A neural network model for mortality prediction in ICU

Scoring the severity of illness of ICU patients can provide evaluation of a patient´s situation and thus help doctors make decisions on what treatment to take. This study aimed to develop an artificial neural network model for patient-specific prediction of in-hospital mortality. Data from PhysioNet Challenge 2012 was used. 12,000 records were divided to a training set, a test set and a validation set, each of which contains 4000 records. Outcomes are provided for the training set. A neural network model was developed to predict the risk of inhospital mortality using various physiological measurements from the ICU. Twenty-six features were selected after a thorough investigation over the different variables and features. A two-layer neural network with fifteen neurons in the hidden layer was used for classification. One hundred voting classifiers were trained and the model´s output was the average of the one hundred outputs. A fuzzy threshold was utilized to determine the outcome of each record from the output of the network. Our model yielded an event 1 score of 0.5088 and an event 2 score of 82.211 on the test data set.