The impact of HIV and antiretroviral therapy on TB risk in children: a systematic review and meta-analysis

Background Children (<15 years) are vulnerable to TB disease following infection, but no systematic review or meta-analysis has quantified the effects of HIV-related immunosuppression or antiretroviral therapy (ART) on their TB incidence. Objectives Determine the impact of HIV infection and ART on risk of incident TB disease in children. Methods We searched MEDLINE and Embase for studies measuring HIV prevalence in paediatric TB cases (‘TB cohorts’) and paediatric HIV cohorts reporting TB incidence (‘HIV cohorts’). Study quality was assessed using the Newcastle-Ottawa tool. TB cohorts with controls were meta-analysed to determine the incidence rate ratio (IRR) for TB given HIV. HIV cohort data were meta-analysed to estimate the trend in log-IRR versus CD4%, relative incidence by immunological stage and ART-associated protection from TB. Results 42 TB cohorts and 22 HIV cohorts were included. In the eight TB cohorts with controls, the IRR for TB was 7.9 (95% CI 4.5 to 13.7). HIV-infected children exhibited a reduction in IRR of 0.94 (95% credible interval: 0.83–1.07) per percentage point increase in CD4%. TB incidence was 5.0 (95% CI 4.0 to 6.0) times higher in children with severe compared with non-significant immunosuppression. TB incidence was lower in HIV-infected children on ART (HR: 0.30; 95% CI 0.21 to 0.39). Following initiation of ART, TB incidence declined rapidly over 12 months towards a HR of 0.10 (95% CI 0.04 to 0.25). Conclusions HIV is a potent risk factor for paediatric TB, and ART is strongly protective. In HIV-infected children, early diagnosis and ART initiation reduces TB risk. Trial registration number CRD42014014276.


Reference searching
We examined the 52 review or overview articles found by our database search 1-52 and examined the articles they listed in their bibliographies. This led to our considering 9 unique new abstracts, all of which were excluded.

Citation searching
The 5 relevant reviews found by our database search with the highest number of citations on Google Scholar were used (see Table 2). The citing articles listed by Google Scholar for these reviews were examined, yielding 87 unique new abstracts to screen. From these 39 full-text articles were examined and 3 additional articles contributed to the TB cohorts. [53][54][55]

Other sources
A further HIV cohort analysis 57 (then under review) was brought to our attention after presenting preliminary results at the Union meeting in Cape Town, 2015.

TB cohorts
A forest plot presenting the HIV prevalence observed in children with TB is shown in Figure  1.

HIV cohorts
A forest plot of TB incidence reported by cohorts of children with HIV is shown in Figure 2.
Color codes study location and point shape the ART status of children.

TB cohorts Relationship between odds ratios and incidence rate ratios
Let h be the HIV prevalence in the general population, and H the HIV prevalence in TB cases. Let ! be the incidence rate ratio (IRR) for developing TB disease if HIV infected. Let ! ! be the contribution to the incidence of TB due to HIV infected persons and ! ! that due to HIV uninfected persons. By definition The OR for HIV given TB is

Case-control meta-analysis
This analysis used the metafor package for R. 58 Funnel plots for the meta-analysis of TB cohorts reported in the main text as presented in Figure 3.

Meta-analysis using UNAIDS HIV estimates
For studies where HIV prevalence in controls without TB were not available, we sought national UNAIDS estimates of HIV prevalence in children aged 0-15. This data was available for all of the 8 studies with their own controls and a further 27 studies without their own controls (see Table 4). The UNAIDS HIV prevalence mid-point estimates and 95% uncertainty ranges were interpreted as equivalent numerator (n) and denominators (N) using standard binomial confidence interval formulae. We undertook a Bayesian meta-analysis of these data where the relationship between the HIV prevalence estimated nationally by UNAIDS and the locally appropriate HIV prevalence for controls was informed by the studies where both control HIV prevalence and UNAIDS HIV prevalence were available.
The hierarchical model used was therefore characterizes the typical relationship between UNAIDS national HIV prevalence estimates, and ! !~! ! ! , ! ! ! characterizes the effect of HIV on TB risk, analogously to the random effects model in the main paper.
Gibbs sampling implemented in the R statistical programming environment 59 was used (together with a normal approximation to the likelihood) to perform Markov chain Monte Carlo (MCMC) sampling from the posteriors of all the study-specific 'random effects' parameters (e.g. ! ! ), and the parameters at the population-of-studies level of the hierarchy (e.g. ml). Chains were run for 5,000 iterations, discarding the first 1,000 as burn in, and thinning to every 1 in 10 to reduce autocorrelation. Five chains with random starting points for the ! ! parameters are shown in Figure 4 and Figure 5. The upper confidence intervals of the Gelman-Rubin statistics were ≤ 1.2 for all study-specific parameters. The posterior mean (95% credible interval [CrI]) for the OR of TB given HIV was 7.0 (95%CrI: 5.7 -8.5) and the estimated OR relating control HIV prevalence to UNAIDS estimates was 7.3 (95%CrI: 5.9 -8.8). The posteriors for ! ! and ! ! exhibited a high degree of correlation (-0.83). A forest plot comparing this Bayesian analysis with a Bayesian metaanalysis analogous to the above but using only the case-control studies controls is shown in the main article. The Bayesian analysis described above additionally estimates the effect sizes for studies without their own controls.

TB incidence by immunological stage
Three studies reported TB incidence stratified by WHO immunological staging (Not significant, Mild, Advanced, Severe), reporting incidence relative to the base category (Not significant). These data are graphed in the main article. We performed a random effects meta-analysis using the metafor package in R 58 , treating each stage as a separate stratum. The pooled relative incidence is shown in the main article figure by a purple dashed line, points and error-bars. The pooled relative incidence in the Severe category was 5.0 (95%: 4.0 -6.0) with I-squared heterogeneity statistic I 2 = 87.1%.

TB incidence by CD4 percentage
Six studies reported TB incidence reported TB incidence stratified by more than one CD4% category, each study using a different set of CD4% categories (see Supplementary data file). Additionally, one study 60 analysed a large cohort to produce an estimate of the factor change, F, in TB incidence associated with a unit increase in CD4 percentage, obtaining 0.94 (95%CI: 0.91 -0.97).
To produce a pooled estimate of the gradient in the logarithmic TB incidence rate ratio with respect to CD4% change, !, we used a Bayesian model where the TB incidence ! !" for the ith study (i=1,…,6) at the j-th CD4% category mid-point ! ! follows: with the ! ! capturing the overall level of TB incidence in each study, and where ! !" ! are derived from the reported confidence interval for each incidence. Working with within-studydifferenced data (and dropping points with zero incidence from the analysis) means we can ignore ! ! : We also introduce an additional term in the data log-likelihood describing the factor, F, from Crook et al., 57 which is assumed to be normally distributed on a log scale with variance ! ! ! derived from the reported confidence intervals: This means that (for ! ≥ 1): Gibbs sampling can then be used since: (where n is the number of i's including 0), and This scheme was implemented in R. Parameters a and b were chosen to correspond to a prior for ! ! ! with a mean of 0.1 and a variance of 10. Chains were run for 5,000 iterations, discarding the first 1,000 as burn in, and thinning to every 1 in 5 to reduce autocorrelation. Five chains with random starting points (see Figure 6); the upper confidence intervals of the Gelman-Rubin statistics were ≤ 1.0 (within rounding errors) for all parameters. The pooled estimate was ! ! = -0.063 (95%CrI: -0.188 -+0.063). The forest plot of estimates for each study are shown in the main article.

TB incidence by time on ART
A figure the main article shows the data on TB incidence by months-since-ART-initiation reported by 9 studies (colored lines, points and error-bars). The data have been aligned for this plot to standardize for overall TB incidence by shifting a study-specific time-series up or down (on the log-scale) so that its first point lies at the value predicted by a linear model fitted to all preceding (i.e. leftward) adjusted data. This allows visualization of the relative TB incidence by time-on-ART by scaling the data each study so that it is in line with the rest.
A non-linear mixed effects regression model was fitted using the lme4 package in R. 61 The original data set was replicated 10 times, with the log incidences perturbed by a Gaussian random noise with variances chosen to reproduce the reported confidence intervals for each point; thus allowing an approximation to the data likelihood. For the i-th study, the TB incidence ! !" at time t is taken to follow: where ! ! , ! ! and ! ! are treated as random effects, and determine the early rate of change and asymptotic level of protection, respectively. The study-specific intercepts ! ! set the overall level of TB incidence in each study and are not of interest to us.
The figure the main article also shows the dashed black curve corresponding to the pooled estimated trajectory with 95% uncertainty estimates in gray (constructed from 1,000 sample trajectories based on normal samples using the variance-covariance matrix). The estimated asymptotic protection was HR = 0.10 (95%CI: 0.04 -0.25) and the saturation timescale corresponded to 4.5 months.

Protection from ART
Six studies reported a hazard ratio for the protective effect of ART against TB incidence (see Table 5). In addition, Abuogi et al. 62 estimated a factor change in TB incidence of 0.91 (95%: 0.86 -0.95) per month on ART. We undertook a random-effects meta-analysis of these studies (see forest plot in the main article). The pooled hazard ratio for TB on ART was 0.30 (95%CI: 0.21 -0.39). The I-squared heterogeneity statistic was I 2 = 79%.

Quality assessment
Quality was assessed by a modified version of the Newcastle-Ottawa scale: the case-control instrument for the TB cohorts and the cohort instrument for HIV cohorts. Five version of the instrument were applied to HIV cohorts to generate a separate quality for each analysis. Not all questions were relevant to each version.

Newcastle-Ottawa Quality Assessment Scale for case-control studies: adaptation for TB cohorts
Note: A study can be awarded a maximum of one star for each numbered item within the Selection and Exposure categories. A maximum of two stars can be given for Comparability.
This assessment was used to evaluate studies that measured the proportion of cases (children who had TB disease) with HIV and the proportion of controls (children without TB disease) with HIV.

Newcastle-Ottawa Quality Assessment Scale for cohort studies: adaptation for HIV cohort analyses
Note: A study can be awarded a maximum of one star for each numbered item within the Selection and Outcome categories. A maximum of two stars can be given for Comparability We regarded the HIV cohorts to being pertinent to 5 analyses. The same studies could be regarded as having different quality for the purposes of these distinct analyses. The analyses were: A. TB incidence in children with HIV. All studies included a reported TB incidence that could be taken as a measure of TB incidence in children with TB. Comparability/control questions were not applicable. B. TB incidence by clinical immunosuppression grade. Three studies were used to estimate the effect of clinical immune grade on TB incidence. This analysis involved within-cohort comparisons and so control questions were not applicable. C. TB incidence by CD4 percentage. Seven studies were used to inform an analysis of the effect of CD4% on TB incidence. This analysis involved within-cohort comparisons and so control questions were not applicable. D. TB incidence by time on ART. Ten studies were used to estimate the effect of time-on-ART on TB incidence. This analysis involved within-cohort comparisons and so control questions were not applicable. E. The effect of ART on TB incidence. Six studies were used to estimate the protective effect against TB incidence of ART.