Please activate JavaScript!
Please install Adobe Flash Player, click here for download


ISCB 2014 Vienna, Austria • Abstracts - Oral Presentations 19Monday, 25th August 2014 • 9:00-10:30 Monday25thAugustTuesday26thAugustThursday28thAugustAuthorIndexPostersWednesday27thAugustSunday24thAugust C06.4 Structure and interpretation of classification and regression trees RJ Marshall1 1 Auckland University, Auckland, New Zealand Since Breiman’s seminal work in 1984 the use of classification and regres- sion trees in many areas of statistical analysis, including medicine, has be- come widespread. Extensions to “forests of trees” and associated B-word refinements (bagging, Bayes, boosting, bootstrapping) have been pro- posed. However, the issue of the nature of tree structures, and classifiers generated by them, is seldom raised. The nature of the tree partitions is generally ignored; emphasis is nearly always placed on performance mea- sures. Yet in medical context the nature of tree structures is important, for the combinations of signs, symptoms, and other prognostic factors that make up a tree node, need to make clinical sense. In this presentation, the nature of tree structures will be discussed. Examples of published trees in the medical literature are presented and examined. It is concluded that the hierarchical nature of trees is restrictive and will invariably output awk- ward and difficult to interpret factor combinations. C06.5 Classification and regression trees for moderator effects in clinical trials R Gueorguieva1 , W-M Tsai1 , R Wu1 , H Zhang1 , SS O’Malley1 1 Yale University, New Haven, United States Classification and regression trees are a powerful and systematic method for ascertaining combinations of predictors of good outcome in medical studies. This approach allows one to consider a large pool of predictor variables, to derive empirically the strongest predictors, to consider inter- active effects and to present the results in the form of decision trees that are easily interpreted by clinicians. Classical applications of this approach focus on identification of predictors of outcome regardless of treatment. However, in treatment studies identification of subject characteristics as- sociated with good outcome on a particular treatment is of primary in- terest. We extend the method for automatic tree growing of Zhang et al. (2011) to identify subgroups of subjects who respond more favorably to one treatment than another based on their baseline characteristics. An automatic pruning step is incorporated, and a novel validation method is also proposed and evaluated. Terminal nodes in the constructed tree are associated with better outcome on a particular treatment, thus can be used to inform personalized treatment decisions. The approach is evalu- ated on simulated data and illustrated with an application from a clinical trial of alternative pharmacological treatments in alcohol dependence. The approach is also compared to alternative methods for identification of subgroups with enhanced treatment effects and advantages and disad- vantages of the different approaches are discussed. Acknowledgement: The project described was supported by Grants R01AA017173, K05 AA014715, and K23 AA020000 from the National Institute on Alcohol Abuse and Alcoholism.  

Pages Overview