CinC Challenge: Predicting in-hospital mortality in the intensive care unit by analyzing histograms of medical variables under Cascaded Adaboost model

In this paper, we develop an effective framework to predict in-hospital mortality (IHM) during an intensive care unit (ICU) stay, on the basis of specific medical variables. This work involves both binary mortality predictions and mortality risk estimates, corresponding to Event-1 and Event-2 of the Computing in Cardiology (CinC) Challenge 2012. Our proposed framework contains 1) feature extraction from medical variables by linear interpolation, histogram analysis, and temporal analysis; and 2) mortality classifier learning under Cascaded Adaboost learning model. A released dataset set-a of ICU medical records is used as training set, where cross validation is performed to evaluate our proposed framework. Our framework achieves Event-1 Score1 0.806 and Event-2 Score2 24.00, which outperform those obtained from SAPS-1 score (Score1 0.296 and Score2 68.39) on the same dataset. Over another dataset set-b, our framework obtains Event-1 Score1 0.379 and Event-2 Score2 5331.15.