Modeling boolean decision rules applied to multiple-observer decision strategies.

Abstract

A model that derives multiple-observer decision strategy ROC curves for boolean decision rules applied to binary decisions of two or three observers is presented. It is assumed that covert decision variables consistent with ROC models of observer performance underlie decisions and that readers' decision criteria are in a fixed relationship. The specific parameters of individual ROC curves and the correlational structure that describes interobserver agreement have dramatic effects upon the relative benefits to be derived from different boolean strategies. A common strategy employed in clinical practice, in which the overall decision is positive if any observer makes a positive decision, is most effective when the readers are of similar ability, when they adopt similar decision criteria, when interreader agreement is greater for negative than for positive cases, and when the individual ROC slope is <<1.0. Different multiple-observer decision strategies can be evaluated using the model equations. A bootstrap method for testing model-associated hypotheses is described.