A modified real AdaBoost algorithm to discover intensive care unit subgroups with a poor outcome

AMIA Annu Symp Proc. 2013 Nov 16:2013:798-803. eCollection 2013.

Abstract

The Intensive Care Unit (ICU) population is heterogeneous. At individual ICUs, the quality of care may vary within subgroups. We investigate whether poor outcomes of an ICU can be traced back to excess deaths in specific patient subgroups, by discovering candidate subgroups, with a modified adaptive decision tree boosting algorithm applied to 80 Dutch ICUs. Genuine subgroups were selected from candidate subgroups when the case-mix adjusted outcomes were poorer than those of the five top performing ICUs. For 59 ICUs we discovered 122 genuine subgroups and most were defined by one to four variables, with a median of three [2-4]. Variables Glasgow Coma Scale and age were used most. There were 29 ICUs with overall poor outcomes, and for 22 our algorithm found all excess deaths. A new method based on adaptive decision tree boosting discovered many subgroups of ICU patients for which there is potentially room for outcomes improvement.

MeSH terms

  • APACHE
  • Aged
  • Algorithms*
  • Area Under Curve
  • Data Mining
  • Decision Trees
  • Hospital Mortality*
  • Humans
  • Intensive Care Units*
  • Middle Aged
  • Netherlands
  • Prognosis
  • Registries
  • Treatment Outcome