Session S91.5

An RPS-NN Approach toward the Prediction of Acute Hypotensive Episodes

MA Mneimneh*, RJ Povinelli

Marquette University
Milwaukee, WI, USA

The 2009 Physionet/Computers in Cardiology Challenge is to develop an automatic technique for the prediction of acute hypotensive episodes. Such episodes if left untreated can lead to organ failure and death. However, acute hypotensive episodes can be treated depending on their cause and if diagnosed in timely manner. This study uses a Reconstructed Phase Space-Neural Network technique for the prediction of the acute hypotensive episodes. The challenge is divided into two parts. The first is to distinguish between patients who have experienced acute hypotensive episodes and patients who have not. The second is to predict acute hypotensive episodes. The records used for this challenge are from the MIMIC II database, which are divided into training and testing. The training set consists of 40 records divided equally into four groups of patients, acute hypotensive episodes treated with pressors, acute hypotensive episodes not treated with pressors, and no acute hypotensive episodes, and acute hypotensive episodes outside the forecast window. The testing set for part one of the challenge consists of 10 records. The number of records in the testing set of the second part of the challenge is 40. Our approach uses a reconstructed phase space (RPS) combined with a neural network predictor to solve both challenges. The blood pressure signals are embedded in an RPS. A neural network is used to learn a model of the embedded blood pressure signal for each case. In the testing phase, a similar process is followed, but now the neural network is used as a predictor. The challenge score is calculated as the fraction of correct classifications. Preliminary results on the challenge are 0.2 and 0.6 for part 1 and part 2, respectively.

(Abstract Control Number: 269)