Keywords: prognostic model, breast cancer, relapse-free survival, overall survival
We developed prognostic models to predict five-year overall survival (OS), ten-year overall survival (OS), and five-year relapse-free survival (RFS) using resources from the well-conducted Breast Cancer Survival Study (SBCSS). SBCSS is a large, population-based cohort study of 4,858 female breast cancer patients aged 20 to 75 years at diagnoses. Patients were recruited between March 2002 and April 2006 and were followed up through December 2012. In addition to patients' demographic, clinical, pathological, treatment information, our model incorporates novel modifiable lifestyle information. Number of events of five-year OS, ten-year OS, and five-year RFS model is 535 (11.0%), 950 (19.6%), and 845 (17.4%). Missing covariates were completed with multiple imputation to reduce bias and increase precision. A multivariable Cox proportional hazard model was developed for each survival outcome. Performance of the models was assessed for both discrimination and calibration. We internally validated our entire model development procedure using the .632 bootstrap method, with which 200 datasets were simulated by randomly sampling from the original dataset independently and with replacement. To further simplify the full models for routine practice, reduced models were derived to approximate the original full models using backward step down method. In conclusion, we built novel prognostic models for breast cancer survivors of Asian ancestry. Comparing to existing prognostic tools, we expanded the model to incorporate modifiable lifestyle predictors, ER, PR and HER2 status. Five-yr OS, 10-yr OS, and 5-yr RFS can be predicted with high accuracy. C-statistics of the full models were 0.759, 0.730, and 0.722, for five-year OS, ten-year OS, and five-year RFS, respectively.