Prognostic modeling in health care has been predominantly statistical, despite a rapid growth of literature on machine-learning approaches in biological data analysis. We aim to assess the relative importance of variables in predicting overall survival among patients with non-small cell lung cancer using a Variable Importance (VIMP) approach in a machine-learning Random Survival Forest (RSF) model for posttreatment planning and follow-up.A total of 935 non-small cell lung cancer patients were randomly and equally divided into 2 training and testing cohorts in an RFS model. The prognostic variables included age, sex, race, TNM stage, smoking history, Eastern Cooperative Oncology Group performance status, histologic type, treatment category, maximum standard uptake value of whole-body tumor (SUVmaxWB), whole-body metabolic tumor volume (MTVwb), and Charlson Comorbidity Index. The VIMP was calculated using a permutation method in the RSF model. We further compared the VIMP of the RSF model to that of the standard Cox survival model. We examined the order of VIMP with the differential functional forms of the variables.In both the RSF and the standard Cox models, the most important variables are treatment category, TNM stage, and MTVwb. The order of VIMP is more robust in RSF model than in Cox model regarding the differential functional forms of the variables.The RSF VIMP approach can be applied alongside with the Cox model to further advance the understanding of the roles of prognostic factors, and improve prognostic precision and care efficiency.