Recursive Partitioning Analysis (RPA) of Prognostic Factors for Overall Survival in Patients with Spinal Metastasis: A New System for Stratified Treatment

BACKGROUND: Accurate survival estimate is necessary when determining the most appropriate treatment modality for metastatic spinal tumor. The main purpose of this study was to identify the prognostic factors of spinal metastasis and establish a decision tree model. METHODS: A consecutive cohort of 507 patients from 3 institutional clinical centers who were treated for metastatic spinal tumor between 2005 and 2015 were retrospectively reviewed. In total, 70% of the participants were randomly selected as a "training sample." The prognostic effect of preoperative factors was evaluated using the "training sample," and a decision tree model was established. Then, the accuracy of the new model, as well as the Tokuhashi and Tomita score, was tested by the "test sample," which consisted of the remaining 30% of participants. RESULTS: A decision tree model was generated based on the significant factors with an order of descending importance on predicting the prognosis. According to the new model, patients were classified into 3 groups, mean survival times of less than 6 months, 6-12 months, and more than 12 months, who were indicated for conservative therapy/palliative operation, palliative operation, and invasive excision, respectively. The newly established model was confirmed to be of high accuracy in predicting overall survival, whereas the Tokuhashi and Tomita scores were of modest accuracy and consistency. CONCLUSIONS: A new decision tree model for prognosis prediction in spinal metastasis was established with a satisfactory accuracy and consistency. However, the Tokuhashi and Tomita systems were presented to be less correlated between the scores and actual survival.