The Survival Tree Method And Its Application In Prognostic Analysis

Tree-structured methods are presently enjoying widespread use. This popularity derives from the conceptual appeal and compelling examples presented in the definitive monograph "Classification and regression trees" (CART) by Breiman et al. Along with it the survival tree method is developed and applied in the follow-up study in medicine. The appealing aspect of this method is its ability to detect nonlinear interactions between two or more covariates, thereby defining new subgroups depending on risk factors and associated survival distributions. This advantage is more obvious when the classical parametric models are not appropriate and performing the data mining of tremendous database. Secondly, it does not depend on common but often unrealistic assumptions, such as the distribution of the survival data, the type of covariates. It is robust with outlier. Third, it can deal with time-dependent covariate also.In my paper, the creation of the tree is expounded. It involvesthree components of splitting, pruning and selecting of subtrees. In the splitting step, a group of patients is split into two subgroups according to values of a selected variable. In each step of analysis, all the possible splits are examined by an algorithm based on the logrank test statistic. The best split is then selected so that the split produces two subgroups of patients with the most different Kaplan-Meier survival curves. In general, such a splitting procedure then generates a large size survival tree. The algorithm prunes the large tree to a desirable size via a bottom-up procedure as described in detail by Segal. Besides, split-complexity measure which was introduced by LeBlance and Crowley is used in pruning tree and selecting subtrees.235 gastric cancer patients were found with complete clinical and pathologic material, including age, sex, tumor size, depth of invasion, lymph node metastasis, Bomman grade, the length of illness and the edge of operation cutting. The survival tree method and Cox proportional hazards model are used in prognostic analysis. Since the prognostic index, which obtain from Cox regression, can identify the subgroups of patients, it may be of interest to decide which one is the best. We can study them from the points of view of their reliability of predictions, predictive power and stability. The results showed that the survival tree method is better than the prognostic index. So we can use the Cox regression to assess the impact of the independent risk factors on patient survival, and use the survival tree method to identify subgroups. The results show that tumor size, lymph node metastasis and theedge of operation cutting are the independent risk factors on patient survival with the RR of 1. 112, 1. 791 and 1. 939 respectively. With the survival tree method, tumor size, lymph node metastasis and age are performed to define the joint effect of risk factors and to separates patients into three subgroups with median survival time of 24, 12 and 5 months. Lymph node metastasis(<3cm) and tumor size(<6. 5cm) found to be the best combination of risk factors to define prognosis.The merit and demerit of the survival tree method, the reliability and the stability of the tree, the resolve method of time-dependent covariate and the software of the survival tree analysis are expounded in discussion.