Survival analysis of breast cancer patients using machine learning models

Breast cancer is a fatal disease. There is no one treatment for breast cancer due to its heterogeneity in terms of response to treatment and prognosis. This study deals with identifying the key covariates responsible for the prognosis of breast cancer patients so that proper treatment can be administered which can improve the overall survival of the patients. The study utilizes the clinical and pathological features from the Molecular Taxonomy of Breast Cancer International Consortium dataset (METABRIC). Three models namely the Cox Proportional hazards (CoxPH) model, random survival forests (RSF) model, and DeepHit were utilized for survival prediction. Both the Random survival forests and DeepHit model gave a Concordance Index (C-Index) of 0.86 and performed better than the Cox PH model which provided a C-Index of 0.85. The most important covariate in the random survival forests model with the maximum absolute value was relapse-free status. Relapse-free status had a high positive correlation of 88% with the survival status of the patient. The Cox model gave four important statistically significant covariates with P < 0.05. They are Age at Diagnosis, Estrogen Receptor (ER) Status, Progesterone Receptor (PR) Status, and tumor stage. Among these ER and PR status have a negative regression coefficient value which reduces the risk of hazard for the patients. Thus, the proposed work helps identify the important prognostic covariates and also aids clinicians in determining the type of treatment to be administered to the patients. Both the Random survival forests model and DeepHit performed the best for survival prediction.