Predicting Probability of Mortality in the Neonatal Intensive Care Unit

Abstract

Artificial neural networks can be trained to predict outcomes in a neonatal intensive care unit (NICU). This paper expands on past research and shows that neural networks trained by the maximum likelihood estimation criterion will approximate the 'a posteriori probability' of NICU mortality. A gradient ascent method for the weight update of three-layer feed-forward neural networks was derived. The neural networks were trained on NICU data and the results were evaluated by performance measurement techniques, such as the Receiver Operating Characteristic Curve and the Hosmer-Lemeshow test. The resulting models applied as mortality prognostic screening tools are presented.

An integrated hybrid data mining systems for preterm birth risk assessment based on a 'semantic web services for healthcare' framework

“…Zhou and Frize's new work using the Canadian Neonatal Network (CNN) database adds risk stratification estimation to the ANN software used for predicting mortality in the Neonatal Intensive Care Unit (NICU) (Zhou & Frize, 2006). Specifically, the MLP ANN with risk stratification estimation incorporates information to stratify patients into groups according to their mortality risk.…”

“…The CNN database contains patient physiological variables, diagnoses, and therapies; traditionally ANNs that include risk stratification estimation include probabilistic information in their input variables. Zhou and Frize explain that frequency analysis can be used to add probabilities to the training set, but that this is laborious and time-consuming (Zhou & Frize, 2006). A preferred method would train the MLP ANN using non-probabilistic variables generally available in clinical databases (e.g.…”

“…Networks trained to minimize the cross-entropy or sum-of-squares cost function can be used to create risk stratification estimation outputs; both closely relate to the maximum likelihood concept (a probability theory method used to estimate the unknown distribution of a population from observed samples) and can be interpreted in a Bayesian learning framework (Zhou & Frize, 2006). The following summarizes the recent work of Zhou and Frize (Zhou & Frize, 2006), explaining the rational for their approach which combines maximum likelihood estimation and neural network learning to approximate risk stratifications for clinical outcomes.…”

Clinical Decision Support System using Real-Time Data Analysis for a Neonatal Intensive Care Unit

“…It is difficult to interpret how an ANN provides the output that it does, since most of the calculations occur in the hidden layers which are not visible to the user of the network. While it is possible to see which inputs contributed most to the final output (Zhou and Frize, 2006), and thus infer which inputs are the most important, it is not easy to explain how the ANN arrived at that conclusion. This method currently requires the Matlab software tool which is proprietary and requires a paid license to use.…”

“…The MIRG at Carleton University and the University of Ottawa in Ottawa, Canada has been actively involved in neonatal medical research. Our research group has developed risk estimation models for a number of clinical events and complications, such as mortality in adult and neonatal intensive care units, length of stay, duration of ventilation, and pre-term birth (Frize et al, 1995Ennett et al, 2004;Zhou and Frize, 2006;Catley et al, 2006;Townsend and Frize, 2008;.…”