Statistical methods to guide clinical decision-making in tuberculosis

Evaluation of new diagnostic tests is complicated by the lack of a perfect reference standard. Often, this is handled by comparison to one or a composite of the available imperfect reference tests. This may, however, yield biased estimates of prevalence and diagnostic test accuracy. Limitations in the diagnostic tools also introduce uncertainty in clinical decision-making. When confronted with diagnostic uncertainty and a decision on whether to start treatment or not, clinicians consider the potential harm and benefit of offering versus withholding treatment. This could result in increased harm of false-positive and false-negative treatment decisions. To minimize the expected harm, treatment decisions may be guided by the concept of Therapeutic Threshold (ThT) i.e., the probability of disease that needs to be exceeded for treatment to be offered. Using Bayesian latent class analysis (LCA), we estimate tuberculosis (TB) prevalence and diagnostic test accuracy in the absence of a perfect reference standard. Standard LCA assumes that (i) given the true unknown TB status, the diagnostic test results are independent, and (ii) the diagnostic test sensitivity and specificity remain constant across different subpopulations. These assumptions are, however, violated for pulmonary TB diagnostic tests. Hence, standard LCA yields biased estimates. We relaxed these assumptions by allowing conditional dependence between certain subsets of diagnostic tests based on similar biological mechanisms. When relaxed, the number of parameters to estimate may exceed the number estimable, rendering the model nonidentifiable. Under the Bayesian framework, we overcame this limitation by working with experts in TB to restrict the underlying dependencies to those known clinically to be of non-negligible magnitude and to elicit informative priors for some model parameters. We also incorporated measured covariates in the model to allow for varying diagnostic test characteristics across the subpopulations defined by measured covariates known to affect TB prevalence and diagnostic test performance. We estimated the ThT for TB against which clinicians can compare post-test probabilities to decide if the patient with certain demographic characteristics, plus symptoms and test results, meets the requirement for TB treatment. We adapted the nominal group technique, a small-group consensus-building exercise, to estimate ThT in TB in the clinical and community settings of Lesotho and South Africa. We compared the findings based on this method to those based on clinical vignettes. The two methods produced similar and reliable estimates of overall ThT. We further developed a clinical decision-making tool to help clinicians estimate and update the post-test probability of TB. Using ThT can minimize the expected harm of false-positive and false-negative treatment decisions.