Bayesian Methods For Evaluation Of Diagnostic Accuracy Of Quantitative Tests And Disease Diagnosis In The Absence Of A Perfect Reference Standard With Examples From Johne's Disease.

摘要

Diagnostic tests are widely used for disease screening and diagnosis, assessment of prognosis and treatment selection. In this work, Bayesian methods and models are developed to improve assessment of the performance of continuous tests without a perfect reference standard.;Methods are developed for Bayesian estimation of the receiver operating characteristic (ROC) curve for normally- or gamma-distributed scores with a limit of detection without a reference standard. A mixture model is proposed for scores of diagnostic tests that are multimodal. Both censoring and truncation models are discussed and further studied in simulation studies. Findings indicated that the methods provided relatively accurate estimation of the area under ROC curve (AUC), except for very high percentage of censoring (≥ 60%) or for tests with poor accuracy (AUC ≤ 0.6). A truncated gamma with a point mass is used to model quantitative real-time polymerase chain reaction (qPCR) assay data for Johne's disease, and the ROC curve and true prevalence are estimated.;Second, to assess the performance of multiple correlated continuous diagnostic tests, a Bayesian ROC-based method is developed for diagnosis based on combined tests with no reference standard. A random-effects model is proposed for correlated scores from multiple tests. Using Bayesian probability modeling, scores from multiple tests are used to create a new diagnostic criterion based on a threshold for predictive probabilities, where the AUC for the diagnostic criterion's ROC curve, namely cAUC, is used as an accuracy metric for combined tests. Simulations indicated that the cAUC is estimated relatively accurately for varying degrees of correlation among scores except in the extreme case where 'highly correlated' simulated scores, generated to be inconsistent with the model, were used to estimate cAUC without reference standard information. The methods are applied to results of three enzyme-linked immunosorbent assays (ELISA) for Johne's disease.;Finally, a Bayesian model is proposed to estimate true aggregate-level prevalence using an imperfect test without reference standard information. The model allows adjustment for variable sensitivity in different latent (unobserved) sub-populations, and is used for estimation of herd-level prevalence of Johne's disease in a study of U.S. dairies in 2007 using composite fecal (environmental) samples.