3.3 Network meta-analysis results
NMA results
We depicted a network graph of 14 kinds of therapies for GBS. The graph were made by R 3.6.1 software and visNetwork package.
Forest graph
There were 22 studies using Disability grade change after 4 weeks as outcome measure, including 13 treatment options, meanwhile, there were 21 studies using rates of improvement by ≥ 1 grades after 4 weeks as outcome measure, including 9 treatment options. The forest figures on results of network pooled comparisons of the clinical effectiveness of each therapy and their calculated ranking probabilities were shown in Fig. 4 and Fig. 5, the forest figure reference were Placebo, PE and IVIg.
In the assessment on efficiency of treatments for GBS, PE, IVIg, PE followed by IVIg, immunoabsorption followed by IVIg, IVIg 1 g/kg daily for 2 days, Half-course of PE were observed to be significantly effective in treating GBS. (outcome①: PE MD=-0.83, 95%Crl[-1.3,-0.38]; IVIg MD=-0.91,95%Crl[-1.5,-0.35]; PE followed by IVIg MD=-1.1, 95%Crl[-1.9,-0.34], immunoabsorption followed by IVIg MD=-1.9, 95%Crl[-3.4,-0.47], IVIg 1 g/kg daily MD=-0.88, 95%Crl[-1.7,-0.068], 2 times of PEs MD=-1.1, 95%Crl[-1.8,-0.35] outcome②:PE OR = 2.7,95%Crl[1.7,4.7], IVIg OR = 3.6,95%Crl[1.9,8.0], PE followed by IVIg OR = 3.7,95%Crl[1.5,10.0]). Both PE and IVIg were available for GBS (outcome①: PE MD=-0.83, 95%Crl[-1.3,-0.38]; IVIg MD=-0.91,95%Crl[-1.5,-0.35];outcome②: PE OR = 2.7,95%Crl[1.7,4.7], IVIg OR = 3.6,95%Crl[1.9,8.0])and all kinds of corticosteroids were be indicated no significant efficiency for GBS (outcome①: MTP MD=-0.18, 95%Crl[-0.66,0.30]; Pred MD = 0.81,95%Crl[0.27,1.3];outcome②: MTP OR = 1.4,95%Crl[0.72,2.6], Pred OR = 0.61,95%Crl[0.24,1.5]). We transferred the base treatment of forest graph for PE and normal dose of IVIg(Fig. 4-b),c) Fig. 5-b),c)), we could find there were no other therapies being more effective with significant difference. We compared different doses of PE and IVIg (IVIg 0.4–0.5 g/kg daily for 4–5 days, 4–5 times of PE, IVIg 1 g/kg daily, IVIg 0.4 g/kg/day for 3 days, 2 times of PE) and found no significant difference between them. For other kinds of therapies, such as IFNb-1a, brain‐derived neurotrophic factor, CSF filtration, Tripterygium Wilfordii Polyglycoside and IVIg 0.4 g/kg/day for 3 days, had no significant difference with placebo.(outcome①: IFNb‐1a MD = 0.095, 95%Crl[-1.5,1.7]; BDNF MD=-0.83,95%Crl[-2.8,1.1]; CSF filtr MD=-0.86, 95%Crl[-1.8,0.12], IVIg 0.4 g/kg/day for 3 days MD=-0.4, 95%Crl[-1.5,0.71] outcome②: IFNb‐1a OR = 1.1,95%Crl[0.13,11.0], BDNF OR = 1.1,95%Crl[0.056,19.0], CSF filtr OR = 2.5,95%Crl[0.49,12.0], TWP OR = 4.6,95%Crl[0.6,47.0]). Regarding to the improvement for GBS among PE, IVIg, and Corticosteroid, the three most conventional treatments, IVIg was the most helpful one (compared with PE (MD 0.073[-0.26,0.41],with methylprednisolone 0.72[-0.01,1.5],with prednisolone 1.7[0.96,2.5])), but there was no significant difference between PE and IVIg. The efficacy of the two hormones was lower than that of PE and IVIg. (outcome①: MTP VS PE 0.66:[-0.017,1.30] Pred VS PE 1.6 [0.96,2.3] MTP VS IVIg 0.73[-0.011,1.5] Pred VS IVIg 1.7[0.97,2.5] PE VS IVIg 0.078[-0.26,0.41] outcome②༚MTP VS PE 0.51:[0.22,1.1] Pred VS PE 0.22[0.075,0.62] MTP VS IVIg 0.38[0.13,0.92] Pred VS IVIg 0.17[0.048,0.50] PE VS IVIg 0.74[0.44,1.2]).
Ranking probability
A clustered ranking plot was generated and presented NMA results visually. To better understand the results, the ranking graph was calculated to evaluate the ranking probabilities of all medications on the outcomes. Results were presented in Fig. 4d) and Fig. 5d). As suggested by ranking probabilities of outcome①, immunoabsorption followed by IVIg had biggest possibility to be a best treatment(P = 0.6), and Half-times of PE, PE followed by IVIg were also likely to be the best treatment. For outcome②, TWP had biggest possibility to be a best treatment(p = 0.4), and IVIg 0.4–0.5 g/kg daily for 4–5 days, PE followed by IVIg followed it. For most probability to be worst treatment, in out come① was prednisone, followed by IFNb-1a, in out come② was prednisone, followed by IFNb‐1a and BDNF.
League Table Heatmap
We use nma.league() Function in BUGSnet package to produce League Table Heatmap. The map would show comparison results of each therapy clearly. (Fig. 6 and Fig. 7)
Gelman-Rubin-Brooks plot
As we carried network meta-analysis based on a bayesian hierarchical framework, we should confirm our simulations have resulted in the convergence of the algorithm, which represented the stability of our results. The plot showed well convergence of the algorithm(Fig. 8.).
The order of SUCRA value
If a treatment always ranks first, then SUCRA = 1, and if it always ranks last, it will have SUCRA = 0.[45] We use the SUCRA function in dmetar package to calculate SUCRA Score and ordered it in descending order.(Fig. 9)We found Conventional dose of IVIg, PE followed by IVIg were in the front of therapies queue. The SUCRA score of PE was lower by IVIg in graphs. Corticosteroids on two outcomes were in the bottom. For the outcome①, BDNF, CSF filtration, PE, Conventional dose of IVIg, Immunoabsorption, IVIg 1 g/kg daily got a similar SUCRA value, which could from the side indicated that there may be no significant difference in efficacy between these treatments.(Fig. 9)
3.4. Consistency analysis and heterogeneity test
We used I2 for consistency checks direct results, From the Fig. 10. and Fig. 12, almost all of the I2 were under 50% which means the heterogeneity of the direct NMA was in a lower rage, the results of direct evidence of the NMA were reliable. We used the fixed effect model for meta-analysis.
The node-splitting method and its Bayesian P value was used to report the inconsistency of our results between direct and indirect results. For the majority of our results, the confidence intervals from direct and indirect evidences were in consistent, with minor differences. In the inconsistency checks, we found that there were some heterogeneity between the four groups in outcome①,(G VS B,M VS B,I VS G,M VS I)༌as there was no obvious heterogeneity in outcome②. To further determine the heterogeneity, we used netheat diagrams in the netmeta package for heterogeneity analysis. The results showed that the heterogeneity was within the acceptable range. We further used the direct.evidence.plot function in the dmetar package (from github) to analyze the sources of direct and indirect evidence, and the results showed that the three groups of results (G VS B,I VS G,M VS I) were dominated by direct evidence. The results of M VS B were mostly based on indirect evidence, but the results of network evidence were consistent with its direct evidence, as we thought direct evidence had more credibility. Based on the above results, we believed In outcome①, there was no significant heterogeneity in the network meta analysis which may influence the result significantly. In the outcome②, we did not find groups with significant heterogeneity. Finally, we could see from the direct.evidence.plot that most of the comparison results in this study are obtained through indirect comparison. Since the indirect results of the meta were calculated based on the bayesian algorithm, they still needed to be verified by a large number of direct comparisons.
3.6 Publication bias.
Funnel plots were used to measure the publication bias. The funnel plot of the improvement in Disability grade change after 4 weeks, and the rates of improvement by ≥ 1 grades after 4 weeks showed potential publication bias of the included RCTs (Fig. 14). It can be seen from the funnel plot that almost all the studies fall within the funnel and the two sides of the funnel were basically symmetrical, so the possibility of publication bias in this study was small.