Integrasi Ensemble Learning dan Explainable AI untuk Prognostik Survival Lima Tahun Pasien Breast Carcinoma

Syifa Naura Milla Celesta (Universitas Sriwijaya, Sumatera Selatan)
Octa Dama Yanti (Universitas Sriwijaya, Sumatera Selatan)
Syifa Alfariani (Universitas Sriwijaya, Sumatera Selatan)
Fathoni Fathoni (Universitas Sriwijaya, Sumatera Selatan)

TIN: TERAPAN INFORMATIKA NUSANTARA
Vol 7 No 5 (2026): October 2026

Five-year survival prediction in breast carcinoma cases has an essential role in supporting clinical decision-making. Nevertheless, limitations related to predictive capability and the interpretability of machine learning models continue to present challenges. In this research, a survival prediction framework was developed by combining ensemble learning methods with Explainable Artificial Intelligence (XAI). The study utilized secondary data from the SEER dataset obtained through Kaggle, covering 4,004 patient records. Several stages were conducted, including exploratory analysis, data preprocessing, and model construction using Random Forest, XGBoost, and CatBoost. These algorithms were subsequently integrated through a stacking mechanism in which a meta-learner was applied. Performance assessment involved accuracy, precision, recall, F1-score, and AUC-ROC measurements, while SHAP and LIME were employed to examine model interpretability. Experimental findings revealed that the stacking approach produced an accuracy of 0.6442, an F1-score of 0.7658, and a recall value of 0.8292, although it did not consistently outperform the standalone models. Among all approaches, CatBoost achieved the strongest predictive results with an accuracy of 0.6604 and an F1-score of 0.7901. In contrast, the stacking model recorded the highest AUC-ROC value of 0.5109 despite only a narrow difference. Further analysis demonstrated that tumor characteristics together with patients’ clinical conditions contributed most significantly to prediction outcomes. The integration of stacking-based ensemble learning and XAI in this study was intended to produce a predictive model that offers dependable performance while remaining more interpretable for clinical use.