A Comparison of Multiple Regression and a Neural Network for Predicting a Medical Diagnosis.
Abstract
Regression and neural network prediction methods were compared using artificial data generated to simulate three types of predictor-criterion relationships: linear, polynomial, and interactive. Analyses of linear data indicated that both methods were comparable on large data sets. On small data sets the neural network tended to overfit the initial data and thus did not generalize as well as the regression equation. Analysis of data with a non- linear component demonstrated the ability of the neural network to fit either a polynomial or interactive term without the user having to model such terms. However, when these effects were modeled, the regression equation permored well. The implications of these results for the development of predictive algorithms were discussed. Algorithms, regression equations, predictive algorithms, diagnostic algorithms, neural network.
Document Details
- Document Type
- Technical Report
- Publication Date
- Oct 01, 1991
- Accession Number
- ADA258366
Entities
People
- David H. Ryman
- William Pugh
Organizations
- Naval Health Research Center