Validation of a clinical prediction score to target viral load testing in adults with suspected first-line treatment failure in resource-constrained settings.

BACKGROUND: Although routine viral load (VL) monitoring currently is too costly for poor countries, clinical failure criteria perform poorly. We previously developed an algorithm combining a clinical predictor score (CPS) with targeted VL testing in a Cambodian patient population (derivation population). We now prospectively validate the algorithm in the same clinical setting (validation population), assess its operational performance, and explore its cost-saving potential.
METHODS: We performed a cross-sectional study in a tertiary hospital in Phnom Penh, Cambodia, applying the CPS in adults on first-line antiretroviral treatment for at least 1 year. Treatment failure was defined as a VL >1000 copies per milliliter. The area under the receiver-operating characteristic (AUROC) curve of the CPS to detect treatment failure in the current study population (validation population) was compared with the AUROC of the CPS obtained in the patient population where the CPS was derived from in 2008 in the same study setting (derivation population). Costs related to VL testing and second-line regimens with the different testing strategies were compared.
RESULTS: One thousand four hundred ninety individuals {56.6% female, median age 38 years [interquartile range (IQR): 33-44]} were included, with a median baseline CD4 cell count of 94 cells per microliter (IQR: 28-205). Median time on antiretroviral treatment was 3.6 years (IQR: 2.1-5.1), 45 (3.0%) individuals had treatment failure. The AUROC of the CPS in validation was 0.75 (95% confidence interval: 0.67 to 0.83), relative to an AUROC of 0.70 in the derivation population. At the CPS cutoff ≥ 2, VL was indicated for 164 (11%) individuals, preventing inappropriate switching to second line in 143 cases. Twenty-four cases of treatment failure would be missed. When applied in routine care, the AUROC was 0.69 (95% confidence interval: 0.60 to 0.77). Overall 1-year program costs with targeted VL testing were 4-fold reduced.
CONCLUSIONS: The algorithm performed well in validation and has cost-saving potential. Further studies to assess its performance, feasibility, and impact in different settings are warranted.