Search

Multilayer neural networks have been faulted for functioning as “black boxes” and for failing to assess the relative importance of the input factors. The aim of this paper is to illustrate how neural networks can classify individuals. The authors investigated the role of weights in the formation of neural networks‘ decision surfaces and decision regions. The data used were from a case-control study. Two strong determinants of case status were used as input “neurons.” Zero, three, and five hidden neurons were used to explore the effect of the number of hidden neurons on the decision surfaces and regions. Mapping of input and output spaces revealed that three hidden neurons were insufficient to fully discriminate cases from controls. Five hidden neurons may be optimal, but at the cost of possible over-fitting. The more complex neural networks were very effective at defining regions of uniform risk in the plane of the initial covariates, and at assigning risk levels. The authors speculate that neural networks will prove useful in epidemiologic problems that require pattern recognition or complicated classification techniques, and that they will be unfavorable in problems that involve distinct effects of distinguishable predictors. Am J Epidemiol 1998; 147: 1112–22. [ABSTRACT FROM PUBLISHER]

The authors developed and cross-validated prediction models for newty diagnosed cases of liver disorders by using logistic regression and neural networks. Computerized files of health care encounters from the Fallon Community Health Plan were used to identify 1,674 subjects who had had liver-related health services between July 1,1992, and June 30, 1993. A total of 219 subjects were confirmed by review of medical records as incident cases. The 1, 674 subjects were randomly and evenly divided into training and test sets. The training set was used to derive prediction algorithms based solely on the automated data; the test set was used for cross-validation. The area under the Receiver Operating Characteristic curve for a neural network model was significantly larger than that for logistic regression in the training set (p = 0.04). However, the performance was statistically equivalent in the test set (p = 0.45). Despite its superior performance in the training set, the generalizability of the neural network model is limited. Logistic regression may therefore be preferred over neural network on the basis of its established advantages. More generalizable modeling techniques for neural networks may be necessary before they are practical for medical research. Am J Epidemiol 1998; 147: 407–13. [ABSTRACT FROM AUTHOR]

A review of the article "The Failure of Academic Epidemiology: Witness for the Prosecution," by C.M. Shy in a previous issue is presented.