Ensemble Feature selection with Adaptive Weighted voting for Robust Breast Cancer Diagnosis

Abstract

Breast cancer is a significant disease prevalent in many women around the globe and one of the most common causes of cancer death worldwide, this underscores the demand for trustworthy and precise computer-aided diagnostic tools. Typically, high-dimensional medical datasets possess redundant and irrelevant features which may adversely affect the classification performance and explainability of models. Moreover, conventional feature selection methods frequently suffer from instability, selecting different feature subsets across training samples, which limits reproducibility in clinical decision-making. To address these challenges, this study proposes an adaptive ensemble feature selection framework that integrates four complementary feature selection methods—namely Mutual Information (MI), ANOVA F-test, LASSO, and Random Forest importance—using three ensemble Voting strategies: hard majority Voting, performance-weighted voting, and rank aggregation. The adaptive weights are estimated using balanced accuracy from 5-fold cross-validation in the training data to ensure data-driven performance and to avoid information leakage during the feature selection process. The proposed adaptive ensemble framework is tested on three publicly available breast cancer datasets based on different clinical domains, namely the Coimbra dataset (blood-based biomarkers), the SEER dataset (epidemiological records), and the Mammographic Mass dataset (BI-RADS characteristics). As a result of the experiment, the framework achieves the balanced accuracies of 88.46%, 78.00%, and 84.77% on the Coimbra, SEER, and Mammographic Mass datasets, correspondingly, and perfect feature stability (Jaccard Index = 1.0) for all datasets in repeated cross-validation. Moreover, the selected features are clinically meaningful, including such diagnostic parameters as glucose, BMI, resistin, tumor characteristics, and BI-RADS characteristics. This means that the adaptive ensemble framework allows obtaining an accurate, stable, and interpretable feature selection algorithm that can be used for decision support in breast cancer diagnosis.