Interactive Computer Program for Selecting Variables in Medical Diagnosis, Based on Information Theory

Abstract

In evaluating the data obtained by physiological experiments (multivariable systems) and clinical observations in intensive care units, criteria have to be established based on both quantitative (instrumental) and qualitative (clinical observations) data to determine which are the most relevant variables among the bulk of results we get. We demonstrate an interactive computer program using an IBM 1800, in an experimental monitoring and a clinical diagnostic system. With the help of information theory we are able to compute the different degrees of relationships between arbitrary combinations of observables which lead to the elimination of redundant variables with the estimate of percentage of information lost by such a reduction. A relevant subset of observables is derived from this method, for each collection of experiments.