Comparative Effectiveness and Safety of 32 Pharmacological Interventions Recommended by Guidelines for COVID-19: A Systematic Review and Network Meta-Analysis Combining 66 Trials

Background: To evaluate comparative ecacy and safety of pharmacological interventions in patients with coronavirus disease 2019. Methods: Medline, Embase, the Cochrane Library and clinicaltrials.gov were searched for randomized controlled trials (RCTs) in patients infected with SARS-COV-2/SARS-COV. Random-effects network meta-analysis within Bayesian framework was performed, followed by Grading of Recommendations Assessment, Development and Evaluation (GRADE) system assessing quality of evidence. The primary outcome of interest includes mortality, cure, viral negative conversion (VNC) and overall adverse events (OAE). Odds ratio (OR) with 95% condence interval (CI) was calculated as the measure of effect size. Results: 66 RCTs with 19,095 patients were included, involving standard care (SOC), 8 different antiviral agents, 6 different antibiotics, high and low dose chloroquine (CQ_HD, CQ_LD), traditional Chinese medicine (TCM), corticosteroids and other treatments. Compared with SOC, signicant reduction of mortality was observed for TCM (OR=0.34, 95%CI: 0.20-0.56, moderate quality) and corticosteroids (OR=0.84, 0.75-0.96, low quality) with improved cure rate (OR=2.16, 1.60-2.91, low quality for TCM; OR=1.17, 1.05-1.30, low quality for corticosteroids). However, increased risk of mortality was found for CQ_HD versus SOC (OR=3.20, 1.18-8.73, low quality). TCM was associated with decreased risk of OAE (OR=0.52, 0.38-0.70, very low quality) but CQ_HD (OR=2.51, 1.20-5.24) and IFN (OR=2.69, 1.02-7.08) versus SOC with very low quality) were associated with an increased risk. Conclusions: Corticosteroids and TCM may reduce mortality and increase cure rate with no increased risk of OAEs compared with standard care. CQ_HD might increase the risk of mortality. CQ, IFN and other antiviral agents could increase the risk of OAEs. The current evidence is generally uncertain with low quality and further high-quality trials are needed. indirectness and


Introduction
As of 3 September 2020, more than 25.8 million people have been infected with severe acute respiratory syndrome coronavirus 2 (SARS-COV-2), which has similar genetics to SARS-COV 1 . This global pandemic coronavirus disease 2019 (COVID-19) has caused 859,130 deaths in 216 countries, which has become a major public health problem and presents unprecedented challenge 1 .
So far, many kinds of drugs in addition to standard of care (SOC) are recommended by different clinical guidelines [2][3][4][5] , including antiviral agents, immune-based therapies [such as corticosteroids (COR), convalescent plasma (CON_PLA) and interferons (IFN)], hydroxychloroquine or chloroquine (CQ), traditional Chinese medicine (TCM) and other adjunctive therapies. However, no speci c drugs were currently proven and nearly all drugs are off-label prescribed 6 . Despite numerous on-going or nished clinical trials, substantial uncertainty about effectiveness and safety still exists in these therapies owing to limited sample size and large variability with insu cient power.
Additionally, these pairwise meta-analyses are not able to provide evidence on the comparative effectiveness and safety for all available treatments. While several network meta-analyses (NMA) are ongoing [18][19][20] , no results for the comprehensive assessment have been reported, and some treatments such as TCM and blood products are not included in these NMA. Furthermore, their network is sparse without the inclusion of similar genetic SARS-COV studies. Besides, placebo and SOC are considered as a single treatment in these NMA which may omit the potential placebo effect 21 .
Therefore, we aimed to collect all RCTs comparing any kinds of pharmacological interventions with placebo or SOC among SARS-COV-2 and SARS-COV patients and conduct an NMA to assess comparative e cacy and safety for these treatments.

Methods
This study was registered on International Prospective Register of Systematic Review PROSPERO, number CRD42020168178. The study was conducted according to PRISMA-NMA checklist.

Data sources and Searches
PubMed, Embase, Web of Science, Cochrane Central Register of Controlled Trials and three Chinese databases including SinoMed, China National Knowledge Infrastructure (CNKI) and WanFang Database were searched from inception to July 20th, 2020 (Appendix 1 for full details about search strategy). In addition, we also checked reference list of all relevant articles to identify additional studies.
Data extraction and quality assessment Data were extracted with respect to trial information (author, publication year, country, virus type, preprint or not, guideline or not, sample size, trial duration, types of intervention and control), population characteristics [mean age ± SD (standard deviation), proportion of female, disease severity], reported outcomes (number of events for dichotomous outcomes and mean with SD for continuous outcomes) and information on methodology. Four investigators (FS, XYZ, JXZ and QXZ) extracted data independently in duplicate. Risk of bias was assessed according to Cochrane risk of bias tool (ROB tool) 22 . Additionally, GRADE (The Grading of Recommendations Assessment, Development, and Evaluation) framework was used to assess quality of evidence contributing to each network estimation, which characterizes the quality of a body of evidence on the basis of study limitations, imprecision, inconsistency, indirectness and publication bias 23 .

Data Synthesis and Analysis
Methods for direct treatment comparisons Pairwise meta-analysis was performed using DerSimonian-Laird random-effects model. Odds ratio (OR) and weighted mean difference (WMD) with 95% con dence interval (CI) were calculated as effect measures for dichotomous and continuous outcomes respectively. The I 2 -statistic was calculated as a measure of proportion of overall variation that is attributable to between-study heterogeneity. For studies with zero-event in both arms, a continuity correction of 0.5 was used 24 . Besides, subgroup pairwise meta-analysis was conducted according to different virus type (SARS-COV-2 and SARS-COV).

Methods for indirect and mixed comparisons
A random-effects NMA within Bayesian framework 25 was then performed by 100, 000 iterations with 20,000 adaptations and thinner equals to 10. OR and WMD with 95%CI were summarized for dichotomous and continuous outcomes respectively. We estimated ranking probabilities for all treatments of being at each possible rank for each intervention. The treatment hierarchy was summarized and reported as surface under the cumulative ranking curve (SUCRA), mean ranks and rank heat plot 26 . Examination of assumptions in NMA (consistency, transitivity and heterogeneity) To check the assumption of consistency in the entire analytical network, a design-by-treatment approach was used 27 . The node splitting method were used to assess inconsistency of model with separating evidence on a particular comparison into direct and indirect evidence.
Heterogeneity was assessed with common tau 2 statistic and predictive interval 28 . The transitivity assumption was evaluated by comparing distribution of clinical variables which could act as effect modi ers across treatment comparisons. Contribution table was used to assess the contribution of each direct comparison to the estimation of each network meta-analytic summary effect 28 . Additionally, comparisonadjusted funnel plot was used to detect potential publication bias for outcomes with at least 10 trials.
All analyses were conducted using R 4.0.2 (gemtc package for NMA and node-split analysis; ggplot2 package for network evidence plot, forest plots and cumulative rank probability graphs; netmeta package for funnel plot; elds package, RColorBrewer package and circlize package for rank heat plot), STATA 13.0 (pairwise meta-analysis, estimation of local heterogeneity and contribution plot) and CINEMA website (https://cinema.ispm.unibe.ch/ for GRADE results).

Patient and public involvement
Patients and the public were not involved in this review.

Methodological quality and risk of bias results
Among the 66 included trials, allocation concealment and blinding of participants and personnel were not clearly reported in 84.8% and 48.5% of the cases, respectively. By contrast, methods for randomization and incomplete outcome data were appropriately described in large majority of studies (56.1% and 86.4%, respectively). 31.8% of trials were open label and 72.7% did not have selective reporting (the remaining 19.7% was unclear due to no related protocol). Additionally, 12.1% of trials were funded by company and 43.9% did not report funding sources (Appendix 3). Overall, risk of bias across evidence network was relatively low.

Results of pairwise meta-analysis
The effects of different drugs on mortality, cure rate, VNC and OAE from pairwise meta-analyses are shown in Appendix 4. TCM and COR were associated with a signi cant reduction in mortality (OR = 0. 33 According to contribution tables of the network (Appendix 6), comparison of SOC versus antiviral agents, COR and CQ had the largest contribution in all 4 entire networks for primary outcomes with 51.8%, 56.5%, 56.9% and 50.0% for mortality, cure rate, VNC and OAE, respectively.
Transitivity, inconsistency and heterogeneity Assessment of transitivity by box plots indicated mean age and proportion of female across treatment comparisons were relatively similar (Appendix 7). The test for global inconsistency did not detect any signi cant difference between consistency and inconsistency models for all the 4 primary outcomes (p = 0.994 for mortality, p = 0.763 for cure, p = 0.952 for VNC and p = 0.773 for OAE, respectively) and 3 secondary outcomes (p = 0.984 for DIA, p = 0.845 for SAE and p = 0.614 for TE, respectively), except for other 6 secondary outcomes(AKI, ARDS, HD, HF, SEI and TFR) could not conduct consistency test due to no loop in the whole network. The test for inconsistency from nodesplitting model showed no signi cant difference in all comparisons across all outcomes (Appendix 8). Most comparisons in all 13 outcomes were with low heterogeneity as indicated in the predictive interval. At visual inspection, funnel plots for all 7 outcomes with number of studies greater than 10 (Appendix 9) were symmetric and did not suggest any signi cant risk of publication bias. GRADE evaluation on quality of evidence According to GRADE, the quality of evidence ranged between very low and moderate (Appendix 11). In terms of TCM versus SOC, the quality was moderate for mortality, low for cure rate, and very low for VNC and OAE. As for COR versus SOC, the quality was low for mortality, cure rate and OAE while very low for VNC. Regarding to CQ_HD versus SOC, the quality was low for mortality and very low for OAE.

Subgroup analyses
In addition, subgroup pairwise meta-analysis by virus type con rmed the bene cial effect on mortality and cure rate of TCM and COR versus SOC, reduction effect on OAE of TCM versus SOC, increased risk of mortality for CQ_HD versus CQ_LD, and increased risk of OAE for CQ_HD, IFN and CQ_LD_AZI versus SOC in SARS-COV-2, which were in agreement with those previous produced (Appendix 12).

Discussion
Considering the global pandemic of COVID-19, increasing attention is being paid to the effectiveness and safety of pharmacological treatments. Our NMA with 66 trials and 19, 095 patients suggested that corticosteroids and TCM could probably reduce mortality and increase cure rate with no increased risk of OAEs compared with SOC. However, CQ_HD might increase the risk of mortality. CQ, IFN and other antiviral agents could increase the incidence of OAEs.
In line with other studies 7,9,18,[31][32][33][34][35] , we did not nd any potential effect on reducing mortality or increasing cure/viral clearance rate for IFN and any antiviral agents, but rather we found CQ_HD was associated with increased mortality. It should be recognized that several sideeffects may be caused according to some observational studies and trials, such as QT prolongation by CQ and FAV, gastrointestinal complications by LPV_RIT and UMI 18,[31][32][33][34][35] . In our study, an increased risk of OAE was detected for CQ and IFN, and diarrhea for LPV_RIT and LPV_RIT_RIB_IFN, although no other signi cant risk of OAE was detected for antiviral agents. Given the potential harms with lack of effectiveness, they are not recommended by several guidelines as treatment for COVID-19, particularly for mild to moderate patients 2-5 . For severe patients who need supplemental oxygen or intensive care, REM is weakly recommended to shorten the time to clinical improvement 5 . However, the evidence was low-or very low-quality with no observed effect on hospitalization duration and mortality.
According to the WHO and other guidelines, routine use of systematic corticosteroids was not recommended for treatment of viral pneumonia, except for patients who require supplemental oxygen and mechanical ventilation [2][3][4] . A systemic in ammatory response may develop in patients with severe COVID-19, which could result in lung injury and multisystem organ dysfunction. Since corticosteroids could decrease in ammatory response, it might lead to fewer intensive care unit transfers, thereby lowering mortality rate. Several studies, including RECOVERY trial 36 and meta-analyses 13, 18 , show similar ndings. However, in mild or moderate patients, this bene t may be outweighed by adverse effects such as delayed viral clearance and increased risk of secondary infection. Results from different studies are not consistent. A meta-analysis 37 of 6458 patients with in uenza pneumonia indicated 75% and 98% increase in mortality and secondary infection risk respectively, while a retrospective study 38 of 201 patients with COVID-19 found 62% decreased risk of mortality for methylprednisolone. Our NMA found improved effects of corticosteroids on mortality and cure rate and no effects on VNC and SEI versus SOC, perhaps due to severe pneumonia of included patients with older age (mean age = 66.2). Further subgroup and meta-regression NMA according to disease severity are essential upon the completion of many other on-going trials, since the inclusion of mild, moderate and severe patients may dilute the effect of corticosteroids.
Our study found TCM as adjuvant therapy achieved signi cant lower mortality and OAE with higher cure rate, which is consistent with previous meta-analysis [14][15][16] . While use of traditional herbs remain controversial in clinical practice, the bene cial effect is biologically reasonable [39][40][41][42][43] . The most commonly used herbs were Radix Glycyrrhizae (Gancao), Astragali Radix (Huangqi), Rhizoma Pinelliae Tematae (Banxia) and Forsythiae Fructus (Lianqiao), which could clear away heat and toxic material, eliminate phlegm-dampness and replenish qi according to TCM theory 39 . Moreover, it has also been con rmed that these herbs have a wide of pharmacological effects including anti-in ammatory, antiviral, antipyretic, antioxidative and immunoregulatory effects [40][41][42][43] . Thus, it could maintain the homeostasis of immune system, inhibit a variety of viruses and thereby effectively block the ranging from mild to critical. However, the dosage, composition, treatment duration and disease severity of COVID-19 cases should be taken into account when considering TCM, since these factors are closely related to safety issues. Despite lower AE rate in TCM compared with SOC, most trials included in our study were unblinded with low quality. Therefore, further well-designed studies are needed to investigate the safety issues of TCM.
Compared with previous relevant meta-analysis, a major strength of our study is the comprehensive search and analysis of effectiveness and safety pro les for all kinds of pharmacological treatments in a whole network with the largest number of studies and sample size.
Furthermore, we included all pharmacological treatments recommended by several guidelines 2-5 , including TCM and CON_PLA as well as other treatments evaluated in previous meta-analysis. Meanwhile, placebo and SOC were separated as two treatment nodes in our evidence network, which could minimize bias due to potential placebo effect 21 . Additionally, we assessed the quality of evidence and incorporate it into explaining the results by the GRADE framework.
Several limitations, however, should be mentioned. First, most comparisons were assessed as low or very low quality in GRADE framework with wide con dence intervals owing to sparse data, which might restrict the interpretation of results. However, these data are still valuable and timely at this stage with no effective speci c drugs for COVID-19. When more data of ongoing trials are available, we will update the analysis. Secondly, methodology of some included trials was poor. Nearly 80% of trials were not performed well in blinding or concealment allocation. Thus, this may introduce bias and results should be interpreted with caution. However, it might be di cult to conduct doubleblind trials for a contagious disease in some clinical situations. Finally, due to sparse data and unavailable access to original trial data, we could not perform detailed NMA subgroup analyses, meta-regression or individual patient data meta-analysis to properly address potentially relevant effect modi ers, such as age, disease severity and treatment duration.

Conclusions
Corticosteroids and TCM may reduce mortality and increase cure rate with no increased risk of OAEs compared with standard care. However, CQ_HD might increase the risk of mortality. CQ, IFN and other antiviral agents could increase the incidence of OAEs. A majority of trials are small-scale trials with important methodological limitations, and no de nitive conclusion could be drawn for most treatments. The current evidence is generally uncertain with low quality and further high-quality trials are needed.

Strengths And Limitations
Most comparisons were assessed as low or very low quality in GRADE framework with wide con dence intervals owing to sparse data.
Methodology of some included trials was poor, particularly in blinding or concealment allocation. Detailed NMA subgroup analyses, meta-regression or individual patient data meta-analysis could not be performed to properly address potentially relevant effect modi ers, such as age, disease severity and treatment duration.
Comprehensive search and analysis of effectiveness and safety pro les for all kinds of pharmacological treatments in a whole network with the largest number of studies and sample size.
We assessed the quality of evidence and incorporate it into explaining the results by the GRADE framework. Figure 1 Flow chart of studies considered for inclusion.