| Home > Publications > PhD theses > Informatics (CW) |
CW 2009_11
Fabian Guiza Grandas
Predictive data mining in intensive care
September 23, 2009
Advisor(s): Maurice Bruynooghe and Hendrik Blockeel
Abstract
Data mining is concerned with the automatic extraction of knowledge from data. Machine learning algorithms perform the data mining process and result in models that can be used for predictive purposes.
Intensive care units are very data-rich environments where the vital functions of patients are continuously monitored. Data from different information sources is collected for each patient, which the physician uses to determine the condition of the patient and to administer the appropriate treatments and medications. Machine learning algorithms can be used to build predictive models based on the intensive care data. These models make use of all the available data to predict the future health-state of the patient, aiding the physician in his choice for better treatments.
In this work we present the use of machine learning algorithms to develop predictive models for the application domain of intensive care medicine.
Because of the particular characteristics of the data available in intensive care, we first introduce some challenging aspects about data mining in this domain. The main part of this dissertation is concerned with the application of machine learning algorithms to several prediction tasks that are performed by the medical staff in intensive care. We first present the use of standard machine learning techniques to predict survival, organ failure and inflammation in intensive care patients. The models learned are found to have similar predictive performance as the medical staff. We then develop a methodology for abstracting the dynamical information of the patient's vital signals and incorporate it in the data mining algorithms. This abstraction method, based on time-series analysis techniques, improves the performance of the predictive models when compared to those that use only static information.
Finally, we present a data mining model that makes use of several information sources in intensive care, including the dynamical aspect of vital signals, in order to predict the moment of a patient's discharge from the intensive care unit. This prediction task is very relevant to physicians because it is indicative of the patient's condition, and to hospital administrators, because it allows for better planning of bed availability and for scheduling of surgical procedures. The data mining model developed results in better predictive performance than the resident medical staff in the intensive care unit.
lirias 237421 / text.pdf ( 11M) / mailto: dtai team