Non-linear feature selection and stacked ensemble learning for breast cancer diagnosis: A fusion of statistical and neural classifiers

Abstract

Breast cancer remains a leading cause of female mortality worldwide; hence rapid, highly accurate diagnosis is essential. This study presents an intelligent stacked-ensemble framework that fuses statistical and neural classifiers to distinguish benign from malignant tumors in the Wisconsin Breast Cancer Diagnostic (WBCD) dataset. Two linear feature-selection methods (Pearson Correlation Coefficient (PCC) and Principal Component Analysis (PCA)) and one non-linear method (Back-Propagation Neural Network (BNN)) are first applied to generate three discriminative feature subsets. Six baseline classifiers (SVM, KNN, Decision Tree, Logistic Regression, Naïve Bayes, and ANN) are then trained on each subset. For every feature-selection route, the two best performers feed a meta-learner via stacked generalization. Performance is assessed with accuracy, precision, recall, specificity, and confusion matrices. The proposed stacking strategy boosts every baseline’s accuracy by ≥ 0.2 percentage points. Notably, the M3 ensemble combining SVM, KNN, and ANN attains 100 % accuracy with zero false positives, demonstrating clinical-grade reliability. Results confirm that stacking heterogeneous learners mitigates individual model biases and enhances diagnostic precision. Future work should explore alternative ensemble paradigms (bagging, boosting) and region-specific datasets such as Nigerian breast-cancer cohorts and extend ensembling to feature-selection stages for further dimensionality reduction gains.