Session S91.2

Forecasting Acute Hypotensive Episodes in Intensive Care Patients Based on a Peripheral Arterial Blood Pressure Waveform

X Chen*, D Xu, G Zhang, R Mukkamala

Michigan State University
East Lansing, MI, USA

We entered the PhysioNet/Computers in Cardiology Challenge 2009 to predict which intensive care patients would experience an acute hypotensive episode (AHE) within a predefined forecast window using various physiologic data prior to the window. We took a pragmatic approach to the challenge. That is, since an AHE was defined by the level of mean arterial pressure (MAP), we investigated the predictive value of the MAP level immediately before the forecast window. We arbitrarily chose MAP over the preceding 5-min interval for AHE prediction. In the training dataset, this simple index performed well in distinguishing between patients who had an AHE (H group) versus those who did not (C group) (receiver operating characteristic area under the curve (ROC AUC) of 0.75) and even better in discriminating between patients who had an AHE with pressor treatment (H1 group) versus those with no documented AHE (C1 group) (ROC AUC of 0.86). For the Event 1 testing dataset, we classified the five patients with the lowest 5-min MAP levels in the H1 group and the remaining five patients in the C1 group. This entry achieved a perfect score of 10/10. For the Event 2 testing dataset, we classified the 16 patients with the lowest 5-min MAP levels in the H group and the remaining 24 patients in the C group. The number of patients in each group here was determined by maximizing the difference between the highest 5-min MAP level in the H group and the lowest 5-min MAP level in the C group while having the required 10-16 patients in the H group. This entry achieved a score of 35/40 (87.5%). We then selected the time duration and weighting of the window for computing MAP by maximizing the ROC AUCs in the training dataset. This optimal index only marginally improved the ROC AUCs and actually degraded the Event 2 testing dataset result. We did also study other indices including cardiac output by peripheral arterial blood pressure waveform analysis, heart rate spectral powers, and premature beats. None of these indices appeared to provide added predictive power in the training dataset. However, some of these indices could offer predictive value in other clinical scenarios.

(Abstract Control Number: 276)