Acute hypotension episode prediction using information divergence for feature selection, and non-parametric methods for classification

Acute hypotension is a critical event that can lead to irreversible organ damage and death. When detected in time, an appropriate intervention can significantly lower the risks for the patient. The objective of this work is to describe an automated statistical method that produces an automated method to predict acute hypotension episodes, using the least data possible. We first detailed the problem of having more features than samples in the PhysioNet/CinC Challenge 2009 training set. We constrained our analysis to the largest common subset of features available for all patients (arterial blood pressure measurements). We then used information divergence (or Kullback-Liebler divergence) between two distributions to identify the most discriminative features. We used these features in each training set to classify the samples in the test sets using a nearest neighbors (NN) algorithm. With this method, we obtained a score of 9/10 for event 1, and 32/40 for event 2 compared to a control method which gives us 10/10 for event 1, and 35/40 for event 2. Our preliminary results showed that our method leads to significantly better than random results, therefore it increases our information about the samples in the test sets.