PRIM versus CART in subgroup discovery: when patience is harmful.

Abstract

We systematically compare the established algorithms CART (Classification and Regression Trees) and PRIM (Patient Rule Induction Method) in a subgroup discovery task on a large real-world high-dimensional clinical database. Contrary to current conjectures, PRIM's performance was generally inferior to CART's. PRIM often considered "peeling of" a large chunk of data at a value of a relevant discrete ordinal variable unattractive, ultimately missing an important subgroup. This finding has considerable significance in clinical medicine where ordinal scores are ubiquitous. PRIM's utility in clinical databases would increase when global information about (ordinal) variables is better put to use and when the search algorithm keeps track of alternative solutions.