Search strategy
We searched PubMed, the Cochrane Library, Embase and ClinicalTrials.gov for RCTs that evaluated treatment responses to pharmacological dosage management in traumatic brain injury patients. The initial search was completed on 7 December 2020. Then, the retrieval results were updated to June 24, 2021 by using the automatic push function of the database on a weekly basis. Syntax and vocabulary were adjusted across databases, including ‘erythropoietin’, ‘Epoetin Alfa’, ‘Darbepoetin alfa’, ‘EPO’, traumatic brain injur*, ‘brain concussion’, ‘brain contusion’, ‘chronic traumatic encephalopathy’, ‘craniocerebral trauma’, penetrating head injur*, ‘basilar skull fracture’, ‘cerebrovascular trauma’, ‘traumatic intracranial hemorrhage’, ‘TBI’, which were used individually or conjunctively. Full details of the search strategies used for this review are listed in Additional file 1. Related articles from the reference lists were also included to search for supplementary articles that may not have been retrieved by the searching strategy prespecified before. All included studies should be written in English. This study was conducted according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA 2009) guideline [23]. Ethical approval was not required since this review did not relate to any individual patient data. A protocol was registered a priori on PROSPERO (CRD42021272500).
Study eligibility criteria
The studies conforming to the following criteria were included in our network meta-analysis: (1) RCTs comparing the effect of erythropoietin with placebo in TBI patients were brought in our study. Nonrandomized trials, retrospective studies, case-control studies, review articles, case reports, and letter to editors were dismissed. (2) Studies investigating the effect of pharmacological treatment of patients hospitalized for TBI management. Mild to moderate patients who have self-limiting disease courses or do not require hospitalization were not eligible for our research. (3) Medication doses should be reported accurately or approximate dose can be calculated from the management description. Studies that did not clearly report drug management and doses were excluded. (4) Significant clinical outcomes and adverse events were reported in the outcome metrics.
Study selection
The initial search records were imported into ENDNOTE X9 literature management software, then the titles and abstracts of records were screened to select appropriate trials according to eligibility criteria by two authors (Q.Y.Z. and D.D). Subsequently, full-text versions of all potentially relevant trials were obtained and examined to ensure eligibility of the study for the network meta-analysis. The most recently reported data were analyzed in the same trial reports at different follow-up periods. Any divergences in regard to eligibility were resolved by discussion with a third reviewer (F.L.X).
Data extraction
Data of interest were collected through a standard data extraction form created using Microsoft Excel 2021 (www.microsoft.com), including eligible studies characteristics (eg, name of the first author, published year, type of research design, follow-up time), characteristics of study participants (eg, mean age, gender, countries), drug administration (eg, route, dose arrangement, total drug dose, time to intervention) and reported clinical outcomes, including couple of adverse events. Any differences on the evaluation of these data were figured out through discussion until consensuses were reached.
Quality evaluation
The risk of bias for individual studies was independently assessed by reviewers employing the identical bias risk assessment tool (RoB2) used for randomized trials from the Cochrane Handbook [24, 25]. A risk of bias graph was generated showing the bias levels as low risk, high risk, and unclear risk, which was demonstrated using Review Manager 5.4(Oxford, UK; The Cochrane Collaboration). The CINeMA (Confidence In Network Meta-Analysis) version 1.9.1 (https://cinema.ispm.unibe.ch/), calculating the NMA contribution matrix by using the Netmeta package of R software, based on grading of recommendations assessment, development and evaluation (GRADE) methodology was applied for assessing quality of evidence, and reported in the results [26]. A comparison-adjusted funnel plot with Egger test was constructed to assess for publication bias [27].
Geometry of the network
We created a network of evidence between the comparisons using STATA (Stata Corp, College Station, Texas, United States of America, version 16.0) [28]. A network plot was drawn to present the interrelationships of comparisons across trials, which showed how each intervention connected to the others through direct or indirect comparisons. In the network, each regimen is represented by a unique node, which means different nodes were designated for different dosages of erythropoietin. Lines indicate direct head-to-head comparison of regimens, and the thickness of line corresponds to the number of trials in the comparison. Size of the node corresponds to the number of studies behind the intervention.
Data synthesis and statistical analysis
ADDIS (IMI GetReal Initiative, EU, version 1.16.8) was used to calculate the comparison results. Parameter setting of ADDIS was follows: number of chains, 4; tuning iterations, 20,000; simulation iterations, 50,000; thinning interval, 10; inference samples, 10,000; variance scaling factor, 2.5. The consistency model was used to pool data regarding to mortality, pulmonary embolism and deep venous thrombosis (DVT). Convergence was assessed using potential scale reduction factor (PSRF) and PSRF closer to 1, indicating the better convergence; generally, PSRF less than 1.2 was acceptable [29].