Interpretable machine learning for survival prediction and risk stratification in elderly patients with breast cancer after breast-conserving surgery

Background: The necessity of adjuvant radiotherapy following breast-conserving surgery (BCS) in elderly patients with early-stage breast cancer remains controversial. Existing studies focus predominantly on population-level benefits without identifying specific prognostic subgroups with different baseline survival probabilities. We aimed to develop interpretable machine learning models to predict survival and establish precise prognostic risk stratification that could inform individualised treatment discussions. Methods: Using the Surveillance, Epidemiology, and End Results database (2016-2022), we included patients aged ≥70 years with T1-2N0M0, oestrogen receptor-positive, human epidermal growth factor receptor 2 (HER2) negative breast cancer who underwent BCS. We developed six machine learning survival models incorporating age, tumour grade, T stage, progesterone receptor status, race, histology, and chemotherapy. Model performance was evaluated using time-dependent area under the curve (AUC) and concordance index. The optimal model was interpreted using SHapley Additive exPlanations (SHAP) framework. Patients were stratified into three risk groups, with survival differences assessed using Kaplan-Meier analysis. Results: A total of 39,872 patients were included (training set: 31,897; test set: 7,975). The eXtreme Gradient Boosting (XGBoost) model demonstrated optimal performance with 1-, 3-, and 5-year AUCs of 0.714, 0.692, and 0.711, respectively. SHAP analysis identified age as the most important predictor, followed by tumour grade and T stage. Risk stratification successfully delineated three distinct prognostic groups: low-risk (37% of patients, 5-year overall survival 88-90%), intermediate-risk (33% of patients, 5-year overall survival 82-84%), and high-risk (30% of patients, 5-year overall survival 65-67%) (log-rank P