Modeling Survival in Breast Cancer: a Large Population Study

摘要

Breast cancer has been recognized as a major threat for women's health around the world. Many efforts have been addressed towards the reduction of high mortality rates by improving survival behavior prediction in patients suffering from this disease. While all cases are important, identifying patterns within the cohort of patients with high risk of dying are of great importance. This study attempts to investigate whether transcriptomic information together with clinical covariates will have the potential to discriminate patients with such risk using survival data and a multi-step data mining approach. The methodology of this work starts with the application of feature selection methods (EBayes, FVI, Info Gain, and ReliefF) to a large population dataset conformed by 1980 samples. Those methods were assessed by implementing three classifiers (GBM, RF, and SVM) where parameter tuning was carried out. Performance metrics as ACC and AUC were used to identify EBayes and FVI as the two top performers, reaching values for ACC up to 75.88% and AUC up to 70.91%. The feature extraction process was able to reduce the number of predictors from ～25,000 to ～200 variables. Lymph nodes, STAT5A, AURKA, and CCNB2 are among the most significant predictors to characterize patient survival status.