Abstract
BACKGROUND: Although meta-analyses have demonstrated survival benefits associated with primary tumor resection in MBC, guidelines lack consensus on the survival benefit of postoperative radiation therapy (RT). METHODS: In this study, we included 1392 patients with de novo metastatic breast cancer (dnMBC) by integrating data from the SEER database (2010-2019) to systematically assess the efficacy of postoperative RT and develop a machine learning-driven prognostic tool. The primary endpoint was overall survival (OS). RESULTS: Propensity score matching (PSM) results showed that postoperative RT significantly improved OS (HR = 0.573, 95 % CI = 0.475-0.693), but this survival gain showed great heterogeneity among different subgroups. It is found that patients with HR-/HER2-or HR+/HER2-subtypes gained significant OS benefit from (p < 0.001) postoperative RT, whereas patients with HER2+ subtype did not gain any survival benefit since the effect of targeted therapy overshadowed the postoperative RT. Further risk stratification by the random survival forest (RSF) model revealed that high-risk patients with T4/N3 stage, high tumor grade and poor response to chemotherapy had significantly prolonged OS after receiving RT (p < 0.001), while low-risk patients showed no additional benefit. The model had excellent predictive efficacy (training set C-index = 0.741, validation set C-index = 0.720) with key predictors including HER2 status, chemotherapy response and tumor grade. The research team developed an interactive web application (https://lee2287171854.shinyapps.io/RSFshiny/) based on this model, which can generate individualized survival risk scores in real-time to guide clinical decision-making. CONCLUSION: This study is the first to propose a risk stratification strategy for postoperative RT in dnMBC, and innovatively integrates machine learning and clinical tools to provide a new paradigm for optimizing precision therapy.