Literature screening was performed according to the preferred reporting items for systematic reviews and meta-analyses (PRISMA) flow chart [10] and the report of the International Society for Pharmacoeconomics and Outcomes Research Task Force on Indirect Treatment Comparisons Good Research Practices [11]. Institutional review board approval was not required.
Search Strategy
PubMed, Embase, ISI-Web of Science and Cochrane Library were searched for RCTs published between 1950 and 2019. The following terms were used in combinations: Gastric cancer, Stomach Cancer, Gastric Neoplasm, Stomach Neoplasm, Open, Open surgery, Open Procedure, Open Surgical Procedure, Hand-Assisted Laparoscopic Surgery, Hand Assisted Laparoscopy, Hand-Assisted Laparoscopy, Hand-Assisted Laparoscopic Surgical Procedures, Robotic Surgical Procedures, Robotic Surgical, Robotic Surgical Procedure, Surgical Procedure, Robotic, Totally laparoscopic, entirely laparoscopic, absolutely laparoscopic, completely laparoscopic, fully laparoscopic, Randomized Controlled Trial, Randomized, and Randomly.
Study Selection
The only criterion for eligibility was the RCTs comparing at least 2 of the following strategies: open, laparoscopic-assisted, totally laparoscopic, and robotic gastrectomy for GC. Duplicate studies were removed using EndNote version X7.7 (Thomson Reuters). Additionally, for eligibility, the patient underwent radical gastrectomy. Studies that fulfilled the eligibility criteria were evaluated in full-text form. We exclude studies that are not RCTs and unavailable data.
Data Collection and Assessment of Risk of Bias
The data were extracted by two reviewers independently using the same standardized collection form. Relevant data were collected, including the first author, the year of publication, country, patient characteristics, treatment approaches, the sample size and outcomes (overall morbidity, the number of HLNs, total blood loss, length of hospital stay, operative time, pulmonary and anastomotic complications). Qualitative assessment was accomplished by two investigators independently, and if there were disagreements, it will be discussed with the third investigator. Qualitative assessment of the articles was conducted using the Cochrane Collaboration tool for assessing the risk of bias in randomized trials [12].
Statistical Analysis
The primary outcome criterion of our NMA was overall morbidity, and the secondary outcome measure was the number of HLNs. The treatment effect of dichotomous data was expressed as OR. The treatment effect of continuous (mean difference) data was expressed as WMD. 95% CIs were used for the direct meta-analysis and Crl for the estimates of the NMA.
Heterogeneity was assessed by the Cochran Q test and measured by the I2 statistic. Interpretation of the I2 values was performed by assigning low, moderate, and high attributes in cases showing values of 0% to 25%, 25% to 50%, and above 75%, respectively.
First, we performed a traditional pair-wise meta-analysis with Review Manager (RevMan. Version 5.3.Copenhagen: The Nordic Cochrane Centre, The Cochrane Collaboration, 2014), synthesizing studies that compared the same surgical approach with a fixed-effect model.
Second, we built a random-effects network within a Bayesian framework using Markov chain Monte Carlo methods in ADDIS 1.16.8 (Drugis.org) [13], three different sets of starting values to fit the model, yielding 100,000 iterations (50,000 per chain) to obtain the posterior distributions of model parameters. For overall morbidity and the number of HLNs, we used 5000 burn-ins and a thinning interval of 50 for each chain. We also estimated the probability that each of the treatments was the best approach, by calculating the OR for each surgical approach compared with an arbitrary common control group, and counting the proportion of iterations of the Markov chain of the OR ranking in treatments.
A variance calculation and a node-splitting analysis provided by the software ADDIS 1.16.8 were applied to assess inconsistency between indirect and direct comparisons in closed loops, or the comparison was used to confirm the inconsistency in a lack of a closed loop. The 95% Crl of the OR and WMD were used for the estimates in the above calculations. Network plot generation and publication bias were performed with Stata 14.2 (StataCorp, College Station, TX). Risk of bias was assessed using the dedicated Cochrane tool of Review Manager.