The effectiveness and safety of treatments used for acute diarrhea and acute gastroenteritis in children: protocol for a systematic review and network meta-analysis

Literature searches will be performed in MEDLINE, EMBASE, Global Health through the Ovid platform, and also in CINAHL, LILACS, and the Cochrane Central Register of Controlled Trials (CENTRAL) using a combination of controlled and free-text terms with various synonyms for the disease and the interventions. The example of the search strategy through Ovid MEDLINE is shown in Appendix A: Table 2. We will use the validated randomized controlled trial (RCT) filters created by the McMaster University Health Information Research Unit for MEDLINE and EMBASE through the Ovid platform. These filters provide a good balance between sensitivity and specificity [25]. In addition, we will use previously validated and used filters for identifying pediatric articles in MEDLINE (translated to the other databases through Ovid), CINAHL, and CENTRAL [26‐28].

Search alerts will be set up for monthly notification, and the search will be repeated before the final manuscript submission. Search strategies were developed with liaison with an experienced librarian. No language, publication status, or date limit will be used.

We will carry out a manual hand search of bibliographies of identified RCTs and guidelines. Additionally, we will perform a grey literature search through (1) trial registries (e.g., clinicaltrials.gov and the WHO International Clinical Trials Registry Platform Search Portal), (2) conference proceedings and abstracts, and (3) dissertation databases (ProQuest Dissertations and Theses database). We will contact authors of unpublished work to ensure eligibility.

The search for studies will be limited to RCTs assessing the efficacy, effectiveness, or safety of different interventions for AD/AGE in children younger than 18 years of age. The diarrhea episode will have to be proven or presumed to be caused by an infectious agent, with a duration of less than 14 days. Diarrhea is usually defined as three or more loose stools in a 24-h period [4, 5].

We will include RCTs and quasi-RCTs that evaluated single and/or combined interventions at any dose and presentation of zinc, vitamin A, micronutrients, smectite, racecadotril, prebiotics, probiotics, synbiotics, fermented milks, and lactose-free milks and formula. Regarding probiotics, we will search for trials using, but not restricted to: Lactobacillus or Bifidobacterium strains, Sacharomyces boulardii, Bacillus clausii or subtilis, Streptococcus thermophylus, and Enterococcus faecium, or mixtures of two or more probiotics. Regarding the prebiotics, we will search for trials studying inulin, fructo- and galacto-oligosaccharides. We will consider a synbiotic when a pharmacological preparation contains one or more probiotics and one or more prebiotics.

Regarding fermented milks, we will include studies using preparations based on S. thermophilus or L. delbrueckii subsp. bulgaricus cultures such as kefir, kumis, and yogurt. Fermented milks supplemented with probiotics will be also included. Finally, as to lactose-free milks, we will include studies evaluating as interventions lactose-free, or diluted (by at least 50 %) milks or formula, including soy-based milks.

We will not include studies testing the effectiveness of antimicrobials for bacterial or parasitic infections because these treatments are not empirically recommended and are only prescribed under certain clinical conditions or for a specific microorganisms’ detection. We will not include neither studies testing bismuth nor diphenoxylate since these drugs are not approved for their use in young children. Lastly, we will exclude studies comparing alternative or non-mainstream interventions, given that we are only interested in widely available treatments approved for their use in children (standard treatments), and such treatments are out of the scope of our question.

The studies will have to compare any of these interventions; among them are different doses of the same intervention, or with placebo, or no intervention (conventional treatment with ORS). In order to be included, the study will have to report the effectiveness of one or more of the described interventions with at least one of the outcomes of interest.

Our primary outcomes are mean diarrhea duration (hours/days) and mortality. The secondary outcomes include: diarrhea lasting >3 or 7 days, stool frequency, diarrhea hospitalizations, treatment failure, and adverse events (vomiting, somnolence/lethargy, ileus, constipation). Table 1 describes the definitions of the outcomes. We will exclude studies in patients with cholera, HIV, or with known or thought to have other causes of diarrhea such as antibiotic-associated, persistent or chronic diarrhea. In case the trials have included children with some of the exclusion criteria, and the authors report the results separately, we will use the specific results for our population of interest. Otherwise, we will include the study if we have information about the proportion of children with the exclusion criteria, and this is less than 30 % of the total sample size.

Paris of reviewers (IDF, CMG, CCG, JMS, AMZ, GPG, JJY, and RA) will perform, independently and in duplicate, the screening of available titles and abstracts to assess its eligibility. As a second step, the full-text articles of the potentially eligible studies will be screened to assess their eligibility. We will include studies where both reviewers agree about their inclusion. In case of disagreement between the reviewers, a third member (IDF, CMG, or JMS) of the team will resolve it. Records of excluded articles in this stage, along with the reason for ineligibility, will be saved for future reference. Reviewers will refer to the eligibility criteria during the screening process. Eligible article citations will be saved in the Endnote X7 library. We will present the PRISMA flow diagram [29], demonstrating the search and screening process. We will contact authors of primary studies during data extraction to obtain any missing information.

The study data will be collected based on a pre-specified instruction extraction form in a Microsoft Excel sheet form. We will extract the following information: study characteristics (design, year, duration of follow-up, sample size per arms, setting, and country), patient characteristics (age, inpatient/outpatient, nutrition status, days of disease, etiology, and hydration status), intervention details (doses, administration forms), and outcome results (number of events, mean and standard deviation or standard errors per arm) at the longest duration of follow-up. All reviewers will test the data extraction form before the review. Eight reviewers (IDF, CMG, CCG, JMS, AMZ, GPG, JJY, and RA) will perform data extraction, working in pairs independently and in duplicate. If a consensus cannot be reached, a third designated reviewer will be involved (IDF, CMG, or JMS).

We will assess all included studies for their ROB. For this purpose, two independent pairs (IDF, CMG, CCG, JMS, AMZ, GPG, JJY, and RA) will use a modified version of the Cochrane ROB tool [30] based on sequence generation, allocation concealment, blinding of participants, personnel and outcome assessors, completeness of follow-up, selective outcome reporting, and other biases. Each criterion will be assigned a score of “definitely low risk”, “probably low risk”, “probably high risk”, or “definitely high risk” [31]. We will try to resolve by consensus the disagreements between two reviewers when assessing the studies. Nevertheless, if a consensus cannot be reached, a third designated reviewer (IDF or GHG) from the team will be involved to resolve it.

We will first describe the results narratively and, where possible, the direct evidence will be pooled. Given that we expect clinical and methodological heterogeneity among the studies (see below in the Rating the confidence in estimates of the effect in NMA section), which in turn will create statistical heterogeneity, we will pool direct evidence for each treatment comparison using a random-effects (RE) model. In comparison to a fixed-effect model (FE), the RE model is conservative in that it considers both within- and among-study variability. The RE models assume that the observed treatment effect for a study is a combination of a treatment effect common to all studies plus a component specific to that study alone [32]. We will pool the outcome data using a Bayesian RE model [33]. Effect estimates along with 95 % credible intervals (CrIs) will be estimated using odds ratio (OR) for binary outcomes, and mean difference for continuous outcomes, if they are reported using the same metrics or standardized mean difference (SMD) otherwise. For studies with binary outcomes, we will add 0.5 to each cell if one arm is zero, whereas we will exclude studies from the analyses with zero events in both arms. We will use non-informative priors for all model parameters apart from the heterogeneity variance parameter, for which we will use the informative prior suggested by Turner et al. [34] and Rhodes et al. [35]. All Bayesian analyses will be performed using the Markov Chain Monte Carlo method.

We will assess heterogeneity by estimating the magnitude of the between-study variance using the empirical distribution as estimated by Turner et al. [34] and Rhodes et al. [35], and by using the I ² statistic to quantify the percentage of variability that is due to true differences between studies rather than sampling error [36, 37]. We will interpret the I ² using the thresholds set forth by the Cochrane Collaboration [30], and it will be used as a criterion for pooling the results and for performing additional subgroups analysis. In case there is important heterogeneity, we will use meta-regression to explain it if we have enough data to do so. Otherwise, we will perform subgroup analyses.

We are proposing six a priori hypotheses to explain variability between studies, therefore, as possible effect modifiers; studies in children [1] in HIC could show a smaller effect of micronutrients in comparison to LMIC [2]; micronutrient deficiencies or malnutrition could show a smaller effect of micronutrients in comparison to well-nourished children [3]; a rotavirus infection could have larger effects than those that do not specify the etiology [4]; inpatients could show more larger effects that in outpatients [5]; trials with a higher ROB could show larger effects than trials with a lower ROB; and [6] studies with majority of children younger than 6 months or above 5 years of age could show a smaller effect, mostly of micronutrients, zinc, or vitamin A, but also of the rest of the interventions. We will perform meta-regression using these hypotheses as the study level covariates, and we will perform a sensitivity analysis based on the studies with high a ROB.

The differences among the studies related to the country setting and the nutrition status are closely related. The previously described differences in the burden of the disease in children of LMIC in comparison with children of HIC are related to the higher prevalence of malnutrition and micronutrient deficiencies among the former than the latter. Thus, zinc, vitamin A, and other micronutrients, and also essential nutrients are expected to have a higher effect on reducing the diarrhea in malnourished children than in eutrophic children. However, the lack of information regarding the baseline nutritional or micronutrient status of the children in the studies is a potential problem; therefore, the country setting is an approximation to this source of heterogeneity. We will analyse differences based on both nutritional/micronutrient status and country setting if we have enough data to do so. We will interpret the results of the heterogeneity in the context of the grading of recommendations, assessment, development, and evaluation (GRADE) approach [38].

We will construct a funnel plot for each outcome to assess the potential publication bias [39], if we retrieve at least ten studies [30]. Visual inspection to determine the funnel asymmetry will be used for this purpose. We will also perform the Begg’s rank correlation [40] and Egger’s regression tests [41] if a similar number of studies is available.

We will perform a random-effects NMA, as we expect between-study heterogeneity. Given that many of the treatment combinations available to treat AD/AGE have not been compared in head-to-head studies, we expect that some of the possible comparisons between the interventions will not have direct evidence. In the absence of direct evidence for a given comparison, the indirect comparison will provide the estimate. In the presence of direct evidence, the NMA will provide a combined estimate (i.e., direct and indirect evidence) [42]. We will combine direct and indirect estimates in a NMA for all interventions on the outcomes if the assumptions of between-study homogeneity and transitivity across treatment comparisons are judged to be justifiable. Violation of the transitivity assumption will cause inconsistency between direct and indirect estimates (loop inconsistency) and/or inconsistency between studies that inform the same treatment comparison but include a different number of treatment arms (design inconsistency). We will apply the design-by-treatment interaction model to evaluate both design and loop inconsistency, and if this suggests inconsistency, then we will apply the loop-specific method to assess local inconsistency [43‐45]. We will perform a network meta-regression or subgroup analysis using the same potential treatment effect modifiers described in the “Direct comparisons and assessment of heterogeneity” section to explore important heterogeneity and/or inconsistency. We will also perform a sensitivity analysis for different heterogeneity priors to assess the robustness of results [33, 34, 35].

The network geometry and the results in probability statements as well as forest plots will be presented; this will ensure interpretability of the results. Effect estimates will be presented along with their corresponding 95 % CrIs. We will rank the probabilities with its 95 % CrIs as well as the surface under the cumulative ranking curve (SUCRA) values and rankograms [46]. It is expected that the best treatments will have high SUCRA values while the worst will have low values. For each paired comparison, we will present the direct, indirect, or network estimates.

We will fit a Bayesian hierarchical model with non-informative priors adjusting for correlation of multi-arm trials, assuming a common-within network heterogeneity variance. A series of 100,000 burn-in simulations will be used to allow convergence and then a further 20,000 simulations to produce the outputs. We will assess model convergence on the basis of the Gelman and Rubin diagnostic test [47]. The analysis will be performed in OpenBUGs (version 3.2.3) [48].

Reviewers (IDF, CMG, CCG, JMS, AMZ, GPG, JJY, and RA), in pairs, will independently assess the confidence in the estimates (quality of evidence) for each reported outcome according to the GRADE approach [49]. We will rate the confidence based on four levels: high, moderate, low, and very low. For the direct comparisons, we will assess and rate each outcome based on the following categories: ROB [50], imprecision [51], inconsistency (which is determined based on the heterogeneity as described above) [52], and publication bias [53].

For the assessment of confidence in the estimates, we will use a recent recommended approach [54]. We will assess and rate the confidence in all the indirect comparisons—if available—obtained from first-order loops following the GRADE categories used for assessing the direct comparisons in addition to the transitivity assessment. Transitivity, also called similarity [55], is the assumption that an indirect comparison is a valid method to compare two treatments, because the studies are sufficiently similar in important clinical and methodological characteristics, or in other words, that they are similar in their distributions of effect modifiers [56, 57]. Then, we will rate the confidence in each NMA effect estimate using the higher rating when both direct and indirect evidences are present.

We will assess and rate the confidence in estimates of effect from the direct comparisons in our pair-wise meta-analyses described previously. For rating confidence in the indirect comparisons, we will focus our assessments on first-order loops (FOLs), which are loops connected to the interventions of interest through only one other intervention. For example, if there are interventions A, B, and C, and there are direct comparisons, e.g., A vs. B (AB) and B vs. C (BC), we could indirectly estimate the effects of A vs. C (AC). The AC indirect estimation will be a FOL. We will choose the FOL with the lowest variances, and thus contributes the most to the estimates of effect, for rating the confidence.

Within a FOL, the indirect comparison confidence will be the lowest of the confidence ratings we have assigned to the contributing direct comparisons. For example, if we find that AB has moderate confidence and BC has high confidence, we will judge the associated indirect comparison, AC, as moderate confidence. We may rate down confidence in the indirect comparisons further if we have a strong suspicion that the transitivity assumption has been violated.

Our overall judgement of confidence in the NMA estimate for any paired comparison will be the highest of the confidence ratings among the contributing direct and indirect comparisons. However, we may rate down confidence in the network estimate if we find that the direct and indirect estimates have inconsistency. For this purpose, the GRADE approach recommends to assess the incoherence (or inconsistency as described in the “The network meta-analysis: the direct and the indirect evidence” section) criteria, which is defined as the differences between direct and indirect estimates of effect [54].