<![CDATA[Breast cancer classification using machine learning has been widely studied, particularly with the Breast Cancer Wisconsin Diagnostic dataset. Therefore, the main issue is not merely to report high accuracy, but to present a reproducible and clinically cautious comparative evaluation that prioritizes malignant-case performance, prevents data leakage, reports model configurations, and provides consistent interpretation. This study compares Logistic Regression, Support Vector Machine with radial basis function kernel, Gradient Boosting, and Random Forest using 569 instances and 30 numerical features extracted from digitized fine-needle aspiration cell nuclei. Standardization for scale-sensitive algorithms was placed inside the cross-validation pipeline. The malignant class was treated as the clinically critical positive class, and the models were evaluated using accuracy, malignant precision, malignant recall, malignant F1-score, specificity, balanced accuracy, Matthews Correlation Coefficient, and AUC. SVM RBF achieved the strongest overall test performance with accuracy of 0.9825, malignant recall of 0.9762, malignant F1-score of 0.9762, balanced accuracy of 0.9812, MCC of 0.9623, and AUC of 0.9977. A Wilcoxon signed-rank comparison across ten folds showed no significant difference between SVM RBF and Logistic Regression, while SVM RBF was significantly better than Gradient Boosting for malignant F1-score. Permutation importance applied to the selected SVM RBF model indicated that worst smoothness, worst texture, worst area, radius error, and worst radius contributed strongly to the predictions. The findings are limited to one curated public dataset and do not establish clinical validity]]>
