Early prediction of circulatory failure in the intensive care unit using machine learning

Analysis of the effect of training set size on model performance by artificially subsampling patients at random and retraining the model. This analysis was performed using the circEWS alarm system evaluation policy. We observed that model performance decreases drastically when subsampling to less than 5% of the original training set size, and that the model did not show obvious saturation effects as we move to the full size of the data. A linear model baseline (logistic regression; "LogReg"), a tree-ensemble based method (based on lightGBM, "GBM"; used to construct circEWS), an individual decision tree (based on lightGBM, "DecTree"), and a recurrent neural network ("LSTM") were compared. The Tree models received identical input as given to GBM.