COMPARATIVE ANALYSIS OF MACHINE LEARNING CLASSIFIERS FOR BREAST CANCER DIAGNOSIS WITH EVIDENTIAL VOTING INTEGRATION (EVI-STACKER): AN INTERDISCIPLINARY BENCHMARK STUDY ON THE WISCONSIN DIAGNOSTIC BREAST CANCER DATASET WITH STATISTICAL VALIDATION AND IMPLI

Abstract

Breast cancer, including hereditary forms driven by pathogenic mutations in the BRCA1 and BRCA2 tumour-suppressor genes, remains the most frequently diagnosed malignancy in women worldwide. Automated classification of tumour malignancy from cytological biopsy morphometry is a clinically important task where machine learning (ML) offers substantial promise. This study presents a rigorous, statistically validated comparative benchmark of nine ML classifiers — Logistic Regression, K-Nearest Neighbours, Naive Bayes, Decision Tree, Support Vector Machine (SVM), Random Forest, XGBoost, Multi-Layer Perceptron, and a proposed stacking ensemble termed EVI-Stacker (Evidential Voting Integration) — on the publicly available Wisconsin Diagnostic Breast Cancer dataset (n=569, 30 nuclear morphometric features). A 70/30 stratified train-test split with 10-fold cross-validation and grid-search hyperparameter optimisation was applied uniformly. Statistical significance of performance differences was assessed using McNemar test and paired cross-validation tests; interpretability was assessed via SHAP analysis. The EVI-Stacker achieved the highest performance across all metrics: Accuracy=98.23%, F1-Score=98.05%, AUC-ROC=0.9963, and Matthews Correlation Coefficient=0.9601. While the single test-set difference versus XGBoost did not reach statistical significance (McNemar p=0.625), the improvement was consistent and significant across cross-validation folds (paired t-test p<0.001; Wilcoxon p=0.002). SHAP analysis identified worst radius, worst concave points, and worst perimeter as the highest-impact features, a pattern consistent with the nuclear pleomorphism reported for BRCA1/2-mutated tumours, although the dataset contains no genomic data and this association is presented as a hypothesis for future validation. The EVI-Stacker offers an interpretable, statistically validated ensemble framework relevant to breast cancer screening. This work is interdisciplinary in nature, bridging machine learning and computational analysis with tumour biology and BRCA-related genomic risk stratification.