A Comprehensive Framework for Breast Cancer Recurrence Prediction Using Class Balancing and Feature Optimization Techniques

Background: Breast cancer is one of the most common malignancies among women worldwide and remains one of the leading causes of cancer-related mortality. Accurate prediction of breast cancer recurrence is clinically important for patient survival, treatment planning, and follow-up strategies. However, recurrence prediction remains challenging due to high-dimensional feature spaces, class imbalance, limited sample size, and noisy variables in medical datasets. Methods: In this study, an integrated machine learning framework combining hybrid ensemble learning, data balancing, and optimization-based feature selection is proposed for breast cancer recurrence prediction. In the proposed approach, different classifiers were comparatively evaluated, and the XGBoost + SVM hybrid model was selected based on its performance in terms of recall and F1-score. To address class imbalance, ADASYN was employed to increase the representation of recurrence cases during training. Feature selection was subsequently performed by combining Optuna and Ant Colony Optimization (ACO) to identify a smaller subset of informative features. Results: The framework was developed and evaluated primarily on the Breast Cancer Wisconsin Prognostic (WPBC) dataset, where it achieved an accuracy of 93.30%, an F1-score of 86.60%, a recall of 91.30%, and an AUC of 95.86%. To examine its applicability to other breast cancer classification tasks, the same modeling approach was also evaluated on the Breast Cancer Wisconsin Diagnostic (WDBC) and Wisconsin Breast Cancer Original (WBC) datasets. The corresponding AUC values were 99.91% and 98.74%, respectively. Conclusions: These findings suggest that the integration of XGBoost + SVM, ADASYN-based class balancing, and Optuna + ACO-based feature selection can contribute to breast cancer recurrence prediction on the WPBC dataset. The additional experiments on WDBC and WBC further indicate that the proposed modeling approach is applicable to breast cancer classification tasks with different feature structures and prediction targets.