An ablation analysis of feature selection strategies for machine learning–based breast cancer recurrence prediction

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Breast cancer recurrence after treatment remains a critical clinical problem that directly affects patient management and survival outcomes. In this study, a machine learning–based classification framework is proposed to predict breast cancer recurrence risk. Experiments were conducted on the Wisconsin Prognostic Breast Cancer and Recurrent Breast Cancer datasets. During the data preprocessing stage, missing values were imputed, categorical variables were transformed into numerical representations, and all features were standardized using z-score normalization. The class imbalance problem was addressed using the Synthetic Minority Over-sampling Technique. To systematically investigate the effect of feature selection on classification performance, several scenarios were evaluated, including no feature selection, embedded feature selection using L1-regularized Logistic Regression, and different metaheuristic optimization-based wrapper feature selection methods. These optimization-based methods included Particle Swarm Optimization, Teaching–Learning-Based Optimization, Whale Optimization Algorithm, Genetic Algorithm, Ant Colony Optimization, and Artificial Bee Colony. The selected feature subsets were evaluated using Support Vector Machine, Random Forest, and Extreme Gradient Boosting classifiers. All experiments were performed using a 10-fold cross-validation strategy, and the Wilcoxon signed-rank test was applied to statistically compare the performance differences among the feature selection methods. The experimental results show that optimization-based feature selection methods can improve classification performance while producing more compact feature subsets. In particular, the Genetic Algorithm-based feature selection approach achieved balanced and competitive results across different datasets and classifiers. The findings indicate that feature selection strategies have an important effect on both model performance and model complexity in breast cancer recurrence prediction. In this regard, the proposed framework provides an applicable and interpretable machine learning–based approach for decision support systems aimed at assessing breast cancer recurrence risk.