EXPERIMENTS TO DETERMINE WHETHER RECURSIVE PARTITIONING (CART) OR AN ARTIFICIAL NEURAL-NETWORK OVERCOMES THEORETICAL LIMITATIONS OF COX PROPORTIONAL HAZARDS REGRESSION

New computationally intensive tools for medical survival analyses incl ude recursive patitioning (also called CART) and artificial neural net works. A challenge that remains is to better understand the behavior o f these techniques in effort to know when they will be effective tools . Theoretically they may overcome limitations of the traditional multi variable survival technique, the Cox proportional hazards regression m odel. Experiments were designed to test whether the new tools would, i n practice, overcome these limitations. Two datasets in which theory s uggests CART and the neural network should outperform the Cox model we re selected. The first was a published leukemia dataset manipulated to have a strong interaction that CART should detect. The second was a p ublished cirrhosis dataset with pronounced nonlinear effects that a ne ural network should fit. Repeated sampling of 50 training and testing subsets was applied to each technique. The concordance index C was cal culated as a measure of predictive accuracy by each technique on the t esting dataset. In the interaction dataset, CART outperformed Cox (P < 0.05) with a C improvement of 0.1 (95% CI, 0.08 to 0.12). In the nonl inear dataset, the neural network outperformed the Cox model (P < 0.05 ), but by a very slight amount (0.015). As predicted by theory, CART a nd the neural network were able to overcome limitations of the Cox mod el. Experiments like these are important to increase our understanding of when one of these new techniques will outperform the standard Cox model. Further research is necessary to predict which technique will d o best a priori and to assess the magnitude of superiority. (C) 1998 A cademic Press.