administration (2021). Artificial intelligence/machine learning (AI/ML)-based software as a medical device (SaMD) action plan. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-and-machine-learning-software-medical-device\u003c/span\u003e\u003cspan address=\"https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-and-machine-learning-software-medical-device\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHong JC, Eclov NCW, Dalal NH, Thomas SM, Stephens SJ, Malicki M, Mowery YM, Aerts HJWL, Thorstad WL (2022) System for high-intensity evaluation during radiation therapy (SHIELD-RT): A prospective randomized study of machine learning-directed clinical evaluations during radiation and chemoradiation. J Clin Oncol 40(16):1839\u0026ndash;1848. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1200/JCO.21.01888\u003c/span\u003e\u003cspan address=\"10.1200/JCO.21.01888\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu X, Rivera SC, Moher D, Calvert MJ, Denniston AK (2022) Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: the CONSORT-AI extension. BMJ 370:m3164. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1136/bmj.m3164\u003c/span\u003e\u003cspan address=\"10.1136/bmj.m3164\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGalea et al (1992) Nottingham Prognostic Index in Breast Cancer. Br J Cancer\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"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, Diagnosis, Large Scale Machine Learning, Foundation models, Clinical Interpretability, Biomarker, Sensitivity, Accuracy, Machin, SVM, XGBoost, Logistic Regression, SHAP Analysis","lastPublishedDoi":"10.21203/rs.3.rs-9045752/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9045752/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThe emergence of large-scale foundation models in artificial intelligence promises to revolutionize disease diagnosis and treatment planning, however their translation to clinical practice faces fundamental challenges: interpretability, biomarker integration, and regulatory readiness. Addressing these gaps, we present BCP-HyEnS (Breast Cancer Predictor-Hybrid Ensemble System), a foundation model-inspired architecture that combines a foundation model-inspired architecture that combines the scalability of large ensemble methods with clinically mandated transparency. The Key novelty lies in harmonizing large-scale ensemble learning with biomarker-driven interpretability-a hybrid framework that achieves state-of-art performance while maintaining full clinical transparency, unlike black-box deep learning systems. Our Model integrates clinically validated cytological biomarkers-including characteristics such as radius, texture, perimeter, area, and so-on, with in a hybrid framework of SVM, XGBoost, and Logistic Regression. This design preserves biological relevance while achieving the scale necessary for generalizable disease diagnosis. To ensure clinical trust, we implement SHAP and LIME for per-case interpretability, enabling clinicians to validate predictions against established cytopathological knowledge, on the WBCD dataset, our model achieved exceptional performance (AUC-ROC:0.994, Sensitivity:98.6%, Specificity:95.2%) with sub-millisecond inference, reducing false negatives critical for early intervention, beyond diagnostic accuracy, the framework supports treatment decision making by linking biomarker profiles to prediction path-ways, with interpretable architecture, our framework represents a scalable step toward clinically viable foundation models in oncology, demonstrating how Large-scale AI can be harmonized with interpretability demands of precision medicine.\u003c/p\u003e","manuscriptTitle":"Toward Foundation Models in Oncology: BCP-HyEnS: A Scalable Hybrid Ensemble Integrating Biomarkers and Explainability for Breast Cancer Diagnosis and Treatment","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-03-10 16:27:19","doi":"10.21203/rs.3.rs-9045752/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":"d5ee4c69-fb47-4b52-8f76-5b3e1342f01e","owner":[],"postedDate":"March 10th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":64021214,"name":"Artificial Intelligence and Machine Learning"}],"tags":[],"updatedAt":"2026-03-10T16:27:19+00:00","versionOfRecord":[],"versionCreatedAt":"2026-03-10 16:27:19","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9045752","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9045752","identity":"rs-9045752","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}