Search strategy

Literature searches were performed systematically by screening for eligible studies published up to 2 October 2020 through PubMed, Scopus, Cochrane Controlled Register of Trials (CENTRAL), MEDLINE (via EBSCO), and Cumulative Index to Nursing and Allied Health Literature (CINAHL) databases, using keywords listed on S2 Table in S1 File. Additional searches for grey literature were conducted through Google Scholar and ProQuest databases, in addition to literature snowballing of references from included studies and similar reviews. Screening and searching of relevant studies were conducted by two independent investigators (GL and KT), and any discrepancies were resolved by a third investigator (ML) in a blinded fashion. No language restrictions were applied, and any title or abstracts deemed potentially eligible for inclusion by either investigator were retrieved for full-text assessments.

Study eligibility criteria

Studies were included in this review if they met the following inclusion criteria: (1) design, interventional or observational studies including but not limited to cohort, case-control, and cross-sectional studies; (2) studies enrolling both HIV-infected and HIV-uninfected DR-TB patients receiving second line anti-TB drugs; and (3) studies reporting AEs based on HIV seropositivity. AEs were defined as any untoward event occurring following the administration of second-line anti-TB and/or antiretroviral (ARV) agents. Conversely, criteria for exclusion were: (1) non-original research, including qualitative research, case studies, reports, or case series with <20 patients; (2) irretrievable full-text articles; or (3) articles not in English.

Considering that the association between HIV infection and AE occurrence is usually studies as part of a greater cohort, the authors made every effort to minimize rejection of studies judged potentially eligible for inclusion but did not dichotomize AEs based on HIV seropositivity, and the corresponding authors of the respective studies were contacted to obtain the additional data. When no response was provided or the authors were unable to retrieve the data, the studies were excluded from this systematic review (see S1 File pg. 8).

Data extraction and quality assessment

The following data were extracted from each included study: (1) first author’s last name and year of publication; (2) patient recruitment period; (3) study characteristics, viz. study design, location, and exposures; (4) patient characteristics, viz. sample size, mean age, frequency and proportion of male patients, proportion of HIV-infected patients, and TB resistance type (see S1 File pg. 6, 7); and (5) outcomes related to AEs. The primary outcome of this study was the frequency of patients experiencing at least one AE, while the secondary outcomes were: (1) frequency of patients experiencing at least one serious AE (SAE) and (2) frequency of specific AEs–as classified according to the Division of AIDS (DAIDS) Table for Grading the Severity of Adult and Pediatric Adverse Events, Corrected Version 2.1 [13]. When the reported outcomes were not listed in the DAIDS catalogue, AEs were categorized as per the Common Terminology Criteria for Adverse Events (CTCAE) Version 5.0 [14], or per authors’ definitions when the reported outcomes were not available in both DAIDS and CTCAE directories. A serious AE was defined as an AE leading to treatment suspension, withdrawal or discontinuation, requiring prolonged hospitalization or immediate interventions to prevent permanent damage, or resulting in significant disabilities, congenital abnormalities, or death [15]. In the case of studies reporting the aforementioned outcomes only through graphical illustrations, the data were digitized with GetData Graph Digitizer ver. 2.26 (www.getdata-graph-digitizer.com).

The included studies were further assessed for methodological quality using the Risk of Bias in Non-randomized Studies of Intervention (ROBINS-I) for non-randomized studies of intervention (NRSI), Newcastle-Ottawa scale (NOS) [16] for cohort and case-control studies and the modified NOS tool for cross-sectional studies [17] (S4 and S5 Tables in S1 File, respectively). Studies assessed with NOS were subsequently classified as yielding low (0–3 stars), moderate (4–6 stars), or high (7–9 stars for longitudinal studies and 7–10 stars for cross-sectional studies) quality. Data extraction and bias assessments were conducted by two independent authors (KT and AW) and any discrepancies were resolved by a third author (GL)–also in an independent manner.

Statistical analysis

Meta-analysis was performed with R ver. 4.0.0 (R Foundation for Statistical Computing, Vienna, Austria) [18] with the additional meta (ver. 4.9–6) [19], metafor (ver. 1.4–0) [20], and robvis [21] packages, while additional analyses were performed with MetaXL software ver 5.3. (www.epigear.com) [22]. Both adjusted and unadjusted estimates were pooled in the meta-analysis; however, adjusted estimates were prioritized for the interpretation of the results [23]. As the covariates adjusted in each study are highly variable [10], we pre-specified a minimum adjustment factors of age and sex for study estimates to be included in the analysis.

We selected relative risk (RR) along with their 95% confidence interval (CI) as the common measure of the association between HIV infection and AEs. Odds ratios (ORs) and hazard ratios (HRs) were converted using the formula provided by the Cochrane Handbook ver. 6.0 [24] and VanderWeele et al. [25], respectively (see S1 File pg. 10). As clinical heterogeneity was anticipated, effect sizes were pooled using random-effects models [26]. Heterogeneity was investigated with Cochran Q statistics (p<0.10 indicated statistical heterogeneity) and I² value–classified as negligible (0–25%), low (25–50%), moderate (50–75%), or high (>75%).

A priori, we determined subgroup and sensitivity analyses only for the pooled adjusted effects. Subgroup analyses were performed to address potential sources of heterogeneity by categorizing studies based on sample size, study design, location, and methodological quality; in addition to posteriori-determined subsets according to AE type, AE seriousness, and DR-TB treatment regimen. Furthermore, analysis based on study-reported effect size was also conducted to explore the impact of data conversion. We also performed additional analyses according to CD4 count, antiretroviral therapy (ART) status, TB resistance type, and AE severity to explore clinical disparities between subsets.

On the other hand, sensitivity analysis was performed by excluding studies with low methodological quality and sequentially omitting one study at a time. When the number of studies were adequate (n≥10) [27], potential publication bias were investigated visually by contour-enhanced funnel plot [28] and quantitatively by Egger’s [29] and Begg’s test [30]. In the case of detected publication bias, trim-and-fill analysis was performed to evaluate the potential sources of publication bias [31].

Study selection and characteristics

Details of the literature search process are illustrated on Fig 1. The initial search retrieved 5316 records, of which 722 were deduplicated and 4496 were excluded following title and abstracts screening, resulting in the full-text assessments of 98 studies. Sixty-one studies were further excluded due to inappropriate settings (31 with insufficient information for reviewers to judge the availability of potential data, six did not include HIV-negative patients, and two did not report the number of HIV-negative patients experiencing AEs), no reported outcome of interests (14 studies), inadequate sample size (three studies), incomprehensible language (three studies), and irretrievable full-text articles (two studies). Therefore, a total of 37 studies were included in this systematic review. Four studies had sufficiently distinct statistical methods; however, we discovered that the four studies involved the same cohort [32–35]–thus we decided to regard them as one study for the remainder of this article. Four studies were later excluded from the meta-analysis due to distinct study designs (one interventional trial [36] and one cross-sectional study [5]) and incompatible effect measures (i.e. only p-values were reported by two studies [37, 38]). Among the included studies (i.e. 34 studies), 32 were cohort studies (22 retrospective and 10 prospective) and one each was a cross-sectional and an interventional study. Most of the included studies were conducted in Africa (30 studies), while the other two each in Asia and America (S6 Table in S1 File).

Fig 1

Diagram flow illustrating the literature search process.

^aSee S1 File pg. 8 for further details. CENTRAL, Cochrane Controlled Register of Trials; CINAHL, Cumulative Index to Nursing and Allied Health Literature.

A total of 8657 patients (range: 27–1390) were included in this review, where 4437 (51.3%) were co-infected with HIV and 2832 (63.8%) were concomitantly treated with ARV drugs. With regards to TB resistance pattern, 7263 patients were diagnosed with MDR-TB, while 605 with XDR-TB. Pyrazinamide was the most extensively used anti-TB drugs (6190 patients, 25 studies), followed by ethionamide (4289 patients, 23 studies) and kanamycin (4241 patients, 23 studies). On the other hand, lamivudine (897 patients, 11 studies), efavirenz (862 patients, 15 studies), and tenofovir (830 patients, 13 studies) were the most widely used antiretroviral drugs. However, ART regimens were not consistently reported in the included studies, hence potentially underestimating the observed prevalence.

Quality assessments of the observational studies revealed a predominant moderate-to-high methodological quality (8 and 24 studies, respectively; S7, S8 Tables in S1 File), while the only NRSI study yielded a serious risk of bias (S1 Fig in S1 File). All studies controlled for age or sex, and most studies satisfied the criteria for the representativeness of exposed and non-exposed cohorts (30 and 31 studies, respectively). On the other hand, half of the included studies failed to demonstrate that outcome of interest was not present at baseline, and 12 studies did not sufficiently adjust for additional confounders. These limitations may potentially be explained by the fact that most of the included studies were retrospective, thus predisposing these studies to selection bias [39].

Outcomes

Potential overlapping populations were observed in ten studies [40–49] (S9 Table in S1 File), of which Hong et al. [43], Perumal et al. [40], Sagwa et al. [48], and Modongo et al. [49] were prioritized for analyses due to larger sample sizes. We discovered that HIV co-infection independently increased the risk of developing AE by 12% with negligible heterogeneity (RR 1.12 [95% CI: 1.02–1.22]; I² = 0%, p = 0.75; Fig 2), which was consistent with our findings on the pooled unadjusted effects (RR 1.18 [95% CI: 1.05–1.32]; I² = 62%, p<0.01; S2 Fig in S1 File). Subgroup analyses revealed no substantial heterogeneity between subgroups (Table 1), thus further ascertaining the consistency of our model. Furthermore, our model remained robust following sensitivity analysis (S3 Fig in S1 File), except when Sagwa et al. [48] or Avong et al. [50] was excluded, which resulted in borderline significance. In addition, we also discovered that injectable DR-TB regimens resulted in a more apparent risk of AE than all-oral DR-TB regimens (RR 1.11 [95% CI: 1.01–1.21] vs RR 1.61 [95% CI: 0.92–2.84]; Table 1), although further studies are needed to confirm our premises due to paucity of studies in the all-oral subgroup.

Fig 2

Pooled adjusted effects on the association between HIV co-infection and the occurrence of adverse events.

^aEffect estimate calculated using the formula provided by Cochrane Handbook ver. 6.0 [24]; ^bEffect estimate calculated using the formula provided by VanderWeele et al. [25] CI, confidence interval; RR, risk ratio.

Table 1

Summary of meta-analysis and subgroup analyses for the pooled adjusted effects on the association between HIV co-infection and adverse events occurrence.

Outcome	Studies	Events/N		Outcome		Heterogeneity
		HIV+	HIV-	RR (95% CI)	P	I2	P
Adverse events	8	772/1333	554/1049	1.12 (1.02–1.22)	0.014	0%	0.75
Subgroup analysis
AE type					0.748
Any AE	5	427/680	367/676	1.10 (0.94–1.28)		2%	0.39
Specific AE	3	345/6533	187/373	1.13 (1.01–1.27)		0%	0.97
AE seriousness					0.618
Any AE	7	566/918	468/894	1.12 (1.03–1.23)		0%	0.67
Serious AE	1	206/415	86/155	0.98 (0.58–1.67)		NA	NA
ES type					0.354
HR	4	433/802	167/283	1.05 (0.88–1.24)		0%	0.47
OR	4	339/531	387/766	1.15 (1.03–1.28)		0%	0.82
Sample size					0.360
<100 patients	2	61/68	46/52	1.40 (0.86–2.29)		0%	0.33
≥100 patients	6	711/1265	508/997	1.11 (1.01–1.22)		0%	0.78
Location					0.488
South Africa	4	412/675	166/277	1.09 (0.96–1.23)		0%	0.44
Others	4	360/658	388/772	1.16 (1.02–1.32)		0%	0.77
Design					0.781
Prospective	3	227/387	81/128	1.10 (0.85–1.42)		19%	0.29
Retrospective	5	545/946	473/921	1.14 (1.03–1.27)		0%	0.87
TB resistance type					0.189
MDR-TB	5	428/716	348/690	1.10 (0.97–1.23)		0%	0.72
XDR-TB	1	33/37	23/26	1.61 (0.92–2.84)		NA	NA
DR-TB treatment regimen					0.199
Injectables	7	739/1296	531/1023	1.11 (1.01–1.21)		0%	0.85
All-oral	1	33/37	23/26	1.61 (0.92–2.84)		NA	NA

AE, adverse event; CI, confidence interval; ES, effect size; HIV, human immunodeficiency virus; HR, hazard ratio; MDR-TB, multidrug-resistant tuberculosis; N, total sample size; NA, not available; OR, odds ratio; RR, risk ratio; TB, tuberculosis; XDR-TB, extensively drug-resistant tuberculosis.

As we were unable to perform analyses to explore clinical disparities between subgroups on the adjusted model, we decided to perform these analyses on the unadjusted model. Unsurprisingly, we discovered that the risk of developing AE was more consistent in patients receiving ART. Moreover, we also discovered that the risk was more pronounced in MDR-TB patients (RR 1.22 [95% CI: 1.04–1.42] vs XDR-TB: RR 0.80 [95% CI: 0.51–1.26]; S4 Fig in S1 File). Assessment of publication bias revealed a symmetrical funnel plot (S5 Fig in S1 File), which was further ascertained by non-significant Egger’s and Begg’s tests (p = 0.160 and p = 0.275, respectively). As no publication bias was detected, a trim-and-fill analysis was not performed. Publication bias assessment of the adjusted model was not conducted due to study paucity (n<10).

Contrary to our findings on the primary outcome, we discovered that HIV co-infection was not associated with the development of serious AE (RR 0.93 [95% CI: 0.58–1.50]; Fig 3). However, it is worth noting that these effects were unadjusted for confounders as we were unable to perform a formal analysis on the adjusted effects due to paucity of studies. Nonetheless, Merid et al. further ascertained our findings, stating that HIV co-infection was not an independent risk factor of serious AE occurrence [51]. Similarly, we discovered that the extent of immunosuppression was not associated with risks of developing AE (S4 Fig in S1 File), an observation which was further supported by Smith et al. who found that severe immunosuppression (i.e. CD4 count <200 cells/mm³) was not independently associated with increased AE risk [44].

Fig 3

Pooled unadjusted effects on the association between HIV co-infection and the occurrence of serious adverse events.

CI, confidence interval; RR, risk ratio.

When analysis was conducted for specific AEs, we discovered that HIV co-infection increased the risk of hearing loss (RR 1.44 [95% CI: 1.18–1.75]), depression (RR 3.53 [95% CI: 1.38–9.03]), and renal impairment (RR 2.45 [95% CI: 1.20–4.98], S10 Table in S1 File). Our findings on depression and renal impairment yielded negligible heterogeneity (both I² = 0%), although the models were relatively imprecise. On the other hand, our findings on hearing loss yielded moderate heterogeneity (I² = 60%; p<0.01). None of the other specific AEs yielded significant results, except for peripheral neuropathy which resulted in borderline significance (RR 1.49 [95% CI: 0.98–2.28]).