TREE-STRUCTURED STATISTICAL METHODS

自《分类与回归》一书出版以来

Since the publication of the book Classification and Regression Trees, the program CART, which embodies tree-structured methods developed in the book, has been used widely and with success. In particular, there have been substantial applications to problems of diagnosis and prognosis in cancer and cardiology. Ideas in the book have been extended to cover censored survival data and have been applied to assessing the relationship of dose intensity and other prognostic factors in diffuse large cell lymphoma. Also, CART-like pruning of large trees to smaller ones has been incorporated into new algorithms for lossy data compression in digital radiography. The present proposal is for research on new capabilities for and further applications of tree-structured methods. The areas of research have a wide range, from achieving supercomputer speed in tree construction by using distributed computing on a workstation network to expanding tree-structured methods from data compression to enhance magnetic resonance images of lymphomatous mediastinal masses. Also, we will have access to a data set gathered from more than 12,000 subjects for purposes of comparing new techniques with old ones in the problem of quickly diagnosing myocardial infarction for. patients who enter the emergency room with chief complaint of acute chest pain. Particular improvements and new capabilities to existing algorithms include but are not limited to time series and curve recognition, multiple response regression, analysis of contingency tables, local risk estimation, s g of trees, more effective linear combination splits, spline splits, polishing trees, density estimation and clustering, distributed computing, and two-step lookahead.