The analytic framework in the Figure shows the target populations, interventions, and intermediate and health outcome measures we examined. The analytic framework was developed in consultation with the USPSTF and was refined after review by 6 content experts. We considered screening to be testing for HIV infection in asymptomatic persons or those with mild, nonspecific symptoms (such as fatigue) that are not predictive because they are so common. We excluded children (<13 years of age) because the prevalence of HIV in this population is low (9.3 per 100 000 population) and because most were infected vertically.3 We excluded other specific populations such as patients who had undergone transplantation, patients with known chronic viral hepatitis, and patients undergoing hemodialysis. In these groups, treatment considerations, adverse effects from treatment, and natural history may differ from those in the general population of HIV-infected persons; such patients are also usually excluded from clinical trials. We excluded patients with occupational exposures and blood donors because of consensus regarding testing for HIV infection in these situations. We excluded studies of HIV-2 infection because it is rare in the United States and its natural history differs substantially from that of HIV-1 infection.
Our review considered the standard screening strategy for HIV-1 infection to be an office-based venipuncture for anti-HIV enzyme-linked immunosorbent assay, followed by confirmatory Western blot for positive test results.46,245 We also considered rapid tests, home-based sampling, and tests using saliva or urine specimens. Viral load plus CD4 cell count testing was considered the standard work-up to determine the stage of infection and eligibility for interventions in infected patients.13,14,16,245
We evaluated recommended HAART regimens, prophylaxis against opportunistic infections, immunizations, Papanicolaou testing, counseling to reduce risky behaviors, and routine monitoring and followup. We excluded interventions not recommended for antiretroviral-naive patients or those not known to be effective. These include enfuvirtide; structured treatment interruptions; sequential initiation of therapy with antiretroviral drugs; induction-maintenance regimens; hydroxyurea; interleukin-2; acyclovir; and prophylaxis against candidiasis, histoplasmosis, coccidioidomycosis, herpes simplex virus infection, or cryptococcosis.13,16 We also did not consider resistance testing in antiretroviral-naive patients to be a routine intervention. Although the presence of primary antiretroviral drug resistance is increasing, resistance testing has mainly been studied in patients in whom a regimen has already failed. In patients with untreated chronic HIV infection, current U.S. guidelines either do not recommend routine resistance testing13 or do not give firm recommendations.247
For outcomes, we were particularly interested in reviewing literature on the benefit of early interventions in asymptomatic, treatment-naive patients. Clinical outcomes that we evaluated were mortality, AIDS-related opportunistic infections, spread of disease, and quality of life or functional status. For counseling, we included rates of sexually transmitted diseases as clinical markers of high-risk behaviors. Intermediate outcomes were loss of detectable viremia, improvement in CD4 cell counts, and changes in risky behaviors. We also reviewed harms from screening, work-up, and treatment. For harms from treatment, we focused on the long-term risk for cardiovascular complications and intolerable (causing discontinuation of therapy with the drug) side effects from HAART. Although interventions for chronic HIV infection, particularly HAART, are associated with many clinically significant short-term side effects, many are tolerable or patients can be switched to effective alternative regimens. In addition, intention-to-treat analyses of clinical outcomes incorporate the effects of intolerable or serious side effects.248 We did not include antiretroviral resistance as a separate outcome because its effects are seen in other intermediate (CD4 cell count, viral load) and clinical outcomes.
We searched the topic of HIV in the MEDLINE® and Cochrane Library databases. Most searches were done from 1983 (the year that HIV was characterized) through 30 June 2004. For searches on antiretroviral therapy, we electronically searched these databases from 1998, the year that HAART was first recommended in U.S. guidelines;249 we supplemented these searches by an electronic search for systematic reviews of antiretroviral therapies from 1983. We performed a total of 13 searches covering the areas of risk factor assessment, screening tests, work-up, and interventions. Appendix B presents detailed electronic search strategies and results. Periodic hand searching of relevant medical journals, the Centers for Disease Control and Prevention Web site, and reviews of reference lists supplemented the electronic searches.
Content experts who reviewed the draft report identified additional citations. For rapid HIV tests, we included unpublished studies reported in manufacturer inserts. Other unpublished material was not included. Abstracts were not included in systematic searches, but major abstracts cited in reference lists or presented at recent conferences were included. We also obtained reviews, policy statements, and other papers with contextual value.
Papers were selected for full review if they were about HIV infection, were relevant to key questions, and met inclusion criteria. We also included cost-effectiveness analyses of HIV screening in outpatient settings in the HAART era. For all key questions, articles were limited to those that evaluated the general adult and adolescent population with chronic HIV infection. We excluded studies that included only overtly symptomatic patients or those with end-stage disease. Although the population of interest was persons with unsuspected HIV infection who would be identified by screening, we included studies of patients with a broad spectrum of chronic HIV disease to get a picture of the effects of screening and treatment in patients with different degrees of immune deficiency. We included studies performed in the United States, Australia, Canada, and countries of western Europe, in which the epidemiology and management of chronic HIV infection are similar. When important studies for a specific key question had been done only in other countries, we included these as well. We excluded studies of nonhuman subjects and those without original data. We considered non-English-language papers if they reported on clinical trials and an abstract was available in English. We searched for relevant systematic reviews for all key questions. A separate report lists additional key question-specific inclusion criteria.17
We used predefined criteria from the USPSTF to assess the internal validity of included systematic reviews, trials, and observational studies, which we rated as "good," fair," or "poor." We also rated the applicability of each study to the population that would be identified by screening. The rating system was developed by the USPSTF and is described in detail elsewhere18 and summarized in Appendix C. For included trials and systematic reviews, we abstracted information about setting, patients, interventions, and outcomes. For intervention studies, when available we abstracted intention-to-treat results in which missing data were classified as treatment failures.248 We rated the overall body of evidence for each key question using the system developed by the USPSTF. We also rated studies evaluating cost-effectiveness of HIV screening in the HAART era using criteria developed by the USPSTF for evaluation of cost-effectiveness analyses (Appendix C).19
Table 5 estimates the outcomes after 3 years from one-time screening for HIV in 3 hypothetical cohorts of 10 000 adolescents or adults. We limited our time horizon to 3 years because longer studies on the clinical benefits from HAART are not yet available. We excluded areas from this table in which reliable data to estimate the clinical magnitude of benefit or harm were not available, such as harms from screening (anxiety, labeling, violence, suicide, partnership dissolution) and decreased transmission from counseling or other interventions. We also had insufficient data with which to estimate the impact of screening on earlier diagnosis of HIV and the proportion of patients qualifying for different interventions. Because short-term adverse events from HAART are usually self-limited, and effective alternative regimens are usually available, we focused on the long-term cardiovascular harms of HAART. We calculated numbers needed to screen and treat to prevent 1 case of clinical progression (new category B or C event) or death and to cause 1 cardiovascular event (myocardial infarction, invasive cardiovascular procedure, or stroke). Data from clinical trials were insufficient to separate clinical outcomes by severity.
Several assumptions made our estimates on the benefits of screening conservative. First, we focused on the effects of HAART. For some interventions (for example, most immunizations, more frequent Papanicolaou testing, routine monitoring and followup, and counseling), data were insufficient to estimate the magnitude of benefit. For others, such as prophylaxis against opportunistic infections, the magnitude of benefit from HAART substantially outweighs the benefit from other interventions, and successful treatment with HAART would also reduce the proportion of patients requiring prophylaxis by increasing CD4 counts. Second, we assumed that only asymptomatic patients with CD4 cell counts less than 0.200 × 109 cells/L would routinely receive HAART because they are at highest risk for clinical progression, evidence for clinical benefits of treatment is strongest in this group, and recommendations are less firm for asymptomatic patients with higher CD4 cell counts. Third, we estimated benefits only for the first 3 years after screening, although HAART is likely to be beneficial beyond that time period.
Because no clinical trials have directly evaluated the relative risk for clinical progression or death associated with HAART (antiretroviral therapy with 3 drugs) compared to no treatment in HIV-infected persons, we calculated this relative risk indirectly from data provided in a systematic review of clinical trials of 1-drug therapy versus no antiretroviral agents, 2-drug versus 1-drug therapy, and 3-drug versus 2-drug therapy in antiretroviral-naive persons.133 Bucher and colleagues250 proposed a method for indirect treatment comparisons to estimate odds ratios from 2 sets of clinical trials; we adapted this method to calculate the relative risk indirectly from the 3 sets of trials. Bucher and colleagues' method has been shown to usually agree with results of direct treatment comparisons.251 For this calculation, let RRMN, RRDM and RRTD denote relative risk for clinical progression or death on 1-drug therapy versus no antiretroviral drugs, 2-drug versus 1-drug therapy and 3-drug versus 2-drug therapy, respectively. The relative risk for clinical progression or death during 3-drug therapy versus no antiretroviral agents (RRTN) is given by:
RRTN = RRMN × RRDM × RRTD. (1)
To calculate the (1 -α)% Cl for RRTN, it is usual to use the natural log scale:
log(RRTN) = log(RRMN) + RRDM + log(RRTD) (2).
The variance of log relative risk is given as:
Var(log(RRTN)) = var(log(RRMN)) + var(log(RRDM)) + var(log(RRTD)). (3)
by assuming independence among log(RRMN), log(RRDM) and log(RRTD). Since log(RRTN) is approximately normally distributed, the (1 -α)% Cl for RRTN are
(RRTN exp(- Ζα/2sqrt(var(log(RRTN)))), RRTN exp(Ζα/2sqrrt(var(log(RRTN))))). (4)
Jordan and colleagues133 reported the rates for clinical progression or death from clinical trials of 1-drug therapy vs. no antiretroviral agents (15 studies), 2-drug vs. 1-drug therapy (16 studies) and 3-drug versus 2-drug therapy (9 studies). In our analysis, we obtained estimates of RRMN and var(log(RRMN)) from a meta-analysis of the 15 trials comparing 1-drug therapy versus placebo. Similarly, we estimated RRDM and var(log(RRDM)) from a meta-analysis of the 16 trials comparing 2-drug versus 1-drug therapy; and we obtained estimates of RRTD and var(log(RRTD)) from a meta-analysis of the 9 studies of 3-drug versus 2-drug therapy. The assumption of independence between should be adequately satisfied because each value was estimated from different trials. We calculated an overall estimate of RRTN and its corresponding 95% CI by plugging these estimates into formulas (1) through (4). For each meta-analysis, tests for heterogeneity indicated statistically significant variation among studies, so we used a random-effects model to combine studies and calculate the estimates of RRMN, RRDM and RRTD. Estimates obtained by using a fixed-effects model, however, were similar to those from a random-effects model. Bucher and colleagues250 used a fixed-effects model to combine studies. Jordan and colleagues133 also used a fixed-effects approach to estimate odds ratios for 1-drug therapy versus placebo, 2-drug versus 1-drug therapy, and 3-drug versus 2-drug therapy.
The background rate (cases per 3 person-years) and relative risk for myocardial infarction and cardiovascular and cerebrovascular events (myocardial infarction, stroke, or invasive cardiovascular procedures) associated with combination antiretroviral therapy after 2 to 4 years compared to no exposure were calculated on the basis of raw data from the Data collection on Adverse events of anti-HIV Drugs (DAD) study (Figure; we used outcomes for no antiretroviral treatment and combined outcomes for 2 to 3 and 3 to 4 years of exposure) according to standard statistical methods.228,229
Calculations of numbers needed to screen for benefit (NNSB) and numbers needed to treat for benefit (NNTB) were based on estimates from different sources in the literature (Appendix Table). The indicated range of estimates and variation associated with estimates were incorporated in the calculations by using Monte Carlo simulations and are reflected by the ranges in the calculated NNSB and NNTB. The sampling distributions of the estimates used in the simulations were either the underlying distribution on which the calculation of 95% CI was based, or one that best approximated the point estimate and CI. For example, if the estimate was a rate or proportion, the logit of the rate or proportion was sampled assuming an approximately normal distribution; it was then transformed back to its original scale. For relative risks, we assumed that the log of relative risk was approximately normally distributed. The log of the relative risk was sampled from the normal distribution and then transformed back to relative risk. In each iteration of the Monte Carlo simulation, one sample of each proportion, relative risk, or other estimate was drawn to calculate the NNSB and NNTB. The point estimates and 95% CI of NNSB and NNTB were based on 1 000 000 samples. Similar calculations were performed to calculate numbers needed to screen for harm (NNSH) and numbers needed to treat for harm (NNTH). A simple program using R statistical language was written to perform simulations and calculate summary statistics.278