gms | German Medical Science

54. Jahrestagung der Deutschen Gesellschaft für Medizinische Informatik, Biometrie und Epidemiologie e.V. (GMDS)

Deutsche Gesellschaft für Medizinische Informatik, Biometrie und Epidemiologie

07. bis 10.09.2009, Essen

Bootstrap based regression trees for patient classification in the Austrian DRG-System

Meeting Abstract

Suche in Medline nach

  • Karl-Peter Pfeiffer - Department f. Med. Statistik, Informatik und Gesundheitsökonomie, Innsbruck
  • Thomas Grubinger - Department f. Med. Statistik, Informatik und Gesundheitsökonomie, Innsbruck
  • Conrad Kobel - Department f. Med. Statistik, Informatik und Gesundheitsökonomie, Innsbruck

Deutsche Gesellschaft für Medizinische Informatik, Biometrie und Epidemiologie. 54. Jahrestagung der Deutschen Gesellschaft für Medizinische Informatik, Biometrie und Epidemiologie (gmds). Essen, 07.-10.09.2009. Düsseldorf: German Medical Science GMS Publishing House; 2009. Doc09gmds120

DOI: 10.3205/09gmds120, URN: urn:nbn:de:0183-09gmds1205

Veröffentlicht: 2. September 2009

© 2009 Pfeiffer et al.
Dieser Artikel ist ein Open Access-Artikel und steht unter den Creative Commons Lizenzbedingungen (http://creativecommons.org/licenses/by-nc-nd/3.0/deed.de). Er darf vervielfältigt, verbreitet und öffentlich zugänglich gemacht werden, vorausgesetzt dass Autor und Quelle genannt werden.


Gliederung

Text

Introduction: In 1997 a regression tree based patient classification system for all inpatients in Austria was introduced. Since this time the so-called “Austrian DRG-system” is regularly updated.

Background: The classification rules are based on regression tree models which terminal nodes should be cost homogenous. However, the resultant regression tree often has to be adjusted manually to be medically reasonable. Despite the possibility of manually adapting the original tree, bootstrapping can be used to search systematically for alternatives.

Material / Methods: The data base are the Austrian DRG-data of previous years (approx. 2.500.500 admissions per year divided into approx. 900 groups according to the main diagnosis or procedures). The dependent variable is the length of stay of the cost of a patient.

Bootstrap-based methods and different model evaluation criteria are used for searching through a space of models.

Results:

The use of bootstrapping assisted in constructing a number of alternative trees which are at least as accurate as the tree constructed by the currently used semi-automatic procedure. In some of the datasets the bootstrap search for alternative models helped to overcome local minima and increased accuracy significantly.

Conclusion: Bootstrapped regression trees are a powerful tool to assist in the development of patient classification systems. They can improve accuracy and offer several models to choose from. Moreover, by the use of bootstrapping often less complex trees can be found. The principles of this tree selection strategy can of course be used for many other applications.