Multivariate approach to predictive diagnosis of bone-marrow iron stores

Abstract

The authors constructed a computer-assisted multivariate data analysis system that finds the subset of a group of laboratory tests (i.e., experimental variables) best able to discriminate between previously identified subpopulations. A Bayesian decision-analysis algorithm was devised for predictive diagnosis by the use of sample means and standard deviations of those identified tests. This approach was used on a set of nine hematologic laboratory tests to find a subset capable of efficiently predicting the presence or absence of bone-marrow iron stores. This identified subset proved to have a diagnostic efficiency of 0.90. The success of this system suggests the feasibility of such an approach in finding a "best" subset of loosely related variables.