|Home | About | Journals | Submit | Contact Us | Français|
To assess and summarize the published literature on the extent to which bacterial vaginosis (BV) may increase the risk of HIV acquisition.
Meta-analysis of published studies.
MEDLINE and other electronic databases were systematically searched for eligible publications. The association between BV and incident HIV was separately analyzed from that between BV and prevalent HIV. The latter were further analyzed stratified by BV diagnostic method, HIV risk profile of the study population and whether or not adjusted estimates were presented.
Twenty-three eligible publications were identified, including a total of 30,739 women. BV was associated with an increased risk of HIV acquisition in HIV-incidence studies (relative risk = 1.6, 95% CI: 1.2, 2.1). All but one of 21 HIV-prevalence studies reported estimates above the null. The latter results were heterogeneous and showed some evidence of funnel plot asymmetry, precluding the estimation of a single summary measure. The association between BV and HIV in prevalence studies appeared stronger for women without high-risk sexual behavior.
BV was consistently associated with an increased risk of HIV infection. High BV prevalence may result in a high number of HIV infections being attributable to BV. More prospective studies are needed to accurately evaluate the role of BV in HIV acquisition in low versus high risk women. Furthermore, randomized clinical trials may be worth considering to determine the effect of BV control measures on HIV acquisition.
Bacterial vaginosis (BV) is the most frequent type of vaginitis in women of reproductive age [1–3]. BV is an imbalance in the ecology of the normal vaginal flora  that is characterized by the depletion of lactobacilli , and the proliferation of anaerobic bacteria such as Gardnerella vaginalis, Morbilincus species, Prevotella species, Mycoplasma hominis and the recently identified Atopobium vaginae [2, 5–7]. It most often manifests clinically as a vaginal pH of >4.5, the presence of thin whitish homogenous vaginal discharge, the detection of “clue” cells and the presence of an amine odor after the addition of 10 percent potassium hydroxide [8, 9]. BV has been shown to increase the risk of adverse gynecological and obstetrical outcomes such as preterm delivery[10, 11] pelvic inflammatory disease (PID) and upper genital tract infections [12–14]. However, the effect of BV on the risk of HIV infection in women has not been clearly quantified.
The magnitude of the association between BV and HIV has varied in epidemiological studies, ranging from the absence of any association to a near four-fold odds of being HIV infected among BV-positive women compared to BV-negative women . BV is estimated to be the most prevalent vaginal infection particularly in countries with high HIV prevalence . If BV is confirmed to increase the risk of HIV infection, the treatment of BV could be a meaningful intervention to prevent HIV acquisition. In a 2001 review of the role of sexually transmitted diseases in HIV acquisition, Rottingen et al estimated that BV was associated with a 40% increase in the risk of HIV based on an analysis of two studies. Obtaining a precise and updated estimate of the strength of the association between BV and HIV from published studies could be useful in predicting the potential impact of the control of BV on HIV incidence rates in a population. This prediction could also be more accurate if factors that modify the strength of the BV and HIV association were identified.
This paper aims to systematically review all published studies of the association between BV and HIV infection. Estimates of the association between BV and HIV are presented for both HIV incidence and prevalence studies, and analyzed for potential modification factors, publication bias, and heterogeneity of study results.
Medline was searched for peer-reviewed publications on the association between BV and HIV using the following terms: (“Vaginosis, Bacterial”[MeSH] OR “bacterial vaginosis” OR (“Bacterial Infections”[MeSH] AND (“Vaginitis” [MeSH] OR vagina*)) OR “Gardnerella”[MeSH]) AND (HIV OR “HIV Seropositivity”[MeSH] OR “HIV”[MeSH] OR “HIV Infections”[MeSH]). This search yielded 281 published articles. The Web of Science™, Popline™ and NLM Gateway™ databases were further searched using the keywords “bacterial vaginosis and HIV”. These searches yielded 193, 83 and 166 publications, respectively, but no additional articles were located beyond the Medline search. Articles were included in this meta-analysis if the magnitude of the association between BV and HIV was presented or could be calculated from the information provided in the article, and if there was a clear description of the diagnostic methods used for ascertaining both BV and HIV infections. When there was evidence of multiple publications of the same study over time, only the article with the largest sample size was included. Although non-English articles were eligible for inclusion, none were found. Conference abstracts and other unpublished manuscripts were excluded, as these could not be systematically reviewed.
For each study, the following information was extracted: author(s), year of publication, study site, year(s) during which study was conducted, study design (HIV incidence or prevalence), HIV risk of study population (low or high), BV diagnosis criteria, HIV ascertainment method, mean or median age of participants, crude and adjusted measures of the association between BV and HIV, and variables for which measures were adjusted. Appropriate measures of association were calculated, when possible, for studies that did not present them.
HIV incidence studies were defined as those that recruited HIV-negative women and prospectively measured incident HIV-infection. HIV prevalence studies were those that assessed BV and HIV status at the same time. Study populations with HIV-risk factors such as female sex workers (FSWs), sexually transmitted infection (STI) clinic attendees and regular partners of HIV infected males were classified as ‘High HIV-risk groups’. Women without any of these characteristics were classified as being in a low HIV-risk group. BV diagnostic criteria included the Nugent scoring system, Amsel clinical criteria and modifications of Amsel criteria. In the Nugent scoring system (BV when score ≥ 7) Gram stained vaginal smears were assessed for the average number of bacterial morphotypes seen per oil immersion field with Lactobacilli being scored from 0–4, Gardnerella/Bacteroides spp from 0–4 and curved gram-variable rods 0–2. Amsel criteria define BV as the presence of any three of: i. abnormal vaginal discharge, ii. vaginal pH>4.5, iii. presence of clue cells, and/or iv. positive amine test with release of fishy odor on addition of 10 percent KOH to vaginal secretions. Modifications of the Amsel criteria included diagnosing BV only when all four of the above criteria are present or when only two of the four elements are present. BV prevalence at baseline or BV prevalence in controls of case-control studies was recorded as an estimate of BV prevalence in the study population.
All abstracted data were double checked to confirm accuracy. Separate meta-analyses were conducted for HIV incidence studies and HIV prevalence studies (STATA™ version 8). Study estimates, relative risks (RR) or prevalence odds ratios (POR), and corresponding 95 percent confidence intervals were plotted for each analysis. Adjusted estimates, when reported, were preferentially used in analyses. For studies that reported estimates using two different methods of BV diagnosis, the estimate based on the Nugent score was used because the definition of BV using this score was more consistent across studies than the use of Amsel diagnostic criteria. Reported estimates were examined for publication bias graphically, using funnel plots, and statistically using the test of Begg and Mazumdar , the test of Egger et al. , and Duval and Tweedie’s “trim and fill” imputation method . Homogeneity test p-values were computed from Cochran’s Q test statistic . To explore sources of heterogeneity, HIV prevalence studies were stratified by HIV risk group, BV diagnostic criteria and whether there was any adjustment for confounders or not. Random effects meta-regression models, using restricted maximum likelihood to estimate the among-study variance , were also conducted to assess the association of average study estimates with study characteristics. BV diagnostic method, HIV risk group and a variable indicating the use of adjusted or unadjusted estimates were the independent variables in the models. The impact of each estimate on the summary estimate was explored using influence analyses (14).
Twenty-three eligible papers were identified altogether including 30,739 women. These reported 29 different estimates of the association between BV and HIV for 25 study populations (two papers each reported estimates for two distinct study populations [16, 21], and four other papers each reported estimates using two different methods of BV diagnosis [15, 22–24]). Four of the 25 study populations were followed prospectively in Kenya , Malawi  and South Africa  to access incident HIV infection. The remaining 21 studies were studies of prevalent HIV (table 1). Five of these were conducted in the USA, two in Thailand and the remainder in sub-Saharan Africa.
BV was diagnosed either using clinical criteria only (in 13/25 study populations), Nugent’s score only (in 8/25 study populations) or both clinical and Nugent’s score (in 4/25 study populations). BV prevalence in these studies ranged from 11.1 percent in women aged 20–35 years in the USA to 70.0 percent in STI symptomatic women in South Africa. The pooled BV prevalence was 33%. BV prevalence was consistently higher using Nugent’s criteria as opposed to clinical criteria in all four studies that used both BV diagnostic methods [15, 22–24].
There was little evidence of funnel plot asymmetry (Begg p=1.0, Egger p=0.2) or heterogeneity in the study estimates (p=0.7, figure 1). BV was associated with an increased risk of HIV acquisition in HIV incidence studies (figure 1) (RR=1.61; 95 percent CI: 1.21, 2.13).
Funnel plot tests gave some evidence of asymmetry (Begg p=0.3, Egger p=0.06). In the ‘trim and fill’ analysis, when a random-effects model was used, one hypothetically missing result was imputed. When a fixed-effect model was used, 9 hypothetically missing results were imputed (Figure 3).
HIV seroprevalence was higher in BV positive women in all but one study (17) (figure 2). The prevalence odds ratios (POR) estimates from these prevalence studies were highly heterogeneous (p<0.0005), and ranged from 0.77 to 3.70. As shown in table 2, this heterogeneity persisted even within strata defined by each of HIV risk group, BV diagnostic method, or estimate adjustment. POR estimates in HIV low-risk groups, tended to be higher than that in HIV high-risk groups. It is worth noting that all but one of the studies classified as low-risk group were studies of pregnant women. In a multivariate meta-regression for prevalence studies, controlling for differences in BV diagnostic method and adjusted versus non adjusted estimates, the POR in the low HIV risk group was still higher, 1.43 fold that in the high HIV-risk group (95 percent CI: 0.94, 2.17). There was little evidence that BV diagnostic criterion (comparing Nugent to clinical criterion) was associated with the magnitude of the association as the ratio of PORs, accounting for variations due to HIV risk groups and adjusted versus non-adjusted estimates, was 0.88 (95 percent CI: 0.61, 1.26). Similarly there was little evidence that estimate adjustment was a substantial source of heterogeneity as the ratio of PORs, comparing adjusted to unadjusted estimates, accounting for variations due to HIV risk groups and BV diagnostic method, was 1.14 (95 percent CI: 0.76, 1.74).
The impact of each study on the summary estimate was evaluated by successively omitting each of the studies and obtaining a summary for all the other studies. For HIV incidence studies the RR varied little, ranging from 1.58 (after excluding the Taha, 1998 study) to 1.93 (after excluding the Martin, 1999 study). The POR estimate from the study by Greenblatt et al. (1999), by far the most precise of the estimates from the prevalence studies, was highly influential. Although the homogeneity test P-value remained <0.0005 upon removing this result, it substantially influenced the symmetry of the funnel plot. When this result was removed, the symmetry tests produced P=0.2 (Begg) and P=0.6 (Egger); the ‘trim and fill’ analysis imputed just one hypothetically missing result, regardless of whether a fixed effect or random effects model was used.
Sensitivity analyses were conducted to determine the impact of the decision to systematically choose Nugent score results over clinical criteria for BV diagnosis for studies with both results available. These sensitivity analyses were conducted by repeating the analyses after replacing, for studies that used two diagnostic methods, estimates based on Nugent scores with those based on Clinical criteria. The sensitivity of the effect estimate to this choice was found to be robust with an overall RR in HIV incidence studies of 1.47 (95 percent CI: 1.10, 1.95). Heterogeneity in prevalence studies remained unchanged with the POR estimate in low-HIV risk groups of 2.30 (95 percent CI: 1.68, 3.15) still being higher compared to the POR in high-HIV risk groups of 1.53 (95 percent CI: 1.29, 1.82).
This is an updated systematic review and meta-analysis, of the association between BV and HIV infection. Overall BV prevalence was high in several populations of women studied, with prevalence rates as high as 70 percent. Our analyses of HIV incidence studies indicate that BV increases the risk of HIV acquisition by approximately 60 percent (95 percent CI 21–113 percent). This was slightly higher that the 40% reported in a previous review of 2 studies . Studies of HIV prevalence tended to find higher HIV prevalence in women with BV. However these prevalence study estimates were heterogeneous and had evidence of funnel plot asymmetry.
The BV-HIV association tended to be weaker in high HIV-risk groups, though the few number of prospective studies limited the confirmation of this trend in HIV incidence studies. A weaker association in high risk women may possibly be due to a depletion of susceptibility to HIV resulting from women in high risk groups having a greater risk of acquiring HIV from causes other than BV. Once HIV infected, they are no longer at risk of acquiring HIV attributable to BV, thus reducing the effect of BV in this group. More data from prospective cohorts are needed to better examine the heterogeneity, by HIV risk group, in the effects of BV on the risk of acquiring HIV. This information could be helpful in identifying specific sub-populations, with a stronger association between BV and HIV, in whom to target BV control measures.
BV results in several changes in the vaginal flora that provide biological plausibility for an increased risk of HIV acquisition in BV positive women. BV is associated with a depletion of hydrogen peroxide-producing lactobacilli that may reduce vaginal defense against microorganisms including HIV[26, 27]. Higher vaginal pH (>4.5) that occurs with BV may also increase the availability of vaginal HIV target cells by increasing CD4 lymphocyte activation and multiplication. High vaginal pH may also increase the adherence and survival of HIV. BV has also been associated with a reduction in vaginal fluid levels of secretory leukocyte protease inhibitor (SLPI), which has been shown to block HIV infection in vitro. Finally, by increasing intravaginal levels of interleukin-10, BV may increase the susceptibility of macrophages to HIV. These changes, combined with the difficulties of successfully eradicating BV, may explain the increased risk observed in most epidemiology studies.
Some methodological limitations to this review need to be considered. The first are concerns on whether a meta-analysis of observational studies can effectively control for confounding and bias. An attempt at reducing these was made by the preferential use of adjusted estimates in the estimation of summary measures. Meta-regression also revealed little difference between the adjusted and unadjusted estimates used in the final analysis. The second limitation has to do with the relatively few prospective studies included in this analysis. The restricted number of HIV incidence studies prohibited any sub-group analysis. However, this did not appear to be necessary as there was no heterogeneity among the estimates from these studies. The limited number of studies also prohibited any reliable analysis of other potential sources of heterogeneity such as pregnancy or age. More prospective studies are needed to accurately evaluate the causal association between BV and HIV. Third, this review was limited to that of published studies. This had little impact on the estimate from HIV incidence studies as there was no evidence of funnel plot asymmetry. The impact of publication bias on HIV prevalence studies was however unclear because of the heterogeneity in the estimates and discrepancies in the results of the various methods of assessing publication bias (Begg’s versus Egger’s methods; Duval and Tweedie’s ‘trim and fill’ random versus fixed models). The funnel plot of the POR estimates was unusual in that the one estimate that was by far the most precise (Greenblatt et al. 1999) fell well outside the range of the other estimates. Given the pronounced heterogeneity among all POR results, this result was not given a very high weight when the trim and fill analysis was conducted using a random-effects model. With a fixed-effect model, however, the exceedingly high inverse-variance weight assigned to this estimate caused the trim and fill analysis to suggest publication bias so profound that one-third of all prevalence results are unreported, all of them on the reduced-prevalence side of the null. Although some publication bias in that direction might have occurred, we are not inclined to believe that it could have been that great. In any event, the prevalence results were much too heterogeneous to warrant aggregating them to produce a single, summary estimate. We found nothing obvious about the Greenblatt et al. (1999) study that should have caused it to produce an estimate so unlike the remainder of the literature. It was one of five studies conducted in the United States and one of four US studies designed to include sizable proportions of HIV-positive and HIV-negative women. Yet it was the only one to produce an inverse association. We are inclined to consider it’s departure from the main thrust of the literature an unexplained anomaly.
Limitations in the original studies included in this meta-analysis could also impact our estimates. With BV being a time-dependent condition, prospective studies of HIV incidence are susceptible to misclassification in the definition of BV status resulting from the use of either BV status at enrollment, or BV status at prior visit as indicators of BV status immediately preceding HIV acquisition. Misclassification could also result from false positive or false negative diagnosis of BV using either clinical or bacteriologic criteria. Both of these mechanisms of misclassification are expected to be non-differential and thus lead to more conservative estimates of the effect of BV on HIV acquisition. Original studies of HIV prevalence by BV status (using case-control or cross-sectional designs), in addition to the aforementioned susceptibility to misclassifying BV status, could also be subject to selection bias when HIV cases are enrolled from high sexual-risk populations in whom BV is more frequent. If not controlled, such a bias would result in an overestimate of the association between BV and HIV. However, we do not think this may have been substantial as most studies did control for atleast one indicator of sexual risk thus attenuating the impact of selection bias. Finally, all the observational studies included in this analysis are subject to residual confounding, which could result in an underestimate or overestimate of the magnitude of the association between BV and HIV.
Despite these limitations, this review was strengthened by extensive search of published literature using multiple databases and references of identified publications. Furthermore, by separating HIV incidence studies from HIV prevalence studies, the effect of BV on incident HIV was separately analyzed. This distinction is important as studies of incident HIV are not liable to reverse causation bias that would result from HIV infected women being more likely to acquire bacterial vaginosis. The analysis of sources of heterogeneity also allowed us to identify HIV risk group, and not method of BV diagnosis, as an important source of heterogeneity in prevalence study results. Finally we refrained from using summary estimates in the presence of heterogeneity. It has been argued that even when a random effects model is used to obtain a summary estimate, the latter is not always conservative and is potentially misleading if interpreted as an average effect.
The high prevalence of BV in certain populations (particularly those most impacted by the HIV pandemic) implies that notwithstanding the relatively modest effect of BV on HIV infection, a high proportion of HIV infection could be attributable to BV. In a population of women having a BV prevalence of 30 percent, with a relative risk of 1.6, the population attributable risk proportions (PARP), the proportion of HIV in a population that is attributable to BV, is estimated at 15 percent. Although other sexually transmitted infections (STIs) have been shown to increase the risk of HIV infection with a higher RR in the order of 2–5, the relatively lower prevalence of these STIs as seen in some of the studies included in this analysis[34–37] may result in similar proportions of HIV infection being attributable to these STIs as to BV.
The potential impact of BV could also be expressed in the number of women who need to have BV for each additional case of HIV. This depends on the baseline risk of HIV amongst women without BV. For instance, with a 2.0% baseline risk of HIV seroconversion among BV-negative women, a relative risk of 1.6 would correspond to an absolute risk increase of approximately 1.2%, or about 1 additional case of HIV for every 80 to 90 women with BV. These data suggest that greater attention needs to be given to BV in the global fight against HIV infection. Randomized clinical trials (RCT) to determine the effect of BV control measures on HIV acquisition may be worth considering. A previous RCT of the effects of mass treatment of STIs on HIV conducted in Rakai (Uganda) used a single dose of oral metronidazole 2g and found no effect on HIV acquisition. However, although 2g of metronidazole can cause short term remission, it is not the recommended treatment[39, 40] thus limiting the inference that can be made on the effect of BV treatment from the Rakai study. Future RCTs assessing this effect will need to use the recommended treatment regimen with a longer duration associated with lower recurrence rates. In addition to the need to evaluate the potential of BV treatment to prevent HIV acquisition and transmission, a better understanding of its risk factors and determinants of BV recurrence is required.
We are grateful to Jean Blackwell for assistance during the literature search. We also thank the authors of the original studies included in this analysis.
JA is a Fogarty AITRP fellow supported by NIH Fogarty grant DHHS/NIH/FIC 5 D43 TW01039-08 AIDS International Training and Research Program to the University of North Carolina, Chapel Hill. CP was supported in part by a grant from the National Institute of Environmental Health Sciences (P30ES10126). JSS was supported by the UNC Center for AIDS Research (CFAR).
JA, PN, AAA and JSS defined the question and designed the study. JA and JSS abstracted and reviewed the data. JA, CP and JSS conducted the analysis. All authors wrote or reviewed the manuscript.
An earlier version of this study was presented at the International Society for Sexually Transmitted Disease Research (ISSTDR) 2005 conference in Amsterdam.