Model-based predictions of HIV incidence among African women using HIV risk behaviors and community-level data on male HIV prevalence and viral suppression

Abstract

Background

Pre-exposure prophylaxis (PrEP) with Tenofovir Disoproxil Fumarate and Emtricitabine (TDF-FTC) has proven highly effective in preventing HIV acquisition and is therefore offered to all participants in the control group as part of the standard of care package in many new HIV prevention studies. We propose a methodology for predicting HIV incidence in a hypothetical “placebo arm” for open-label studies or clinical trials with active control among African women. We apply the method to an open-label PrEP study, HIV Prevention Trials Network (HPTN) 082, which tested strategies to improve PrEP adherence in young African women all of whom were offered PrEP.

Methods

Our model predicted HIV infection risk for female study cohorts in sub-Saharan Africa using baseline behavioral risk factors and contemporary HIV prevalence and viral suppression in the local male population. The model was calibrated to HIV incidence in the Vaginal and Oral Interventions to Control the Epidemic study (VOICE).

Results

Our model reproduced the annual HIV incidence of 3.2%−4.8% observed over one year of follow-up in the placebo groups of four completed clinical studies. We predicted an annual HIV incidence of 3.7% (95% CI 3.2–4.2) among HPTN 082 participants in the absence of PrEP and other risk reduction interventions.

Conclusions

We demonstrated the potential of the proposed methodology to provide HIV incidence predictions based on assessment of individual risk behaviors and community and time-specific HIV exposure risk using HIV treatment and viral suppression data. These estimates may serve as comparators in HIV prevention trials without a placebo group.

Introduction

Women in sub-Saharan Africa continue to face very high risk of HIV infection, with 4% annual HIV incidence reported in recent studies^1,2. Multiple randomized controlled trials (RCTs) have demonstrated HIV infection risk can be reduced by daily oral pre-exposure prophylaxis (PrEP) consisting of a two-drug combination (tenofovir disoproxil fumarate, TDF, and emcitritabine, FTC)^3–7. However, two placebo-controlled trials of daily oral TDF-FTC PrEP among young African women did not show efficacy, which was attributed to low adherence^8,9.

Given the high level of evidence for TDF-FTC PrEP efficacy with approval by the FDA and guidance from the CDC and WHO, it is recommended that HIV prevention studies now offer TDF-FTC PrEP to all participants in the placebo group as part of standard of care. Thus, new PrEP products are compared to TDF-FTC alone so their effectiveness relative to no intervention is no longer directly assessable in an RCT. For example, recently reported results from HPTN 083 study demonstrated that long-acting Cabotegravir is non-inferior to TDF-FTC but its biological efficacy was not estimated¹⁰. In open-label studies in which HIV prevention, such as oral PrEP, STI diagnosis and HIV treatment, and counseling, are offered to all participants, model-based simulations of HIV acquisition among hypothetical intervention-free cohorts, closely resembling trial populations, may provide a valid counterfactual, i.e. an estimate of HIV incidence under standard of care (SOC) that does not include any intervention, which can be used to derive the effectiveness of the tested intervention

Mathematical models have a history of simulating “no intervention” scenarios when assessing the potential population-level impact of interventions but they usually consider only few predefined risk groups¹¹. Statistical models have been used to create counterfactual estimates for open-label studies, following RCTs, based on adjustments of the HIV incidence observed in the placebo arm of the completed RCT^12–14.

In this paper we capitalize on data from five previous clinical trials to develop and validate a methodology, based on mathematical and statistical modeling, to predict HIV incidence clinical trial cohorts of sub-Saharan African women in the absence of any prevention interventions for open-label studies and RCTs with active control arms. Such incidence estimates may be used to evaluate the intervention effectiveness in these studies. Our methodology integrates baseline sexual behavior data collected at study enrollment, HIV risk associations calibrated to HIV incidence observed in completed RCTs, and contemporary viral suppression and HIV prevalence data in men from the communities where a study conducted.

We apply our modeling method to predict the non-intervention HIV incidence in HPTN 082 which was an open-label study of PrEP uptake and adherence among young women in South Africa and Zimbabwe. HPTN 082 was designed to assess the characteristics of women who accept TDF-FTC PrEP and evaluate the difference in PrEP adherence between groups randomized to either standard of care adherence support or enhanced adherence which included drug level feedback as well as two-way text messages, peer adherence support groups, and brief counseling in the control arm¹⁵. All participants in HPTN 082 were offered TDF-FTC PrEP and 95% accepted PrEP, so that HIV acquisition in the absence of intervention is unknown.

Data from the placebo-controlled PrEP RCTs such as the Vaginal and Oral Interventions to Control the Epidemic (VOICE) trial conducted in South Africa, Uganda, and Zimbabwe from 2009 to 2012⁹, has been used to inform the associations between HIV risk and demography, social status, and sexual behavior risk factors. These data were used to develop an HIV risk score (the VOICE risk score) consisting of six risk factors associated with higher HIV acquisition risk including relationship status, alcohol use, and the presence of an STI at baseline¹⁶. Our methodology refines the association between these risk factors and HIV risk by integrating the associations between risk factors and HIV risk observed in VOICE with associations between these same risk factors and sexual risk behavior reported at study enrollment in HPTN 082.

Trial participants in HPTN 082 were pre-screened for PrEP interest and higher than average HIV risk, based on a score of ≥ 5 on the VOICE risk score. As a result, the HIV incidence among the HPTN 082 cohort was expected to be higher than in the general female population. However, the rates of HIV exposure and acquisition are likely lower for HPTN 082 than VOICE trial participants due to increased HIV testing, anti-retroviral therapy (ART) coverage, and viral suppression among their male partners, which is integrated in our model.

Methods

HPTN 082 enrolled 427 HIV-uninfected women ages 16–25 from two sites in South Africa (Cape Town and Johannesburg) and one site in Harare, Zimbabwe between 2016–2018. Participants were offered open-label PrEP and followed for up to one year. The 427 women who accepted PrEP were randomly assigned in a 1:1 ratio to receive standard adherence support or enhanced adherence support¹⁵. At baseline, participants reported a median of four vaginal sex acts in the past month (IQR 2–8), with 21% reporting at least one anal sex act, and 39% having a curable STI. Thirty four percent reported multiple sexual partners within the prior month and 40% reported using condoms during their most recent sex act.

Model Framework

Sexual behavior and HIV risk for each female participant were informed by the computer-assisted self-interview behavioral questionnaires completed at enrollment into the study of interest, as well as community-specific information on the male population from which participants select sexual partners, particularly men’s HIV prevalence, ART coverage and viral suppression (Figure 1A).

Figure 1. Model Framework.

A: HIV risk of each individual is simulated over 365 days. Site specific epidemic data and responses to the risk score surveys are translated by the behavioral model into parameters of the transmission model, which in turn is used to predict incidence. B: HIV status and treatment cascade states of the sexual partners in the model. Main partners may transition over time following the arrows, with larger arrows representing more rapid rates. STI=Sexually Transmitted Infection

These inputs were used to predict sexual parameters such as frequency of sex, condom use, and probability of selecting HIV positive partners using a generalized linear model with the VOICE risk factors as regressors, the behavioral model. These parameters were assumed to remain fixed for the duration of follow up.

Using the parameters calculated with the behavioral model, we estimated the annual HIV incidence among the women by simulating the HIV risk of each participant over 365 days via a Markov chain model (the transmission model). Women had two types of sexual partners selected at random from the local male population: casual partnerships which are sampled independently for each act and main partnerships which persist for several months. Over the course of the follow up, individuals could have multiple main partners, although no more than one at a time, with partners’ HIV and treatment status changing over time (Figure 1B). The daily risk of infection is dependent on the HIV and treatment status of the main and casual partners, as well as whether anal sex is practiced, given the higher transmission risk associated with anal sex¹⁷.

A complete description of the behavioral and transmission models as well as the multiple data sources used to inform them is included in the Supplemental Digital Content 1.

Community-specific adjustments

The HIV prevalence and treatment status among the local male population are important determinants of HIV acquisition risk in the simulated female cohort. To generate accurate estimates of incidence, we account for prevailing HIV prevalence as well as the scale-up of ART programs specific to the time and community modeled using data from national HIV surveys (Table 1).

Table 1.

Characteristics of the male population at each site. HIV incidence and prevalence is computed for males aged 15–49, whereas ART coverage and viral suppression is for males aged 15+. The VLS column indicates whether the treatment coverage represents fraction of HIV infected individuals on ART or who are virally suppressed. Data sources:^21–25

Study	Site	HIV Incidence (per 100 person years)	HIV Prevalence (%)	ART Coverage (%)	Viral Suppression (%)
VOICE	Durban	1.4 (1.0–2.1)	17 (13–22)	22 (17–28)	N/A
	Harare	0.9 (0.5–1.5)	11 (9–14)	16 (12–20)	N/A
	Johannesburg	1.1 (0.7–1.6)	13 (10–18)	22 (17–28)	N/A
	Kampala	0.5 (0.3–0.6)	4 (4–5)	20 (7–43)	N/A
	Klerksdorp	1.3 (0.8–2.1)	16 (10–23)	22 (17–28)	N/A
HPTN 082	Capetown	0.4 (0.2–0.7)	9 (5–16)	N/A	45 (39–52)
	Harare	0.5 (0.3–0.7)	10 (7–13)	N/A	72 (46–89)
	Johannesburg	0.5 (0.3–0.7)	11 (7–16)	N/A	43 (37–50)

We predicted the HIV incidence in the placebo arms of six clinical studies, including VOICE (model calibration), four studies (HPTN 035, ASPIRE, FEM-PrEP, ECHO) used for model validation, and HPTN 082 (main results). HPTN 035 and FEM-PrEP were clinical trials of, respectively, microbicidal gels and daily oral tenofovir for HIV prevention among African women that showed no significant HIV effectiveness^8,18. The prevalence of each risk factor included in the VOICE risk score as a predictor of HIV incidence were used in a validation of the risk score based on behavioral and HIV incidence data from HPTN 035 and FEM-PrEP¹⁶. The ASPIRE trial, which took place between 2012 and 2017, evaluated the effectiveness of a dapirivine vaginal ring for reducing HIV infection in sub Saharan Africa and demonstrated 27% efficacy¹⁹. The ECHO trial, which was conducted between 2015 and 2018 in Eswatini, Kenya, South Africa, and Zambia, evaluated the difference in HIV acquisition risk among young women using different forms of contraception¹.

These studies were conducted at a total of thirty-four study communities in Sub-Saharan Africa, after combining sites in the same metropolitan area, such as Soweto and Johannesburg, and removing one site from FEM-PrEP that accrued very little data (description of the studies and their communities in the Supplementary Digital Content 2). For all 34 communities, we adjust male partner HIV incidence and prevalence and ART uptake to local epidemic data,

VOICE Risk Score

Previously, Balkus and colleagues¹⁶ developed and validated a risk score (ranging from 0 to 11) that predicted HIV acquisition risk among women in sub-Saharan Africa, based on data from the VOICE trial and accounting for the heterogeneity in HIV incidence by site with a majority of HIV infections observed from sites in KwaZulu-Natal, South Africa. In our analysis we used the hazard ratio for HIV infection associated with each factor included in the VOICE score (Figure 2). Only women with a score of at least 5 on a modified scale (excluding HSV-2 testing which was part of the VOICE risk score but not done in HPTN 082) were enrolled in HPTN 082. As a result, the prevalence of each risk factor was greater in HPTN 082 than in VOICE (Figure 2).

Figure 2. VOICE Risk Score.

The risk factors that make up the VOICE risk score and their prevalence in the VOICE trial and HPTN 082. The hazard ratios of infection associated with each risk factor were calculated from the VOICE trial data in 15.

The prevalence of each risk factor in a specific study population^1,16,20 was the second set of inputs to our model (Figure 1A). If a study did not ask participants about a particular risk factor, the population level risk was uniformly sampled from the range observed in the remaining studies. Within each study, the distribution of risk factors was assumed to be the same at all sites.

To simulate the HIV acquisition risk of a female cohort, we ran the transmission model for an individual with each of the 2⁶ = 64 possible combinations of risk factors, then weight the results using the distribution of survey answers which is either derived from the prevalence of each risk factor, or, for HPTN 082, directly from baseline questionnaires.

Calibration, validation, and simulation

We calibrated our model to 1) the overall incidence and odds ratio associated with each risk factor in the VOICE trial and 2) the relationship between the VOICE risk factors and sexual behavior recorded in the HPTN 082 cohort at baseline. We used Markov chain Monte Carlo (MCMC) to generate 100 parameter sets consistent with the calibration data. For full details on the calibration procedure see the Supplementary Digital Content 1.

We run our model with each of the 100 parameter sets generated by MCMC to predict HIV incidence in each of the epidemic scenarios used for model validation (HPTN 035, ASPIRE, FEM-PrEP, ECHO) and HPTN 082 (main results). Since we simulated one year of sexual behavior, we compared our results to the data censored at one year^16,20. We made an exception for the ECHO study simulation, which is simulated for the full 18-month follow-up period as one-year censored data were not available. Study level estimates were derived averaging over all participating communities in each study.

Results

Association of risk factors with HIV infection and sexual behavior

Model calibration identified associations between risk factors, included in the VOICE risk score and sexual behavior parameters, primarily based on associations within the baseline data of HPTN 082 (Figure 3). Cohabitation with a main partner is most influential on sexual behavior (Figure 3A). Women who did not live with a main partner were more likely to have multiple partners (OR=3.9, 3.0–4.6) but have fewer sex acts with their main partner (RR=0.43, 0.39–0.45) and use condoms more frequently (OR=4.2, 3.1–5.1). They were also more likely to engage in anal sex (OR=1.6, 1.4=2.2) and their casual partners were substantially more likely to be HIV positive (OR=3.6, 2.7–4.5). Overall, the lower sex frequency and higher condom use, decreased the hazard ratio associated with non-cohabiting women from a prior of 1.80¹⁶ to 1.64 in the final calibration. Other notable associations, suggested by our analysis, are presented in Figure 3B and the complete list is given in the Supplementary Digital Content 3.

Figure 3. Relationship between VOICE risk factors and sexual risk behaviors as estimated by the behavioral model parameters (βs).

The calibrated parameters (green dot and line representing 95% confidence interval) are superimposed on either the values from the HPTN082 self-reported baseline data (blue bars representing 95% confidence interval) or other literature sources (grey bars) used for calibration. A) Behavioral effects of not living with main partner. B) Other strong associations in the behavioral model SE: Main partner may have other sex partners, ST: STI at enrollment AG: Age <25. For a complete parameterization refer to the Supplemental Digital Content 3

Model Validation: Estimated HIV incidence in completed RCTs

Using the associations between risk factors and sexual behavior parameters derived from our calibration, we applied our risk-based methodology to estimate the HIV incidence in four other clinical trials, accounting for community-specific ART scale-up and VL suppression and HIV prevalence during each study as well as the behavioral risk characteristics of the participants at baseline (Figure 4).

Figure 4. HIV incidence estimates.

Predicted HIV incidence female cohorts under different scenarios. Model predictions are superimposed over observed incidence (grey bars) where available.

In FEM-PrEP, our projected incidence of 4.8 (95%CI: 3.5–6.1) is the same as the observed incidence of 4.8 (95% CI: 3.7–6.1) infections per 100 person years. In the ASPIRE trial, our HIV incidence estimate of 3.2 (95%CI: 2.6–4.5) infections per 100 person-years is close to the observed incidence of 3.7 (95% CI: 3.0–4.5) in the control arm over the first 12 months of follow-up.

For HPTN 035, our model predicted an HIV incidence of 4.3 (95% confidence interval (CI); 3.4–5.6) infections per 100 person years which is higher than the incidence of 3.4 (95% CI 2.7–4.1) observed in the trial. The difference is possibly due to misalignment in risk factors associations between Malawi and Zimbabwe, where the most of HPTN 035 participants were recruited and, South Africa where most VOICE trial participants were recruited.

For the ECHO trial, our model estimated HIV incidence of 3.2 (95%CI: 2.6–4.1) per 100 person-years when simulating an 18-month follow-up is slightly lower than the HIV incidence of 3.8 (95%CI: 3.5–4.2) observed in the trial.

HIV Incidence estimates for HPTN 082

Our model predicted an overall HIV incidence of 3.7 (95% CI: 3.1–4.3) infections per 100 person-years in HPTN 082 in the absence of the HIV prevention package provided in HPTN 082, including PrEP, STI diagnostic testing and treatment and counseling offered to trial participants. The highest transmission risk was projected in Johannesburg (5.2 per 100PY, 95% CI 4.2%−6.0%), followed by Cape Town (4.2 per 100PY, 95% CI: 3.0%−5.7), while the incidence projections for Harare was substantially lower (1.7 per 100PY, 95% CI: 1.4–2.4) (Figure 4). The differences in HIV incidence between communities are due to estimated differences in HIV prevalence, ART coverage, and viral suppression among the male partners (Table 1). Our overall incidence estimate for HPTN 082 is lower than the incidence reported in the VOICE trial (6.0 per 100PY, 95%CI: 5.3–6.7), despite the higher VOICE risk scores of HPTN 082 participants. This can be largely attributed to the substantial improvement of the HIV treatment cascade among male partners in the years between VOICE and HPTN 082.

We also used our model to isolate the effect of ART scale-up since the VOICE trial. If the high-risk HPTN 082 female cohort is simulated with partners from the time and communities where the VOICE trial was conducted (Figure 4, scenario 7), the HIV incidence projection more than doubles to 11.6 infections per 100 person-years (95%CI 10.2–13.2) which is comparable to the incidence of 12.4 observed in high-risk subgroups within VOICE¹⁶. Conversely, if the VOICE cohort, which enrolled participants with lower overall HIV risk, were conducted at HPTN 082 sites during the time of HPTN 082 enrollment (Figure 4, scenario 8), their HIV incidence was expected to reduce more than three-fold to 2.0 per 100PY (95%CI 1.7–2.2).

Discussion

Given the demonstrated efficacy of TDF-FTC PrEP for HIV prevention, studies which evaluate HIV prevention delivery methods and adherence support need to offer TDF-FTC PrEP to all participants due to the ethical mandate to provide proven prevention methods. Therefore, in open label studies of existing and new PrEP products among young African women, it may be desirable to estimate HIV incidence in the absence of PrEP.

In this study we propose a methodology for estimating the unobserved, counterfactual annual HIV incidence within female cohorts in the absence of interventions beyond standard of care based on their reported baseline sexual behavior, the prevalence of HIV infection, ART coverage and viral suppression among men in their communities. Our model predicts an anticipated counterfactual HIV incidence of 3.7% for the HPTN 082. which was not observed as discussed further below.

Our analysis demonstrates that tools such as the VOICE risk score may be useful when estimating HIV incidence only if considered in their spatial and temporal context. The probability of HIV exposure and acquisition depends on the likelihood that the sexual partner is HIV positive while not being virally suppressed, which is a consideration generally not included in HIV risk score. Local sexual networks also may influence HIV risk, facilitating repeated exposure to HIV and elevated transmission potential if anal sex is practiced. This creates a heterogeneity in HIV exposure which can affect HIV incidence estimates^26,27.

A key feature of our modeling framework is its ability to combine recent data representative of the local HIV epidemic at a specific time with a well-informed risk profile of the population to produce calibrated community- and time-specific HIV incidence projections. More specifically, it allows us to connect the local HIV prevalence and treatment cascade data with the relative HIV risk distribution informed by the factors included in the VOICE risk score. The availability of population level HIV data from representative samples of population in African countries including by sex, age groups and subnational levels enables inclusion of such data in this and future models²³. The critical importance of both local epidemic and risk assessment data for current and more precise HIV incidence estimates was clearly demonstrated by the variability in the modeling predictions when populations with similar risk distributions are simulated in different community conditions, or when populations with different risk profiles are simulated under the same epidemic conditions.

To validate our methodology, we applied the same modeling approach to predict HIV incidence over the first year of follow-up in recently completed studies which enrolled women from sub-Saharan African countries. Although our method showed good agreement with the observed trial incidence, the validation revealed potential limitations of our analyses. Our predictions tended to be slightly better for South African communities, which formed the majority of the calibration data, compared to Zimbabwean communities, even after accounting for the lower HIV prevalence and higher viral suppression rates in Zimbabwe. Another potential limitation is that our methodology only uses participant behavior reported at baseline. Every HIV prevention study includes vigorous risk reduction activities such as detection and treatment of STIs that may increase susceptibility to HIV²⁸. Therefore, HIV risk is likely to decrease during many studies in a manner that is independent of pharmaceutical interventions. We have implicitly assumed in our risk model that this effect is consistent across all trials, matching the reduction that occurred in VOICE, which was used for calibration. Explicitly including changes in sexual behavior over time may provide refined estimates and is an avenue for future work. Also, the expansion of HIV prevention coverage (including oral PrEP) may affect future estimates. However, the extremely low current PrEP coverage which for South Africa is estimated at 0.2% of the 15–49 year-old population²⁹, makes it highly unlikely for the participants to initiate PrEP outside the study during follow up.

Our model was calibrated and validated using one year of follow up and may need adjustments when used for multi-year trials. For example, our projection underestimated the HIV incidence in ASPIRE when the full multi-year follow-up was included suggesting that it may be necessary to account for changes in risk behavior over time. Finally, the dichotomized age assignment used in the VOICE risk score may not be enough to capture the relationship between age and HIV risk especially for adolescent girls, as was demonstrated in HPTN 068³⁰.

Because we account mechanistically for changing epidemic conditions, our method is most useful in cases where a reference population in a similar place and time is not available. We expect this to become increasingly common as HIV studies offer PrEP to all participants. If the HIV infection status and viral suppression of the women’s main partners is known, then other methods may be used. For example, Kahle and colleagues³¹ developed a risk score for HIV transmission in HIV serodiscordant couples, including factors such as partner’s viral load and age, which they validated using data from the Partners PrEP placebo-controlled efficacy trial⁴. The risk score was used to give a counterfactual estimate of HIV incidence and assess PrEP effectiveness in the subsequent Partners PrEP Demonstration Project which enrolled a similar cohort of African HIV serodiscordant couples^{12, 13}. However, even in these cases the risk of infection from other partners will be unknown and highly dependent upon the nature of the cohort enrolled as well as the epidemic conditions in the trial community.

Our HIV incidence estimate of 3.7 per 100 person-years in absence of the HPTN 082 intervention package was significantly higher than the observed incidence in the study of 1.0 per 100 person-years, which suggests that the package of PrEP with counseling, group support, SMS reminders may have been particularly effective. The impact of these interventions on sexual behavior rather than the effect of PrEP needs to be taken into account as the low observed HIV incidence could be partly due to lower likelihood of HIV exposure during study follow-up. In addition, the sample size of 427 participants and low number of infections in HPTN 082 implies large uncertainty in the point estimate of HIV incidence. A comprehensive analysis of the behavioral and PrEP usage data over time from HPTN 082 is under way to address these questions and provide more robust estimates of the intervention effectiveness.