Informatics in Medicine Unlocked · Published 2026-07-25 · DOI 10.1016/j.imu.2026.101792
Anusha Ihalapathirana, Pekka Siirtola, Taneli T. Mattila, Gunjan Chandra, Satu Tamminen, Juha Röning, Outi Laatikainen
Background:: Individualized treatment optimization remains a major challenge in healthcare, and its applications in breast cancer remain limited. This study introduces a computational framework that integrates survival analysis with explainable artificial intelligence (XAI) to support individualized treatment selection based on time-to-event outcomes in patients with breast cancer. Methods:: A retrospective cohort of breast cancer patients was analyzed using survival-based machine-learning models to estimate individualized survival probabilities under two treatment strategies: surgery alone (S-Only) and multimodal therapy (MultiTx). The models included Random Survival Forests (RSF), DeepSurv, DeepHit, Cox proportional hazards (CoxPH), and XGBoost with an accelerated failure time (XGBoost-AFT). Predictive performance was assessed using the concordance index (C-index) and the integrated Brier score (IBS), whereas treatment selection was evaluated by comparing Kaplan–Meier survival estimates between model-concordant and model-discordant patient groups. Results:: DeepHit achieved the highest predictive performance with 20 features (C-index, 0.85; IBS, 0.072), whereas RSF demonstrated the strongest discrimination at the 10-year survival horizon and produced statistically significant stratification of outcomes in both treatment groups. In contrast, the standard CoxPH models recommended MultiTx for all patients, reflecting its limited capacity to capture heterogeneity in treatment benefit. SHAP-based model explanations identified routinely collected clinical features, including tumor stage, histology, and comorbidity burden, as key predictors. These findings were consistent with established clinical knowledge and revealed treatment-specific patterns. Conclusion:: Explainable survival-based machine learning can effectively support individualized treatment selection in breast cancer and provide a foundation for time-to-event–driven, patient-centered clinical decision support.
Abstract from DOAJ. Public domain (CC0 1.0).
Read the article at the publisher →
Ihalapathirana, A., Siirtola, P., Mattila, T., et al. (2026). Towards explainable AI-driven individualized treatment optimization for breast cancer time-to-event outcomes. Informatics in Medicine Unlocked. https://doi.org/10.1016/j.imu.2026.101792