Predicting factors and top gene identification for survival data of breast cancer

Abstract

For high-throughput research with biological data-sets generated sequentially or by transcriptional micro-arrays, proteomics or other means, analytic techniques that address their high dimensional aspects remain desirable. The computation part basically predicts the tendency towards mortality due to breast cancer (BC) by using several classification methods, i.e., Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), Linear Discriminant Analysis (LDA) and Decision Tree (DT), and compared the models' performances. We proceed with the RF method since it provides better results than any other underlying models based on accuracy. We have also demonstrated some traditional and competing risk models, illustrated the models with real data analysis, depicted their curves' natures and also compared their fits using prediction error curves and the concordance index. Furthermore, two different survival splitting rules are used by using separate Random Survival Forest (RSF) methods and also constructing the ranking of risk factors due to breast cancer. The results show that high-level grade and diameter are the most important predictors for mortality progression in the presence of competing events of death, and lymph nodes, age and angiography are other vital criteria for this purpose. We have also implemented RSF backward selection criteria, which enables top gene selection related to mortality progression due to breast cancer. This method identifies c-MYB, CDCA7, NUSAP1, BIRC5, ANGPTL4, JAG1, IL6ST, and remaining genes that are mainly responsible for mortality progression due to breast cancer. In this work, R software is used to obtain and evaluate the results.