Background: HIV viral load (VL) monitoring is recommended globally for patients on antiretroviral therapy (ART). However, cost and infrastructure in resource limited settings (RLS) still limit widespread implementation of this guideline.
VL pooling reduces the number of assays needed to ascertain individuals’ VL. However, existing pooling methods ignore information from routinely-collected clinical markers (RCMs), such as CD4 count and percent. RCMs are correlated with VL and provide valuable information that can be utilized to improve the efficiency of pooling by reducing the number of assays needed for deconvolution.
Methods: We develop a deconvolution algorithm that uses RCMs to determine an optimal individual sample testing sequence for resolving a positive pool (e.g. having a detectable VL). We first use RCMs in a prediction model to estimate risk of viral failure for each sample in a positive pool. Individual samples are then assayed sequentially in a descending order of their risk scores, following a stopping algorithm that dictates whether subsequent assays of remaining samples ‘in the pool’ are needed. Using simulation models informed by clinical data, we examine the potential advantages of our approach by comparing it to (1) individual VL testing; (2) standard pooling (VL assay of pooled samples followed by VL assays of individual samples for a positive pool); and (3) algorithm-based pooling (VL assay of pooled samples, followed by sequential VL assays of individual samples and a stopping rule that depends on updated VL values for samples remaining in the pool [May et al. JAIDS 2010]). Simulations were based on data from 597 women at the Miriam Hospital HIV Clinic (Providence, USA). Markers used for risk prediction included CD4 count, CD4%, and their 6-month changes. Viral failure was defined using three VL thresholds: 150, 400, and 1000 copies/mL.
Results: The cohort had a median VL 75 copies/mL (IQR 75~400), CD4 count 407 cells/μL (IQR 254-576), and CD4% 23 (IQR 17-30). The Table depicts the expected number of VL assays required for each method and the percent reduction in VL tests compared to individual testing. Standard deconvolution may need more assays depending on the VL threshold. Algorithm-based deconvolution reduces the number of VL tests by 18-35%; our rank-based algorithm further reduces VL assays by 26-45%.
Conclusions: Rank-based deconvolution based on RCMs can substantially increase the capacity of comprehensive VL monitoring, suggesting its potential utility in RLS.