Report Number: CS-TR-96-1564
Institution: Stanford University, Department of Computer Science
Title: Medical Applications of Neural Networks: Connectionist Models
of Survival
Author: Ohno-Machado, Lucila
Date: March 1996
Abstract: Although neural networks have been applied to medical
problems in recent years, their applicability has been
limited for a variety of reasons. One of those barriers has
been the problem of recognizing rare categories. In this
dissertation, I demonstrate, and prove the utility of, a new
method for tackling this problem. In particular, I have
developed a method that allows the recognition of rare
categories with high sensitivity and specificity, and will
show that it is practical and robust. This method involves
the construction of sequential neural networks. Rare
categories occur and must be learned if practical application
of neural-network technology is to be achieved. Survival
analysis is one area in which this problem appears. In this
work, I test the hypotheses that (1) sequential systems of
neural networks produce results that are more accurate (in
terms of calibration and resolution) than nonsequential
neural networks; and (2) in certain circumstances, sequential
neural networks produce more accurate estimates of survival
time than Cox proportional hazards and logistic regression
models. I use two sets of data to test the hypotheses: (1) a
data set of HIV+ patients; and (2) a data set of patients
followed prospectively for the development of cardiac
conditions. I show that a neural network model can predict
death due to AIDS more accurately than a Cox proportional
hazards model. Furthermore, I show that a sequential neural
network model is more accurate than a standard neural network
model. I show that the predictions of logistic regression and
neural networks are not significantly different, but that any
of these models used sequentially is more accurate than its
standard counterpart.
http://i.stanford.edu/pub/cstr/reports/cs/tr/96/1564/CS-TR-96-1564.pdf