Tuesday 2 May 2006 at 14h00 in Celestijnenlaan 200 room S01.04
Predictive Data Mining in Intensive Care
By Fabian Guiza (DTAI)
It is crucial in intensive care to detect clinical problems early enough so that preventive or curative treatments can be applied in time. In practice, an intensivist analyses all the patient related data in order to foresee a change in the patient's condition and administer an appropriate treatment. An average ICU patient is estimated to be described by more than 250 different parameters, so it is likely that there is more information in the data than what is currently being extracted from it by humans. Accordingly, data mining could assist clinicians by analysing the ICU data and detecting problems earlier than an experienced intensivist would, and could also be used to generate models that would assist the intensivist in deciding for the best treatment for a specific clinical problem.
Here, we describe an application of data mining methods for different prediction tasks in an intensive care unit. Some of the challenging aspects of performing data mining in this domain are highlighted.
We present a study on an ICU dataset with daily registered data, for which we consider several prediction tasks, some of which have not been previously addressed by machine learning techniques. Inflammation and organ failure are clinical conditions that severely endanger the progress of a patient during his ICU stay. In this study we consider the tasks of predicting: Kidney-dysfunction, Inflammation, Severe-Inflammation and Inflammation-Shock. We used four different data mining algorithms on these tasks: Decision trees, First Order Random Forests, Naive Bayes and Tree Augmented Naive Bayes.
The applied methods result in models with good performances within medical standards.