Diagnostic likelihood ratio - the next-generation of diagnostic test accuracy measurement.

Abstract

To select, interpret, and assess the fitness-for-purpose of diagnostic tests, we need to compare the likelihoods of test results being true vs. false across both infected and non-infected individuals. Diagnostic sensitivity (DSe) and specificity (DSp) report the accuracy of classification in infected and non-infected individuals separately and do not compare these likelihoods directly. Positive and negative predictive values combine these likelihoods, but they also heavily depend on the prevalence in the tested populations and, therefore, cannot be generalised. We propose the adoption of the diagnostic likelihood ratio (LR), which balances the likelihoods of true vs. false results and is population independent. As a relative measure, LR ignores the absolute accuracy of tests, and two tests with different accuracy profiles may have the same LR. This can be easily mitigated by using listed complementary measures of accuracy, including DSe and DSp, or ancillary selection criteria. Overall, LR is a more relevant and universal measure of diagnostic test accuracy, which makes it the logical next-generation measure to adopt. We illustrate the applications and benefits of LR using three assays certified by the World Organisation for Animal Health as serological tests for bovine tuberculosis.