assessing diagnostic accuracy of doctors without a gold standard using bayesian networks and k-modes clustering algorithm

Background & aim: the diagnostic accuracy of a test is the ability to discriminate accuratelybetween patients who have and do not have the target disease. a common problem in assessing thediagnostic accuracy of doctors is the unknown true disease status which in the literature is referredas “absence of a gold standard”. methods & material: in this article, a naïve bayesian network with hidden class node and a clusteringbased algorithm for categorical data named k-modes are proposed for estimating the diagnosticaccuracy of 5 physicians in diagnosing diabetic retinopathy. also to assess and compare the efficiencies of these models, a simulation study with two different scenarios is conducted. results: simulation study indicates that for naïve bayesian network and the non-rare disease, say forprevalence 0.1 and 0.2, as the sample size increases so the coverage probability. but for high prevalencevalues, say 0.5, coverage probabilities are not as good as those of non-rare disease. k-modes algorithm's efficiency decreases by the increase in the number of records, but it achieves betterresults when there are a small number of records, prevalence is approximately 0.3 and sensitivitiesare high. results of the real data set reveal that sensitivities for all physicians except one, were higher than 85% and all specificities were higher than 90%. also the estimated prevalence happensto be 0.32. conclusion: through simulations and data analysis we show that this new approach based on naïve bayesian networks provides a useful alternative to traditional latent class modeling approaches usedin this setting.