A0770
Title: Comparison of diagnostic tests in multireader multicase ROC studies with missing data: A rank-based approach
Authors: Guangyong Zou - Western University (Canada) [presenting]
Abstract: Comparison of diagnostic tests can be done by quantifying differences between areas under the receiver operating characteristic curves (AUCs) using the multireader multicase (MRMC) design, where each case undergoes each of several diagnostic tests and the resulting images are interpreted by each of several readers who are blinded to the true disease status. This design is required for the clinical evaluation of computer-aided diagnostic devices and imaging diagnostic modalities by regulatory agencies. Procedures for confidence interval estimation are under-developed for MRMC, especially in the presence of missing data. We propose a rank-based approach to confidence interval estimation for AUCs and their difference. We first transform each observation into a case-specific accuracy value based on the difference between its rank among all observations in the reader-test combination and its rank within its disease status group. We then analyze these case-specific accuracies using regression models for clustered data to obtain point and variance-covariance estimates of the AUCs. Asymptotic confidence intervals for test-specific AUCs and their difference are constructed under logit- and inverse hyperbolic tangent (artanh)-transformation, respectively. Simulation results based on real studies suggest that our method performed very well in terms of coverage and empirical power, even in the presence of missing data.
