A Unifying Paradigm in Predictive Oncology Through Hybrid Voting Ensembles for Optimising Breast Cancer Detection

Abstract

Machine learning (ML) offers substantial promise for improving cancer diagnosis, yet many existing approaches rely on single-classifier models. This study proposes and evaluates a hybrid weighted voting ensemble to optimise predictive accuracy for breast cancer classification by combining a diverse set of ML classifiers. Using the Wisconsin Diagnostic Breast Cancer (WDBC) dataset, five base learners were developed: a multi-layer perceptron (MLP), a Classification and Regression Tree (CART), and three support vector machine (SVM) models with linear, radial basis function (RBF), and polynomial kernels. A comparative analysis of individual model performance was conducted against the proposed ensemble. The weighted voting model achieved the best overall performance, with an accuracy of 0.9737, F1-score of 0.9641, and AUC of 0.9952, surpassing the strongest individual classifiers (MLP and SVM-RBF, each with an accuracy of 0.9591). Permutation feature importance highlighted "worst" measurements, such as concave points_worst and radius_worst, as the most informative predictors, consistent with established malignancy indicators. The proposed hybrid voting ensemble provides a robust and reliable framework for breast cancer prediction, demonstrating clear advantages over single-classifier models and supporting more precise decision-making in oncology.