This protocol follows the Preferred Reporting Items for Systematic Review and Meta-analysis Protocols (PRISMA-P) 2015 statement (see Additional file 1, PRISMA Checklist)[45]. And we will report the findings of this study in accordance with the PRISMA extension for network meta-analyses[46]. This protocol is registered with PROSPERO (CRD42021244230). We will describe any amendments to this protocol in an update to the PROSPERO registration.
Eligibility criteria
Population
We will include adults living with HIV (18 years of age or older) who are at least identified with mild depression based on standardized diagnostic criteria (e.g., the Diagnostic and Statistical Manual of Mental Disorders or the International Classification of Diseases) or any validated rating scale for depression (including self-rating scales)[47–51]. For the clinical difference, we will exclude randomized controlled trials (RCTs) in which 20% or more of the participants are suffering from bipolar or psychotic depression but not involving patients with other comorbid psychiatric disorders (e.g., anxiety disorder). We will exclude trials in which participants are diagnosed with treatment-resistant depression. Studies in which participants are women with peripartum depression will be excluded for the clinical difference. We will also exclude studies in which participants with a serious concomitant medical illness.
Interventions
We will only consider non-pharmacological interventions. We will exclude interventions combined with drugs (e.g., antidepressants). We will exclude comparative studies between different intensity levels or different subtypes in the same category. If necessary, we will classify them individually. All the eligible interventions will constitute a synthetic comparison set. In principle, all eligible participants are equally likely to be randomly assigned to any interventions in the synthetic comparison set.
Comparators
The types of eligible comparator conditions may include standard of care, enhanced standard of care, waitlist, psychological placebo, or active non-pharmacological control groups. We will exclude studies using a specific drug or in combination with drugs as the comparator.
Outcomes
Primary outcomes
1. Efficacy (continuous variable), measured by the overall mean change scores (from baseline to endpoint) on the standardized and validated depressive symptom scales, such as Hamilton Depression Rating Scale[52], Beck Depression Inventory I or II[53, 54], or the Center for Epidemiologic Studies—Depression Scale[55]. We will include studies with depression as a primary or secondary outcome. But we will exclude scores that combined depression and other symptoms. If researchers used more than one depression scale simultaneously to measure depression scores in a study, we would apply a pre-defined hierarchy to extract the most appropriate data. This hierarchy will be based on the evidence from systematic reviews (Table 1)[56–60]. For scales not included in this table, we will follow the following rules: (1) Prefer self-reports rather than scales assessed by clinical staff; (2) Prioritize scales specifically targeted to measure depression symptoms over scales with a broader scope; (3) Prefer the most commonly reported scale across studies[61].
2. Acceptability of treatment—treatment discontinuation (dichotomous variable) is defined as the proportion of participants who withdrew for any reason during the delivery of the intervention.
Table 1
Hierarchy of depression symptom severity measurement scales
Hierarchy
|
Depression symptom severity measurement scales
|
Abbreviation
|
1
|
The Patient Health Questionnaire-9
|
PHQ-9
|
2
|
Hamilton Depression Rating Scale
|
HAMD
|
3
|
Montgomery Asberg Depression Rating Scale
|
MADRS
|
4
|
the Beck Depression Inventory
|
BDI
|
5
|
the Hospital Anxiety and Depression Scale
|
HADS
|
6
|
the Center for Epidemiologic Studies of Depression Scale
|
CES-D
|
Study designs
Any RCTs of non-pharmacological interventions to reduce depression among PLWH will be included, including cross-over trials and cluster randomized trials. We will exclude quasi-randomized trials. We will only consider peer-reviewed original articles. We will exclude protocols, unpublished articles, conference abstract, and the thesis. And we will retain systematic reviews associated with our topic to search for potential, qualified studies in their reference lists.
Information sources and search strategy
With the assistance of a senior librarian, a systematic search strategy has been developed (see Additional file 2. Search strategy). To avoid research waste, we will preliminarily search the Cochrane Database of Systematic Reviews, PROSPERO, and JBI databases for similar ongoing and published systematic reviews or protocols. Then, we will search through PubMed, Embase, Cochrane Central Register of Controlled Trials, PsycINFO, and CINAHL for original articles. We will use the following keywords and Medical Subject Headings (MESH) for searching articles: HIV, Acquired Immunodeficiency Syndrome, HIV Infections, Depression, and Depressive Disorder. We will adapt the search strategy for each database. Additionally, the bibliographies of published systematic reviews and meta-analyses will also be hand-checked to identify additional studies. There is no restriction by language and publication year. And we will search again before the final analysis.
Data collection and analysis
Study selection
All studies will be imported into Endnote X9 to remove duplicates, and two investigators (T, H) will independently screen all titles and abstracts identified in the searches. Afterward, the same two investigators will separately access the full texts of the remaining articles for eligibility. If any disagreement occurs between the investigators, a third reviewer will provide arbitration. The PRISMA 2020 flow diagram will be used to outline the study selection process and the reasons why full-text articles were excluded[62]. If a study was published in duplicate, we will include the most informative and complete one.
Data extraction
Using a pre-prepared structured data extraction sheet in Microsoft Excel 2020, two investigators (T, H) will independently read each study and extract the following characteristics: study characteristics (first author, year of publication, and journal), participants characteristics at baseline (mean age, sex, country, setting (high-income countries (HICs) / LMICs), years living with HIV, the severity of depression, and baseline CD4 T-cell counts), study design (the measurement tool for depression, study type, number of arms, total sample size, and sample size for each group), interventions details (type, format (group vs individual), delivery form (mHealth vs telephone vs face-to-face), frequency, the length and number of sessions, duration, and follow-up period), specification of the control group (type, format, delivery form, frequency, the length and number of sessions, duration, and follow-up period) and the following outcomes: dropouts for each group, pre-intervention mean, end point mean, and both with the corresponding standard deviations (SDs). When SDs are not reported, we will first use the standard errors (SEs), t-statistics, p values, and so on to estimate or transformed. Suppose the data reported in the studies are insufficient for estimation or conversion using the above methods. In that case, we will contact the original study authors to retrieve relevant missing data and seek clarifications. All outcomes will be extracted at the end of the study period. We will extract the results for intention-to-treat analyses preferentially. We will exclude studies without relevant available data.
For cross-over studies, we will only extract data from the first period because of carry-over effects. For cluster randomized trials, we will extract data used to illustrate the clustering in the results. The two extractors will resolve any disagreement in data extraction by consulting a third reviewer.
Risk of bias and quality assessment
Two investigators (T, H) independently assessed methodological quality in the included studies using the Cochrane Collaboration Risk of Bias tool[43], which includes six domains: (1) selection bias(sequence generation and allocation concealment); (2) performance bias(blinding of participants); (3) detection bias(blinding of personnel); (4) attrition bias(incomplete outcome data); (5) reporting bias(selective reporting); and (6) other bias. According to the Cochrane Handbook version 6.1.0 (Collaboration, 2020)[43], We will grade methodological quality as low risk, high risk, or unclear risk of bias. Any disparities between the two investigators will be settled by consulting the third reviewer. We will provide a summary table of the risk of bias for each eligible study.
Data synthesis
Characteristics of included studies and information flow in the network
We will first generate the descriptive statistics of each variable and the characteristics of eligible studies. A network diagram will be conducted in STATA (version 15.1) to present the available evidence and describe the network’s structure. The size of the nodes in the network diagram represents the corresponding sample size of each depression intervention, and the thickness of the lines might reflect the number of studies directly compared. To understand the most influential pairwise comparisons in the network and how direct and indirect evidence influences the final summary data, we will use the contribution matrix to describes the percentage contribution of each direct comparison to the entire body of evidence[63].
Standard pairwise meta-analyses and network meta-analyses
All data will be double-entered into the database to ensure accuracy. We will synthesize data to obtain the summary standardized mean differences (SMD) for continuous outcomes and the summary odds ratios (ORs) for dichotomous outcomes using pairwise and network meta-analysis, both with corresponding 95% credible intervals (95% CrI). We will first compare all available direct evidence using pairwise meta-analyses. Heterogeneity will be estimated using Cochran’s Q test and I2 statistic. P<0.10 means the heterogeneity is statistically significant. And I2 reflects the degree of variability. If I2 >50%, which indicates substantial heterogeneity, we will consider subgroup analysis.
When standard pairwise meta-analysis is completed, we will employ network meta-analysis to synthesize all the available evidence. We will use arm-based data and do our analyses within the Bayesian framework with hierarchical models. Gibbs sampling and Markov chain Monte Carlo methods will be employed for Bayesian inference[64]. We will use the binomial likelihood for dichotomous outcomes and the normal likelihood for continuous outcomes. For the correlations derived from multi-arm studies, we will use multivariate distributions to illustrate.
We will fit our models using OpenBUGS (version 3.2.3). We will use uninformative prior distributions for the intervention acceptability and a minimally informative prior distribution for the universal heterogeneity parameter. All meta-regression coefficients will be assumed to have no prior information. We will set the OpenBUGS program to run 100 000 simulations. The first 10 000 simulations will be discarded as burn-in. And we will ensure the convergence of models through visual inspection of three chains and Brooks-Gelman-Rubin diagnostic.
We will consider both fixed-effects models and random-effects models. The deviance information criterion (DIC) will be used for model selection. We will give priority to the model with lower DIC. If the DIC difference between the two models is greater than 3, we will think it is significant. And we will calculate residual deviation to access the fit of a single model. Total residual deviation should approximate the number of data points (the sum of the arms of all included studies) for a good fit. We will choose the model with the best fit as the primary analysis model.
We will assume a universal network-specific heterogeneity parameter. We will also estimate the prediction intervals to assess the relative impact of common heterogeneity on the extra uncertainty anticipated in a future study. A forest plot will be used to present the relative effect and the 95% CrI of the efficacy of each intervention. We will also show the ranking probability distribution of each intervention using the distribution of the ranking probabilities, and the surface under the cumulative ranking curves (SUCRA).[65]
Assessment of transitivity
Transitivity requires the different sets of randomized trials are similar in characteristics (effect modifiers) that may affect the two relative effects. We will evaluate the transitivity by comparing the distribution of clinical and methodological variables that can act as effect modifiers across the different comparisons. Mean baseline age, setting, year of HIV diagnosis, depressive severity at baseline, baseline CD4 T-cell counts, intervention format, intervention duration, and study quality will be all considered as the potential effect modifiers[22, 25, 40, 66, 67].
Assessment of inconsistency
Both local and global approaches will be used to evaluate inconsistency, defined as a disagreement between different sources of information for a particular relative effect. Specifically, for global methods, we will use the design-by-treatment test to assess inconsistency from all possible sources in the network. And for local methods, we will use the node splitting approach to compare the inconsistencies between direct estimates and indirect evidence from the entire network.
Subgroup and sensitivity analyses
Subgroup and meta-regression analyses will be conducted to explore the possible sources given important potential heterogeneity. We will consider the following characteristics: mean baseline age, setting (HICs vs. LMICs), depressive severity at baseline (mild vs. moderate vs. severe), baseline CD4 T-cell counts, intervention format (group vs. individual), delivery form (mHealth vs. telephone vs. face-to-face), intervention duration (short-term (1-6 months) vs. min-term (7-12 months) vs. longer-term (13-24 months)), and measurement tool for depression. We will do sensitivity analyses for our primary outcomes by analyzing studies where the low risk of bias rating has been assessed to test the robustness of results and determine whether one or several studies dominate the estimation of the summary effect size of the interventions.
Small study effects and publication bias
If ten or more studies report a particular outcome, we will use the contour-enhanced funnel plot to investigate whether the asymmetry observed in the funnel plot within each pairwise comparison may be explained by publication bias[68]. We will also use the comparison-adjusted funnel plot to evaluate whether the network has small-study effects[63].
Confidence in cumulative evidence
We will determine the overall strength of evidence using the Grading of Recommendations Assessment, Development, and Evaluation (GRADE) system for network meta-analysis[69]. Overall, the quality of evidence could be rated on four scales: high, moderate, low, and very low. The approach will start by assuming that the evidence is of high quality and then rates down the evidence based on five criteria of evidence (risk of bias, inconsistency, indirectness, imprecision, and publication bias). We will use the Confidence In Network Meta-analysis (CINeMA) software during the evaluation[70].