Session SA1.5

Predicting Acute Hypotensive Episodes

F Jousset*, M Lemay, JM Vesin

Ecole Polytechnique Fédérale de Lausanne
Lausanne, Switzerland

The goal of the challenge 2009 is to predict which patients will experience acute hypotensive episode (AHE) within a forecast window. AHE may result in dangerous complications and even death. Predicting the onset of AHE would be of clear benefit to a patient's outcome and reduce its stay in intensive care unit (ICU). The challenge comprised two topics, the first focuses on distinguishing between two groups of ICU patients who are receiving pressor medication: patients who experience an AHE (H1), and patients who do not (C1). These two groups represent extremes of AHE-associated risk. The second topic aims to address the broad question of predicting AHE within the forecast window in a population in which about one third of the patients experience AHE (H2 and C2).
In all cases, classification was performed using a support vector machine (SVM) with a linear function kernel, trained with features selected on the training sets. To classify records from training set A belonging to the subgroup H1, the features leading to the lowest error rate were kept. The two features used were the median and the median absolute deviation of the mean arterial blood pressure (MABP) over the two hours preceding the forecast window. Using a leave-one-out technique, 29 out of 30 training records were correctly classified using those two robust statistics. Additional features such as heart rate, respiration rate, systolic or diastolic ABP did not increase the classification rate. To classify records from training set B belonging to the subgroup H2, an additional feature was introduced: the MABP slope over the last 20 minutes. This increased the classification of the training set from 20 to 22 correct classifications out of 30.
When applied to the test sets, we obtained a correct classification of 10 out of 10 for test set A and 29 out of 40 for test set B.
Robust statistics were found to perform better in both events than standard ones. The limited performance in event 2 may be explained by the large variability of the data.

(Abstract Control Number: 267)