A hierarchical ensemble machine learning framework for precision survival prediction in breast cancer using SEER data

• Developed a hierarchical ensemble model for breast cancer survival prediction. • Combined bagging, boosting, and stacking with logistic regression meta-learner. • Identified 8 key prognostic features from 322 SEER clinical variables. • Achieved 95.2% accuracy and robust precision, recall, and F1 performance. • Supports precision oncology through individualized treatment outcome prediction. Breast cancer remains one of the most frequent cancer diagnoses internationally and a leading cause of cancer related deaths among women. Accurate survival prediction of breast cancer is thus crucial, since it can guide personalized care and thus improve patient outcomes. Moreover, it can influence health policy by throwing light on the factors which most influence the survival rate. However, the nature of medical data brings with it challenges like clinical heterogeneity and high dimensionality. These issues make the task of reliable prediction difficult. This study proposes a hierarchical ensemble machine learning approach using the Surveillance, Epidemiology, and End Results (SEER) dataset to predict patient survival. Following comprehensive preprocessing and feature selection using Analysis of Variance (ANOVA) F-test and Recursive Feature Elimination (RFE), we identified 8 optimal prognostic features from 322 variables. Selected features included clinically relevant factors, such as age at diagnosis, histological stage, HER2 status, and survival time. We evaluated six baseline algorithms and proposed a two-tier hierarchical ensemble combining bagging, boosting, and stacking techniques with a logistic regression meta-learner. The ensemble model achieved 95.2% accuracy and 95% precision, recall, and F1-score, outperforming the individual algorithms, including XGBoost and Gradient Boosting. The results demonstrate the effectiveness of hierarchical model integration in improving predictive stability and precision for heterogeneous clinical datasets. This framework offers a promising direction toward data-driven, individualized survival prediction in precision oncology and can be extended to other cancer prognostic modeling applications.