Data synthesis  
The characteristics of the included studies were tabulated and reviewed to exclude those studies that might result in intransitivity. Network meta-analysis was done by Bayesian approach using a random effects model with Markov chain Monte Carlo simulation with vague priors (GEMTC, BUGSnet) using the R-software (Version-R 3.6.2)17,18. Generalised linear models with 4 chains, burn-in of 50,000 iterations followed by 100,000 iterations with 10,000 adaptations was used18.The geometry of the networks were assessed using network plots with the size of the nodes being proportional to the number of subjects included in the intervention and the thickness of the arms connecting the different intervention nodes corresponding to the number of studies included in the comparison. Model convergence was assessed using Gelman-Rubin plots as well as by analysing the trace and density plots19. Inconsistency was assessed by node-splitting20. Pair-wise meta-analysis evaluating the direct evidence for the different NIV modalities was also done and heterogeneity was assessed using I2 statistic and Cochran Q test. The results of the network meta-analysis were expressed as risk ratios (RR) with 95% credible intervals in league matrix tables and forest plots. The league matrix tables display the RR of the outcome parameter for the intervention in the row versus that in the column in the lower triangle and vice versa in the upper triangle. The comparison of direct and indirect evidence using node-splitting are expressed as odds ratios (ORs) with 95% credible intervals. Surface under the cumulative ranking curve (SUCRA) was used to rank interventions for all the outcomes. SUCRA is an index with values from 0 (least effective intervention) to 1 (best intervention)21. SUCRA should always be interpreted with 95% credible intervals as well as the quality of the evidence. The confidence in the final estimates for all the outcomes were assessed using GRADE approach as recommended by the GRADE working group22.