Surveillance for TB drug resistance using routine rapid diagnostic testing data: Methodological development and application in Brazil

Abstract

AbstractEffectively responding to drug-resistant tuberculosis (TB) requires accurate and timely information on resistance levels and trends. In contexts where use of drug susceptibility testing has not been universal, surveillance for rifampicin-resistance—one of the core drugs in the TB treatment regimen—has relied on resource-intensive and infrequent nationally-representative prevalence surveys. The expanded availability of rapid diagnostic tests (RDTs) over the past decade has increased testing coverage in many settings, however, RDT data collected in the course of routine (but not universal) use may provide biased estimates of resistance. Here, we developed a method that attempts to correct for non-random use of RDT testing in the context of routine TB diagnosis to recover unbiased estimates of resistance among new and previously treated TB cases. Specifically, we employed statistical corrections to model rifampicin resistance among TB notifications with observed Xpert MTB/RIF (a WHO-recommended RDT) results using a hierarchical generalized additive regression model, and then used model output to impute results for untested individuals. We applied this model to case-level data from Brazil. Modeled estimates of the prevalence of rifampicin resistance were substantially higher than naïve estimates, with estimated prevalence ranging between 28-44% higher for new cases and 2-17% higher for previously treated cases. Our estimates of RR-TB incidence were considerably more precise than WHO estimates for the same time period, and were robust to alternative model specifications. Our approach provides a generalizable method to leverage routine RDT data to derive timely estimates of RR-TB prevalence among notified TB cases in settings where testing for TB drug resistance is not universal.Author SummaryWhile data on drug-resistant tuberculosis (DR-TB) may be routinely collected by National TB Control Programs using rapid diagnostic tests (RDTs), these data streams may not be fully utilized for DR-TB surveillance where low testing coverage may bias inferences due to systematic differences in RDT access. Here, we develop a method to correct for potential biases in routine RDT data to estimate trends in the prevalence of TB drug resistance among notified TB cases. Applying this approach to Brazil, we find that modeled estimates were higher than naïve estimates, and were more precise compared to estimates produced by the World Health Organization. We highlight the value of this approach to settings where testing coverage is low or variable, as well as settings where coverage may surpass existing coverage thresholds, but that could nonetheless benefit from additional statistical correction.