Peeling finds interpretable prognostic rules based on a sequence of box shaped regions in the predictor space indexed by the fraction of patients in the prognostic group. Initially the entire data set is considered and then a fraction α of the data is removed from either extreme of a variable distribution among all variables. This process, called peeling, is repeated until only small fraction of the data remains. The data corresponding to a rule with sufficiently extreme outcome is removed and the remaining data are peeled again. The end result is a logical rule representing an extreme outcome group.   The schematic of sequence of decisions for two predictor variables is presented.

Please send request for R-functions to

mleblanc [at] fhcrc.org