Survival prediction of ICU patients using knowledge intensive data grouping and selection

Intensive care unit (ICU) patients are closely monitored by vital signs and other continuous and discrete measures (e.g. lab test). It has been established in prior research that it is possible to predict probability of survival (or death) of an ICU patient ahead of time using vital signs, which allows caregivers extra time and measures to save the patient's life. In this work we are focusing on predicting the survival of ICU patients using lab test results. We have identified several challenges associated with this task, and propose an efficient and scalable data driven solution. Specifically, we propose two complementing techniques for feature reduction and selection. The first one is a knowledge intensive patient grouping scheme, which helps in forming homogeneous groups of data. The second technique uses multiple criteria for selecting the best features from a dataset based on the feature coverage, individual predictive ability as well as inter-dependency among features. We combine these two techniques into a unified framework that strengthens the individual contribution of each technique. We have evaluated our proposed technique on a real ICU patients database and achieved notable success in reducing 89% or more of the feature vector, while improving the prediction accuracy upto 6% and achiving upto 7 times speedup.