Back to Program | 2010 Program and Abstracts Overview | 2010 Posters

Purpose : We define prognostic model that predicts survival of patients with colorectal cancer after a radical potentially curative resection. This study uses a data-driven approach to identify highly predictive cancer characteristics that involve TNM as well as non-TNM factors. A novel prognostic methodology, random survival forest, accounts for the complex interplay among clinical and histologic features. Methods : Survival data of 2,534 colon and 2,380 rectal cancer patients undergoing a radical resection between 1969 and 2003 were analyzed by novel random forest technology. We examined the role of TNM and non-TNM factors such as histologic grade,lymph node ratio, type of surgery, ASA, and age in staging and prognosis. A forest of 1,000 random survival trees was grown using log-rank splitting. Risk-adjusted random survival forest methodology maximized survival prediction and produced variable importance measure. Results : Risk-adjusted 5-year survival after resection of colon and rectal cancer was dominated by pT and lymph node ratio. Survival of colon cancer was modulated by ASA grade, whereas type of resection modulated survival of rectal cancer. Risk-adjusted predicted survival of patients ordered by increasing mortality and divided into 10 groups demonstrated decreasing survival. The degree to which lymph node ratio as one of the dominant predictors for colon and rectal cancer relates to proposed stage grouping is shown in Figure 1. Conclusions : The novel data-driven methodology for survival data was a useful tool for predicting survival for patients with colorectal cancer and identifying patterns of cancer characteristics. Important prognostic values from new model include lymph node ratio, ASA, and type of surgery. Specifically, higher predictive power of lymph node ratio as compared with traditional pN classification was observed. These observations may lead to stage groupings that redefine a simple, logical arrangement of TNM.