Purpose: The recursive partitioning analysis (RPA) classes for malignant glioma patients were previously established using data on over 1500 patients entered on Radiation Therapy Oncology Group (RTOG) clinical trials. The purpose of the current analysis was to validate the RPA classes with a new dataset (RTOG 90-06), determine the predictive power of the RPA classes, and establish the usefulness of the database norms for the RPA classes.
Patients and methods: There are six RPA classes for malignant glioma patients that comprise distinct groups of patients with significantly different survival outcome. RTOG 90-06 is a randomized Phase III study of 712 patients accrued from 1990 to 1994. The minimum potential follow-up is 18 months. The treatment arms were combined for the purpose of this analysis. There were 84, 13, 105, 240, 150, and 23 patients in the RPA Classes I-VI from RTOG 90-06, respectively.
Results: The median survival times (MST) and 2-year survival rates for the six RPA classes in RTOG 90-06 are compared to those previously published. The MST and 2-year survival rates for the RTOG RPA classes were within 95% confidence intervals of the 90-06 estimates for Classes I, III, IV, and V. The RPA classes explained 43% of the variation (squared error loss). By comparison, a Cox model explains 30% of the variation. The RPA classes within RTOG 90-06 are statistically distinct with all comparisons exceeding 0.0001, except those involving Class II. A survival analysis from a prior RTOG study indicated that 72.0 Gy had superior outcome to literature controls; analysis of this data by RPA classes indicates the survival results were not superior to the RTOG database norms.
Conclusion: The validity of the model is verified by the reliability of the RPA classes to define distinct groups with respect to survival. Further evidence is given by prediction of MST and 2-year survival for all classes except Class II. The RPA classes explained a good portion of the variation in survival outcome in the data. Lack of correlation in RPA Class II between datasets may be an artifact of the small sample size or an indication that this class is not distinct. The validation of the RPA classes attests to their usefulness as historical controls for the comparison of future Phase II results.