How to best test suspected cases of COVID-19: an analysis of the diagnostic performance of RT-PCR and alternative molecular methods for the detection of SARS-CoV-2

As COVID-19 testing is rolled out increasingly widely, the use of a range of alternative testing methods will be beneficial in ensuring testing systems are resilient and adaptable to different clinical and public health scenarios. Here, we compare and discuss the diagnostic performance of a range of different molecular assays designed to detect the presence of SARS-CoV-2 infection in people with suspected COVID-19. Using findings from a systematic review of 103 studies, we categorised COVID-19 molecular assays into 12 different test classes, covering point-of-care tests, various alternative RT-PCR protocols, and alternative methods such as isothermal amplification. We carried out meta-analyses to estimate the diagnostic accuracy and clinical utility of each test class. We also estimated the positive and negative predictive values of all diagnostic test classes across a range of prevalence rates. Using previously validated RT-PCR assays as a reference standard, 11 out of 12 classes showed a summary sensitivity estimate of at least 92% and a specificity estimate of at least 99%. Several diagnostic test classes were estimated to have positive predictive values of 100% throughout the investigated prevalence spectrum, whilst estimated negative predictive values were more variable and sensitive to disease prevalence. We also report the results of clinical utility models that can be used to determine the information gained from a positive and negative test result in each class, and whether each test is more suitable for confirmation or exclusion of disease. Our analysis suggests that several tests exist that are suitable alternatives to standard RT-PCR and we discuss scenarios in which these could be most beneficial, such as where time to test result is critical or, where resources are constrained. However, we also highlight methodological concerns with the design and conduct of many included studies, and also the existence of likely publication bias for some test classes. Our results should be interpreted with these shortcomings in mind. Furthermore, our conclusions on test performance are limited to their use in symptomatic populations: we did not identify sufficient suitable data to allow analysis of testing in asymptomatic populations.