Interpretable Ensemble Learning for Breast Cancer Prediction: From Cytological Features to Clinical Decision Support: A Review

Abstract

Ensemble learning has become a default choice for breast-cancer prediction from cytological features, but combining multiple models raises a specific concern for clinical decision support: a prediction assembled from several classifiers can be harder to explain than a prediction from one. This review synthesizes five sources to examine whether interpretability has kept pace with the field's move toward ensembles. A systematic review of 50 mammogram-classification studies finds that conventional machine learning, deep learning, and hybrid/ensemble approaches have each reached near-ceiling accuracy on their own benchmarks, and concludes that future work must balance accuracy against interpretability and resource efficiency rather than optimizing accuracy alone. Four applied studies, remarkably, all evaluate their models on the identical 569-sample Wisconsin Diagnostic Breast Cancer cytological dataset, making their results directly comparable: a Random Forest model reaches 96% accuracy with a perfect AUC of 1.00; a three-way rotating stacking ensemble of Support Vector Machine, Naïve Bayes, and K-Nearest Neighbors reaches 95% accuracy; a six-model comparison built explicitly around SHAP-based interpretability finds its Stacking Classifier reaching 98.25% accuracy with zero false positives, identifying tumor border irregularity and size as the most influential features; and a hybrid ensemble combining three classical models with three deep-learning architectures reaches 99.76% accuracy on the same data. Taken together, these four independently conducted studies on one shared benchmark show ensemble accuracy converging near a ceiling the underlying dataset may simply not allow further improvement past, while only one of the four builds interpretability into its methodology as a first-class design goal rather than an afterthought. This review argues that the field's next problem is not squeezing further accuracy from an already-saturated benchmark, but ensuring that the explanation delivered alongside an ensemble's prediction is as reliable, and as clinically actionable, as the prediction itself.