Akademik Veri Yönetim Sistemi
INTERNATIONAL CONGRESS ON INTERDISCIPLINARY SCIENCE AND TECHNOLOGY, Kars, Türkiye, 27 - 28 Mart 2026, ss.385-393, (Tam Metin Bildiri)
Breast cancer recurrence refers to the reappearance of the disease after treatment, and the early and accurate prediction of recurrence risk is critically important for patient follow-up, identification of highrisk individuals, and personalized treatment planning. In this study, Bidirectional Gated Recurrent Unit (BiGRU) and BiGRU with Attention (BiGRU+Attention) based deep learning models were developed for breast cancer recurrence prediction, supported by Optuna-based feature selection and hyperparameter optimization.Experiments were conducted on the widely used Wisconsin Prognostic Breast Cancer (WPBC) and Recurrent Breast Cancer datasets. To ensure a fair comparison, a single feature mask strategy was adopted, and both models were evaluated using the same selected feature subset and common hyperparameter settings. The class imbalance problem was addressed only within the training folds using the Adaptive Synthetic Sampling (ADASYN) method. Model performance was assessed through 5-fold Stratified K-Fold cross-validation using accuracy, precision, recall, specificity, F1-score, and Receiver Operating Characteristic–Area Under the Curve (ROC-AUC) metrics.The experimental results indicate that on the WPBC dataset, the BiGRU model achieved an accuracy of 93.55%, while the BiGRU+Attention model obtained an accuracy of 88.52%. On the Recurrent Breast Cancer dataset, BiGRU reached an accuracy of 85.54%, whereas BiGRU+Attention achieved 77.11% accuracy. Overall, the findings suggest that, under the optimized feature subset and controlled experimental settings, the BiGRU architecture demonstrates more stable and consistent generalization performance compared to the BiGRU+Attention model. These results indicate that the proposed BiGRU-based approach provides a competitive and reliable framework for breast cancer recurrence risk prediction.