Identification of biomarker signatures has gained substantial attention recently in developing personalized medicine for cancer patients. Usually, there are high-dimensional candidate biomarkers in discovery process due to the development of modern technique. In this project, we consider an innovative approach to identify (prognostic or predictive) biomarker signatures under the framework of Cox proportional hazard regression with high-dimensional covariates, where the penalized technique (i.e., Lasso) is adopted for variable selection. The advantage of this work is further to incorporate the idea of deriving the adjusted p-values by multi-splitting algorithm and bootstrapping method to overcome the issue that the traditional p-values are not valid due to restrictions in the penalized approaches. Extensive simulations are conducted to evaluate the performance of our proposal, and show that the family-wise error rate can be well controlled, and also the model selection accuracy is improved. Finally, we apply our method to a breast cancer study.