The growing availability of high-dimensional clinical datasets has enabled the development of intelligent systems for early breast cancer diagnosis. However, standalone machine learning models often suffer from feature redundancy, overfitting, and limited generalization. To overcome these challenges, this study proposes an optimal feature engineering and ensemble stacking framework designed to maximize predictive accuracy while ensuring statistical robustness and interpretability. The methodology incorporates comprehensive preprocessing, including missing-value imputation, Z-score normalization, and Synthetic Minority Over-sampling Technique (SMOTE) for class balancing. Mutual information–based feature selection is employed to identify the most discriminative biomarkers and reduce dimensionality. The refined features are used to train an ensemble stacking architecture comprising an optimized Support Vector Machine (RBF kernel), Random Forest classifier, and lightweight neural network. A logistic regression meta-learner integrates their probabilistic outputs to generate the final prediction. Experiments conducted on the Breast Cancer Wisconsin Diagnostic dataset (569 instances) using 10-fold cross-validation demonstrate superior performance of the proposed framework, achieving 98.67% accuracy, 99.1% sensitivity, 98.2% specificity, and a ROC–AUC of 0.992. Statistical validation using paired t-tests confirms significant improvement over baseline models (p < 0.05). Additionally, SHAP-based analysis enhances interpretability by identifying key biomarkers influencing malignancy prediction. The proposed hybrid framework provides a reproducible, statistically validated, and clinically relevant solution for highprecision breast cancer analytics, demonstrating strong potential for deployment in decision-support systems.
Gayathri et al. (Thu,) studied this question.