HIV Testing After a First Positive Rapid Diagnostic Test: A Role for Nucleic Acid Testing?

We developed an open-access, Excel-based model simulating currently recommended and alternative algorithms for adult HIV testing as a preliminary investigation of trade-offs between accuracy and costs. Despite higher costs, simpler HIV testing algorithms incorporating point of care nucleic acid testing may improve outcomes and thus merit additional research and field testing.

Toward the goal of diagnosing 90% of people with HIV by 2030, the World Health Organization (WHO) outlined an adult HIV testing algorithm based on rapid diagnostic tests (RDTs). However, uptake of the current WHO algorithm is problematic—less than 20% of country-level HIV testing protocols follow WHO recommendations [1, 2], likely due to the algorithm’s complexity and number of required assays [3]. For example, up to 8 RDTs may be needed (Supplementary Figure 1: discordant RDT results, low prevalence settings).

Algorithm accuracy is also concerning: program audits report many false-positive (range, 0%–10%) [4-8] and false-negative results (7%) [4, 9]. Incorrect results have a substantial impact on individuals and lead to errors in estimation of program requirements and costs. Point of care (POC) nucleic acid testing (NAT) may facilitate antiretroviral therapy (ART) initiation by permitting same-day HIV diagnosis but may have higher per-assay costs. As a preliminary investigation of accuracy and costs, we modeled simpler alternative algorithms incorporating POC-NAT, compared with the current RDT-based WHO HIV testing algorithm [1].

METHODS

Using an Excel-based model, we simulated 3 currently recommended and novel alternative algorithms for adult HIV diagnostic testing (Supplementary Figure 1): current RDT-based WHO guidelines; POC-NAT to resolve discordant RDT results (NAT-resolve); and POC-NAT to confirm a first positive RDT result (NAT-confirm). In the online Excel tool, model-users may specify HIV prevalence; assay sensitivity, specificity, and costs; and the lifetime per-person cost of HIV care and ART (“care/ART cost”) to reflect a range of settings. We assumed conditional independence across tests (ie, that there is no reason 1 false-positive will result in another) [3, 10]. In practice, patients who receive a final “inconclusive result” should repeat the entire algorithm; in the model in these cases, we applied an average per-person algorithm assay cost again and assigned results for repeat testing based on true infection status. We assumed that the NAT-resolve and NAT-confirm algorithms would not include any pre-ART retesting, based on low uptake of retesting recommendations in practice [3].

In the base case, we reflected high-burden, low-income settings: 15% HIV prevalence and lifetime care/ART costs of $4000, estimated from model-based analyses (Supplementary Table 1) [11]. We assumed that after false-positive diagnosis, patients incurred lifetime care/ART costs equivalent to after true-positive diagnosis; after false-negative diagnosis—with later risk of costly opportunistic infections or transmission of HIV without ART—patients incurred HIV care costs equal to 75% of lifetime care/ART costs [12]. Assay sensitivity/specificity/cost were derived from averaging WHO Prequalification Reports and from UNICEF (RDT: 99.8%/99.6%/$1.47; POC-NAT: 95.0%/100%/$27.92) [13, 14]. Outcomes included accuracy ([true-positive + true-negative results]/all results), number and costs of assays required, and care/ART costs. We excluded all clinical outcomes, including morbidity, mortality, onward transmissions and related costs, associated with true- and false-positive and -negative results.

In scenario analyses, we simulated combinations of prevalence and cost to reflect diverse settings: high-prevalence (15%) and low-prevalence (3%), with middle-income ($20000) and low-income ($4000) lifetime care/ART cost. We also simulated a field-based setting, with lower test sensitivity/specificity, as might occur when implemented in the field (RDT: 89.0%/82.9%; POC-NAT: 94.0%/99.4%) [5, 15].

In sensitivity analyses, we examined a range of parameters, including the cost of false-negative diagnoses (up to 7-fold true-positive costs) and specific assays (up to 30-fold base case). Additional model inputs, assumptions, and testing algorithms (including distinct WHO algorithms for low-prevalence [<5%] or high-prevalence [≥5%] settings) are available in the Supplementary Appendix.

RESULTS

Overall Accuracy

In the base case, NAT-resolve had similar accuracy as the WHO algorithm (99.97%); NAT-confirm was least accurate (99.23%) (Table 1).

Table 1.

Base Case and Setting Analyses: Modeled Outcomes of 3 HIV Testing Strategies

False-Negative Results

NAT-confirm led to the most false-negative diagnoses (774/100 000) due to low POC-NAT assay sensitivity (94.00%). When POC-NAT was instead used to resolve discordance between 2 RDTs (NAT-resolve), the effect of the NAT’s lower sensitivity was mitigated by the greater sensitivity of the second RDT (99.83%): 27/100 000 were false-negative results, similar to 26/100 000 for the WHO algorithm (Table 1).

False-Positive Results

NAT-confirm led to 0% false-positive results. Among the other strategies, false-positive results were similar (1–3/100 000 false-positive results) (Table 1).

Assay Utilization and Costs

On average, the WHO algorithm required more assays/person compared with all other strategies (1.5 vs 1.2) (Table 1). Costs attributable to the assays alone were greatest for NAT-confirm ($5.75/person tested), whereas assay cost for NAT-resolve ($1.80/person tested) was closer to, but still lower than, the WHO algorithm ($2.15/person tested). The total algorithm cost (including lifetime care/ART) for NAT-resolve was similar to the WHO algorithm, whereas NAT-confirm cost less than the WHO algorithm by $4/person tested.

Sensitivity Analyses

Holding all other parameters equal to the base case, NAT-confirm became more costly than the WHO algorithm when the POC-NAT cost was >$55 (Supplementary Table 2) or when lifetime HIV care cost after false-negative diagnosis was equal to or exceeded lifetime care/ART cost after true-positive diagnosis (Supplementary Figure 2A).

Alternative Prevalence and Income Settings

In low-prevalence settings, NAT-confirm and NAT-resolve still required fewer assays/person than the WHO algorithm: 1.0 vs 1.2 assays/person tested (Table 1). In both high- and low-income settings, NAT-resolve had accuracy and total algorithm cost similar to the WHO algorithm, and NAT-confirm was less accurate and less expensive. For all algorithms, overall costs varied widely from lifetime ART/care cost ($121–$3001).

Field-Based Settings

Substituting lower field-based sensitivities/specificities for all assays, NAT-confirm became the most accurate algorithm (97.60% vs 93.10%–95.72%), in contrast to the base case. It was also the least expensive ($590/person vs $691–$698/person) (Table 1). NAT-confirm’s greater overall accuracy included fewer false-positive results (87 vs 4938/100 000) but more false-negative results (2318 vs 1982/100 000) than the WHO algorithm. NAT-confirm remained less expensive than the WHO algorithm, except when the POC-NAT cost was >15-fold higher than the base case ($275) (Supplementary Table 2) or when lifetime HIV care cost after false-negative diagnosis was >7-fold higher than lifetime ART/care cost after true-positive diagnosis (>$30 000) (Supplementary Figure 2B). NAT-resolve became costlier than the WHO algorithm when the POC-NAT cost was >3-fold higher than the base case ($85) (Supplementary Table 2); NAT-resolve was less expensive than the WHO algorithm at all examined costs of false-negative diagnosis (Supplementary Figure 2B) in the field-based setting.

DISCUSSION

We find that compared with the WHO algorithm, POC-NAT-based algorithms (NAT-confirm and NAT-resolve) may require fewer assays and have similar or lower costs. Using WHO Prequalification Report sensitivity/specificity, NAT-resolve had equal accuracy and lower lifetime costs than current WHO algorithms; using field data, however, NAT-confirm was more accurate and less expensive than WHO algorithms.

Despite concerns about assay cost, this was not the most influential parameter: POC-NAT assay cost had to increase 2–15-fold for NAT-based algorithms to become costlier than the WHO algorithm. Importantly, the cost savings with both NAT-based strategies reflect 2 opposing effects: permitting false-negative results (less expensive because treatment is not immediately provided, but risking higher morbidity and mortality) and averting false-positive results (avoiding unnecessary care/ART costs for people without HIV, potentially for life). NAT-confirm is most cost-saving in the base case, but its low cost is driven primarily by failure to identify HIV-infected people (false-negative results)—clearly a harmful outcome of this approach. This focused model excludes key long-term clinical, transmission, and cost outcomes of incorrect diagnoses, and thus is intended to inform further research rather than provide definitive guidance.

Our findings highlight 2 important gaps in the currently available data: First, reproducible estimates of assay sensitivity and specificity in the field are needed. The base case and field scenario results differ due to differences in reported sensitivity/specificity, which are the most influential determinants of model results. Notably, the field-based modeled results for the WHO algorithm closely match published proportions of false-positive (5%) and false-negative (2%) results (Table 1) [4-8]. When applied at scale, decimal differences in sensitivity/specificity have substantial implications: in the high-prevalence base case, for every 10 000 people tested, a 1% decrement in sensitivity causes 14 additional false-negative results, and a 1% decrement in specificity causes 3 additional false-positive results. WHO guidelines assume assay sensitivity/specificity comparable to values in the WHO Prequalification Report—at least 99/98% [1], whereas published ranges reach far lower (89.0/82.86%) (Supplementary Table 1) [5, 8, 13, 15, 16]. Second, clinical studies and detailed simulation modeling of incorrect diagnoses are needed to more fully capture resulting clinical and economic outcomes. In the absence of these data, we excluded important clinical outcomes and their associated costs, but this information will be critical to inform algorithm selection.

In addition to excluding clinical outcomes, this model makes key simplifying assumptions: First, for NAT-based algorithms, we exclude some potential costs (eg, training or equipment costs for implementing and maintaining a new testing algorithm or pre-ART retesting) or additional benefits (eg, increasing access to POC-NAT for early infant diagnosis or enabling same-day confirmed HIV diagnosis and linkage to care and ART). Second, the model assumes perfect algorithm fidelity; as a result, the benefits of complex WHO algorithms may be overestimated. Finally, assays may not be conditionally independent as modeled [3], and some countries may use assays with widely differing sensitivity and specificity for sequential or confirmatory testing, which may change the overall accuracy and costs (Table 1) [1, 8].

High assay costs are often cited as a rationale for not exploring NAT-based algorithms. We find that assay cost has minimal impact on overall algorithm costs relative to assay sensitivity and specificity and the costs of lifetime HIV care and ART. Detailed assessment of assay performance and cost in the field, as well as of the clinical harms and costs associated with false-positive and false-negative results, is needed to inform future guideline development. Simpler HIV testing algorithms incorporating POC-NAT may improve outcomes and thus merit additional research and field testing. In the meantime, we demonstrate that HIV testing algorithms integrating POC-NAT should not be dismissed simply because of their perceived high per-test cost before algorithm fidelity, accuracy, and downstream costs are also critically examined.

Supplementary Data

Supplementary materials are available at Open Forum Infectious Diseases online. Consisting of data provided by the authors to benefit the reader, the posted materials are not copyedited and are the sole responsibility of the authors, so questions or comments should be addressed to the corresponding author.

Click here for additional data file.