Background. We previously demonstrated that a mathematical technique called
recursive partitioning analysis (RPA), when applied to the Radiation Thera
py Oncology Group Head and Neck Cancer database, created rules that formed
subgroups ("classes") having unique outcomes. We sought to learn if the app
lication of RPA-derived rules to a new head and neck database would create
classes that were similarly associated with outcome and thereby validate th
is technique.
Methods. The rules derived from recursive partitioning analysis of the prev
ious database were used to subgroup an independent, new head and neck cance
r database (RTOG 85-27), created as part of a phase III trial of the hypoxi
c-cell radiosensitizer, Etanidazole. The resulting classes were compared wi
th each other and with the classes formed from the previous database.
Results. The rules derived by RPA from our previous data-base correctly gro
uped the tumors in the new database into unique classes of similar outcome.
RPA could successfully use either survival or local-regional control of di
sease as the measure of outcome. As judged by comparison of the 95% confide
nce intervals, the outcome of the classes in the new database is essentiall
y indistinguishable from the outcome of the classes in the previous databas
e.
Conclusion. RPA-derived rules provide a reliable method to assort head and
neck tumors into unique classes that are predictive of outcome. These rules
can be successfully applied to new databases that were not used in the cre
ation of the rules and thereby validate the methodology. (C) 2001 John Wile
y & Sons, Inc.