Adding probability to neural network prediction of NICU mortality

Abstract

It has been shown that neural networks can be trained to predict clinical outcomes of a neonatal intensive care unit (NICU). This thesis expands past research and presents a neural network approach that predicts the probability of NICU mortality. The resulting neural network models are able to not only classify the dichotomous outcome (i.e. survival or death) but also estimate the probability of death. The general conditions necessary for the neural network to estimate probabilities are discussed in the thesis. The neural network estimation of mortality probability includes the following steps: modeling the neural network cost function with the likelihood function; deriving the gradient ascent training algorithm to perform the maximum likelihood estimation; developing the neural network models with the NICU data; evaluating performance by the Receiver Operating Characteristic Curve and Hosmer-Lemeshow test. These neural network based probability estimation models applied as the mortality prognostic tools are presented. For this purpose, two approaches for improving the models' sensitivity are suggested: adjustment of the cost function and the cutoff point. Both of them were tested and results are discussed.