Evolving classification of intensive care patients from event data.

Abstract

This work aims at predicting the patient discharge outcome on each hospitalization day by introducing a new paradigm-evolving classification of event data streams. Most classification algorithms implicitly assume the values of all predictive features to be available at the time of making the prediction. This assumption does not necessarily hold in the evolving classification setting (such as intensive care patient monitoring), where we may be interested in classifying the monitored entities as early as possible, based on the attributes initially available to the classifier, and then keep refining our classification model at each time step (e.g., on daily basis) with the arrival of additional attributes. An oblivious read-once decision-tree algorithm, called information network (IN), is extended to deal with evolving classification. The new algorithm, named incremental information network (IIN), restricts the order of selected features by the temporal order of feature arrival. The IIN algorithm is compared to six other evolving classification approaches on an 8-year dataset of adult patients admitted to two Intensive Care Units (ICUs) in the United Kingdom. Retrospective study of 3452 episodes of adult patients (≥16years of age) admitted to the ICUs of Guy's and St. Thomas' hospitals in London between 2002 and 2009. Random partition (66:34) into a development (training) set n=2287 and validation set n=1165. Episode-related time steps: Day 0-time of ICU admission, Day x-end of the x-th day at ICU. The most accurate decision-tree models, based on the area under curve (AUC): Day 0: IN (AUC=0.652), Day 1: IIN (AUC=0.660), Day 2: J48 decision-tree algorithm (AUC=0.678), Days 3-7: regenerative IN (AUC=0.717-0.772). Logistic regression AUC: 0.582 (Day 0)-0.827 (Day 7). Our experimental results have not identified a single optimal approach for evolving classification of ICU episodes. On Days 0 and 1, the IIN algorithm has produced the simplest and the most accurate models, which incorporate the temporal order of feature arrival. However, starting with Day 2, regenerative approaches have reached better performance in terms of predictive accuracy.