Abstract
Baseline survival estimates do not capture evolving prognosis in breast cancer with de novo bone metastases. We aimed to develop and internally evaluate a prognostic framework integrating baseline and conditional cancer-specific survival (CSS) estimates. Using the Surveillance, Epidemiology, and End Results (SEER) database, we retrospectively analysed 4,013 women diagnosed in 2010–2015 and randomly allocated 70% to training and 30% to internal validation. Random survival forest-assisted variable selection and Cox proportional hazards regression were used to construct baseline and conditional survival nomograms. Seven predictors were retained: age, primary tumour surgery, nodal stage, molecular subtype, and concomitant brain, liver, and lung metastases. The validation concordance index was 0.662 (95% confidence interval: 0.642–0.681), indicating moderate discrimination. Model-averaged probabilities of surviving to 10 years conditional on 5-year survival were 43.0% and 43.7% in training and validation, respectively, compared with unconditional CSS estimates of 17.2% and 17.7%. Landmark calibration was satisfactory for predicting subsequent 3-year CSS among 3-year survivors; other conditional scenarios were not empirically validated. The framework may support updated prognostic assessment once initial treatment status is known. However, 10-year estimates remain exploratory because of limited late follow-up and non-proportional hazards, and external validation is needed before clinical application.