Network Meta-analysis of Surgical Interventions

We performed a network meta-analysis using the Bayesian framework employing gemtc and rjags packages in R4.2.0 (https://www.r-project.org/). Simultaneously, the meta package was used for pairwise analysis. Network meta-analysis results provided more accurate estimates and ranked various interventions to provide clinical recommendations compared to results from traditional pairwise analyses [21 (link), 22 (link)]. We uniformly used random effects models as conservative estimates, generating a risk ratio (RR) or mean difference (MD) with a 95% confidence interval (CI) to represent the efficacy of each intervention. We compared the consistent and inconsistent models using the deviance information criterion (DIC) [23 (link)]. A difference of the DIC less than 5 implies that the model has good goodness of fit, and there is no global inconsistency. In addition, we assessed the local inconsistency of the model using the node-splitting method [24 (link), 25 (link)]. If the value of P > 0.05, the direct comparison was considered to be in good agreement with the indirect comparison. We also evaluated the heterogeneity between studies using the I-squared statistics (I²) [26 (link), 27 (link)]. The range of I² values was 0–100%, where 0–49% was low heterogeneity, 50–74% moderate heterogeneity, and 75–100% high heterogeneity. By calculating the surface under the cumulative ranking curve (SUCRA), we compared and ranked the safety and efficacy of various interventions. Higher ranking grades indicated lower perioperative complication rates or better bowel function. Due to the large variation in sample sizes of the included studies, sensitivity analyses on anastomotic leakage were performed to assess the reliability of the results, which included only studies with sample sizes greater than or equal to 20 in a single arm. To assess the publication bias of studies in the network meta-analysis, we used STATA 16.0 (Stata Corporation, College Station, TX, USA) to generate a comparison-adjusted funnel plot and thus explore the impact of publication bias or other small-sample studies [28 (link)]. In the absence of publication bias, the estimates for all comparisons were symmetrically distributed around the null hypothesis.