According to the previous protocol, we performed a systematic literature search for clinical trials and observational studies on the following databases and registers: Embase, PubMed, ClinicalTrials.gov, WHO COVID-19 Global literature on coronavirus disease, Cochrane Central Register of Controlled Trials, Scopus, Web of Science, medRxiv, bioRxiv and ArXiv. The searching strategy used subject and free-text terms covering (“intravenous immunoglobulin” or “immunoglobulin” or “intravenous Ig” or “IVIg” or “IgIV”) and (“COVID-19” or “2019-nCoV disease” or “Coronavirus Disease 2019” or “SARS-CoV-2”) in each database. The publish date of studies was restricted between December 30, 2019 and February 17, 2022, with no language limitation.

The inclusion criteria included (1): Randomized controlled clinical trials and cohort-controlled trials that reported the efficacy of IVIg; (2) Hospitalized patients with lab-confirmed SARS-CoV-2 infection. The exclusion criteria included: (1) Review articles, editorials, opinions, case report, case series; (2) Single-arm study with IVIg intervention only; (3) Studies that examined the efficacy of IVIg on pediatric patients; (4) Studies which applied the non-standard IVIg preparation (e.g., hyperimmune IVIg or IgM-enriched IVIg, et al.). Two authors (LF.L and XD.L) conducted the literature screening independently, titles, abstracts of citations, and full-text, if appropriate, were accessed to determine whether they met the eligibility criteria above. Citations with duplicates or disagreements were solved by another author (XS.L). Further uncertainties about inclusion were resolved by two senior authors (W.C and TS.L).

The primary outcome of interest of this study was the overall mortality among hospitalized COVID-19 patients. The secondary outcome of interest included the length of hospital stay, the needs for mechanical ventilation, the incidence of adverse events (AEs) and serious adverse events (SAEs).

Data on first author name, publish date, type of study, the severity of disease, number of participants, countries of study of each study were extracted. According to the Cochrane Handbook, the Revised Cochrane tool for assessing the risk of bias in randomized trials (RoB 2.0) was applied to evaluate the risk of bias among included RCT studies (33). The Risk of Bias in Non-randomized Studies-of Interventions tool (ROBINS-I) and the Newcastle-Ottawa Scale (NOS) were applied to evaluate the risk of bias among included cohort studies (34, 35). The NOS scale includes 8 items in 3 domains, with a total of 9 scores. Studies with scores ≥ 6 are considered to be of high quality and low risk. The level of evidence was judged according to the Cochrane’s Grading of Recommendations, Assessment, Development and Evaluation (GRADE) assessment handbook (36).

Meta-analysis was conducted on at least two studies with combinable outcome indicators. Dichotomous variables (including all-cause mortality, the need for mechanical ventilation, the incidences of AE and SAE) were calculated using relative risk (RR) with 95% confidence interval (CI), and continuous variables (e.g., the length of hospital stay) were calculated using mean difference (MD) with 95% CI. Continuous variables reported with median and interquartile range (IQR) were converted into mean and standard deviation (SD) using the online tool (https://www.math.hkbu.edu.hk/~tongt/papers/median2mean.html) (37–39). If the unadjusted and adjusted RRs were reported in cohort studies, the adjusted effects were used for data pooling. The random-effects model was used to calculate effect sizes and the summarized results were shown as forest plots, respectively. The Q test and I² index were used to calculate the heterogeneity of the integrated results (Not significant heterogeneity: I²< 40%; moderate heterogeneity: 40%-70%; substantial heterogeneity: I²> 70%). For evaluating the risk of publication bias, the inverted funnel-plot analysis and Egger’s test were conducted.

Subgroup analyses and meta-regression were used to explore the source of heterogeneity. Subgroup analyses based on study design and disease severity (critically ill/severe/non-severe) were conducted. The definitions of disease severity followed the WHO clinical management guideline (40). Critically ill COVID-19 was defined by the presence for ARDS, sepsis, septic shock, or need for mechanical ventilation (40). Severe COVID-19 was defined by any of the following indicators not in conformity with critically-ill: respiratory distress (respiratory rate ≥ 30 breaths/min), oxygen saturation (SpO₂) ≤ 90% at rest on room air, or signs of pneumonia and severe respiratory distress (40). Non-severe COVID-19 was defined as the absence of any symptoms for critically ill and severe type (40).

The maximum likelihood random-effects meta-regression was conducted to explore the association of mortality with sample size, age, sex, the incidence of hypertension or diabetes, the percentage of corticosteroids usage, the daily dosage of IVIg (high-dose [defined as 0.4-1.0 g/kg/day] or low-dose [less than 0.4 g/kg/day], except for one study that applied 15 g/day as the threshold (20), and the duration of IVIg therapy of the participants from each study. Regression coefficient (Coef.) was calculated to represent the correlation. Variables with P-value < 0.1 in univariable analysis were selected into further multivariable analysis. The trial series analysis (TSA) was performed to determine the two-sided conventional test boundaries, O’Brien-Fleming statistical significance boundaries, futility boundaries. The control event proportion was obtained from included studies. All P-values in this study were two-tailed, and statistical significance was set at P < 0.05. All analyses were performed using Review Manager (RevMan, version 5.3, the Cochrane Collaboration, UK), STATA (version 16.0, StataCorp LLC, USA) and the TSA software (version 0.9.5.10 Beta; Copenhagen Trial Unit, Copenhagen, Denmark).

As shown in Figure 1 , a total of 53,039 records from electronic literature databases were identified through the initial searching process. At the first eligibility check stage on title and abstract, 147 articles were considered as potentially eligible and were assessed for full-text for further screening. After screening, 17 studies (including five RCT (18, 19, 22, 41, 42) and 12 cohort studies (20, 21, 43–52) that were considered to meet the selection criteria and included in the meta-analysis. Different form the latest study (28), four studies were newly identified and included in this meta-analysis (48–50, 52).

A total of 4,711 hospitalized COVID-19 patients were included in this study, and 1,925 (40.9%) of them were treated by IVIg ( Table 1 ). Among the 17 studies, six studies were conducted in China, four were conducted in Iran, two were conducted in India or Turkey, respectively. The dosage and duration of IVIg were greatly variable between studies. Of note, the average age of participants ranged from 37 to 68 years old ( Table S2 ). Also, the percentage of male ranged from 33.0% to 94.9%, the incidence of hypertension or diabetes ranged from 20.2% to 58.5% and 11.5% to 53.0% across different studies, showing the internal discrepancies in demographic characteristics and comorbidities.

The reported outcomes of included studies were summarized in Table 2 . Overall, the application of IVIg did not reduce the mortality in RCT studies (RR= 0.57 [0.19, 1.68], P= 0.30; I² = 72%), or cohort studies (RR= 0.93, [0.63, 1.36], P= 0.71; I² = 78%) ( Figure 2A ). As there were five cohort studies that reported more than one adjusted effect estimated with different statistical methods (20, 44, 45, 47, 49), we additionally calculated the pooled RR of cohort studies as sensitive analysis (adjusted RR= 0.91, [0.67, 1.26], P= 0.58; I² = 81%) ( Figure S1A ). These data suggested no significant improvement of IVIg on mortality of overall COVID-19 patients, and the certainty level was low due to the inconsistency in design bias and outcome measurement of studies. ( Table 2 ).

Since the clinical states caused great differences in prognosis, we conducted subgroup analysis on different disease severity. Of note, there were two studies examining the effects of IVIg in patients with severe and critically ill type (20, 50), and one was excluded in this analysis for not providing the results of subgroups (50). Results showed that IVIg did not reduce the mortality in different groups (Critically ill type: RR= 1.16 [0.71, 1.91], P= 0.56, I² = 80%; Severe type: RR= 0.65 [0.33, 1.26], P= 0.20, I² = 71%; Non-severe type: RR= 1.52 [0.10, 24.35], P= 0.77, I² = 30%; P for interaction= 0.363) ( Figure 2B ). We also conducted subgroup analysis based on the differentially reported adjusted RR from four cohort studies (44, 45, 47, 49), with similar results ( Figure S1B ).

Effects of IVIg on the length of hospital stay were reported in 11 studies, including five RCT studies (18, 19, 22, 41, 42) (MD= -2.60 [-8.98, 3.77] days, P= 0.42; I² = 97%) and six cohort studies (20, 43, 44, 46) (MD= 2.87 [-0.69, 6.44] days, P= 0.11; I² = 88%) ( Figure S2A ). The integrated MD was 0.29 [-3.40, 6.44] days (P= 0.88; I² = 96%; low quality of certainty) ( Table 2 ). We found that the IVIg treatment might prolong the hospital stay of severe patients (MD = 2.62 [0.25, 4.99] days, P= 0.03; I² = 79%), and low certainty level. ( Figure S2 ; Table 2 ).

Effects of IVIg on the need for mechanical ventilation were reported in eight studies, including three RCT studies (19, 41, 42) (RR= 0.77 [0.49, 1.21], P= 0.26; I² = 39%) and five cohort studies (21, 44, 46, 48, 50) (RR= 1.02 [0.76, 1.37], P= 0.89; I² = 66%) ( Figure S3 ). One RCT study was excluded in this analysis because all patients had applied mechanical ventilation before allocating to receive IVIg treatment (22). The pooled RR was 0.93 ([0.73, 1.19], P= 0.31; I² = 56%), indicating no significant improvement of IVIg on the use of mechanical ventilation ( Figure S3 ; Table 2 ).

Safety of IVIg on the incidence of AE was reported in six studies, including three RCT studies (22, 41, 42) and three cohort studies (43, 44, 47). The overall incidence of AE and SAE were similar between groups [AE RR= 1.15 [0.99, 1.33], P= 0.06; I² = 20%; low quality of certainty; SAE (RR= 1.56 [-0.89, 2.73], P= 0.12; I² = 0%; low quality of certainty] ( Figure S4 ; Table 2 ). Notable, the overall AE was might be underestimated since two studies only reported the incidence of acute kidney injury, instead of other AEs (43, 47). Meanwhile, the incidence of specific AE was unable to be further analyzed as not all studies reported the detailed list of AEs.

Considering the internal discrepancies among studies, the meta-regression was further conducted to explore the source of heterogeneity. In the univariable meta-regression analyses, IVIg daily dosage (high or low) was considered as the risk factor in impacting mortality [Coef.= -0.34 (-0.57, 0.01), P= 0.004] ( Table 3 ). The further multivariable meta-regression showed that IVIg daily dosage and duration accounted for 66.7% of heterogeneity and the daily dosage of IVIg remained statistically significant in multivariable meta-regression ( Table 3 ).

Subgroup analysis focusing on the impact of IVIg daily dosage on the mortality of patients with different disease severity was conducted. Three studies were not been included in the subgroup analysis for lacking a report on the specific dosage and the relevant mortality (43, 46, 49). Results showed that the overall mortality among patients with severe COVID-19 was reduced in high-dose IVIg subgroup (RR= 0.33 [0.13, 0.86], P= 0.02, I² = 68%; Very low certainty) ( Figure 3 ). Additional analyses were conducted based on the differentially reported adjusted RR from four studies (44, 45, 47, 49), with similar results ( Figure S5 ). Interestingly, we noticed that severe patients treated with high-dose IVIg group did not have a longer length of hospital stay (MD = 1.54 [-1.34, 4.42] days, P= 0.29; I² = 80%) ( Figure S6 ). The prolonged hospital stay might be associated with the application of inadequate IVIg ( Figure S6 ). Further TSA analysis showed that the cumulative Z curve of estimate effects crossed neither the futility, nor the TSA boundaries, it also not exceeded the required information size ( Figures S7, 8 ).

The overall methodological judgments of RCT studies were conducted with ROB 2.0 tool ( Tables S3, S4 ). Only one study was considered with a low risk of bias (22), while the rest four studies were considered with a high risk of bias due to the open-labeled design (18, 19, 41, 42). The NOS scores of cohort studies were ranging from 4 to 7, with an average of 5.8 ( Table S5 ). Four studies were considered with a low risk of bias (43–45, 47), and the others were considered with the high risk of bias. The further ROBINS-I analysis showed that three studies were judged to be at moderate risk of overall bias (44, 48, 49), while the rest were considered with a serious risk of bias ( Tables S5, S6 ). The funnel plots and results from Egger’s regression test showed that no significant publication bias was observed (intercept= 0.90 [-2.68, 0.88], P= 0.30) ( Figure S9 ).