PhysioNet 2012 Challenge: Predicting mortality of ICU patients using a cascaded SVM-GLM paradigm

Abstract

The focus of the PhysioNet/CinC Challenge 2012 is to develop methods for patient-specific prediction of inhospital mortality using general descriptors recorded at the time of admission to the ICU and up to 37 time-series measurements collected during the first 48 hours after admission. We developed an algorithm that uses both general descriptors and time-series measurements to predict the in-hospital death (IHD) of ICU patients in Event 1, and to provide a probability estimate of IHD in Event 2. Both aggregated variables and general descriptors were used as features of quadratic Support Vector Machine (SVM) classifiers. Six SVMs were trained using, for each one, all the positive examples plus, in turn, one sixth of the negative examples in the training set. Finally, a Generalized Linear Model with probit link was used to predict the probability of IHD for Event 2 using the raw outputs of the six SVMs as regressors. A positive binary prediction of IHD for Event 1 was made when the probability estimate was higher than an optimized threshold. Official final results of the challenge reported that our entry achieved an Event 2 score of 17.88, which is the best score out of the total 23 submissions, and Event 1 score of 0.5345 (second best score). © 2012 CCAL.