Receiver Operating Characteristic (ROC) analysis is recognized widely as the best way of measuring and specifying the accuracies of diagnostic procedures, because it is able to distinguish between actual differences in discrimination capacity, on one hand, and apparent differences that are due only to decision-threshold effects, on the other. Key methodological needs remain to be satisfied before ROC analysis can address all of the practically important situations that arise in diagnostic applications, however. This project employs signal detection theory and computer simulation to address several of those needs, by: (1) refining and continuing distribution of software developed previously by the applicants for fitting ROC curves and for testing the statistical significance of differences between ROC curve estimates; (2) developing and evaluating new algorithms for ROC curve-Fitting and statistical testing, based on their recently-developed """"""""proper"""""""" binormal model, that should provide more meaningful results in experimental situations that involve small samples of cases; (3) investigating the usefulness of a form of ROC methodology that is based on mixture distributions in order to rduce the need for diagnostic truth in ROC experiments; (4) investigating the effect of case-saple difficulty on the statistical power tests for differences between ROC curves, in order to determine the optimal difficulty of cases that shouldbe studied on rank diagnostic systems; and (5) developing methods for training artificial neural networks (ANNs) to maximize diagnostic accuracy in terms of ROC analysis and signal detection theory.
