A systematic review and network meta-analysis of current and investigational treatments for active ankylosing spondylitis

NMAs were conducted using Bayesian methods. In a Bayesian NMA, a posterior distribution for each treatment effect is derived by combining a prior probability distribution (i.e., a prior belief of the relative efficacy of a treatment) with a likelihood (i.e., a statistical model) for the effect estimate. A non-informative prior, which assigns an equal probability to any theoretically plausible effect estimate, was used to avoid undue influence of an overly informative prior distribution. Given the data and likelihood, a posterior distribution consisting of 15,000 effect estimates (i.e., mean difference (MD) for continuous endpoints or odds ratio (OR) for binary endpoints) was estimated using Markov Chain Monte Carlo re-sampling methods. This distribution was then used to obtain the median effect estimate (i.e., a point estimate) and the 95% credible interval (CrI), which includes all posterior values between the 2.5th and 97.5th percentiles of the posterior distribution.

Bayesian NMAs were conducted for ASAS20 response, change from baseline in BASFI, and change from baseline in CRP using a random effects (RE) model. ASAS20 response was defined as a ≥ 20% relative improvement from baseline and absolute improvement from baseline of ≥ 10 units on a 0–100mm scale in ≥ 3 of the following domains: patient global assessment, spinal pain assessment, function (BASFI), and inflammation (the last two questions of BASDAI) [20]. Pairwise comparisons between interventions were reported as median OR with 95% CrIs for ASAS20 and as MD with 95% CrIs for change from baseline in BASFI and change from baseline in CRP. The Surface Under the Cumulative Ranking curve (SUCRA) values, reported as percentages, were calculated to reflect the relative probability of an intervention being among the best options. Each NMA was performed in accordance with the methodology recommended by the National Institute for Health and Care Excellence (NICE). To adjust for differences in placebo response across studies, a baseline risk-adjusted sensitivity analysis was conducted using code reported in the NICE Decision Support Unit (DSU) Technical Support Document (TSD) 3. Deviance information criterion (DIC) and posterior residual deviance were reported to assess model fit for each analysis. The estimated regression coefficient (β) and 95% CrIs were reported to assess model fit for baseline risk-adjusted analyses. All analyses were performed using WinBUGS [21] and R [22] with burn-in and sampling durations ≥ 50,000 iterations. Comparisons of baseline risk-adjusted and unadjusted NMAs were used to assess risk of bias from between-study heterogeneity. Inconsistency models were developed and compared to consistency models to assess inconsistency in the networks.

A total of 1466 published manuscripts were identified from database searches and 104 from supplementary searches. A total of 1570 records were screened at the title and abstract stage and 276 were included for full-text review. Of those reviewed in full text, 244 records were excluded with reason and 32 records, representing 30 unique studies, were included in the final review (Appendix S1).

The 30 unique RCTs enrolled a total of 6711 patients. A summary of study and patient baseline characteristics is presented in Appendix S2. The mean age of patients ranged from 27.4 to 48.0 years and the mean disease duration ranged from 5.2 to 23.0 years. The proportion of male patients ranged from 52.6 to 100% and the proportion of HLA-B27-positive patients ranged from 65.0 to 97.4%. Mean baseline scores for Bath Ankylosing Spondylitis Disease Activity Index (BASDAI) ranged from 4.4 to 7.6 cm, for BASFI ranged from 3.2 to 7.4 cm, and for CRP ranged [23] from 6.2 to 32.0 mg/l. Of the 30 studies, 23 allowed concomitant use of conventional synthetic disease-modifying antirheumatic drugs (csDMARDs), four prohibited concomitant use, and three did not report this information.

The risk of bias was generally low across trials for random sequence generation, allocation concealment, incomplete outcomes, and selective reporting (Appendix S4).

An NMA was conducted for each outcome of interest: attainment of ASAS20 response, change from baseline in BASFI, and change from baseline in CRP. The network diagram for ASAS20 is provided in Fig. 1. Each intervention is represented by a node, and randomized comparisons are shown as links between the nodes. Results of each NMA are presented in Figs. 2, 3, and 4 for the models with the best fit. An odds ratio greater than one or a mean difference less than 0 implies that the reference treatment was better than the comparator. Model fit statistics for each analysis are provided in Appendix S6. A significant β for a baseline risk-adjusted NMA suggested that the adjusted model was a better fit than the unadjusted model.

Twenty-six studies were included in the comparison of ASAS20 at weeks 12 to 16 (Fig. 1). The baseline risk-adjusted model had the best fit because the 95% CrI for β excluded 0 (Appendix S6). Tofacitinib 5 mg had superior ASAS20 response compared to GOL IV 2 mg/kg and TOF 10 mg, and both TOF 5 mg and GOL IV 2 mg/kg were superior to TOF 2 mg, PBO, RIS 90 mg, SEC SC 75 mg, UST 90 mg, RIS 180 mg, APR 30 mg, and UST 45 mg (Fig. 2). Tofacitinib 5 mg and GOL IV 2 mg/kg were of greater or equal efficacy compared to the other treatments. Rankings based on SUCRA values for ASAS20 response were highest for TOF 5 mg (93%), GOL IV 2 mg/kg (90%), and FIL 200 mg (86%) (Fig. 5).

Twenty-three studies were included in the comparison of the change from baseline in BASFI at weeks 12 to 16 (Appendix S5A). The unadjusted model had the best fit because the 95% CrI for β excluded 0 (Appendix S6). Golimumab IV 2 mg/kg was ranked highest among all treatments, followed by IFX 5 mg/kg. Both interventions had superior reductions from baseline for BASFI compared to PBO and UST 45 mg, and IFX 5 mg/kg was also superior to UST 90 mg (Fig. 3). SUCRA values for the change from baseline in BASFI were highest for GOL IV 2 mg/kg (81%), IFX 5 mg/kg (80%), and GOL SC 100 mg (69%) (Fig. 5).

Nineteen studies were included in the comparison of the change from baseline in CRP at weeks 12 to 16 (Appendix S5B). The unadjusted model had the best fit because the 95% CrI for β excluded 0 (Appendix S6). Infliximab 5 mg/kg was ranked highest among all treatments, followed by GOL IV 2 mg/kg. Infliximab 5 mg/kg showed superior reduction from baseline in CRP compared to UST 90 mg, UST 45 mg, and PBO and GOL IV 2 mg/kg was superior to PBO (Fig. 4). SUCRA values for the change in CRP NMA were highest for IFX 5 mg/kg (90%), GOL IV 2 mg/kg (82%), and IXE 80 mg Q4W (76%) (Fig. 5).

Between-study heterogeneity was assessed for each outcome using the I² statistic for pairwise comparisons to placebo with ≥ 2 independent studies (Appendix S10) [23]. The I² statistic was low (< 40%) or moderate (30 to 60%) for most comparisons, suggesting that heterogeneity was unlikely to bias the results. Substantial heterogeneity (I² > 60%) was observed only in pairwise comparisons of treatments with low SUCRA values which suggests heterogeneity is unlikely to alter conclusions. A comparison of unadjusted and baseline risk-adjusted analyses for each outcome showed that results were similar between models overall (Appendices S6 and S7). For ASAS20, the baseline risk-adjusted NMA yielded smaller heterogeneity compared to the unadjusted NMA. For change in BASFI and CRP, the unadjusted and baseline risk-adjusted models yielded similar results. A sensitivity analysis including studies reporting results at 10 weeks was conducted for each outcome. This sensitivity was conducted to align with the most recently published NMA, Wang et al. [16], which included studies reporting results between 10 to 16 weeks. Inclusion of studies reporting results at 10 weeks resulted in the addition of one new study (Marzo-Ortega 2005 [24]) in the ASAS20, change from baseline in BASFI, and change from baseline in CRP networks. Results of these analyses were similar to those from the reference analyses on all three outcomes (Appendix S7).

To the best of our knowledge, this study is the most recent and comprehensive SLR and NMA comparing both currently available and investigational biologic and OSM treatments for active ankylosing spondylitis. Prior NMAs published in this therapeutic area have been restricted to TNF inhibitors. This study uniquely provides data on the comparative efficacy of treatment options with differing mechanisms of action to better inform treatment decisions. Given the recent advances in research for the treatment of ankylosing spondylitis, this NMA also provides data on the comparative efficacy of newly studied treatments such as GOL IV, FIL, IXE, RIS, and UST. The analyses conducted aligned with best practices for NMAs [25, 26] and used publicly available code to ensure transparency and reproducibility. Furthermore, an adjusted analysis accounting for baseline risk differences between trials was also conducted to ensure baseline differences were not confounding the results. The analysis reported here focuses on published summary-level data from RCTs. There are indirect treatment comparison methodologies, such as matching-adjusted indirect comparisons (MAIC) and propensity score matching (PSM), available that can leverage individual participant data (IPD) from RCTs to adjust for potential heterogeneity between studies. Although MAICs and PSMs can better adjust for heterogeneity than NMA by leveraging IPD, there was limited heterogeneity identified in the current analysis. Further, MAICs and PSMs only permit comparison of one treatment at a time. In contrast, NMA provides a more holistic view of the body of evidence in ankylosing spondylitis. This is the first NMA to simultaneously compare current and investigational treatments.

There were limited clinical trial data for each treatment to inform the comparative efficacy of biologic and OSM therapies for active ankylosing spondylitis. This is evident from the presence of many single study connections in each network, which limits the strength of the analysis. Insufficient clinical data also limited the choice of outcomes presented in this analysis. All three outcomes reported represent measures of change in a patient’s status; however, these measures do not show the absolute status of a patient. Ankylosing Spondylitis Disease Activity Score (ASDAS) inactive disease, defined as ASDAS < 1.3, was investigated as a measure of the absolute status of a patient in a limited number of studies. Unfortunately, limited reporting on ASDAS inactive disease among the identified studies prevented assessment of this outcome. Similarly, most studies did not report change in active inflammatory lesions (e.g., short-tau inversion recovery, gadolinium-enhanced T1-weighed imaging), though CRP was reported broadly and was investigated as a marker for change in active inflammation. Another limitation of this analysis was that data collected between weeks 12 to 16 were pooled because of the lack of consistent reporting across studies. Given that ankylosing spondylitis is a chronic disease, it was expected that results at these timepoints would be comparable. Furthermore, safety outcomes such as adverse events, infection, and toxicity were not examined in this study. Since safety events are typically uncommon, RCTs often lack sufficient follow-up and sample size to detect differences in such outcomes. Other NMAs examining adverse events have pooled indications to overcome this limitation [27]. Finally, almost all indications of superior efficacy of GOL IV, IFX, and TOF were in comparison to failed investigational therapies (e.g., UST, APR, RIS) and PBO.