Robust prediction of patient mortality from 48 hour intensive care unit data

The aim of this study was to develop a new algorithm to predict individual patient mortality with improved accuracy with respect to established methods from data collected over the first 48 hours of admission to the Intensive Care Unit. A binary classifier was developed to participate in Event 1 of the PhysioNet/Computing in Cardiology Challenge 2012. The algorithm development was undertaken using only posterior knowledge from the training dataset (Set-A), containing 41 demographic and clinical variables from 4000 ICU patients. For each variable a feature was defined as the average (across all available measurements of the given variable) likelihood of being part of the “survivors” group. To select features with highest discrimination ability (“survivors” vs. “non-survivors”), a forward sequential selection criterion with logistic cost function was adopted and repeated for cross-validation on N (=10) “leave Mout” (M=50%) random partitions of Set-A. Features that were selected in more than one partition were considered (#Feat = 32). A logistic regression model was used for classification. The score was defined as the lowest between sensitivity and positive predictive value in classification. The proposed method scored 54.9% on Set-A and 44.0% on the test set (Set-B), outperforming the established method SAPS-I (29.6% on Set-A, 31.7% on Set-B).