Association between tocilizumab, sarilumab and all-cause mortality at 28 days in hospitalized patients with COVID-19: A network meta-analysis

Objective: To estimate pairwise associations between administration of tocilizumab, sarilumab and usual care or placebo with 28-day mortality, in COVID-19 patients receiving concomitant corticosteroids and non-invasive or mechanical ventilation, based on all available direct and indirect evidence. Methods: Eligible trials randomized hospitalized patients with COVID-19 that compared either interleukin-6 receptor antagonist with usual care or placebo in a recent prospective meta-analysis (27 trials, 10930 patients) or that directly compared tocilizumab with sarilumab. Data were restricted to patients receiving corticosteroids and either non-invasive or invasive ventilation at randomization. Pairwise associations between tocilizumab, sarilumab and usual care or placebo for all-cause mortality 28 days after randomization were estimated using a frequentist contrast-based network meta-analysis of odds ratios (ORs), implementing multivariate fixed-effects models that assume consistency between the direct and indirect evidence. Results: One trial (REMAP-CAP) was identified that directly compared tocilizumab with sarilumab and supplied results on all-cause mortality at 28-days. This network meta-analysis was based on 898 eligible patients (278 deaths) from REMAP-CAP and 3710 eligible patients from 18 trials (1278 deaths) from the prospective meta-analysis. Summary ORs were similar for tocilizumab [0.82 [0.71-0.95, P=0.008]] and sarilumab [0.80 [0.61-1.04, P=0.09]] compared with usual care or placebo. The summary OR for 28-day mortality comparing tocilizumab with sarilumab was 1.03 [95%CI 0.81-1.32, P=0.80]. The P value for the global test for inconsistency was 0.28. Conclusion: Administration of either tocilizumab or sarilumab was associated with lower 28-day all-cause mortality compared with usual care or placebo. The association is not dependent on the choice of interleukin-6 receptor antagonist.


Introduction 52
Following the recent publication of results from a prospective meta-analysis 1 and an updated guideline from the 53 WHO 2 , the interleukin-6 receptor antagonists, tocilizumab and sarilumab, have been recommended alongside 54 corticosteroids for the routine treatment of hospitalized patients requiring oxygen support for COVID-19. 55 56 Findings from the prospective meta-analysis, which unlike standard meta-analysis is planned whilst trials are 57 ongoing, preceding any knowledge of trial results and therefore less prone to biases sometimes associated with 58 standard meta-analysis of aggregate data 3 , showed that the interleukin-6 antagonists were associated with 59 lower all-cause mortality 28 days after randomization than standard care alone. In a prespecified analysis 60 stratified by individual interleukin-6 receptor antagonists, whilst there was a clear association between reduced 61 mortality and tocilizumab (based on the results of 8048 patients from 19 randomized trials), the evidence 62 supporting the use of sarilumab (based on 2826 patients from 9 randomized trials) was less certain. In further 63 pre-specified analyses, a stronger association between the interleukin-6 antagonists and reduced mortality was 64 observed among patients receiving concomitant corticosteroids at randomization than those not receiving 65 corticosteroids, and the proportion of patients receiving concomitant corticosteroids at randomization was lower 66 in sarilumab trials than tocilizumab trials. Possibly related to interpretations of these findings, there is some 67 emerging evidence that clinicians and healthcare providers have tended to favor the use of tocilizumab, leading 68 to potential problems with access to treatment for these patients. 69 70 The best way to resolve this uncertainty is to compare the relative effectiveness of tocilizumab with sarilumab. 71 However, because the prospective meta-analysis set out to compare interleukin-6 antagonists with standard of 72 care, trials that directly compared individual agents were excluded. Therefore, only indirect comparisons 73 between tocilizumab and sarilumab, summarized as a ratio of odds ratios, were possible. An indirect 74 comparison of the two agents, in patients receiving corticosteroids as part of usual care, suggested similar 75 associations for both agents with 28-day all-cause mortality (Ratio of odds ratios, 0.77 [95%CI 0.44-1.33, 76 4 immune modulation therapy domain of the trial, pre-defined triggers for equivalence between tocilizumab and 83 sarilumab were met. The investigators reported beneficial effects of both tocilizumab and sarilumab on the primary outcome, organ support-free days, as well as on all pre-planned secondary outcomes including in 85 hospital survival; 90-day survival; and both intensive care unit and hospital discharge. Furthermore, they 86 reported that in their Bayesian analysis, the probability that sarilumab was non-inferior to tocilizumab was 87 98.9%. 88 89 Therefore, to further inform clinical practice and to clarify the evidence regarding these two treatments, we 90 planned a network meta-analysis, bringing together the relevant data on tocilizumab and sarilumab from all 91 randomized clinical trials. Eligible randomized trials that aimed to compare tocilizumab or sarilumab with standard care in the treatment of 98 hospitalized patients with COVID-19 were identified from the searches conducted by the same authors for a 99 recently published systematic review and prospective meta-analysis 1 . Full details of the methods used have 100 been previously reported 1 , and are included in the prospectively registered protocol (CRD42021230155) 6 . 101 102 For this network meta-analysis, we also carried out searches of trial registers (Clinicaltrials.gov and the EU 103 Clinical Trials Register) to identify any randomized trials in addition to REMAP-CAP that directly compared 104 tocilizumab with sarilumab in a similar population, using the search terms sarilumab, tocilizumab, random* and 105 COVID. Patients were eligible for inclusion in this network meta-analysis if they were included in any of the 106 eligible randomized trials and received either NIV (including high-flow nasal canula), IMV or ECMO at 107 randomization. Furthermore, because the prospective meta-analysis demonstrated that corticosteroid use 108 modifies the association of interleukin-6 antagonists with mortality, patients also needed to have received 109 corticosteroids as part of usual care to be eligible. 110 111 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted August 28, 2021. ; https://doi.org/10.1101/2021.08.26.21262523 doi: medRxiv preprint The primary outcome was all-cause mortality up to 28 days after randomization. Data on all eligible patients 112 included in the prospective meta-analysis were extracted from the summary data supplied. We requested data 113 using bespoke data collection forms (developed for the prospective meta-analysis) for any trials identified as 114 having made a direct comparison between tocilizumab and sarilumab. 115 116 All included trials secured institutional review board approval, and informed consent for participation in each trial 117 was obtained, consistent with local institutional review board requirements. Approval was not required for these 118 secondary analyses as all data were published either as part of the prospective meta-analysis and/or in 119 individual trial reports. 120 121

Risk of bias 122
Risk of bias for each trial included in the prospective meta-analysis had already been assessed for all-cause 123 mortality 28 days after randomization as part of the prospective meta-analysis and was not repeated here. We 124 planned to similarly assess risk of bias for any additional eligible trials identified for the network meta-analysis 125 for this outcome using version 2 of the Cochrane Risk of Bias Assessment Tool 7 . 126 127

Contemporaneous randomization in REMAP-CAP 128
Because REMAP-CAP is a multi-arm trial with an adaptive non-parallel design, for the purposes of this analysis 129 it is represented as three independent observations in the model (tocilizumab vs usual care or placebo, 130 sarilumab vs usual care or placebo, sarilumab vs tocilizumab). A small group of patients (21, 4 deaths by 28 131 days) were randomized to usual care or placebo contemporaneously with both treatment arms. We re-allocated 132 these patients (and events) to the tocilizumab vs usual care or placebo and sarilumab versus usual care or 133 placebo observations in proportion to their total counts and events, and thereafter assumed independence 134 between these observations. 135 136

Statistical analysis 137
Pairwise associations between tocilizumab, sarilumab and usual care or placebo were estimated using a 138 network meta-analysis of odds ratios (ORs), using a frequentist contrast-based approach implemented in 139 multivariate fixed-effects meta-analysis models 8 . These models assume consistency between 'direct evidence' 140 (associations estimated in trials directly comparing the pair of interventions) and 'indirect evidence' 141 (associations estimated through the network). The 'net evidence' from the network meta-analysis is a weighted 142 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted August 28, 2021. ; 6 average of the direct and indirect evidence. Inconsistency between direct and indirect evidence was examined 143 locally using symmetrical node-splitting 9 and globally using a design-by-treatment interaction model 8,10 .
Borrowing of strength statistics were calculated using the score decomposition method 11 to illustrate the 145 proportion of information for each net estimate that is due to indirect evidence. Treatment rankings were also 146 calculated and are summarized according to the surface under the cumulative ranking curve (SUCRA) value, 147 which represents the re-scaled mean ranking 12,13 . Following the approach in the prospective meta-analysis 1 , we 148 report precise p values and do not set a threshold for statistical significance. All analyses were conducted in 149 Stata statistical software version 16.1 [StataCorp, USA] using the 'network' user-written command suite 14 . 150 151

Study selection and description of eligible trials 153
Of the 27 trials included in the prospective meta-analysis, in nine trials patients were randomized prior to 154 guidance to include corticosteroids as part of routine care, or patients requiring non-invasive or mechanical 155 ventilation were not included, or an interleukin-6 agent other than tocilizumab or sarilumab or a combination of 156 these were used. Thus, no patients in these trials are eligible for the network meta-analysis ( Figure 1). Of the 157 nine trials, a similar number compared tocilizumab (5 trials, 848 patients) to usual care or placebo as compared 158 or sarilumab (4 trials, 815 patients) to usual care or placebo, although the sarilumab trials comprised a far 159 higher proportion of the total sarilumab patients in the prospective meta-analysis. The remaining 18 trials 160 contained at least one eligible patient and are included in this network meta-analysis, with 5 published 4,15-18 , 3 161 reported on pre-print servers 19,20 and 10 unpublished (see Table 1 for trial registration numbers). These 18 trials 162 had compared tocilizumab (13 trials) or sarilumab (4 trials) or both (1 trial) with usual care or placebo. From 163 these studies, 3710 patients (40%, 1278 deaths by 28 days) were eligible for inclusion in the network meta-164 analysis as they received corticosteroids and either non-invasive or mechanical ventilation (Figure 1 12th, 2021 using a standardized outcome data collection form (developed for use in the prospective meta-171 analysis) and finalized data were subsequently verified by the trial team prior to inclusion in this analysis. Of 172 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted August 28, 2021. ; https://doi.org/10.1101/2021.08.26.21262523 doi: medRxiv preprint eligible for inclusion in the network meta-analysis. 175 176

Risk of bias within studies 177
Detailed risk of bias assessments for the 18 included trials that contributed to the prospective meta-analysis 178 have already been reported 1 . In summary, 12 were assessed as low risk of bias (1003 deaths by 28 days, 65% 179 of total deaths); five were judged to have some concerns (257 deaths by 28 days, 17% of total deaths) largely 180 as small numbers of patients who did not receive their assigned interventions were excluded. One trial (18 181 deaths by 28 days, 1% of total deaths) was judged as high risk of bias as the usual procedures to ensure 182 concealment of the allocation sequence were not in place; however, concealed allocation did appear to have 183 been implemented as intended. Risk of bias for the additional REMAP-CAP direct comparison was judged as 184 low risk of bias. Thus, in total, 12 trials (14 comparisons, 1281 deaths by 28 days, 82% of total deaths) were 185 judged as low risk of bias.   In this network meta-analysis of patients receiving both corticosteroids and either NIV, IMV or ECMO at 208 randomization, both tocilizumab and sarilumab were associated with lower all-cause mortality 28 days after 209 randomization compared with usual care or placebo. The associations of these agents with all-cause mortality 210 appeared similar, consistent with the direct findings from the REMAP-CAP trial 5 in which tocilizumab and 211 sarilumab met the criteria for equivalence. More generally, these results confirm a clear association of 212 interleukin-6 receptor antagonists with lower all-cause mortality in this patient population. 213

214
The comparison of tocilizumab versus usual care was based mainly on direct evidence from the prospective 215 meta-analysis, with only limited additional information from the network. By contrast, for sarilumab versus usual 216 care or placebo, the direct comparison was limited to fewer than 500 patients from the prospective meta-217 analysis. Therefore, the indirect evidence (arising from the association of tocilizumab with reduced all-cause 218 mortality compared to usual care or placebo and the direct comparison of tocilizumab with sarilumab) has a 219 substantial impact on the net estimate for this comparison. In the absence of any other direct comparisons of 220 tocilizumab with sarilumab, this network meta-analysis provides the strongest evidence in support of the 221 hypothesis that both agents are similarly associated with lower all-cause mortality at 28-days in this patient 222 population. is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted August 28, 2021. ; https://doi.org/10.1101/2021.08.26.21262523 doi: medRxiv preprint Therefore, variability of the population within the network and resulting inconsistency was reduced, and 234 interpretability was increased. This was only possible through the prospective and collaborative approach 3 we 235 adopted as part of the prospective meta-analysis 1 , collecting detailed data on both oxygen support and 236 corticosteroid use subgroups. This enabled us to make decisions often only available in an individual participant 237 data network meta-analysis and resulted in increased consistency and harmonization. Furthermore, over 80% 238 of the total events included in this network were from trials judged to be at low risk of bias. ventilation and corticosteroids at randomization, confirms that administration of tocilizumab or sarilumab, 250 compared with usual care or placebo, is associated with similarly lower 28-day all-cause mortality. 251 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted August 28, 2021. ; https://doi.org/10.1101/2021.08.26.21262523 doi: medRxiv preprint Table 2: Summary of direct, indirect and net evidence for the associations of tocilizumab, sarilumab and usual 285 care or placebo with all-cause mortality 28 days after randomization for patients receiving corticosteroids and 286 either NIV, IMV or ECMO at randomization 287 Table 3: Ranking of interventions (% probability) and SUCRA values for all-cause mortality 28 days after 288 randomization 289 is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted August 28, 2021. ;

357
These patients are events are re-allocated in proportion to their total counts and events.

CC-BY 4.0 International license
It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted August 28, 2021.
after randomization for patients receiving corticosteroids and either NIV, IMV or ECMO at randomization 360 is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted August 28, 2021. is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted August 28, 2021. is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted August 28, 2021. is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted August 28, 2021. is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted August 28, 2021. is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted August 28, 2021.