Since the FDA approved the OraQuick In-Home HIV test in 2012, the use of HIV self-testing has grown considerably in the United States, amplified in part by the COVID-19 pandemic and reliance on telehealth and telemedicine services, such as COVID-19 home tests [1]. In comparison to non-FDA approved HIV self-collection kits that collect a dried blood spot (DBS) sample, oral swab self-testing gives test results within minutes and does not require a sample to be sent to an external laboratory [2]. HIV self-testing offers an innovative strategy to help expand HIV testing across the United States, increase testing among populations most affected by HIV, and reach persons with undiagnosed HIV who are not being reached by traditional testing programs, including those in laboratory settings [3], [4].There is currently only one FDA-approved HIV self-test in the US [5], [6].

Several studies have demonstrated the effectiveness of self-testing programs. In [4], a trial among gay, bisexual and other men who have sex with men (collectively referred to as MSM) showed that distribution of oral self-tests led to a 47% increase in total HIV testing over the observed period. The availability of self-testing was not found to have any impact on STI acquisition and performed similarly to laboratory-based testing for all HIV risk indicators. A trial utilizing internet-based recruitment, in which participants were mailed self-tests [7], found that this approach led to increases in testing levels especially among MSM who had never had a previous HIV test [3]. This is generally consistent with existing literature, which suggests that overall testing increases with improved availability of self-testing [8], [9]. While the cost of the current FDA-approved self-test has been reported to be a deterrent from further adoption of self-testing in the United States [10], [11], recent analyses have found that self-testing is cost saving [12].

Despite encouraging findings in trials, concerns about the real-world impact of scaling-up self-testing programs remain. The FDA-approved self-tests (like all oral-swab tests including CLIA waived rapid tests) are less sensitive than laboratory-based tests, particularly in the acute phase of HIV infection, estimated from 45 to up to 90 days after acquiring infection and referred to as the detection period (or window period) [5]. While the instructions for the OraQuick self-test state that the test should not be used as a substitute for laboratory-based testing, some researchers have posited a concern that this could happen in practice. In addition, in [4], [7], [8], participants in controlled settings were carefully instructed on the proper use and interpretation of such tests. However, fidelity to proper use may vary with larger-scale implementation in community settings.

Given the inherent difficulty in measuring self-testing behaviors outside of a controlled setting and evaluating the effectiveness of scaling up self-testing programs, researchers have resorted to mathematical models to evaluate the effects of self-testing on the wider population [13], [14]. Findings in these two studies have been mixed, with self-tests having a positive or negative net-effect on HIV incidence based on the degree to which they replace laboratory-based tests and how much their introduction causes overall testing rates to increase. Both studies employed area-specific models of MSM; a network-based model was used in [13] and a compartmental model in [14]. In some instances, these studies considered demographic pools that already had high rates of laboratory HIV testing, limiting the potential benefit of any additional testing-based intervention, including self-testing.

Our analysis expanded the mathematical modeling of self-testing by precisely quantifying the relationship between increased HIV self-testing and future HIV incidence and awareness of HIV status in the MSM population. This was accomplished through the introduction of a streamlined mathematical model, that allowed us to vary levels of HIV self-testing, levels of overall HIV testing, and any delays in obtaining a formal HIV diagnosis after a positive self-test, in a straightforward manner. We then applied the model to the US MSM PWH population and identified the conditions in which self-testing could reduce HIV incidence and increase awareness of HIV status. We found threshold levels, the amount by which overall HIV testing must increase to offset the reduced sensitivity and delays in obtaining a formal diagnosis associated with increased HIV self-testing. While previous models identified scenarios in which self-testing both reduced and increased HIV incidence, they did not identify the specific conditions in which self-testing leads to HIV incidence reductions.

In this section, we apply the model to evaluate three different scenarios for HIV self-testing among MSM in the United States:

Perfect supplementation: If self-testing is added to existing testing levels and there is no replacement of non-self-tests with self-tests among PWH, we examine how much self-testing can reduce HIV incidence and increase awareness of status among the MSM population in the United States.

In each analysis, we parameterize the model using values (or methods) reported in Table 1a–1c. For the mortality and force-of-infection terms, we use a data-driven procedure to parameterize the model directly from public, national-level surveillance data for MSM in the United States provided on NCHHSTP ATLAS Plus [26]. These data are provided and summarized in Table 2, and the parameterization procedure is outlined in Supplement B.

The simulation period occurs in three distinct stages: a 7-year dry run, a validation period (2017–19), and the intervention period (2020–30). The 7-year dry run serves to initialize the model and reduce sensitivity to initial conditions. The validation period (years 2017–19) follows the dry run. During the validation period, we compare model outputs with surveillance data to ensure that the model accurately reproduces surveillance data. Figure 3 compares simulation values to the corresponding values computed from surveillance data, spanning the years 2017–19. Transmission rate is computed as NewHIVinfectionsTotalPWHpopulation, while mortality rate is computed as PWHDeathsTotalPWHpopulation. Self-testing is introduced during the intervention period at the beginning of the year 2020; note this means that all simulations are identical prior to the intervention period. The outcomes collected over the intervention period are used to evaluate the effect of HIV self-testing.

We are interested in how self-testing affects both HIV incidence and awareness of status among MSM. We report the percent change in incidence as the relative difference in cumulative incidence I from 2020–30 as compared to a baseline scenario with no self-testing χ=0. Mathematically, this is defined as: %IncidenceChange=∫20202030Iself-testt-Ibaselinetdt∫20202030Ibaselinetdt.

For awareness of status, we report the percentage of MSM PWH who are diagnosed at the end of the modeled period (year-end 2030).

Before further discussion of our results, we briefly note that we restricted our analysis to MSM populations for two key reasons. First, MSM comprised 67% of new HIV diagnoses and 65.7% of estimated new HIV transmissions in the U.S. in 2021; thus, they are a priority population for expanding HIV prevention efforts and reducing HIV incidence in the U.S. [36], [37]. Second, given the simple, four-compartment, single-population structure of the introduced model, restricting the analysis to this important subgroup helps simplify important questions regarding differences in awareness levels, testing rates and transmission behaviors across different populations.

To demonstrate the potential favorable effects of HIV self-testing on HIV incidence and awareness of status among MSM, we first consider the special case of perfect supplementation. In this scenario, self-tests are assumed to not be a replacement of non-self-tests among undiagnosed PWH, and 100% of the increase in the overall testing rate among undiagnosed PWH is attributable to adding self-testing to the existing non-self-testing efforts. Mathematically, this corresponds to the condition: γ=χ1+χ.

At the programmatic level, this can be interpreted as an optimistic situation, in which self-testing serves to increase the overall testing rate among MSM, while not affecting the underlying rate of non-self testing.

We vary the time period of formal diagnosis and engagement following a positive self-test as tself→diag.=1, 2, 3, 6, and 12 months. As the amount of non-self-testing does not change in these scenarios, we expect that no incidence increases will occur. However, the amount that incidence decreases will depend on both the extent of supplementation and how quickly formal diagnosis and engagement in care occurs, post-self-test.

The perfect supplementation scenario shown above illustrated the best possible case for expanding HIV self-testing, in which testing is increased among MSM who do not test regularly as a result of the added self-testing, while regularly-testing MSM do not change their testing behavior. In reality, expanded self-testing may result in some replacement of non-self-tests with HIV self-tests among MSM PWH. Some persons who test regularly may switch to using self-tests with no change in the overall testing rate (replacement). However, simultaneously, some MSM who rarely or never test may begin to test more often, and further, some persons who test regularly increase their rate of testing, due to expanded self-testing (supplementation). This analysis sought to understand how much supplementation among MSM PWH must occur to offset the effects of replacement and ensure that self-testing does not result in incidence increases in that population – we call this the threshold level of supplementation.

We considered 25,000 random samples of in which the change in overall testing rate χ and percentage of tests that are self-tests γ were varied independently from 0 to 100%. For each sample of 25,000 cases, the time to formal diagnosis after a positive self-test tself→diag. was varied as 1, 2, 3, 6, and 12 months. We assumed a baseline scenario as one with no self-testing (γ=0), and the overall calibrated MSM PWH testing rate over 2017–2019 χ=0. Note that with no self-testing (γ=0), time to formal diagnosis tself→diag. has no effect (this parameter does not influence testing outside of self-tests), and hence the baseline scenario was identical for all tself→diag.. For each tself→diag.=1,2,3,6,12 months, we sought to determine the threshold level of supplementation for each given level of self-testing prevalence.

We now evaluate the effect of self-test sensitivity κu,sself, and detection period σa→u. In the previous section, following the FDA guidelines, we considered self-test sensitivity κu,sself=.92, and that the period necessary for self-tests to become accurate (detection period) was governed by σa→u=60 days. These values may be conservative, and some literature suggests that detection period σa→u may be as low as 42 days, or as high as 90 days, and that self-test sensitivity κu,sself may be as high as .95 [38], [39], [40], [41].

To analyze the effects of these parameters on our conclusions, we performed a sensitivity analysis. We considered the same underlying setup as in the preceding scenarios. We let self-test sensitivity κu,sself=.92,.94,.96, and, for each κu,sself, we performed simulations for detection period σa→u=30,60,90 days, in line with the range of estimates in the literature. As in previous scenarios, we performed simulations for tself→diag=1, 2, 3, 6 and 12 months.

We introduced a parsimonious 4-compartment model to quantify the effect of self-testing on HIV transmission and awareness of HIV status in the United States. By focusing on the pre-diagnosis period in detail and considering the post-diagnosis continuum of care only implicitly, we reduced model complexity, while still describing our problem of interest.

We then applied the model to examine the case of perfect supplementation, in which self-testing strictly adds to existing testing among MSM with undiagnosed HIV, with no replacement. We found that in the supplementation scenario, self-testing can decrease incidence by as much as 10%. While formal diagnosis and engagement-in-care occurring promptly after diagnosis ensures the largest possible incidence reductions in the perfect supplementation scenario, this effect is small (a 2% difference in incidence).

We then performed an analysis in a more realistic scenario in which we allowed replacement of non-self-tests with self-tests among MSM with undiagnosed HIV. This study showed that self-testing can provide consistent decreases in HIV incidence, even if some replacement occurs, provided formal diagnosis and engagement in care after a positive self-test occur quickly. Longer delays may necessitate larger increases in the testing rate to ensure incidence reduction. Further, ensuring such delays are short also leads to larger decreases in incidence.

Finally, we examined how possible differences in self-test performance, in terms of self-test sensitivity and self-test detection period, may change our conclusions. Improvements in either or both areas would further decrease incidence and reduce the possible negative impact of replacement. However, these effects were small.

Our study was not directly comparable to those performed in [13], [14] as we considered a national level model, rather than individual cities. Nonetheless, our analysis was more optimistic about self-testing than [13], [14]. The analysis in [14] showed increases in incidence of nearly 50% for strict substitution of non-self-tests with self-tests, and potentially as much as 300% for long delays in formal diagnosis after a positive self test tself→diag.. In [13], the case of 25% self-tests with no increase in overall testing rate increased incidence 0.28% in Atlanta and 3.75% in Seattle, while 50% self-tests and no increases in overall testing increased incidence 3.4% in Atlanta and 8.75% in Seattle. In contrast, our analysis never showed increases of more than 4.0% in incidence, even for the most extreme case of 100% self-testing, no increase in overall testing rate, and a 12-month delay in formal diagnosis and engagement in care. We also found that the potential reductions in incidence were also greater (as much as 10%) in our analysis compared to the 1% and 5% reductions found in [13]. Further, we found that the increases in overall testing levels necessary to offset any replacement effects among PWH with undiagnosed HIV are small, especially for rapid formal diagnosis and engagement in care after a positive self-test. We note that our study specifically considers changes in testing among MSM with HIV, rather than among the general MSM population. As such the modeled changes in testing behavior may not be directly comparable, and this may explain some or all the differences in our findings.

However, our analysis was consistent with the findings in [13], [14] regarding two important points: strict substitution of non-self-tests with self-tests had a negative effect on HIV incidence, and that ensuring that time to formal diagnosis and engagement in care after a positive self-test was the most important factor in guaranteeing that self-testing reduced incidence.

We acknowledge several shortcomings of our analysis and modeling approach. The introduced compartmental model does not further stratify the population of interest into sub-groups within each compartment. This allows us to keep the necessary number of input parameters low, and to easily obtain parameter values directly from surveillance data. However, testing patterns vary among individuals [4], [7], [8], [9], [10] and CDC recommends different testing intervals for groups based on background prevalence and individual characteristics [42], [43]. To help alleviate some of these concerns, we restricted our analysis to the MSM population. However, accounting for further behavioral differences within the MSM population, as was done in [13], [14], may improve model performance, provided reliable data can be obtained. This is particularly important for modeling interventions that prioritize groups known to have lower testing levels. We believe self-testing benefits would be even more likely among PWH populations or jurisdictions with lower background testing rates, as overall testing rates are more likely to increase when baseline testing rates are lower. Our analysis shown here, which exhausted the overall parameter space, does generally account for such scenarios, such as the perfect-supplementation analysis. However, it is difficult to define and study more precisely defined interventions without expanding the modeling framework. Extending the model resolution to account for such scenarios is unlikely to change our quantitative conclusions. However, such analyses may be useful for intervention planning.

Our model did not consider behavioral changes resulting from negative self-tests. As a result, this analysis does not address any potential concern arising from, for example, a false-negative self-test leading to increased transmission risk behavior. Additional data from HIV self-testing implementation studies, including data on harms, would help to refine this model.

Linearizing the transmission in our model also has its advantages, most notably the fast simulation (order of seconds) of thousands of scenarios and the way in which we may parameterize from surveillance data (see Supplement B). These characteristics enabled us to easily characterize outcomes over the entire parameter space and precisely identify threshold conditions. However, this linearization is made possible by assuming that HIV prevalence is relatively stable, as data for the US MSM population suggest [15], [16], [17], allowing us to assume that changes in the susceptible fraction of the population are small enough to be ignored. Our formal analysis (Supplement A) shows that this assumption is valid, and that any error incurred by the linearization can be bounded in terms of model parameters. Given the epidemiologically relevant ranges of these parameters, our bounds suggest that these assumptions are acceptable when considering the US MSM population, whose transmission dynamics are largely stable [18]. While this modeling approach is likely generalizable to some degree to entire countries/jurisdictions/populations with a similar HIV profile, it may be unsuited for analyses of other countries/jurisdictions/populations. in smaller sub-populations with potentially greater HIV transmission rates, such as rural populations of persons who inject drugs [22], [23], the epidemic does not exhibit the relatively stable growth patterns observed in the US MSM population if HIV were to be introduced into this sub-population.

These modeling assumptions also affect the practical interpretation of our results. Since non-PWH are not explicitly considered, all modeled changes in testing behavior refer to changes in testing among PWH with undiagnosed HIV. While most current HIV modeling studies either only explicitly model the undiagnosed PWH population, as was done here, or define relationships between the testing rates of undiagnosed PWH and non-PWH, this model does not define, or attempts to define, any such relationship. This model could consider testing among the general population by assuming some functional relationship between testing among undiagnosed PWH and non-PWH without significant structural modifications. Similar approaches have been employed in several models of HIV transmission in the United States [27], [44], [45], [46]. Incorporating testing among non-PWH is necessary to explore many important issues related to HIV testing, including, for example, the effects of false positive HIV tests on HIV transmission, or the influence of social dynamics on factors affecting HIV transmission, including testing behaviors [47], [48].

The model also may be limited in its direct programmatic application. From the programmatic point of view, two distinct HIV testing strategies may lead to the same changes in testing rates within a particular undiagnosed population or community. For example, one testing strategy may increase existing testing levels 10% among the larger population homogenously, including both persons with and without HIV across the board, implying a 10% increase among those undiagnosed PWH within that population as well. However, a carefully designed HIV testing strategy may specifically aim to prioritize testing among those persons, communities, or settings where undiagnosed HIV may be more prevalent, which could result in a 10% testing increase among the undiagnosed PWH, but a much lower testing increase among non-PWH. The scope, design, and cost of the two HIV testing strategies would be different; however, the effect on overall awareness of status among MSM PWH, and future incidence, would be the same. The modeling approach used here, in the absence of any additional information, cannot distinguish between the two HIV testing strategies.

From a practical point of view, the analysis should be extended to Ending the HIV Epidemic in the U.S. (EHE) priority jurisdictions [49], [50]. We considered a national-level model of MSM as a proof-of-concept to demonstrate the overall potential benefit of HIV self-testing on HIV incidence among MSM in the United States. However, different jurisdictions vary greatly in underlying levels of incidence, prevalence, HIV testing among PWH, and awareness of status. Accounting for populations besides MSM may also require that the model introduce additional stratifications. These changes may result in varying conclusions; self-testing may be expected to provide a larger or smaller effect on incidence, or different levels of testing increases may be necessary in different jurisdictions or among different populations. Any such potential differences would have important implications for future intervention efforts. This is an important direction for future work. We note that the modeling framework introduced here is well-suited for such a task, as it allows for easy, efficient parameterization from a small amount of easily obtained surveillance data. An important next step is to apply this model to particular local jurisdiction scenarios where the local dynamics are taken into consideration and more direct recommendations could be made regarding how best to design an HIV self-testing program.

While we considered HIV self-testing in this work, we note that the modeling framework introduced here may be applied generically to compare any two (or more) different types of tests. To this end, such an approach may be applied to evaluate the efficacy of other types of tests compared to the gold standard laboratory-based tests, such as rapid CLIA-waived tests in community settings.

Finally, in the present work, we considered a simplified testing situation, in which all non-self-tests were assumed to be laboratory tests. More realistic models should account for the heterogeneous testing landscape outside of self-testing and incorporate community-based and point-of-care rapid testing programs. From the modeling perspective, such effects would be reflected in differences in the sensitivity and time-to-diagnosis parameters for non-self-tests. Relative levels of laboratory, community-based, and point-of-care testing vary by jurisdiction [30], and incorporating such effects will likely require a jurisdiction-specific approach. In general, we may expect such effects to lead to reduced sensitivity for non-self-tests and increased delays in diagnosis after a positive result. Provided adequate input data, the modeling framework introduced here can account for these effects. Future work applying this model, especially at the jurisdiction-specific level, should consider these factors. However, we note that accounting for these heterogeneities would, in general, reduce self-testing threshold levels. Thus, the results shown here are conservative; actual incidence decreases may be even larger, and any necessary testing increases even smaller, than those shown here. Thus, the current results likely understate the positive impacts of expanding self-testing, particularly regarding the threshold levels.

Self-testing offers clear advantages over other forms of testing in terms of convenience and availability. Our modeling study of MSM in the United States assessed the concern that, given their lower sensitivity, self-tests could perhaps lead to overall increases in HIV incidence if self-tests replace laboratory-based testing. We found these concerns to be minimal, although some small risk remains in certain scenarios. For these reasons, it is important to expand HIV self-testing programs in a purposeful way that minimizes the possible risks while maximizing the potential benefits. From our analysis, we found that the most important factor in developing an effective self-testing program is ensuring that confirmatory laboratory-based testing and engagement in care occur promptly after a positive self-test. Additionally, to offset any potential replacement effects, programs must ensure that the expansion of self-testing results in increases in the underlying population testing rate.. Focusing self-testing programs in groups with lower background testing rates would likely reach more people who have never or rarely ever tested, which would improve the overall testing rates. In summary, focusing self-testing programs in jurisdictions or among populations with greater undiagnosed HIV infection and lower background testing rates, while also ensuring prompt engagement in care, would maximize the chances of reducing incidence.Finally, continued research on developing more sensitive HIV self-tests, particularly for recently acquired infection, should be pursued. Provided such conditions can be assured, expanded HIV self-testing can become an even more effective tool in the fight against HIV.