Comparison of the Radiation Therapy Oncology Group recursive partitioning classification and Union Internationale Contre le Cancer TNM classification for patients with head and neck carcinoma

Abstract

Background: Prognostic models need to be tested in external validation studies to assess generalizability. Recursive partitioning analysis (RPA), a prognostic system based on the creation of a classification tree, has been proposed as a classification method in patients with head and neck carcinoma. The aim of this study was to compare the RPA and Union Internationale Contre le Cancer (UICC) TNM classification systems in patients with head and neck carcinoma treated consecutively in a single center.

Methods: A total of 2166 patients with carcinomas of the oral cavity, oropharynx, hypopharynx, and larynx was classified according to both the RPA and the TNM classification systems, and the results were compared. The endpoints considered were observed survival and survival free of locoregional tumor. The two methods of classification were evaluated objectively by use of measures of intrastage homogeneity (hazard consistency), interstage heterogeneity (hazard discrimination), predictive power (outcome prediction), and patient distribution between stages (balance).

Results: When the endpoint considered was observed survival, there were no clinically relevant differences between the two classifications. However, when the endpoint was locoregional control, the RPA system was sensitive to the type of treatment used, and it was not generalizable.

Conclusions: To evaluate generalizability, new classification proposals need external validation studies that objectively measure the quality of the model. The performance of the RPA system was not reproducible in our cohort of patients when the endpoint evaluated was locoregional control.