Comparing Several Methods for Estimating Sensitivity and Specificity in Medical Devices

• Activity Number: 388
• Type: Topic Contributed
• Date/Time: Wednesday, August 10, 2005 : 10:30 AM to 12:20 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Wenji Pu*+ and Feng Tang and Duo Zhou
• Companies: Medtronic, Inc. and Medtronic, Inc. and Medtronic, Inc.
• Keywords: Sensitivity ; specificity ; Generalized Estimating Equations ; marginalized conditional proportion estimate ; random-effects hierarchical model ; Bayesian estimate

Abstract

Sensitivity and specificity are two commonly used statistics to quantify the quality of a detection algorithm in medical devices, where observed data are often correlated binary responses (such as indicators telling whether episodes are correctly classified by the algorithm) within each subject with unbalanced numbers of observations across subjects. Currently, Generalized Estimating Equations (GEE) approach, taking the correlation between binary responses within a subject into consideration, has been widely applied to estimate and construct confidence interval for sensitivity and specificity for correlated binary data. However, this approach might be less efficient in the common situations in medical devices with heavy heterogeneity in the number of observations per subject and inignorable variation in the correlation structure for episodes across patients. As alternatives, we propose several other approaches in estimating sensitivity and specificity, including a weighted estimate with weights related to the number of observations per subject, a marginalized conditional proportion estimate from a generalized linear mixed effect model (GLMM), and a Bayesian estimate.