Implementation and Validation of an Explainable Ensemble Learning Framework for Breast Cancer Classification Using Cytological Features

Abstract

Explainable machine learning on cytological features is a promising route to reliable, interpretable breast-cancer diagnosis from fine-needle aspiration (FNA). Building on the comparative analysis of explainable models for cancer classification using cytological features by Siddiqui et al. [1], this study implements their comparative approach as the first layer of an explainable ensemble framework and validates the framework rigorously. Six base learners (logistic regression, support vector machine, k-nearest neighbours, random forest, extra trees and gradient boosting) and two ensembles (soft voting and stacking) were trained on the Wisconsin Diagnostic Breast Cancer dataset (569 FNA samples, 30 nuclear features). Validation combined 5 × 10-fold repeated cross-validation on a 75% development set with corrected resampled t-tests, a 25% held-out test set with bootstrap confidence intervals, calibration analysis and a screening-oriented decision threshold. Explanations were produced with SHAP for every model and evaluated for cross-model agreement and stability across folds. The soft-voting ensemble reached a cross-validated MCC of 0.947 ± 0.041 and AUC of 0.994 ± 0.009, and on the held-out test set an accuracy of 0.979, sensitivity of 0.943, specificity of 1.000 and AUC of 0.997 (95% CI 0.990–1.000). It was significantly better than random forest in MCC (p = 0.032) but statistically indistinguishable from logistic regression and the support vector machine, while showing the smallest fold-to-fold variability. A threshold of 0.17, chosen on development data, raised test sensitivity to 0.981 with specificity 0.978. SHAP rankings agreed well between the ensemble and its members (Spearman ρ 0.76–0.92) and 7 features led by worst perimeter and mean and worst concave points — were among the ten most important in every fold, consistent with cytopathological markers of malignancy. The framework provides a validated, interpretable and stable classifier, and its validation shows that a well-specified linear model remains a strong baseline that ensembles should be required to match.