Development of an algorithm for determination of the likelihood of virological failure in HIV-positive adults receiving antiretroviral therapy in decentralized care

Background: Early identification of virological failure (VF) limits occurrence and spread of drug-resistant viruses in patients receiving antiretroviral treatment (ART). Viral load (VL) monitoring is therefore recommended, but capacities to comply with this are insufficient in many low-income countries. Clinical algorithms might identify persons at higher likelihood of VF to allocate VL resources. Objectives: We aimed to construct a VF algorithm (the Viral Load Testing Criteria; VLTC) and compare its performance to the 2013 WHO treatment failure criteria. Methods: Subjects with VL results available 1 year after ART start (n = 494) were identified from a cohort of ART-naïve adults (n = 812), prospectively recruited and followed 2011–2015 at Ethiopian health centres. VF was defined as VL≥1000 copies/mL. Variables recorded at the time of sampling, with potential association with VF, were used to construct the algorithm based on multivariate logistic regression. Results: Fifty-seven individuals (12%) had VF, which was independently associated with CD4 count <350 cells/mm(3), previous ART interruption, and short mid-upper arm circumference (<24cm and <23cm, for men and women, respectively). These variables were included in the VLTC. In derivation, the VLTC identified 52/57 with VF; sensitivity 91%, specificity 43%, positive predictive value (PPV) 17%, negative predictive value (NPV) 97%. In comparison, the WHO criteria identified 38/57 with VF (sensitivity 67%, specificity 74%, PPV 25%, NPV 94%). Conclusions: The VLTC identified subjects at greater likelihood of VF, with higher sensitivity and NPV than the WHO criteria. If external validation confirms this performance, these criteria could be used to allocate limited VL resources. Due to its limited specificity, it cannot be used to determine treatment failure in the absence of a confirmatory viral load.