The databases searched included Cochrane Library databases, EMBASE, Medline, and clinicaltrials.gov until April 2021. No date or language restrictions were set for published and unpublished studies. Supplementary Table S2 shows the detailed search strategies used. We also searched the website of the U.S. Food and Drug Administration (FDA) to identify very recently approved drugs in April 2021 (https://www.fda.gov/).

Studies were eligible for our network meta-analysis if they met the following criteria: 1) population: participants with vision better than hand motion and a history of noninfectious intermediate uveitis, posterior uveitis, or panuveitis; 2) intervention: surgical injection of intravitreal therapeutic agents: DEX implant or FA implant or IVTA or IVB or IVR; 3) controls: at least one arm with sham injection or observation treatment or standard of care (SOC) or a different comparative intravitreal therapeutic regimens. For the SOC group, subjects were administrated with systemic therapy following expert panel guidelines (Jabs et al., 2000). Prednisolone or an equivalent corticosteroid alone was used, or an immunosuppressive agent was added to the therapy (Jabs et al., 2000). For the sham injection group (placebo), procedures were performed identically as in the injection group except for the intravitreal therapeutic agent administration into the vitreous cavity; 4) outcome: at least one outcome concerning efficacy or safety; 5) study design: randomized controlled trial (RCT).

Two individuals (WTL, XJF) independently screened the titles and abstracts found in the various databases and identified potential eligibility by retrieving the full-text articles. The final eligibility was independently decided by these two individuals. If they disagreed on a candidate article, a third person was consulted (ZYZ).

Data from selected studies were extracted by two independent individuals and included numbers of participants, interventions, study design, quantitative results of interventions, clinical endpoints, and risk of bias. In the case of missing data that could not be extracted directly from the article, we sent an email to the authors and asked for raw data. GetData GraphDigitizer (http://getdata-graph-digitizer.com) was also used to obtain digital information from figures. Cochrane Risk of Bias Tool was used to assess the quality of RCTs (Barcot et al., 2019).

Efficacy was the primary outcome and referred to a best-corrected visual acuity (BCVA) change from baseline, percentage of eyes achieving a vitreous haze grading of zero, uveitis recurrence rate, and retinal thickness change from baseline. Safety was a secondary outcome referring to incidence of cataract formation or progression as well as use of IOP-lowering medications after baseline.

A pairwise meta-analysis was first used to estimate direct comparison between two interventions. We estimated risk ratios (RRs) for dichotomous outcomes and mean deviations (MDs) for continuous variables in both pairwise meta-analysis and Bayesian network meta-analysis. A random-effects model was applied to synthesize effect sizes.

Heterogeneity, generally defined as variations in the estimated effect between studies, was estimated through Higgins I-squared inconsistency statistics (Trikalinos and Ioannidis, 2001). A large degree of heterogeneity (e.g., I² > 50%) was considered as statistically significant (DerSimonian and Kacker, 2007).

All our models were fitted in a Bayesian network meta-analysis (Li et al., 2011; Higgins et al., 2012). In our study, Bayesian modeling relied on the Markov chain Monte Carlo (MCMC) methods under noninformative priors in favor of R (version 3.6.3 with JAGS) to generate Bayesian probability estimates (Lu and Ades, 2004; Ades et al., 2008; Song et al., 2012). We assume that the number of chains equals four, tuning iterations equal 20,000, and simulation iterations equal 50,000. Convergence of models was checked by observation of Brooks–Gelman–Rubin diagnostic plot (Supplementary Figure S1), trace, and density plot (Supplementary Figure S2) (Brooks and Gelman, 1998; Coleman et al., 2012; Jansen and Naci, 2013). The probabilities of the best, second, third, and so on until the least effective treatment are presented, ranging from 0% to 100%, and values closer to 100% are higher probabilities (Salanti et al., 2011).

We further assess inconsistency (the difference of estimates of effect between direct comparison and indirect comparison) of evidence in network meta-analysis using node-splitting models (Dias et al., 2010). If p value testing the inconsistency between direct and indirect evidence in this network meta-analysis is significant (p < .05), then inconsistency is denoted indicating a violation of the network analysis assumption.

Sensitivity analysis was performed by deleting studies that resulted in high heterogeneity in pairwise meta-analysis (I-squared >50%). A Bayesian network meta-analysis was repeated after omitting studies leading to high heterogeneity.

All analyses were conducted using the “gemtc” and “rjags” packages of R, version 3.6.3 (R Foundation) (Supplementary Table S3).

Using the data from the studies mentioned, we assessed BCVA change from a baseline of 11 treatments (Lowder et al., 1960; Soheilian et al., 2010; Rahimi et al., 2012; Shin and Yu, 2015; Staurenghi et al., 2018; Lai et al., 2018; Thorne et al., 2019; Jaffe et al., 2020) (Figure 2A), percentage of eyes achieving a vitreous haze grading of zero of four treatments (Lowder et al., 1960; Jaffe et al., 2020) (Figure 2B), uveitis recurrence rate of five treatments (Pavesio et al., 2010; Sangwan et al., 2015; Jaffe et al., 2020) (Figure 2C), and retinal thickness change from baseline of seven treatments (Lowder et al., 1960; Soheilian et al., 2010; Rahimi et al., 2012; Shin and Yu, 2015; Thorne et al., 2019; Jaffe et al., 2020) (Figure 2D) in network meta-analysis. We compared incidence of cataract of seven treatments (Lowder et al., 1960; EMC, 2012; Shin and Yu, 2015; Thorne et al., 2019; Jaffe et al., 2020) (Figure 2E) and use of IOP-lowering medications after a baseline of seven treatments (Lowder et al., 1960; Pavesio et al., 2010; EMC, 2012; Jaffe et al., 2020) (Figure 2F) in network meta-analysis.

In pairwise comparison, a significant difference (p < .05) was detected in these five comparisons. DEX implant, 700 µg (MD 7.10, 95% CI 1.68–12.52), FA implant, 0.2 µg/day (MD 5.85, 95% CI 0.94–10.76), and IVTA, 4 mg (MD 4.85, 95% CI (3.47–6.23) were associated with greater improvement from baseline BCVA compared with placebo at 6 months, and FA implant, 0.59 mg, was associated with greater improvement from baseline BCVA compared to FA implant, 2.1 mg (MD 5.50, 95% CI 3.44–7.56), and SOC (MD 0.82, 95% CI 0.55–1.09) at 12 months. Compared with placebo, IVR was associated with a significantly higher improvement of BCVA at 2 months (MD 5.68, 95% CI 1.00–10.36) (Table 1).

In Bayesian network meta-analysis, there was no significant difference in efficacy of improving BCVA among those treatments (Table 1). We compared BCVA change from baseline at 6 months of six treatment regimens with that of the placebo, and no significant difference was detected. Two FA implants revealed no significant difference compared with SOC in BCVA change from baseline at 12 months. FA implant, 0.2 µg/day, and IVR, 0.5 mg, also showed no significant difference compared with placebo in BCVA change from baseline at 12 months. Compared with placebo, IVR was associated with a significant efficacy of improving BCVA at 2 months (MD 5.63, 95% CI 0.92–12.66) (Figure 2A).

The ranking probabilities of an intravitreal therapeutic agent at any possible position are presented in Table 3. The probabilities, whereby FA implant, 0.2 µg/day, ranked as the first, second, third, fourth, fifth, sixth, and seventh most efficacious drug for improving BCVA at 6 months was 31%, 17%, 15%, 14%, 12%, 7%, and 4%. The rank of interventions improving BCVA at 12 months was presented in Table 3.

In pairwise meta-analysis, DEX implant, 350 µg (RR 2.18, 95% CI 0.68–1.79), DEX implant, 700 µg (RR 1.97, 95% CI 1.03–3.78), FA implant, 0.2 µg/day (RR 1.35, 95% CI 1.03–1.77) at 6 months were associated with an increased rate of achieving a vitreous haze grading of zero compared with placebo with statistical significance (p < .05) (Table 1). In Bayesian network meta-analysis (Table 1), an identical trend was detected, but the difference was not statistically significant among drugs. We compared the efficacy of three intravitreal implants with that of the placebo, and there was no significant difference among drugs in RCTs (Figure 2B). The probabilities of DEX implant, 350 µg, ranked as the first, second, third, or fourth effective treatment for improving vitreous haze at 6 months was 51%, 34%, 11%, and 4%, respectively (Table 3).

In pairwise meta-analysis, patients in the FA implant, 0.2 µg/day, group were associated with a lower risk of uveitis recurrence than those in the placebo group at 6 months (RR 0.36, 95% CI 0.25 to 0.50, p < .05), and FA implant, 0.59 mg, was associated with lower risk of recurrence than the SOC group at 24 months (RR 0.29, 95% CI 0.17 to 0.49, p < .05) (Table 1). Considerable heterogeneity was detected in the comparison between FA implant, 0.2 µg/day, and placebo (I-squared = 87.7%) or between FA, 0.59 mg, and FA, 2.1 mg (I-squared = 67.2%).

In Bayesian network meta-analysis, there was no significant difference in uveitis recurrence at 24 months among drugs in RCTs. Uveitis recurrence rate at 6 months of FA implant, 0.2 µg/day, was significantly lower than that of placebo (RR 0.36, 95% CI 0.14 to 0.90, p < .05) (Figure 2C). The chances of the FA implant, 0.59 mg, being ranked as the first, second, and third most effective clinical intervention for reducing the relapse of noninfectious uveitis at 24 months was 60%, 38%, and 2% (Table 3).

In pairwise meta-analysis, a statistically significant difference in the change of retinal thickness was found when comparing IVTA, 4 mg, versus placebo (MD −46.30, 95% CI −52.64 to −39.66, p < .05) and IVTA, 4 mg, versus IVB, 1.25 mg (MD −7.54, 95% CI −12.54 to −2.54, p < .05) at 6 months (Table 1). Bayesian network meta-analysis showed no significant difference in the change of retinal thickness among seven treatments at 6 months (Table 1). We compared the efficacy of six intravitreal therapeutics agents with that of placebo, and no significant difference was observed (Figure 2D). The probability of IVTA, 2 mg, ranking as the first to seventh best intervention for lowest retinal thickness at 6 months were 70%, 11%, 5%, 4%, 3%, 3%, and 1% (Table 3).

In pairwise comparison, there were statistically significant differences when comparing the incidence of cataract in FA implant, 0.59 mg, versus SOC (RR 4.33, 95% CI 2.97 to 6.33, p < .05) at 24 months or FA implant, 0.2 µg/day, versus placebo (RR 2.15, 95% CI 1.08 to 4.25, p < .05) at 12 months (Table 2), and no heterogeneity was detected in both comparisons (I-squared = 0.0%).

In Bayesian network meta-analysis (Table 2), a statistically significant result in the incidence of cataract was detected between FA implant, 0.59 mg, and SOC at 24 months (RR 4.41, 95% CI (1.51–13.13, p < .05). We compared the incidence of cataract of four intravitreal therapeutics regimens, and no significant difference was detected. There were no significant differences in the treatments compared with placebo in the incidence of cataract (Figure 2E). The probabilities of DEX implant, 350 µg, ranked as the first to fifth intervention that associated with lowest risk of cataracts were 18%, 21%, 24%, 23%, and 14% (Table 3).

In pairwise comparison (Table 2), patients in the FA implant, 0.59 mg, group were associated with increased risk of using IOP-lowering medications at 24 months than those treated with SOC (RR 2.42, 95% CI 1.94 to 3.01, p < .05). Heterogeneity was within acceptable limits (I-squared = 13.5%). Comparison of IOP rising at 6 months between DEX implant, 700 µg, and IVTA, 4 mg (RR 1.80 95% CI 1.34 to 2.42, p < .05) showed a statistically significant difference.

In Bayesian network meta-analysis (Table 2), IVTA, 4 mg, at 6 months is shown to be associated with a lower risk of a high intraocular pressure compared with FA implant, 0.2 µg/day, at 36 months (RR 3.43 95% CI 1.12 to 11.35, p < .05). We compared the IOP rising of four intravitreal therapeutic agents with that of placebo, and IVTA, 4 mg, used significantly less IOP-lowering medications than that of placebo (RR 0.32 95% CI 0.11 to 0.91, p < .05) (Figure 2F). FA implant, 0.59 mg, caused significantly more IOP rising than SOC at 24 months (RR 2.53 95% CI 1.14 to 6.25, p < .05) (Figure 2F). The probabilities of IVTA, 4 mg, ranking as the first to fourth intervention showing the lowest incidence of an elevated IOP was: 90, 7, 7, 2, 1, and 1% (Table 3).

The node-splitting approach (Higgins et al., 2012; Dias et al., 2013) was used to assess inconsistency and demonstrates that all p-values were higher than .05 and varied from .55 to .80 (Tables 1, 2). Therefore, no significant inconsistency was detected, and the test of inconsistency was able to apply when direct head-to-head evidence was available.

Ranking depended on point estimates in pairwise comparison between an intravitreal therapeutics agent and placebo or SOC (Figure 2). (Rücker and Schwarzer, 2015) We created a ranking on the safety of IOP rising in RCTs. IVTA, 4 mg, might be the best intravitreal therapeutic regimen associated with low risk of IOP rising (RR 3.13 95% CI 1.10–9.00) (Figure 2F). Additionally, because there were no significant difference observed in other efficacy and safety outcomes of drugs, we were unable to obtain a ranking of treatment based on data currently available.

Sensitivity analysis was performed by removing studies that showed high heterogeneity in the pairwise meta-analysis (Tables 1, 2). For uveitis recurrence at 24 months, high heterogeneity was found in the comparison between FA implant, 2.1 mg, and FA implant, 0.59 mg. After conducting the Bayesian network meta-analysis in the remaining studies, no significant change was revealed in network meta-analysis, indicating the reliability of our study. High heterogeneity was also detected in comparison between FA, 0.2 µg/day, and placebo for uveitis recurrence at 6 months. However, all included studies provide overwhelming evidence that FA, 0.2 µg/day, was associated with a lower recurrence rate. Thus, omission of any included study would not alter the result, indicating the robustness of the statistical analysis.

To ensure that the indirect comparisons done in this study were based on the inclusion of comparable patient populations and using similar designs, we limited our review to a set of homogeneous trials with rigorous criteria, including DEX implant, 350 μg; DEX implant, 700 μg; FA implant, 0.2 µg/day; FA implant, 0.59 mg; FA implant, 2.1 mg; IVB, 1.25 mg; IVR, 0.5 mg and IVTA, 2 mg; and IVTA, 4 mg; which are novel intravitreal therapeutic agents for the treatment of noninfectious uveitis. This network meta-analysis focused on nine intravitreal therapeutic regimens and systemic therapy involving 1806 patients by deriving data from 13 RCTs. Overall, no specific treatment regimen showed a statistically significant advantage or disadvantage over another regimen with regard to efficacy of BCVA improvement, vitreous haze improvement, uveitis recurrence, and change of retinal thickness. In our study, we assessed the safety profile of the treatments by evaluating cataract occurrence and drug-induced IOP rise. FA implant, 0.59 mg, caused more side effects than SOC in general (Figures 2E,F). IVTA, 4 mg, is shown to cause less elevated IOP than other intravitreal therapeutic agents (Figure 2F).

In the present network meta-analysis, FA implants tend to be more effective in long-term, targeted control of inflammation as compared with SOC, and the difference is close to significance, which caused fewer side effects of cataracts and elevated IOP (Figures 2C,E,F). However, during the long-term use of systemic immunosuppressive drugs, the side effects were not limited to the eye, which typically included hepatotoxicity, renal impairment, severe gastrointestinal upset, and nephrotoxicity (Jabs et al., 2000; McCluskey et al., 2000). In the management of chronic noninfectious uveitis, intravitreal steroid implants are a steroid-sparing agent to control inflammation and are able to reduce serious and intolerable side effects of SOC.

As far as we know, this is the first comparison of the major different intravitreal therapeutic regimens for noninfectious uveitis using a Bayesian network meta-analysis. Lacking multiarm trials and difficult-to-conduct, high-quality, head-to-head RCTs especially among these newly developed intravitreal therapeutic agents made it necessary to apply a network meta-analysis. A Cochrane review (Brady et al., 2016) compared a fluocinolone acetonide implant with SOC and included two trials (625 eyes), whereas our study summarizes 13 studies with a total of 1806 eyes included. Unlike past studies that focus on comparing intravitreal implants for noninfectious uveitis only (Brady et al., 2016; Vieira et al., 2020), the present study thoroughly ranks intravitreal therapeutic regimens, including IVTA and intravitreal anti-VEGF. This distinction is important because 2 or 4 mg of TA is one of the most widely used intravitreal steroids in the treatment of noninfectious uveitis (Ganapathy et al., 2018), and anti-VEGF agents as nonsteroid intravitreal therapeutics are currently undergoing evaluation for efficacy of controlling uveitic complications (Thomas and Lin, 2020), whereas anti-VEGF agents were not approved by the FDA as an intravitreal treatment option for noninfectious uveitis yet.