A comparison of ICU mortality prediction using the Apache II scoring system and artificial neural network

Abstract

Compares the calibration and predictive power of the Acute Physiological and Chronic Health Evaluation II (APACHE II) mortality predictive system with an artificial neural network (ANN) constructed using the same input physiological variables. The study was conducted using the database collected for the UK APACHE II study from October 1987 to April 1989. The patient records were randomly split into three data sets: a training set which was used to train the ANN; a training test set which was used to optimise the performance of the ANN under construction; and a validation test set which was used to validate the performance of the final ANN. The input variables consisted of the 12 physiological variables as well as age, post-emergency operation status, chronic health history and the Glasgow coma score. The worse values among the initial, highest and lowest values of the APACHE II data records collected were used as input values to the network. The classification output of the ANN was the mortality status on discharge from hospital. The backpropagation ANN with 16 input nodes; 35 nodes in the first hidden layer; 10 nodes in the second hidden layer and 1 output node, using extended delta-bar-delta learning rule and sigmoid transfer function gave the best overall results.