Immunosuppressive and monoclonal antibody treatment for myasthenia gravis: A network meta‐analysis

Summary Background We intended to compare and rank all the immunotherapies including immunosuppressant agents or monoclonal antibodies for myasthenia gravis (MG). Methods A network meta‐analysis was performed to synthesize the direct evidence and indirect evidence. Quantitative MG score (QMGS) was defined as the primary outcome. The secondary outcomes included the glucocorticoid reduction and hazard ratios (HR) from the counts of adverse events (AEs). Results We identified 14 studies including 808 MG patients. For the primary outcome, cyclosporine A (CsA) was hierarchically the best with statistical significances of −1.18 (−1.81, −0.59) vs placebo (PLA), −0.98 (−1.72, −0.23) vs mycophenolate mofetil, and −0.77 (−1.57, −0.032) vs tacrolimus (TAC). When the intervention periods were controlled, both eculizumab (ECZ) of −1.50 (−2.81, −0.18) and CsA of −1.23 (−1.81, −0.64) vs PLA reached a statistical significance. Belimumab and ECZ ranked the most tolerable therapies while CsA of 2.41 (0.58, 10.01) ranked the last vs PLA. Conclusion These findings demonstrated that ECZ represented the most effective and tolerable therapeutic alternative to be recommended for refractory MG. TAC may be a beneficial therapy to treat MG extensively while the efficacy of CsA and cyclophosphamide may be limited by their multiple or severe AEs.

Thymectomy for MG may be an elective procedure while plasmapheresis (PP), immunoadsorption, and intravenous immunoglobulin (IVIg) produce rapid improvement in weakness. Recently, monoclonal antibodies are emerging therapeutic alternatives for MG. Monoclonal antibodies to CD20 (MabThera TM ), C5b9 (Soliris TM ) are beneficial drug to treat severe generalized MG. 5,6 So far, traditional meta-analysis has evaluated the efficacy and safety of the above single or double intervention. [7][8][9] However, none of them had included adequate assessment of comparative effectiveness of immunotherapies. Therefore, we performed a network meta-analysis (NMA) of all relevant immunotherapies to comprehensively compare and rank strategies for MG treatment.

| Data extraction and outcome measures
Information regarding study design, participant information, intervention or comparison method, and outcome measures was F I G U R E 1 PRISMA chart of selection procedure and reasons for withdrawal from the network meta-analysis. CENTRAL, Cochrane Central Register of Controlled Trials extracted when available. The extracted participant information included sample size, follow-up months, age at onset, disease duration, number of thymectomy, number of thymoma, and number of anti-AChR antibody serostatus.

Data were extracted by two reviewers (Liang Wang and Xiao
Huan) with a standardized data extraction form and were inspected by another reviewer (Jian-Ying Xi). We examined the published data provided in the original studies and contacted the corresponding authors for additional unpublished data. Any contradictory data were discussed and reached a consensus finally. For studies lacking change in standardization (SD), a method of single imputation with correlation coefficient was utilized. 10 In this NMA, MG Foundation of America (MGFA) quantitative MG score (QMGS) 11 was defined as the primary outcome. The secondary outcome included the steroid-sparing effect measured by GC reduction. Another key secondary endpoint was the safety measured by drug-related adverse events (AEs) in the follow-up months.
The MG activities of daily living (MG-ADL), 12 serum anti-AChR antibody titer, and improvement rate were not included for the limited number of eligible studies.

| Quality assessment-risk of bias
The included studies were graded using the Oxford hierarchy of evidence 2011. 13 The Cochrane risk of bias tool was employed to assess their risk of bias. 14 Two reviewers (Liang Wang and Xiao Huan) evaluated the quality of studies and risk of bias independently. The criteria for evaluating the methodological quality included random sequence generation, allocation concealment, blinding of participants, blinding of observers, incomplete outcome data, selective reporting, and other bias. Comparison-adjusted funnel plot was utilized to test the small-study effect including publication bias.

| Statistical analysis
Network meta-analysis was performed with Bayesian Markov chain Monte Carlo model. 15 We estimated relative treatment efficacy of the competing interventions by employing standardized mean differences (SMD) for continuous variables with 95% confidence intervals (CI). Besides, traditional pairwise meta-analysis of random effects was conducted using "metan" command in each intervention, respectively. Random effects Poisson model was used to evaluate hazard ratios (HR) for count variables with 95% CI. We applied burn-in phase of 40 000 iterations after 20 000 annealing algorithm to evaluate convergence. And surface under the cumulative ranking curve (SUCRA) was employed to acquire the efficacy hierarchy of competing interventions. Finally, we performed network meta-regression with controlled intervention periods to figure out the genuine efficacy.
To check for the existence of inconsistency, the method of nodesplitting model by Dias 16

| Study characteristics
There were 1581 studies identified through database searching and other sources. Studies of two or more treatment arms with sufficient data were included in the quantitative synthesis. Considering the placebo effects, data from three trials without controlled placebo [18][19][20] were excluded in the quantitative analysis. The data from the two studies by Sanders 21,22 were not pooled for different participant cohorts, neither were the data from the four studies by Howard 23,24 or Tindall 25,26 for different time points. Finally, 14 studies 21-34 were included in the NMA of our study.
In total, 808 MG patients from the combined datasets were included. The clinical and demographic characteristics were listed in Table 1. Different regimens and intervention periods were also summarized across the trials. The median sample size was 39 patients

| Efficacy comparison on the QMGS
There were 12 studies involving eight interventions including immunosuppressive agents and monoclonal antibodies evaluating the reduction of QMGS. The network plot was shown in Figure 2A, and estimated SMDs of the relative efficacy are shown in Table 2 with median values and 95% CI. With traditional pairwise mean-analysis, statistical significances were calculated in

| Efficacy comparison on the reduction of GC
Eight studies evaluating the reduction of GC with seven immunosuppressive agents were included in this NMA. Figure 2B revealed the network plot while Table 3 listed the estimated SMDs of the relative efficacy with median value and 95% CI, agent by agent. Besides, Figure 4B exhibited the absence of small-study effects for GC reduction. We further employed network meta-regression to control the intervention periods. However, compared with PLA, the statistical differences were not significant in any immunosuppressive agents.

| Safety comparison of AEs
Adverse events were counted during the intervention combined with the number of participants, respectively. Relative median values with 95% CI were exhibited using HR with random effects Poisson model to control the time and number (Table 4). BLM and

| Risk of bias
All of the included studies were graded as level 2 for their randomized design. However, it is difficult to assess the risk of bias for the lack of detailed reporting. The overall quality of included studies shown in Figure 5 was moderate to low. Studies were not identified in random sequence generation and selecting reporting with definite high risk of bias. One study had evidence of allocation concealment, and two studies exhibited high risk for incomplete outcome data. Two studies were not definitely blinding in participants while three were not clearly defined in detection blinding. Besides, we identified eight studies with definite high risk of other bias including carryover effect, recruitment bias, and publication bias.  orally. This could also explain the reason why AEs occurred more frequently in CsA (2.11, 0.38 to 12.59) than TAC. Some people in the CsA group was withdrawn from the study due to the nephrotoxicity while this event was relatively mild in the TAC group. In addition, both TAC and CsA exhibited steroid-sparing effect but the difference was not significant in our study mainly due to the short-term follow-up.

| D ISCUSS I ON
As an alkylating agent, CTX is widely applied in severe autoimmune disorders including MG. 42 The hematopoietic and immune system could be repopulated with endogenous stem cells after intermittent high-dose CTX. 6 Although QMGS was not performed, muscle strength was improved with statistical significance. Significant reduction of GC was acquired from the study but not in this NMA or traditional pairwise meta-analysis. The incidence of AEs between CTX and PLA did not reveal statistical significance. However, compared with PLA, CTX usually induces more severe AEs including diarrhea, hemorrhagic cystitis, infection, and infertility with carcinogenic, teratogenic potential. 6 F I G U R E 4 A, Comparison-adjusted funnel plot for the primary outcome of quantitative myasthenia gravis score. B, Comparisonadjusted funnel plot for the secondary outcome of glucocorticoid reduction. The red line is representative of the null hypothesis that the study-specific effect sizes are not different from the respective comparison-specific summary estimates. The two black dashed lines are representative of a 95% CI for the difference between comparisonspecific pool effect estimates and study-specific effect sizes. y ixy is the noted effect size for x vs y in study i. μ xy is the comparison-specific summary estimate that compares x with y. was combined with GC than GC alone. 30 The inconsistency from design-by-treatment interaction model (P = 0.032) above may be originated from the different follow-up months. A long-term AZA intervention was needed to induce significance. Although its AEs between AZA and PLA did not reveal statistical significance in this NMA, the potential for myelosuppression and hepatotoxicity is not uncommon. 43  Unfortunately, rituximab (RTX), a mouse/human monoclonal antibody against CD20 antigen, was not included to evaluate its therapeutic effect in this NMA for the lack of RCTs. Intravenous RTX was recommended in refractory MG. 40 One clinical controlled trial demonstrated the probability of favorable outcomes increases with a significant steroid-sparing effect in anti-MuSK antibody-positive MG patients. 45 A meta-analysis of RTX concluded the overall effective rate was 83.9%, and incidence of AEs was rather low. 7 There were several limitations from the evidence and NMA pro-

CO N FLI C T O F I NTE R E S T
The authors declare that they have no conflict of interest.