Prediction of ICU in-hospital mortality using a deep Boltzmann machine and dropout neural net

The capability to predict in-hospital mortality of patients in intensive care units will be of paramount importance. We explore state-of-the-art machine learning techniques to estimate the in-hospital mortality probability of a patient using various physiological measurements taken within the first forty-eight hours of patient admission. A generative model, a deep Boltzmann machine, is trained using a set of recently developed techniques to automatically extract features from the patient data, and then used to initialize a feed-forward neural network. The neural network is then discriminatively fine-tuned using an efficient approximation to an ensemble of neural networks, dropout, to prevent overfitting on the limited number of labeled training examples.