The Application Of Two-part Bayesian Model In ROC Analysis With Covariates When There Is No Gold Standard

Diagnostic tests are important component of clinical researches. Based on screening tests, diagnostic tests can distinguish the patients or the suspects from other illness further. What's more, therapeutic evaluation, prognosis estimation and others are dependent on diagnostic tests outcome at a certain extent. Therefore, searching a way of evaluating diagnostic test effectively can not only provide reliable information for diagnosis and affect doctors' medical plan for patients, but also be the key of therapeutic evaluation and prognosis estimation. A perfect diagnostic test can not just diagnose illness early, but utilize resources efficiently and then avoid resources wasting for unnecessary inspection. As the key of diagnostic tests research, methods about the accuracy evaluation of diagnostic tests have attracted attention of medical researchers recently. The researches on either evaluating only one variable or the joint evaluation of diagnostic tests have been reported. However, few studies have taken more than one covariate into account. Research on the methods of accuracy estimating in diagnostic tests with covariates when there is no gold standard is even a vacancy in China.ROC analysis has been widely recognized as the best composite indicator to evaluate the accuracy of diagnostic tests by medical workers in recent years. Meanwhile, it is one of subfab methods in methods analysis. For the specific characteristic of medicine phenomena, the accuracy of the evaluation in diagnostic tests always can inevitably be affected by various covariates. If we want to evaluate the accuracy of diagnostic tests correctly, it is necessary to consider the influence of covariates. An important assumption of traditional ROC analysis with covariates is that a gold standard test reference test with perfect sensitivity and specificity is available for a test data set. However, a gold standard test may not exist or may be too expensive or impractical to administer. It is difficult or impossible to establish a definitive diagnosis. Therefore, imperfect diagnostic tests are frequently used in ROC analysis. Then the accuracy of the test will often be either under- or overestimated. To find a correct method in estimating ROC curves in such situation, the theory and application of two-part Bayesian model are studied in-depth in this paper.A simulation study is conducted to explore the accuracy of parameter estimations by two-part Bayesian model. The outcome shows: Firstly, the accuracy of parameter estimations improves as the sample size increases, and it is close to the true value when the sample size reaches to 150. Secondly, although there is a little difference between the mean and median of the Bayesian estimate, the median is closer to the true value. Thirdly, comparing the true D and the estimated D, we get the D (%). When the size is less than 100, D (%) is less than 85%. It adds to 90% when the sample size reaches to 150. As the size is 300, D (%) is near to the true value. In conclusion, the parameter estimations of two-part Bayesian model is accurate when we estimate the ROC curves with covariates if there is no perfect reference test for diagnosis. A sample size with more than 150 subjects ensures a higher accuracy.When two-part Bayesian model is used in diagnostic test, such as the coronary heart disease diagnostic test, it can not only deal with the situation that without gold standard test we can not finish the ROC analysis, but also explore the independent or synergistic effect of accuracy which the covariates affect. Combining with the medical practice request, it can control the factors stratified. Furthermore, it is convenient to draw the ROC curve and calculate the corresponding area under the ROC curve, and thus to evaluate and explain the accuracy of a diagnostic test. According to the ROC curve, the optimal operating point can be determined; it can give useful advice to the diagnosis, medical plan establishing, therapeutic evaluation and prognosis estimation and provide theory evidence for the method about accuracy evaluation of diagnostic tests.Two-part Bayesian model is a composite evaluation method that can not only concern the effect of covariates on the diagnosis if there is no perfect reference test but also solve the problem of diagnosis, grouping and therapeutic evaluation. It can be modeled flexibly and interpreted reasonably. It is easy for MATLAB software to program, and not rigidly adhere to the fixed model. This meets the needs of the research very well. Because the model is not specific to the dependent variable distribution, different link function can be used with different distribution. Also the accuracy of a diagnostic variable can directly evaluate by drawing ROC curve and calculate the area with covariates being controlled stratified. Meanwhile, the diagnosis can be implemented by the optimal operating point determination. Thus, it is important to apply two-part Bayesian model in diagnosis.