Determining prognostic factors for gastric cancer using the regression tree method

Abstract

Background. The regression tree method is a useful statistical technique that has been little used in the analysis of prognosis. Methods. The prognostic factors of gastric cancer were investigated, using the regression tree method, in 555 patients who had undergone curative resection for serosa-negative gastric cancer and who were enrolled in a randomized controlled trial of postoperative adjuvant chemotherapy (JCOG [Japan Clinical Oncology Group] 8801 study). Results. By the regression tree method, the first divided prognostic factor (the most important factor) was lymph node metastasis; in particular, extent of lymphatic spread had the greatest impact on prognosis. In addition, age, tumor size, depth of invasion, and individual dose intensity were found to be significant prognostic factors, whereas sex, tumor location, macroscopic tumor type, and extent of lymph node dissection were not. The resulting tree structure consisted of nine terminal nodes with different prognostic factors, and four clusters were obtained by the merging of terminal nodes that showed a similar prognosis. The cluster which showed the best survival rate (5-year survival rate, 0.986) consisted of two terminal nodes: node 12, which contained NOT1 patients who were younger than 62 years and had a tumor size of less than 7.5cm, and node 14, which contained N1 patients who were younger than 46 years. Conclusion. In serosa-negative gastric cancer, lymph node metastasis was the most important prognostic factor. Utilization of the regression tree method enabled visual interpretation of the results of statistical analyses through the graphic representation of prognostic factors. It allowed the identification of the optimal combination of these prognostic factors that defined several groups of patients with distinct prognoses and may serve as a useful reference for the individualization of treatment strategy.