Temporal change in population-level prevalence of detectable HIV viremia and its association with HIV incidence among key populations in India: a serial cross-sectional study

SUMMARY

Background:

Population-level prevalence of detectable HIV viremia (PDV) has been proposed as a metric for monitoring the population-level effectiveness of HIV treatment as prevention. We aimed to characterize temporal changes in PDV among people who inject drugs (PWID) and men who have sex with men (MSM) in India and evaluate community and individual-level associations with cross-sectional HIV incidence.

Methods:

Cross-sectional, baseline (10/2012 to 12/2013) and follow-up (08/2016 to 05/2017) respondent-driven sampling surveys were conducted among MSM (10 community sites) and PWID (12 community sites) aged ≥18 years in India. Annualized HIV incidence was estimated using validated multi-assay algorithms. PDV was calculated as the percentage of persons with detectable HIV RNA (>150 copies per mL) in a community site. Community-level associations were determined by linear regression. Multivariable multi-level Poisson regression assessed associations with recent HIV infection.

Findings:

We recruited 21,990 individuals and 21,726 individuals in the baseline and follow-up surveys, respectively. The median community-level HIV incidence estimate increased from 0·9% (range=0·0%-2·2%) at baseline to 1·5% (range=0·5%-3·0%) at follow-up among MSM and from 1·6% (range=0·5%-12·4%) to 3·6% (range=0·0%-18·4%) among PWID. At the community-level, every 1 percentage point increase in baseline PDV and temporal change in PDV between surveys were associated with higher annualized HIV incidence at follow-up (β for baseline PDV= 0·41 [95%CI=0·18, 0·63] and β for change in PDV=0·52 [95%CI=0·38, 0·66]). After accounting for individual-level risk factors, every 10 percentage point increase in baseline PDV and temporal change in PDV were associated with higher individual-level risk of recent HIV infection at follow-up (aRR for baseline PDV=1·85 [95%CI=1·44–2·37] and aRR for change in PDV=1·81 [95%CI=1·43–2·29]).

Interpretation:

PDV was temporally associated with community- and individual-level HIV incidence. These data support the scale-up of treatment as prevention programmes to reduce HIV incidence and the programmatic use of PDV to monitor community HIV risk potential.

INTRODUCTION

‘Treatment as prevention’ (TasP) serves as a foundational component of the global UNAIDS 95–95-95 fast-track targets, which aim to have 86% of all HIV-positive persons virally suppressed by 2030. The population-level effectiveness of TasP is based on the premise that reducing the number of individuals with onward HIV transmission risk potential in a given population will reduce the probability of an HIV-negative person having an infectious contact from whom they can acquire HIV. Accordingly, the proportion of people in the entire population with a detectable HIV viral load (VL) (i.e., prevalence of detectable viremia [PDV]) (in a particular geography) has been proposed as a measure to monitor the effectiveness of TasP and as a potential indicator of population-level HIV incidence.^1,2 Although PDV has been associated with population-level HIV incidence and the individual-level risk of HIV acquisition in generalized epidemic settings,^3–5 the temporal relationship between PDV and HIV incidence has not been examined in key populations.

While declines in population-level HIV incidence have been observed in generalized epidemic settings corresponding with TasP scale-up,^6–8 trends in HIV incidence among key populations have been highly variable.^9–11 In 2019, key populations and their sexual partners accounted for 62% of new adult HIV infections, globally.¹² The challenges of HIV epidemic control among key populations, such as men who have sex with men (MSM) and people who inject drugs (PWID), are compounded in low- and middle-income countries (LMIC) where they face many barriers to accessing HIV testing and treatment services as well as direct preventive services.^12,13 In 2012–2013, we observed high HIV prevalence and incidence among community-based samples of MSM and PWID across multiple Indian states,^14,15 despite continued declines among heterosexual populations.^16,17

Here, we characterize temporal changes in measures of HIV burden, including PDV and HIV incidence, among PWID and MSM using data from two serial, population-based, cross-sectional surveys conducted in 22 sites across 21 Indian cities. To assess the potential utility of PDV as an indicator of population-level HIV incidence, we examine the association of PDV and temporal change in PDV with community-level HIV incidence. To assess whether PDV is a determinant of HIV transmission, we examine the temporal association of PDV and temporal change in PDV with individual-level HIV risk.

METHODS

Study Design and Participants

Data were obtained from cross-sectional baseline and follow-up surveys conducted as a part of a cluster-randomized trial among MSM and PWID in India (ClinicalTrials.gov no.: NCT01686750).^18,19 The trial assessed the effectiveness of integrated care centers on improving population-level uptake of HIV testing, prevention, and treatment among MSM and PWID in 22 sites across 21 cities (10 MSM sites and 12 PWID sites); one city (New Delhi) had both an MSM and PWID site. The intervention was associated with improved uptake of HIV testing but this difference was not statistically significant and no associations were observed with secondary outcomes including HIV incidence.¹⁸ The baseline survey was conducted prior to the intervention between October 2012 and December 2013 and the follow-up survey was conducted after the intervention between August 2016 and May 2017. While the National AIDS Control Organization scaled HIV testing and ART during the study period, access to pre-exposure prophylaxis (PrEP) was limited.

Using identical methodology in both survey rounds, participants were recruited via respondent-driven sampling (RDS), a chain-referral strategy for “hidden” populations where no sampling frame exists.²⁰ Two or three individuals ( “seeds”) considered to be highly influential among key population members in the local city were given two coupons to recruit members of their key population networks. Each eligible recruit enrolled was given two coupons to recruit others. Recruitment was continued until a target of ~1000 participants was reached at each site. Eligible individuals were aged ≥18 years, provided informed consent, and possessed a valid RDS referral coupon. Eligible MSM self-identified as male and self-reported oral/anal sex with another man in the prior year. Eligible PWID self-reported injection drug use in the prior two years. A biometric (fingerprint) system was used to prevent duplicate enrollment in a given survey round. We considered key population members (MSM or PWID) sampled in a given site as a community.

This study was approved by institutional review boards at Johns Hopkins Medicine, Johns Hopkins Bloomberg School of Public Health, and the YR Gaitonde Centre for AIDS Research and Education. Participants provided oral consent. The study protocol is publicly available.¹⁹

Procedures

Study procedures have been previously described.^18,19 Briefly, trained interviewers administered face-to-face structured surveys. A blood sample was collected and all participants received on-site rapid HIV testing and counseling per Indian guidelines. Participants were reimbursed 250 Indian rupees. Blood samples were shipped to a lab in Chennai for processing, storage at −80°C, and additional testing.

Among HIV-positive participants, we determined CD4⁺ T-cell counts using the FlowCARE PLG CD4 assay (Beckman Coulter, CA), plasma HIV-1 RNA using the RealTime HIV-1 assay (Abbott Park, IL), and HIV antibody avidity (%) using the JHU-modified Bio-Rad Avidity assay based on the Genetic Systems HIV-1/HIV-2 PLUS O EIA kit (Bio-Rad Laboratories, Hercules, CA). We tested sera from HIV-positive participants by the Aware BED EIA HIV-1 incidence test (Calypte Biomedical Corporation, Portland, OR) in the 2012–2013 survey and the Limited Antigen (LAg)-Avidity EIA (Maxim Biomedical Inc, Rockville, MD) in the 2016–2017 survey.

In the 2012–2013 survey, HIV-positive participants were considered recently infected if they had a CD4⁺ T-cell count >200 cells per μL, HIV RNA >400 copies per mL, a BED-CEIA normalized optical density (OD-n) value <1·0, and a Bio-Rad HIV antibody avidity index (AI) <80%.²¹ In the 2016–2017 survey, HIV-positive participants were considered to be recently infected if they had a CD4⁺ T-cell count >50 cells per μL, an HIV VL >400 copies per mL, a LAg-Avidity OD-n value <2·9, and a Bio-Rad HIV antibody avidity index <80%.²² HIV-1 subtype C is the predominant subtype in India and both recent infection testing algorithms (RITA) have a similar mean duration of time individuals with HIV subtype C infection are classified as recently infected (mean window period [μ] = 0·56 and 0·52 years, respectively). Both RITAs have been validated against directly observed seroconversions in longitudinal cohorts in subtype C epidemic settings.^21–23

Procedures for other laboratory measures are provided in the appendix (p2).

Statistical Analysis

RDS process measures suggested satisfactory performance in terms of achieving equilibrium during recruitment and homophily by HIV status in 2012–2013 and 2016–2017.^14,15,18 Primary analyses were unweighted because the primary purpose of the study was to draw inferences from relative associations within the sample, and because there is no validated method to use RDS weights in calculating biomarker-based HIV incidence estimates.²⁴ However, we conducted sensitivity analyses using RDS-II weights where possible, and these methods are provided in the appendix (p2).

For each key population (MSM and PWID), we examined temporal changes in HIV prevalence, PDV (HIV-positive persons with >150 HIV RNA copies per mL / total population), prevalence of HIV VL non-suppression (HIV-positive persons with >150 HIV RNA copies per mL / all HIV-positive persons regardless of awareness of status), and annualized HIV incidence overall and stratified by age and gender. For prevalence measures, we calculated Clopper-Pearson 95% confidence intervals (CI) using cluster-robust standard errors. Age- and gender-adjusted prevalence ratios (aPR) were calculated to estimate the effect of study period on prevalent HIV measures using Poisson regression models with robust standard errors and a random-intercept for each site with a fixed effect for intervention status. Annualized HIV incidence rates (% per year) was calculated using the following estimator:

$$\left|I=\frac{w}{n\mu }\right.$$

where w is the number of HIV-positive individuals classified as having a recent infection by a given RITA, n is the number of HIV-negative individuals, and μ is the mean window period in years for the RITA. We estimated 95% CIs for HIV incidence rates by delta-method approximation and accounted for the uncertainty in the mean window period (μ) using the Assay-Based Incidence Estimation toolkit.²⁵ We conducted similar analyses stratified by site.

To evaluate the potential role of PDV as an indicator of community-level HIV incidence, we conducted several analyses with site as the unit of analysis (n=22). First, we examined community-level correlations by Spearman rank correlation coefficients; corresponding 95% CIs were drawn from 10,000 bootstrap replicates of paired measurements. Using linear regression, we assessed: 1) the cross-sectional association between PDV and community-level incidence within each survey round; 2) the temporal association between PDV in 2012–2013 and HIV incidence in 2016–2017; 3) the association of a temporal change (i.e., absolute difference) in PDV between 2012–2013 and 2016–2017 with HIV incidence in 2016–2017; and 4) associations of a temporal change in PDV with a temporal change in HIV incidence between 2012–2013 and 2016–2017. The latter two models included adjustment for baseline PDV in 2012–2013 and all but the baseline cross-sectional model included adjustment for site intervention status. We conducted sensitivity analyses omitting influential observations as determined by the DFBETA statistic (cut-off: ±0.43 [2/√22]).

To examine the association of PDV in 2012–2013 with individual-level recent (i.e., incident) HIV infection in 2016–2017, we conducted multi-level Poisson regression with robust variance and a random-intercept for each site. We estimated risk ratios (RR) of recent HIV infection (vs. HIV-negative participants); ‘non-recent’ (prevalent) HIV-positive participants were excluded. We conducted separate analyses modeling PDV continuously and categorically (<10%, 10–20%, and >20%). To account for differences in individual-level risk, multivariable models included adjustment for key population type (MSM or PWID), a log-transformed population type-specific, individual-level HIV risk score (continuous), and an interaction term between these variables. Estimation of the HIV risk scores and model performance characteristics are described in the appendix (p3). Separately for each key population, individual-level risk scores of recent HIV infection (vs. HIV-negative participants) were calculated using a logistic regression model with individual-level covariates determined a priori. The PWID risk score model accounted for age, gender, marital status, educational attainment, incarceration in the past 6 months, sharing injection equipment in the past 6 months, transactional sex in the past 6 months, and hepatitis C virus IgG antibody serostatus. The MSM risk score model accounted for age, income, injection drug use in the past 6 months, number of male partners in the past 6 months, any sex with a woman in the past 6 months, and herpes simplex virus type-2 IgG antibody serostatus. Given the long lag between surveys, we used similar methods to assess the cross-sectional association between PDV and recent HIV infection within each survey round. Except for the baseline cross-sectional analysis, multivariable Poisson models also included adjustment for intervention status.

We conducted a separate multilevel analysis to assess the association between a temporal change in PDV between 2012–2013 and 2016–2017 and risk of recent HIV infection in 2016–2017 using similar methods. The first model included adjustment for baseline PDV in 2012–2013. The second model additionally included adjustment for key population type, the log-transformed HIV risk score, and their interaction, as well as intervention status.

We performed subgroup multilevel analyses stratified by PWID and MSM; however, recent HIV cases in the MSM stratum were limited. For PWID, we constructed additional multivariable models that adjusted for self-reported community-level coverage of harm reduction service utilization in the past 6 months (i.e., medication for opioid use disorder overall and syringe service programs among active PWID).

We included adjustment for community-level HIV prevalence in a sensitivity analysis. We also repeated primary analyses after re-estimating HIV incidence in the 2016–2017 survey using two alternative RITAs previously validated in HIV subtype C epidemic settings (appendix p2). Unless specified otherwise, analyses were done using Stata/MP, version 15·1 and R statistical software, version 3·6·1.

RESULTS

We recruited 21,990 participants (9,997 MSM and 11,993 PWID) between Oct 1, 2012 and Dec 19, 2013, and 21,726 participants (10,005 MSM and 11,721 PWID) between Aug 1, 2016 and May 28, 2017. Among PWID, the median age was 29 (IQR=24–35) in 2012–2013 and 30 (IQR=24–36) in 2016–2017. Most PWID were men in 2012–2013 (n=11,299 [94·2%]) and 2016–2017 (n=11,221 [95·7%]). Among MSM, the median age was 26 (IQR=22–34) in 2012–2013 and 29 (IQR=23–37) in 2016–2017. Community-level characteristics are shown in the appendix (pp4-5).

Overall, age-, and gender-specific data on HIV prevalence, prevalence of VL non-suppression and PDV are provided for each population type and survey period in the appendix (p6). Notably, there was a significant increase over time in HIV prevalence among MSM overall (10·9% [1,086/9,995] vs. 17·6% [1,763/10,003]; aPR=1·34 [95%CI=1·07–1·69]), and particularly among young MSM aged 18–24 years (2·4% [98/4,088] vs. 5·9% [183/3,104]; aPR=2·53 [95%CI=1·52–4·23]). In both survey periods, younger age was associated with a higher prevalence of HIV VL non-suppression among HIV-positive PWID and MSM. Temporal changes in prevalent HIV measures varied substantially within and between sites; changes in PDV in a given site reflected the direction of change in both HIV prevalence and the prevalence of VL non-suppression among HIV-positive persons (Table 1). For instance, among MSM in New Delhi, there was limited change in PDV over time (aPR=0·87 [95%CI=0·60–1·26]) despite a reduction in HIV VL non-suppression (aPR=0·51 [95%CI=0·40–0·64]), because there was an increase in HIV prevalence (aPR=1·63 [95%CI=1·27–2·09]). RDS-adjusted population-level prevalence estimates were similar to crude estimates (appendix [pp7-12]).

Table 1.

Temporal changes in prevalent measures of HIV burden stratified by site.

Site	Prevalence of detectable HIV viremia (among all participants with and without HIV)			Prevalence of HIV infection			Prevalence HIV viral load non-suppression (among participants with HIV)
	2012–2013, % (95% CI)	2016–2017, % (95% CI)	2016–2017 vs. 2012–2013, aPR (95% CI)	2012–2013, % (95% CI)	2016–2017, % (95% CI)	2016–2017 vs. 2012–2013, aPR (95% CI)	2012–2013, % (95%CI)	2016–2017, % (95%CI)	2016–2017 vs. 2012–2013, aPR (95% CI)
PWID
Aizawl	23.3 (20.7–26.0)	23.2 (20.6–25.9)	1.06 (0.90–1.24)	30.6 (27.8–33.6)	28.7 (25.9–31.6)	0.99 (0.87–1.13)	76.1 (71.0–80.8)	80.8 (75.8–85.2)	1.04 (0.96–1.14)
Amritsar	20.1 (17.7–22.7)	17.9 (15.6–20.4)	0.84 (0.70–1.01)	22.6 (20.1–25.3)	22.1 (19.6–24.8)	0.92 (0.78–1.08)	88.9 (84.1–92.7)	81.0 (75.2–85.9)	0.91 (0.85–0.98)
Bilaspur	14.2 (12.1–16.5)	17.9 (15.6–20.4)	1.27 (1.04–1.56)	15.3 (13.1–17.7)	19.9 (17.5–22.5)	1.32 (1.09–1.60)	92.8 (87.5–96.4)	89.9 (84.9–93.8)	0.97 (0.90–1.04)
Chandigarh	10.7 (8.9–12.8)	6.2 (4.8–7.9)	0.59 (0.43–0.79)	12.1 (10.2–14.3)	8.1 (6.5–10.0)	0.68 (0.52–0.88)	88.4 (81.3–93.5)	76.5 (65.8–85.2)	0.87 (0.76–0.995)
Churachandpur	14.1 (12.0–16.4)	10.0 (8.2–12.0)	0.71 (0.56–0.91)	27.9 (25.1–30.8)	23.9 (21.3–26.7)	0.85 (0.74–0.98)	50.5 (44.5–56.6)	41.8 (35.5–48.4)	0.91 (0.76–1.11)
New Delhi	15.7 (13.5–18.1)	38.2 (35.2–41.3)	2.40 (2.03–2.83)	16.6 (14.4–19.1)	39.9 (36.8–43.0)	2.38 (2.03–2.80)	95.7 (91.4–98.3)	95.7 (93.3–97.5)	0.99 (0.95–1.03)
Dimapur	9.3 (7.6–11.3)	7.3 (5.8–9.1)	0.81 (0.60–1.09)	19.7 (17.3–22.3)	14.4 (12.3–16.7)	0.71 (0.59–0.86)	47.2 (40.1–54.4)	51.0 (42.6–59.5)	1.16 (0.93–1.43)
Imphal	18.2 (15.9–20.7)	9.2 (7.5–11.2)	0.49 (0.39–0.61)	31.9 (29.0–34.9)	30.4 (27.6–33.4)	0.88 (0.78–0.98)	57.1 (51.4–62.6)	30.3 (25.1–35.8)	0.59 (0.48–0.71)
Kanpur	33.8 (30.9–36.8)	23.5 (20.9–26.3)	0.72 (0.62–0.83)	35.3 (32.3–38.4)	26.1 (23.4–28.9)	0.76 (0.67–0.87)	95.8 (93.1–97.6)	90.7 (86.5–93.9)	0.95 (0.91–0.99)
Ludhiana	19.4 (17.0–22.0)	12.6 (10.6–14.9)	0.64 (0.52–0.79)	22.8 (20.2–25.5)	20.2 (17.7–22.8)	0.90 (0.76–1.06)	85.1 (79.8–89.4)	64.8 (57.6–71.5)	0.77 (0.69–0.86)
Lunglei	3.9 (2.8–5.3)	2.8 (1.9–4.0)	0.64 (0.39–1.03)	10.2 (8.4–12.3)	10.4 (8.6–12.5)	0.85 (0.66–1.10)	38.2 (28.8–48.4)	26.9 (18.7–36.5)	0.77 (0.51–1.15)
Mumbai	8.3 (6.7–10.2)	8.6 (6.7–10.9)	0.98 (0.71–1.34)	9.4 (7.7–11.4)	10.8 (8.6–13.3)	1.09 (0.82–1.45)	88.3 (80.0–94.0)	80.5 (69.9–88.7)	0.92 (0.81–1.05)
MSM
Bengaluru	6.4 (5.0–8.1)	7.6 (6.0–9.4)	1.15 (0.82–1.62)	9.1 (7.4–11.1)	13.4 (11.3–15.7)	1.35 (1.05–1.75)	70.3 (59.8–79.5)	56.7 (47.9–65.2)	0.80 (0.66–0.97)
Belgaum	4.4 (3.2–5.9)	4.2 (3.0–5.6)	0.87 (0.58–1.31)	6.0 (4.6–7.6)	10.2 (8.4–12.2)	1.55 (1.14–2.10)	75.9 (62.8–86.1)	41.2 (31.5–51.4)	0.54 (0.41–0.72)
Bhopal	5.1 (3.8–6.7)	11.2 (9.3–13.3)	1.56 (1.08–2.24)	6.1 (4.6–7.8)	16.3 (14.0–18.7)	1.72 (1.27–2.33)	83.6 (71.9–91.8)	68.7 (61.0–75.7)	0.85 (0.73–0.98)
Chennai	2.4 (1.5–3.6)	3.2 (2.2–4.5)	1.00 (0.55–1.79)	6.6 (5.1–8.3)	6.1 (4.7–7.8)	0.61 (0.42–0.87)	36.4 (24.9–49.1)	52.5 (39.3–65.4)	1.26 (0.86–1.86)
Coimbatore	9.0 (7.3–11.0)	6.3 (4.9–8.0)	0.73 (0.54–0.99)	15.9 (13.7–18.3)	14.2 (12.1–16.5)	0.96 (0.78–1.17)	56.6 (48.5–64.4)	44.4 (36.0–52.9)	0.78 (0.63–0.96)
New Delhi	6.0 (4.6–7.7)	6.7 (5.2–8.4)	0.87 (0.60–1.26)	8.0 (6.4–9.9)	19.4 (17.0–22.0)	1.63 (1.27–2.09)	75.0 (64.1–84.0)	34.5 (27.9–41.7)	0.51 (0.40–0.64)
Hyderabad	13.2 (11.2–15.5)	8.5 (6.8–10.4)	0.58 (0.45–0.75)	18.8 (16.5–21.4)	20.8 (18.3–23.4)	0.93 (0.79–1.09)	70.6 (63.5–77.0)	40.9 (34.1–47.9)	0.70 (0.58–0.84)
Madurai	4.7 (3.5–6.2)	6.3 (4.9–8.0)	1.22 (0.85–1.77)	13.0 (11.0–15.2)	18.6 (16.2–21.2)	1.23 (1.03–1.48)	36.4 (28.1–45.4)	33.9 (27.1–41.2)	0.95 (0.70–1.30)
Vijayawada	10.4 (8.6–12.5)	16.2 (14.0–18.6)	1.46 (1.15–1.85)	13.5 (11.4–15.8)	32.0 (29.1–35.0)	2.04 (1.70–2.45)	78.2 (70.2–84.9)	50.6 (45.0–56.2)	0.71 (0.61–0.81)
Visakhapatnam	4.9 (3.6–6.4)	9.4 (7.7–11.4)	1.90 (1.37–2.65)	11.6 (9.7–13.7)	25.3 (22.6–28.1)	2.01 (1.67–2.41)	42.2 (33.1–51.8)	37.9 (31.8–44.3)	0.85 (0.66–1.11)

Note: Estimates are unweighted percentages with corresponding Clopper-Pearson 95% confidence intervals. To assess changes in prevalent HIV measures by survey period, adjusted prevalence ratios were estimated from multivariable Poisson regression models with robust standard errors. The multivariable models included adjustment for age and gender for PWID sites and age for MSM sites. Separate models were used for each community site.

Abbreviations: aPR, adjusted prevalence ratio; CI, confidence interval; MSM, men who have sex with men; PWID, people who inject drugs; VL, viral load.

Overall, there were 197 participants (47 MSM and 150 PWID) and 310 participants (62 MSM and 248 PWID) with a recent (i.e., incident) HIV infection in the 2012–2013 and 2016–2017 surveys, respectively. Among PWID, overall annualized HIV incidence was 2·8% (95%CI=2·3–3·4) in 2012–2013 and 5·2% (95%CI=4·2–6·1) in 2016–2017 (Figure 1). Among MSM, HIV incidence was 0·9% (95%CI=0·7–1·2) in 2012–2013 and 1·4% (95%CI=1·0–1·9) in 2016–2017. In 2016–2017, persons aged 18–24 years had the highest age-specific HIV incidence among PWID and among MSM. At the community-level, the median HIV incidence increased from 1·6% (IQR=0·9%-4·3%) in 2012–2013 to 3·6% (IQR=1·0%-8·6%) in 2016–2017 among PWID and from 0·9% (IQR=0·6%-1·4%) to 1·5% (IQR=0·8%-2·0%) among MSM (Wilcoxon signed-rank P=0·0844 for PWID and P=0·0367 for MSM). There were notable increases in HIV incidence among PWID in Aizawl, New Delhi, and Ludhiana, and among MSM in Bhopal (Figure 2; appendix pp13).

Figure 1. Temporal changes in annualized HIV incidence estimates stratified by age group and gender.

Error bars indicate 95% confidence intervals.

Figure 2. Temporal changes in annualized HIV incidence estimates stratified by community site.

Error bars indicate 95% confidence intervals.

At the community-level, there were significant cross-sectional and temporal correlations between PDV and community-level HIV incidence (Figure 3). Higher PDV in 2012–2013 was associated with higher community-level HIV incidence in 2016–2017. After adjustment for baseline PDV and intervention status, a one percentage point increase in PDV from 2012–2013 to 2016–2017 was associated with higher community-level HIV incidence in 2016–2017 (β=0·52 [95%CI=0·38, 0·66]) (Figure 4), as well as a 0·45 [95%CI=0·26, 0·65] percentage point increase in community-level HIV incidence over the same period. These associations were robust after excluding influential sites (appendix pp14), with the exception of the association between a temporal change in PDV and a temporal change in community-level HIV incidence excluding Aizawl, New Delhi, and Kanpur (β=0·04 [95%CI=-0·13, 0·21).

Figure 3. Community-level correlation between the prevalence of detectable HIV viremia and annualized HIV incidence.

95% confidence intervals for spearman correlation coefficients were estimated from 10,000 bootstrap replications. Community site-level linear regression was used to assess the association between the prevalence of detectable HIV viremia and annualized HIV incidence (β refers to percentage point change in annualized HIV incidence per percentage point increase in prevalence of detectable HIV viremia); the model predicted line of fit is shown in black (panels A, B, C). Models examining HIV incidence in 2016–2017 included adjustment for community intervention status from the primary trial (panels B and C).

Figure 4. Community-level association of the temporal change in prevalence of detectable HIV viremia with annualized HIV incidence and temporal change in annualized HIV incidence.

Panel A shows the correlation between observed difference in PDV between surveys and annualized HIV incidence in 2016–2017. Panel B shows the adjusted fitted line for the predicted annualized HIV incidence in 2016–2017 as a function the difference in PDV between surveys. Panel C shows the correlation between observed difference in PDV between surveys and observed difference in annualized HIV incidence between surveys. Panel D shows the adjusted fitted line for the predicted change in annualized HIV incidence between surveys as a function of the change in PDV between surveys. Both linear regression models (Panels B, D) included adjustment for community intervention status and the baseline prevalence of detectable HIV viremia in 2012–2013. The prediction bands reflect pointwise 95% confidence intervals.

In cross-sectional and temporal analyses, higher PDV was associated with increased risk of recent HIV infection independent of intervention status and individual-level HIV risk factors (Table 2). Additionally, for every 10-percentage point increase in the PDV between 2012–2013 and 2016–2017, there was an 81% increased risk of recent HIV infection in 2016–2017 after accounting for intervention status, baseline PDV in 2012–2013, and individual-level HIV risk factors (aRR=1·81 [95%CI=1·43–2·29]). These positive associations were robust to additional adjustment for community-level HIV prevalence and temporal change in community-level HIV prevalence in sensitivity analyses (appendix pp15).

Table 2.

Association between the community prevalence of HIV viremia and the individual-level risk of recent HIV infection.

Predictor variable	Incident HIV Infection in 2012–2013			Incident HIV Infection in 2016–2017
	No. Recent /No. HIV-Negative	RR (95% CI)	aRR (95% CI)	No. Recent /No. HIV-Negative	RR (95% CI)	aRR (95% CI)
PDV in 2012–2013 (per 10% increase		2.61 (1.93–3.54)	1.90 (1.41–2.56)	-	2.64 (1.85–3.77)	1.85 (1.44–2.37)
PDV in 2012–2013 (categorical)
<10%	39/9,839	Ref.	Ref.	58/9,163	Ref.	Ref.
10–20%	88/6,404	3.30 (1.85–5.91)	2.12 (1.27–3.52)	136/6,047	2.76 (1.22–6.22)	1.25 (0.69–2.25)
>20%	70/2,114	6.99 (2.26–21.63)	3.21 (1.11–9.30)	116/2,230	8.42 (4.72–15.01)	2.89 (1.25–6.68)
PDV in 2016–2017 (per 10% increase)	-	-	-	-	2.82 (2.04–3.88)	2.04 (1.53–2.71)
PDV in 2016–2017 (categorical)
<10%	-	-	-	62/10,731	Ref.	Ref.
10–20%	-	-	-	95/4,657	3.45 (2.22–5.36)	2.26 (1.36–3.76)
>20%	-	-	-	153/2,052	12.34 (8.45–18.01)	6.81 (3.81–12.17)
Change in PDV between 2012–2013 and 2016–2017 (per 10% increase	-	-	-	-	2.20 (1.70–2.85)*	1.81 (1.43–2.29)*

Separate models were used for each relationship shown. This analysis excluded people with HIV that were considered to have a prevalent long-term (i.e., non-recent) HIV infection by the given recent infection testing algorithm. Risk ratios (RR) of recent (i.e., incident) HIV infection were estimated from Poisson regression models with a site-level random-intercept and robust standard errors. Adjusted RRs (aRR) were estimated from multivariable models that were adjusted for a log-transformed population type-specific individual-level HIV risk score, population type (MSM/PWID), and an interaction term between population type and the HIV risk score to account for individual-level risk factors. Multivariable models assessing incident infections at the follow-up evaluation survey in 2016–2017 were additionally adjusted for community intervention status from the primary trial.

^*Additionally adjusted for baseline prevalence of detectable viremia (PDV) in 2012–2013.

The positive temporal association between PDV and recent HIV infection was also observed among PWID, including after additional adjustment for community-level harm reduction service coverage (aRR=1·90 [95%CI=1·38–2·61]); appendix pp16). Among MSM, there was a crude temporal association between PDV and recent HIV infection (RR=1·84 [95%CI=0·99–3·40]), but this association was attenuated in multivariable analysis (aRR=1·22 [95%CI=0·54–2·73]) (appendix pp17). Similarly, a temporal change in PDV was associated with recent HIV infection among PWID but not MSM. Primary overall inferences were similar when using alternative RITAs to estimate HIV incidence in the 2016–2017 survey (appendix pp18-19) and when incorporating RDS-II weights (appendix pp20).

DISCUSSION

In this multi-city study in India, we observed substantial increases in HIV prevalence and incidence among PWID and MSM over a four-year period. We further demonstrate that PDV, which combines information from HIV prevalence and HIV VL non-suppression among HIV positive persons, and change in PDV over time were associated with community-level and individual-level HIV incidence, providing strong support for its role as an indicator of transmission risk. Our study particularly highlights the heightened vulnerability of young PWID and MSM—groups that had high HIV incidence and HIV VL non-suppression overall—despite decreases in HIV VL non-suppression over time in many communities.

This study builds upon our prior work demonstrating a strong positive cross-sectional correlation between PDV and community HIV incidence among MSM and PWID.² Tanser and colleagues and the UTT Trials Consortium have also shown that PDV is associated with community-level HIV incidence and individual-level HIV risk in generalized epidemic settings in sub-Saharan Africa.^3,4 Here, we show that PDV is positively associated with community-level HIV incidence and individual-level HIV risk among key populations living in regions where general population HIV prevalence is low. Our study also uniquely demonstrates that a temporal change in PDV in a given community is a significant predictor of HIV incidence at the community and the individual-levels with some suggestion of an association with a temporal change in community-level HIV incidence as well. Collectively, these data support the use of PDV as an indicator for regional HIV transmission risk in routine surveillance.

Additionally, the positive temporal association between PDV and increased individual-level HIV risk supports the scale up of TasP to reduce HIV incidence; however, the degree to which decreased population viremia will reduce HIV incidence remains unclear. One potential explanation for the lack of effectiveness in the four community-randomized trials of universal testing and treatment in generalized epidemic settings is heterogeneity in uptake of the intervention across risk groups; ^26–29 for example, it is possible that those who engage in behaviors associated with increased risk of transmission such as injection drug use are also less likely to be on ART.³⁰ The present analysis demonstrates there was substantial heterogeneity in transmission risk between sites, key populations and demographic characteristics. In particular, young PWID and MSM had higher HIV incidence as well as higher HIV VL non-suppression among those living with HIV. To maximize the benefits of TasP, innovative models of care are needed to increase HIV VL suppression among key populations that are most at risk including scale-up of rapid ART initiation, community-based ART and PrEP for those who are HIV-negative.

This study leveraged a large sample of key populations; however, there are limitations. First, although the multi-city design captured heterogeneous HIV epidemics in India, the sites were purposively selected and the study populations may not be nationally representative. Second, despite recruiting a large sample using RDS, there was no available sampling frame in any city. While we could use the RDS-II estimator to report weighted population-level prevalence measures of HIV burden in sensitivity analyses, methods for RDS-weighted HIV incidence estimates are underdeveloped.²⁴ Third, when examining associations with PDV, we assumed there was no interaction with individuals outside those who were recruited in a given city. Fourth, ecologic analyses were limited to a relatively small sample size and results regarding the temporal change in PDV and temporal change in HIV incidence were driven by a few influential sites, given that most sites did not observe substantial changes in HIV incidence. This particular relationship should be further explored in larger data sets with shorter periods between sampling.

Moreover, despite using a multi-level approach to assess the association of PDV with HIV risk, the study findings may be subject to ecological fallacy and residual confounding. We used a risk-score approach to account for well-established individual-level determinants of HIV transmission in this population; however, the majority of these variables relied on self-report. Although our results were robust to adjustment for community-level coverage of harm reduction use among PWID, these estimates may only reflect coverage in the study population (not the population-at-large) and we were unable to conduct similar analyses for MSM given limited cases.

While it can be considered a limitation that we did not directly observe HIV seroconversions, the use of RITAs may be preferable in key populations who are susceptible to changes in behavior owing to study participation (i.e., the Hawthorne effect) and/or prone to loss-to-follow-up. Reassuringly, the RITAs applied in this study were previously validated against closed longitudinal cohorts.^21–23 Although it is a limitation that the particular RITA used in each survey round differed due to a change in assay availability, both RITAs had a similar window period.^21,22 Indeed, estimating HIV incidence using RITAs can be challenging to implement given that they often change and often require whole blood sample collection for CD4 estimation or other assays (e.g., LAg-Avidity) in addition to VL testing, thereby complicating sample collection and transportation. In contrast, PDV, which relies on VL testing alone, can be estimated from a dried blood spot and potentially a point-of-care assay with minimal infrastructure or training (e.g., Cepheid GeneXpert). Given that VL testing is also becoming more routine, globally, monitoring PDV may be a more practical approach to simultaneously monitor HIV transmission risk and estimate program effectiveness in resource-limited settings.

Here, we show that HIV burden remains unacceptably high and is potentially increasing among PWID and MSM in India. HIV epidemic control in key populations will likely be further challenged during the COVID-19 pandemic, owing to potentially greater social marginalization and reduced service access. PDV, a measure that can be estimated with relative ease compared to other biomarkers of HIV incidence, serves as a marker of both HIV transmission as well as program effectiveness. Incorporation of PDV into routine surveillance efforts could assist to rapidly and efficiently detect HIV outbreaks and prioritize reponse, while simultaneously monitoring progress towards HIV elimination goals.

RESEARCH IN CONTEXT.

Evidence before this study

We did a search in PubMed on November 12, 2020, using the search terms (“HIV incidence”[All Fields] OR “HIV transmission”[All Fields] OR “HIV acquisition”[All Fields] OR “HIV seroconversion”[All Fields] OR “incident HIV”[All Fields]) AND (“viremia”[All Fields] OR “viremic”[All Fields] OR “non-suppression”[All Fields] OR “unsuppressed”[All Fields] OR “detectable viral load”[All Fields] OR “detectable HIV viral load”[All Fields] OR “community viral load”[All Fields] OR “population viral load”[All Fields]), which returned 264 articles. No date or language restrictions were used.

Traditional measures of community HIV viral load as an index for population-level HIV transmission risk potential have been criticized for failing to account for the underlying prevalence of HIV in the population. Thus, the prevalence of detectable viremia (PDV) in the overall population has been advocated as a potential measure to monitor the effectiveness of HIV treatment as prevention. We previously conducted a cross-sectional, respondent-driven sampling study of key populations in India and found an ecological association between PDV and biomarker-based estimates of HIV incidence. While this study was informative in demonstrating that PDV was a better predictor of community HIV incidence than traditional measures of community viral load, it was limited by the lack of temporality and potential for ecologic fallacy. A population-based, prospective cohort study subsequently demonstrated a temporal association between community PDV and individual-level risk of HIV infection in a generalized epidemic setting in South Africa. The temporal association between PDV and HIV incidence has not been examined using community- or individual-level data in key populations.

Added value of this study

This study provides further evidence for the use of PDV as a meaningful indicator for HIV incidence in community-level surveillance. Specifically, the study adds to the already available evidence of a cross-sectional association between PDV and incidence in key populations and a temporal association in a generalized epidemic context by demonstrating a strong temporal association in a concentrated epidemic setting. Moreover, this study demonstrates an association between temporal changes in PDV and subsequent HIV incidence as well as temporal changes in HIV incidence. To be specific, this study uniquely demonstrates that a temporal change in community PDV is associated with HIV incidence at the individual- and community-levels. Furthermore, we show that a large temporal increase in community-level HIV incidence can be detected by an increase in PDV.

Implications of all the available evidence

The data available support the use of PDV as a metric to assess the population-level effectiveness of TasP and rapidly monitor regional HIV risk potential. Given that VL testing is being more commonly conducted even in LMIC and VL testing is unlikely to appreciably change in sensitivity over time, monitoring PDV may be a more practical approach to monitor HIV transmission risk over time as part of integrated biobehavioral surveillance approaches, especially in resource-limited settings. Higher community-level PDV was associated with higher HIV incidence at the individual- and community-levels, and the sites with the largest absolute change in PDV were also the sites with largest absolute change in community-level HIV incidence. Collectively, these findings also support and confirm the role of TasP programmes in reducing HIV incidence at the population-level among key populations.