Cancer staging is an evolutionary process. The Tumor-Node-Metastasis (TNM) esophageal cancer staging, first introduced in 1968, rapidly developed, but unfortunately then stagnated for decades. T classifications had not changed since 1988, N classifications for thoracic esophageal cancer since 1977, and M classifications since 1997. The principal hindrance to evolution was the long held concept of stage groupings of esophageal cancer which was incorrectly based on a simple, orderly arrangement of increasing anatomic T, then N, then M classifications. This assumption was consistent with neither cancer biology nor survival data. Worldwide collaboration1 has provided data for a unique, modern machine-learning analysis2 that has produced data-driven staging for cancer of the esophagus and esophagogastric junction (EGJ).3 This new system is the basis for the 7th editions of the AJCC and UICC Cancer Staging Manuals.4,5 It is more representative of and consistent with the survival following esophagectomy of patients with esophageal cancer. The changes address problems of empiric stage grouping and prior disharmony with stomach cancer staging. In addition, TNM classifications have been reviewed and revised where data analysis and consensus demonstrated a need for change. For the first time, nonanatomic cancer characteristics, primary cancer site (location), histologic grade (grade), and histopathologic type (cell type) are incorporated in esophageal cancer staging.
At the request of the AJCC, the Worldwide Esophageal Cancer Collaboration was inaugurated in 2006. Thirteen institutions from five countries and three continents (Asia, Europe, and North America) submitted de-identified data by July 2007. A database of 4627 esophagectomy patients who had no induction or adjuvant therapy was created.1
Multiple previously proposed revisions of esophageal cancer staging have examined goodness of fit or p values to test for a statistically significant effect of stage on survival. Instead, staging for the 7th edition used Random Forest (RF) analysis, a machine-learning technique that focuses on predictability for future patients.2 RF analysis makes no a priori assumptions about patient survival, is able to identify complex interactions among variables, and accounts for nonlinear effects. RF may be viewed as a “backward” analysis which determines the anatomic classifications (TNM) and nonanatomic cancer characteristics that are associated with specific survival groups.
RF analysis first isolated cancer characteristics of interest from other factors influencing survival by generating risk-adjusted survival curves for each patient. Unlike previous approaches that began by placing cancer characteristics into proposed groups, RF analysis produced distinct groups with monotonically decreasing risk-adjusted survival without regard to cancer characteristics. Then, anatomic and nonanatomic cancer characteristics important for stage group composition were identified within these groups. Finally, homogeneity within groups guided both amalgamation and segmentation of cancer characteristics between adjacent groups to arrive at the final stage groups.3-5
7th Edition TNM Classifications: Changes and Additions