Variable selection using the Lasso-Cox model with Bayesian regularization

• Published in: IEEE Conference on Industrial Electronics and Applications (Online) pp. 924 - 927
• Main Authors: Lu, Wenxin, Yu, Zhuliang, Gu, Zhenghui, Huang, Jinhong, Gao, Wei, Zhou, Haiyu
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.05.2018
• Subjects: Adaptation models; Analytical models; Bayes methods; Bayesian regularization; Hazards; Input variables; Lasso-Cox regression; Numerical models; Prognostic gene; Survival analysis; Tumors
• ISSN: 2158-2297
• DOI: 10.1109/ICIEA.2018.8397844

Selection of prognostic genes associated with tumor has been a subject of considerable research in recent years. In order to solve the high-dimensional gene expression profiles, the Lasso-Cox model has been proposed and widely used in survival analysis. Based on the sparse regression algorithm, the regularization parameter must be carefully tuned by cross-validation to optimize performance. In this paper, we introduce an algorithm based on simple Bayesian approach to replace the process of parameter selection, and the regularization parameter is determined adaptively in training. Simulation results show that variable selection of Bayesian-Lasso (BLasso) can be more accurate than that of Lasso method. We also apply our algorithm to a real dataset DLBCL, and the selected genes have been proven to have close relationship with the tumor.