We developed the Progression and Transmission of HIV/AIDS (PATH) model to estimate the quality-adjusted life expectancy and costs of persons diagnosed with HIV infection at various stages of the disease. PATH is an individual Monte Carlo simulation health state transition model that tracks index patients through different phases of HIV from infection until death. It also includes transmission and follows the infected partners of the index patients until death. The model was developed in Microsoft Excel (Version 2003, Microsoft Corporation, Redmond, WA) with Visual Basic Applications (Version 6.3, Microsoft Corporation, Redmond, WA). Distributions, random numbers, and simulations were generated with @Risk (Version 4.5.7, industrial edition, Palisade Corporation, Ithaca, NY). The unit of time progression is a three-month period representing a quarter of a calendar year, with costs, quality-adjusted life years (QALYs) lost, and other outcomes computed for each quarter. A summary of key input parameters for the model is presented in Table 1. The assumptions, technical details of the model, schematic flowcharts, and the full set of input parameters are presented in Appendix S1, The Progression and Transmission of HIV/AIDS (PATH) Model.

We created three scenarios — one each for routine screening in STD clinics, routine screening in hospital EDs, and diagnoses made in inpatient settings. We ran the PATH model for 10,000 iterations for each scenario. Each iteration represented an individual, or an index person, whom we tracked from infection until death. The three scenarios differed only in CD4 cell count at diagnosis, undiagnosed seropositivity rate, associated screening costs, and assumptions made about the proportion of newly diagnosed persons who were linked to care.

For STD clinics, we assumed that persons visit a clinic for sexual education, health examinations, tests, and treatments, and that screening with a rapid HIV test is routinely conducted as a part of a program for STD prevention. We based our analysis on clinics located in an urban area with a large MSM population in which many persons are tested frequently. For diagnosis in an ED, we assumed that people visit an ED facility because they need urgent or emergency medical care and are routinely screened with a rapid test. For HIV diagnosis in inpatient facilities, we assumed that physicians conduct diagnostic testing (e.g., order HIV tests based on the clinical manifestations of patients) using conventional testing with an HIV enzyme immunoassay (EIA) of serum obtained by venipuncture. In all three settings, positive EIA and rapid tests were assumed to be confirmed with a Western blot.

For CD4 cell count at diagnosis in STD clinics, we used data from the One-on-One program of the Public Health – Seattle & King County (PHSKC) STD Clinic in Washington state from January 2006 to June 2008 (Table 2).(Written communication, M Golden, Public Health-Seattle & King County STD clinic and the Center for AIDS & STD, University of Washington, Seattle, May, 2009. See also [22]) The One-on-One program refers people diagnosed with HIV at Seattle and King County public health clinics to treatment. This clinic is representative of a testing program in an urban area with a large MSM population where the clients are tested with increasing frequency.[23] For the MSM tested, the median CD4 cell count at diagnosis was 429 cells/µL (range 5–1,287 cells/µL).

For CD4 cell count at diagnosis in EDs, we used the results from a program of expanded HIV screening and on-site rapid testing primarily among adult Hispanic and non-Hispanic black patients in an urban academic ED in Oakland, CA (Table 2) (median CD4 count 356 cells/µL; range 4–1,020 cells/µL).[19]

For CD4 cell count at diagnosis in inpatient facilities, we used data on inpatients discharged with a new diagnosis of HIV or AIDS at two academic medical centers in Boston, MA (Table 2) (median CD4 cell count at diagnosis 36 cells/µL; range 2–847 cells/µL).[24] These data are consistent with other studies.(Written communication, D. Rimland, VA Medical Center, Decatur, GA, August, 2009. See also [25])

We assumed that all patients diagnosed in the inpatient setting were linked to care in the quarter following diagnosis. For patients diagnosed in the ED and STD settings, we assumed that 65% were linked to care in the quarter following diagnosis, and that an additional 15% were linked to care by the time their CD4 cell count decreased to 200 cells/µL. The remaining 20% were assumed to be diagnosed as inpatients and were linked to care when their CD4 cell count decreased to 36 cells/µL, the median CD4 cell count at diagnosis in inpatient facilities.[24] These assumptions are consistent with data from studies of linkage to care in various settings.[26]–[29]

We included the following phases of HIV infection as health states in the PATH model: acute infection, asymptomatic HIV infection, symptomatic HIV infection or acquired immunodeficiency syndrome (AIDS), and death. CD4 cell count and HIV viral load were the key determinants of disease progression in this model. When individuals were linked to care according to the above assumptions, they became eligible for HAART when their CD4 cell count decreased to a threshold of either 350 or 500 cells/µL to model previous and current treatment guidelines.[6], [30] Persons linked to care with higher CD4 cell counts did not initiate HAART until their CD4 cell count decreased to these thresholds, whereas persons linked to care with CD4 cell counts under the thresholds initiated HAART in the quarter following diagnosis. The PATH model included up to four suppressive HAART regimens followed by a single salvage non-suppressive regimen.

We predicted the probability of death during each quarter following diagnosis based on probabilities related to age and CD4 count at initiation of antiretroviral treatment.[31], [32] We assigned a utility weight ranging from 0 (for death) to 1 (for perfect health) to an individual's health state for each quarter survived, based on that individual's CD4 cell count during that quarter and whether the individual had an opportunistic illness (OI) such as Pneumocystis jiroveci pneumonia (PCP). We used quality-of-life weights from Tengs and Lin[33] and aggregated them over the person's life. We then subtracted this sum from the QALYs associated with an HIV-uninfected person, assuming a QALY of 1 (good health) for the entire life expectancy from the age of HIV infection,[34] to estimate the QALYs lost due to HIV infection. Measuring QALYs lost due to infection resulted in consistent quality of life estimates when transmissions to partners were included in the model. A decrease in QALYs lost in one setting compared with QALYs lost in another represents a gain in QALYs in the first setting.

We included both treatment costs and program costs in 2009 dollars estimated from the provider perspective. Treatment costs, derived from lifetime cost estimates by Schackman et al.,[35] included the costs of health care resource utilization, antiretroviral therapy, laboratory monitoring (i.e., CD4 cell count, HIV viral load determination, and genotypic antiretroviral resistance testing), diagnosing and treating OIs, and costs incurred during the last month of life.

To estimate program costs, we calculated only the marginal costs associated with testing and counseling in a particular setting for both HIV-infected and uninfected persons. We assumed that the settings evaluated already had HIV testing ability, so fixed and start-up costs were not included in our calculations. For inpatient facilities, we estimated the laboratory costs for conventional HIV testing and post-test counseling costs only for HIV-infected persons. For the ED and STD clinic settings, we included additional costs associated with rapid HIV testing, such as the costs for collecting specimens, the test kits, and post-test counseling for infected persons. We did not include the cost of administrative overhead and other costs that would have been incurred in the absence of a screening program. These program costs were varied in the sensitivity analysis, particularly for STD clinics, to reflect the repeat testing by MSM that often occurs in these settings.[23]

To include the costs of persons who are tested but are not HIV infected, we computed program costs per HIV-infected person identified using the following formula for each setting: {[p * Cost _HIV+ + (1−p) * Cost _HIV−]/p} where Cost _HIV+  =  cost of testing an HIV-infected person, Cost _HIV−  =  cost of testing an uninfected person, and p  =  the undiagnosed HIV seropositivity rate in that particular setting. For example, in inpatient settings the total cost per HIV-infected person derived from Table 2 data equals [(0.143) * 62.4+(1−0.143) * 5.3)]/0.143 = $94.1.

Cost _HIV+ and Cost _HIV− for a particular setting were derived from estimates by Farnham et al.[36] (Table 2). We discounted future costs and QALYs at a rate of 3% per year[37] from the quarter of infection.

We used a quarterly probability of HIV transmission per infected individual derived from an annual transmission rate to add HIV transmission from index patients to the model (Table 1). We estimated transmission probabilities on the basis of a model, first developed by Pinkerton[38] and later updated by Prabhu et al.[39], which is explained in Appendix S1. Transmission probabilities were derived for those acutely infected and unaware of their infection, those non-acutely infected and unaware, and for those non-acutely infected who were aware and either not on or on a HAART regimen. We used separate rates for sexual and injection drug use (IDU) transmission, and we assumed that 12.9% of the index persons were IDU in all settings.[40]

We evaluated secondary transmissions for a single generation of transmissions, i.e., transmission of HIV from index persons to their partners. We assumed that all partners who acquired infections from an index patient were diagnosed at a CD4 cell count of 500 cells/µL and were linked to care based on assumptions similar to those for persons diagnosed in ED and STD clinics. We standardized the linkage to care and treatment approach for infected partners because our primary interest is assessing the timing of diagnosis, linkage to care, and initiation of treatment of index patients on the cost-effectiveness of HIV diagnosis in different settings.

We estimated the costs (treatment and program) and QALYs lost to infection for each of the 10,000 index patients for each setting, and we computed the mean costs () and mean QALYs lost to infection () for each setting. We then used the ratio of the differences in mean costs and differences in mean QALYs to compute the incremental cost-effectiveness ratio (ICER). To calculate ICERs with transmission effects, we included the costs and QALYs lost to infection for the index persons and their infected partners. We calculated 95 percent confidence intervals for mean and incremental costs and QALYs.

Incremental cost-effectiveness ratios may be negative, indicating cost-savings resulting from an increase in QALYs gained and a decrease in costs, or positive, showing that additional costs are required to achieve the gain in QALYs. For the latter, $100,000 per QALY gained represents a reasonable current estimate of the amount society is willing to pay to gain a QALY, although this amount may be even higher.[41]–[43]

The base case simulation of the model for 10,000 iterations used point estimates for all variables in the model. A simulation of 10,000 iterations was necessary as the outcomes of the model reflected the probabilities of the occurrence of different events, such as the development of an opportunistic illness or the probability of dying during the quarter after HAART had been initiated, for each of the 10,000 index persons. Values of the model variables were drawn directly from the literature as noted previously and in Appendix S1. We present base case results both excluding and including transmission and with the assumption of patients initiating HAART at a CD4 count of either 350 or 500 cells/µL.

We then performed one-way sensitivity analyses of the impact of changes in selected variables on the STD-ED ICERs in the base case, assuming initiation of HAART at a CD4 count of 350 cells/µL. These variables included the undiagnosed HIV seropositivity rate in the different testing settings, overall program costs, STD clinic program costs, HIV treatment costs, age at infection, the probability of viral load suppression, and the transmission probabilities. The differences between testing in the STD and ED settings were analyzed in more detail by varying the CD4 count at diagnosis in the STD setting. The impact of linkage to care was examined by assuming that all index patients and their partners were immediately linked to care.

To reflect the overall uncertainty in decision analytic models, we also ran a probabilistic sensitivity analysis by assigning distributions around the point estimates of key variables based on accepted conventions.[44] These variables included: age at infection; CD4 count when infected; CD4 count at diagnosis; set point viral load; the levels of suppressed, rebound, and salvage therapy viral load; and the decline in CD4 count at a specific viral load stratum. Normal distributions were used for most variables. Given the importance of CD4 count at diagnosis for this analysis and the small sample sizes in the studies reporting these values, we used the cumulative distribution based on the minimum, maximum, and inter-quartile values for these variables in an attempt to most accurately use the available data.

The results of the one-way sensitivity analyses in Table 5 comparing screening in STD clinic settings with ED screening showed that the base case results in Table 3 were robust with respect to changes in key variables in the analysis. Variations in undiagnosed HIV seropositivity rates, program costs, HIV treatment costs, age at infection, the probability of viral load suppression, and transmission rates had little impact on the STD-ED incremental cost-effectiveness ratios. These ICERs remained undefined when the transmission effects were excluded, given the zero change in QALYs between the two settings. Screening in the STD clinic setting remained cost-saving compared with ED screening when the benefits of reduced transmission were included. When the CD4 count at diagnosis in the STD setting was varied by increments of 20 cells/µL from 356 cells/µL (equal to the base case value for the ED setting) to 436 cells/µL, STD screening remained cost-saving compared with ED screening even for a difference as small as 20 cells/µL when transmission benefits were included in the analysis. Assuming 100 percent linkage to care for both index patients and partners also did not change the results of the analysis.

In sensitivity analysis (data not shown), the ED-inpatient ICERs were all in the same range as for the base case. Thus, the results for all the incremental cost-effectiveness ratios were robust in the sensitivity analysis.

When the model was run with a probabilistic sensitivity analysis around key variables (Table 6), excluding transmission effects and assuming treatment with HAART at a CD4 count of 350 cells/µL, the ED-inpatient incremental cost-effectiveness ratio was approximately the same as in the base case (Table 3). However, the STD-ED ICER was $44,000/QALY gained compared with the undefined STD-ED ICER in the base case (Table 3). When transmission benefits were included in the analysis, initiating treatment with HAART at a CD4 count of either 350 or 500 cells/µL (Tables 6 and 7) was cost-saving for both the ED-inpatient and STD-ED comparisons (except at the upper bound of the 95% confidence interval).

Our work is subject to a number of limitations. Data regarding disease status (CD4 cell count and HIV viral load at diagnosis) for the different HIV testing settings are very limited. In particular, the data we used for CD4 cell count at diagnosis were drawn from observations at a small number of locations. We, therefore, may not be able to generalize our findings to all EDs, STD clinics, and inpatient settings. Our analysis indicates that more data, particularly on CD4 count at diagnosis by setting, would be useful, given the differences between our base case results and those in the probabilistic sensitivity analysis where we allowed the CD4 count at diagnosis to vary around the median in each setting. On the other hand, our main finding, that diagnosing persons living with HIV at higher CD4 counts is cost-effective, is robust even with the limited data.

We may have under-estimated the costs for screening in STD clinics because we did not include any fixed costs and because many STD settings include clinics that strongly encourage repeat testing among their MSM clients. However, it would be inconsistent to use average costs (that include fixed costs) for STD clinics and marginal or incremental costs (that exclude fixed costs) for the ED and inpatient settings. Although repeat testing would increase STD clinic costs, we showed in the one-way sensitivity analysis that increasing STD screening costs by 100 percent did not change the results of the analysis. In a separate simulation (results not shown), we increased STD screening costs ten-fold from their base case value, and this variation also did not change the overall cost-effectiveness results of the analysis.

Data on linkage to care are sparse and may vary by subgroups in the population. The assumptions in this model are consistent with the existing literature, and our sensitivity analysis did not show any impact of changes in these assumptions. However, better data, particularly on linkage to care in different settings, will improve future modeling efforts.

The PATH model does not incorporate any measure of ongoing transmission beyond the first generation partners. Thus, we may underestimate the cost-effectiveness of early diagnosis as some additional secondary transmission might also be averted. On the other hand, some infections we consider to be averted might only be delayed. Use of a dynamic transmission model in an economic analysis could improve the estimates of the cost-effectiveness of different HIV screening programs, but would introduce additional complexity and uncertainty related to sexual mixing patterns, which are not well defined. Our estimates of the number of transmissions per index partner are consistent with those in the literature.[10], [46] Decreasing the transmission probabilities in the sensitivity analysis reduced the number of transmissions per index partner but did not affect the overall cost-effectiveness results.

Our analysis with the PATH model showed that identifying persons with HIV while their CD4 counts are high is cost-effective and potentially cost-saving, when the effects of early diagnosis on transmission are considered. Although inpatient testing based on clinical manifestations of disease should always be undertaken, our results should prompt additional HIV case-finding efforts, particularly in venues such as STD clinics and emergency departments, where persons are likely to have higher CD4 counts at the time of diagnosis. The results can help guide decisions about implementing HIV screening and should be used to encourage the collection of additional data on CD4 count at diagnosis to identify more settings where persons are likely to be tested early in the course of disease. Our model also showed that initiating treatment with HAART earlier in the course of infection is cost-effective, making early diagnosis even more beneficial.