Session SA1.1

Predicting the Occurrence of Acute Hypotensive Episodes: The PhysioNet Challenge

F Chiarugi*, I Karatzanis, V Sakkalis, I Tsamardinos,
M Foukarakis, A Dermitzakis, G Vrouchos

Foundation for Research and Technology
Heraklion, Greece

Acute hypotensive episodes (AHE) are serious clinical events in intensive care units (ICU) since they could result in multiple organ failure and eventually in death. The AHE prediction would enable physicians to respond in a timely manner and thus improve patient outcome. In this paper we describe the approach adopted for predicting the AHE occurrence using the dataset made available by the Physionet Challenge 2009.
Each record of the learning set provided by PhysioNet includes clinical data, high resolution signals and time series. The clinical data contain medical encounters performed inside and outside the ICU, while the high resolution signals and the time series contain vital signs sampled respectively at 125 and 1/60 Hz. Each record has a one-hour forecast window (FW) starting at a time T0 and is assigned to a group H or C and to a subgroup H1, H2, C1, or C2. Records in group H contain an AHE starting in the FW with patients in subgroup H1 receiving pressor medication and patients in H2 not. Records in group C contain no AHE within the FW with patients in subgroup C1 without any AHE during their hospital stay and patients in C2 with AHE outside the FW. Being that the high-resolution data is not always homogeneous, nor present, a first analysis was based on the exploitation of the always existing time series (systolic, mean and diastolic BP, and heart rate) in the available (max 24) hours before T0. Linear interpolation was used to fill in the empty gaps and a median filter was applied to smooth the artefacts. Several features were extracted and a small subset, based only on the BP time series, was successfully used to build the H1/C1 classifier (event 1). The first version of the H/C classifier (event 2) was built using a larger set of features but obtaining less accuracy than the H1/C1 classifier.
In the event 1 learning set (15 H1, 15 C1), all cases were correctly classified. The first try on 16/04/09 obtained a score of 80% on the event 1 test set (10 records). Further adaptations in the algorithm, maintaining the same selected features and performances on the learning set, produced on the second try on 27/04/09, a perfect score on the test set. In the event 2 learning set (30 H, 30 C) 83.3% of the cases were correctly classified. The algorithm assessment on the event 2 test set (40 records) obtained, at the first try on 27/04/09 and for the time being, a score of 72.5%.
The perfect score obtained on event 1 has focused us on event 2 which was more difficult since the records are less separated in the feature space than in event 1 where classification is between records with severe AHE and without any AHE. In order to improve the performance, these action lines will be considered: enlarge the set of features extracted by the time series, consider the clinical data for patient profiling, assess features extracted from the high-resolution data (where available), use techniques for feature selection and optimal classification.

(Abstract Control Number: 273)