This study had been prepared under the guidance of the Preferred Reporting Items for Systematic Review and Meta-Analysis (PRISMA) guidelines (in Attachment 1) (Page et al., 2021). The study was prospectively registered on the PROSPERO platform (https://www.crd.york.ac.uk/prospero/) with an assigned registration number CRD42021244587.

Studies were considered eligible if they met the following criteria: 1) randomized controlled trial (RCT). 2) Adults aged 18 years or older. 3) All included patients were diagnosed with SP according to the “Adult Community-Acquired Pneumonia (CAP) Guidelines for Diagnosis and Treatment” issued by the American Thoracic Society/American Society of Infectious Diseases (AST/IDSA) in 2007 (Mandell et al., 2007) or the “Guidelines for the Diagnosis and Treatment of Community-Acquired Pneumonia in Chinese Adults” developed by the Respiratory Branch of Chinese Medical Association in 2016 (Qu and Cao, 2016). 4) All patients received treatment with WM such as anti-infectives, phlegm reduction medicines, mechanical ventilation, nutritional support, and so on. Based on this, the experimental group received one of the included CHIs, and the control group received another or only WM. The duration of treatment ranged from 7 to 14 days. 5) At least one of the following seven outcomes with clear definitions were evaluated: the clinical effective rate, ICU length of stay, the time of mechanical ventilation, the level of C-reactive protein (CRP), the level of procalcitonin (PCT), leukocyte (WBC), and adverse drug reactions (ADRs)/adverse drug events (ADEs). The clinical effective rate was calculated by the following formula: (number of cured patients + number of improved patients)/total number of patients × 100%. Patients were regarded as cured when their clinical symptoms and the objective indicators disappeared and the patients returned to normal. Patients were regarded as improved when their clinical symptoms and the objective indicators were alleviated. If the clinical symptoms and objective indicators were either unchanged or aggravated, the patients were identified as having an invalid effectiveness status.

Studies were excluded if any of the following criteria were met: 1) If they described data about only a specific population (patients with tumor, pulmonary fibrosis, tuberculosis, using immunosuppressant, secondary respiratory failure in other systems, etc.). 2) The full text was not available or only with abstracts. 3) Data were incorrect, incomplete, or unavailable.

A comprehensive literature search was performed using the electronic databases of PubMed, the Cochrane Library, Embase, Web of Science, China National Knowledge Infrastructure (CNKI), Wanfang Database, and the Chinese Scientific Journal database (VIP) from their inception up to April 1, 2021. The medical subject headings (MeSH) and free text words were used. Language restriction did not exist in this study. Furthermore, we manually searched the reference lists of all retrieved studies. Six different kinds of CHIs were included in this NMA: TRQ, XYP, RDN, XBJ, SF, and SM. Full details of the search strategy are shown in Attachment 2.

Two researchers (LQ Niu and L Xiao) independently screened the studies according to the inclusion criteria. After checking for duplicate studies, the researchers eliminated reviews and irrelevant studies by reading the titles and abstracts. Then, full texts were read to select studies that meet the pre-specified inclusion criteria. Inconsistencies were resolved by extensive discussion or the third researcher (XZ Liu). A data spreadsheet was developed with Microsoft Excel 2019 to collect relevant information. The information including eligible study characteristics (e.g., first author and year of publication), participant characteristics (e.g., gender, age, and sample), details of interventions (e.g., duration and frequency of drugs), outcome data, and factors to evaluate risk of bias were extracted and entered into the spreadsheet.

The methodological quality of each included study was evaluated with Revised Cochrane risk-of-bias tool for randomized trials (RoB 2) (Sterne et al., 2019). The domains include the following: 1) randomization process, 2) deviations from intended interventions, 3) missing outcome data, 4) measurement of the outcome, and 5) selection of the reported result. There are some signaling questions required to be answered by “yes (Y),” “probably yes (PY),” “probably no (PN),” “no (N),” or “no information (NI)” for each domain. After that, the risk of bias is categorized into three levels: high risk, some concerns, and low risk. These domain-level judgements will inform an overall risk of bias judgment for the outcome. The quality assessments were performed by two independent reviewers (LQ Niu and L Xiao), and disagreements were resolved by consensus or a third opinion.

OpenBUGS 3.2.3 (Lunn et al., 2009) and STATA 14.0 software (Stata Corporation, College Station, TX, USA) were employed to compute calculations and prepare graphs. For binary outcomes, the combined results were calculated as odds ratios (ORs) with 95% credible intervals (CIs). For continuous outcomes, standardized mean differences (SMD) with 95% CIs were used. When the 95% CIs of ORs did not include one and the 95% CIs of the SMDs did not contain zero, the differences between the groups were considered statistically significant. If there were multi-arm trials, we split them into two-arm trials. Chi-squared test was employed to assess heterogeneity between different studies (Zheng et al., 2019). If with homogeneity (p ≥ 0.1, I ² ≤ 50%), a fixed-effect model was adopted; if with obvious heterogeneity (p < 0.1, I ² > 50%), a random-effect model was applied. If closed loops existed, we employed the inconsistency factor (IF) to examine the consistency between direct and indirect evidence. If 95% CIs of IF values were truncated at zero, it indicated that the two sources are in agreement (Salanti et al., 2011). Due to non-closed loops in this NMA, the assumption of consistency between direct and indirect evidence was not utilized.

The Markov chain Monte Carlo method was performed by using the OpenBUGS software to carry out the NMA. In OpenBUGS software, the number of iterations was set to 300,000, and the first 100,000 iterations were used for the annealing algorithm to eliminate the impact of the initial value. The network graph was constructed using STATA software to show a comparative relationship between different interventions. Surface under the cumulative ranking curve (SUCRA) probability values were applied to rank the examined treatments, and the SUCRA values of 100% and 0% were assigned to the best and worst treatments, respectively (Salanti et al., 2011; Riley et al., 2017). Furthermore, a clustering analysis was utilized to compare the effect of CHIs between two different outcomes. After that, publication bias were reflected by funnel plots (Chaimani et al., 2013; Salanti et al., 2014).

A total of 4,762 studies were identified from the search at first. After removing duplicates, 2,725 remained. By screening titles and abstracts, 2,224 studies were excluded because they were reviews, irrelevant studies, and animal experiments. Afterward, 501 relevant studies were reviewed for eligibility by full-text evaluations. Finally, 64 studies that met the inclusion criteria were included in our Bayesian NMA. Four hundred thirty-seven records were excluded for the following reasons: 1) observational studies (n = 33); 2) the use of non-intravenous injections or irrelevant drugs (n = 63); 3) the disease did not meet the diagnostic criteria or studies did not report the established outcomes (n = 327); 4) incomplete data (n = 6); and 5) duration of therapy was not satisfied (n = 8). The literature selection process is illustrated in Figure 1.

The Bayesian NMA was performed using 64 RCTs with a total of 5,904 adult patients and their sample sizes varying from 24 to 675 participants. All RCTs were conducted in China and published between 2011 and 2021. Six CHIs were incorporated, including XBJ (n = 46), TRQ (n = 10), RDN (n = 1), XYP (n = 2), SF (n = 3), and SM (n = 2). The control groups have been treated with WM such as anti-infectives, apophlegmatisant, mechanical ventilation, vasopressor, and nutritional support. The anti-infectives mainly include fluoroquinolones, β-lactam, carbapenems, and linezolid as monotherapy or combination therapy. On the basis of the control group, the intervention of the experimental group was one of the included CHIs. The duration of treatment ranged from 7 to 14 days. The details of the study characteristics are depicted in Table 1, and the compared connections among each intervention for each outcome are displayed in Figure 2.

We used the RoB 2 to conduct a quality evaluation. Three studies were assessed as “low risk” (Zhu and Liu, 2014; Kong, 2015; Song et al., 2019) and two studies were “high risk” for the randomization process because of their incorrect method of random sequence generation (Tian et al., 2019; Deng and Chen, 2021). All trials were rated to have low risk of bias for deviations from intended interventions, missing outcome data, and selection of the reported result. Nineteen studies were evaluated as “low risk” in measurement of the outcome (Han et al., 2012; Li, 2013; Lin et al., 2013; Chen et al., 2014; Gao et al., 2014; Zhu and Liu, 2014; Song and Chai, 2015; Zhang and Wang, 2015; Zhuang, 2016; Diao et al., 2017; Lei et al., 2017; Lv et al., 2017; Chen, 2019; Sheng et al., 2019; Song et al., 2019; Zhao et al., 2019; Yang, 2020; Yang et al., 2020; Yuan and Zhao, 2020), and the remaining were assessed as “some concerns.” In general, most studies were classified as “some concerns.” Further details of the risk of bias assessment are shown in Figures 3 and 4. In addition, although blinding was not implemented and applicable for subjects in most studies, the lack of blinding was unlikely to have influenced the assessment of primary outcome indicators.

The clinical effective rate was deemed as the primary outcome. A total of 46 RCTs (Qi et al., 2011; Wu and Ru, 2011; Sun et al., 2012; Wu, 2012; Gao et al., 2014; Yang et al., 2014; Zhang et al., 2014; Huang, 2015; Kong, 2015; Wu, 2015; Ma, 2016; Wu and Ren, 2016; Xi et al., 2016; Xie, 2016; Li, 2017; Niu et al., 2017; Pan et al., 2017; Wang et al., 2017; Xia and Xie, 2017; Xu, 2017; Zhou et al., 2017; Fan et al., 2018; Meng and Xie, 2018; Wang, 2018; Zhou, 2018; Chen et al., 2019; Liu and Qin, 2019; Shang and Zhang, 2019; Tian et al., 2019; Wang, 2019; Wang, 2019; Wang et al., 2019; Yang and Zhang, 2019; Yin, 2019; Cai, 2020; Ding and Lu, 2020; Sun et al., 2020; Wang et al., 2020; Wei and Huang, 2020; Xiao et al., 2020; Xin, 2020; Zhang et al., 2020; Zheng et al., 2020; Deng and Chen, 2021; Qiu, 2021; Zhang, 2021) of six CHIs reported the clinical effective rate. In the pairwise meta-analysis, there was no significant heterogeneity in the pooled analysis of all included studies (p = 1.000, I ² = 0.0%). The results of the heterogeneity are shown in Table 2. According to the network of comparisons in Table 2, there were 21 comparisons, and XBJ, TRQ, RDN, and SM combined with WM improved the clinical effective rate more significantly than WM alone (XBJ + WM: OR = 0.24, 95% CI: 0.19, 0.30; TRQ + WM: OR = 0.22, 95% CI: 0.12, 0.37; RDN + WM: OR = 0.29, 95% CI: 0.04, 0.94; and SM + WM: OR = 0.27, 95% CI: 0.08, 0.63). However, the results showed no significant differences between XYP and SF combined with WM than WM alone (XYP + WM: OR = 0.30, 95% CI: 0.03, 1.09 and SF + WM: OR = 0.47, 95% CI: 0.09, 1.36).

In Table 8 and Figure 5, the SUCRA values suggested that TRQ + WM was the optimal treatment, XYP + WM was the second, and RDN + WM was the third.

As the other dominating outcome, WBC was estimated in 26 RCTs (Qi et al., 2011; Sun et al., 2012; Li, 2013; Gao et al., 2014; Yang et al., 2014; Zhang et al., 2014; Huang, 2015; Kong, 2015; Wu and Ren, 2016; Xi et al., 2016; Xie, 2016; Zhuang, 2016; Diao et al., 2017; Lei et al., 2017; Lv et al., 2017; Pan et al., 2017; Xu, 2017; Fan et al., 2018; Chen, 2019; Chen et al., 2019; Liu and Qin, 2019; Shang and Zhang, 2019; Yang and Zhang, 2019; Cai, 2020; Sun et al., 2020; Zhang, 2021). The heterogeneity results of the pairwise meta-analysis are shown in Table 3. According to Table 3, the CHIs like XBJ and TRQ combined with WM (XBJ + WM: SMD = −0.87, 95% CI: 1.37, −0.36 and TRQ + WM: SMD = −1.53, 95% CI: 2.20, −0.87) had better clinical effective rate than WM alone, and the differences among the above interventions were statistically significant. However, the rest of the CHIs including RDN, XYP, SF, and SM combined with WM did not perform more outstanding than WM alone. In addition, in terms of Table 3, there were no significant differences between each comparison of different types of CHIs.

Based on the ranking analysis, XYP + WM attained the first rank. TRQ + WM was the second, and XBJ + WM was the third (Table 8 and Figure 5).

CRP was tested in 30 RCTs (Qi et al., 2011; Sun et al., 2012; Gao et al., 2014; Zhang et al., 2014; Kong, 2015; Zhang and Wang, 2015; Xi et al., 2016; Xie, 2016; Zhuang, 2016; Diao et al., 2017; Lei et al., 2017; Lv et al., 2017; Pan et al., 2017; Xia and Xie, 2017; Xu, 2017; Fan et al., 2018; Chen, 2019; Chen et al., 2019; Liu and Qin, 2019; Shang and Zhang, 2019; Song et al., 2019; Tian et al., 2019; Yang and Zhang, 2019; Cai, 2020; Ding and Lu, 2020; Sun et al., 2020; Xiao et al., 2020; Yang, 2020; Yuan and Zhao, 2020; Zhang, 2021) involved seven interventions. The heterogeneity results of the pairwise meta-analysis are shown in Table 4. Four of them were noticeably better than WM treatment alone for decreasing the level of CRP, as XBJ + WM (SMD = −1.62, 95% CI: 2.06, −1.18), TRQ + WM (SMD = −2.15, 95% CI: 2.93, −1.36), RDN + WM (SMD = −3.72, 95% CI: 5.19, −1.35), and SF + WM (SMD = −1.52, 95% CI: 2.87, −0.17) were remarkable among them compared with WM alone. What is more, based on WM, TRQ and RDN had a more excellent performance in decreasing CRP than XBJ (XBJ + WM vs TRQ + WM: SMD = −0.53, 95% CI: 1.43, −0.53 and XBJ + WM vs RDN + WM: SMD = −1.66, 95% CI: 3.63, −1.66), and the results between the rest of the comparisons of different CHIs showed no significant differences.

The SUCRA mentioned above was also affirmed, and RDN + WM was the best choice, followed by TRQ + WM and XBJ + WM (Table 8 and Figure 5).

The potency of decreasing the level of PCT was assessed. The heterogeneity results of the pairwise meta-analysis are shown in Table 5. Four interventions with 22 RCTs (Chen et al., 2014; Gao et al., 2014; Kong, 2015; Wu and Ren, 2016; Pan et al., 2017; Xia and Xie, 2017; Fan et al., 2018; Chen, 2019; Liu and Qin, 2019; Sheng et al., 2019; Song et al., 2019; Tian et al., 2019; Wang, 2019; Yang and Zhang, 2019; Zhao et al., 2019; Cai, 2020; Xiao et al., 2020; Xin, 2020; Yang, 2020; Yang et al., 2020; Yuan and Zhao, 2020; Zhang, 2021) had data in contrast with WM, shown in Table 5. The results revealed that all CHIs involved combined with WM were advantageous in decreasing PCT compared to WM alone (XBJ + WM: SMD = −2.19, 95% CI: 2.67, −1.70; TRQ + WM: SMD = −2.12, 95% CI: 3.27, −0.96; SF + WM: SMD = −2.84, 95% CI: 4.83, −0.86; and SM + WM: SMD = −1.69, 95% CI: 3.11, −0.27), and the differences among the abovementioned combinations were statistically significant.

Treatment ranking based on SUCRA values, from largest to smallest, were as follows: SF + WM, XBJ + WM, TRQ + WM, SM + WM, and WM. The details are depicted in Table 8 and Figure 5.

Twenty-two RCTs (Han et al., 2012; Li, 2013; Lin et al., 2013; Chen et al., 2014; Gao et al., 2014; Yang et al., 2014; Zhang et al., 2014; Zhu and Liu, 2014; Song and Chai, 2015; Ma, 2016; Lei et al., 2017; Lv et al., 2017; Pan et al., 2017; Wang, 2018; Zhou, 2018; Chen et al., 2019; Sheng et al., 2019; Wang et al., 2019; Zhao et al., 2019; Wei and Huang, 2020; Xiao et al., 2020; Deng and Chen, 2021) with five treatments reported the ICU length of stay. The heterogeneity results of the pairwise meta-analysis are shown in Table 6. As shown in Table 6, XBJ + WM (SMD = −1.02, 95% CI: 1.33, −0.72), TRQ + WM (SMD = −1.24, 95% CI: 2.16, −0.31), and XYP + WM (SMD = −1.78, 95% CI: 3.07, −0.49) were more effective than WM alone. However, the results showed no significant difference in most cases.

Based on the ranking analysis, XYP + WM attained the first rank. TRQ + WM was the next, and XBJ + WM was the third (Table 8 and Figure 5).

In terms of the time of mechanical ventilation, five treatments with 17 RCTs (Wu and Ru, 2011; Han et al., 2012; Li, 2013; Chen et al., 2014; Gao et al., 2014; Yang et al., 2014; Zhu and Liu, 2014; Song and Chai, 2015; Ma, 2016; Lv et al., 2017; Pan et al., 2017; Wang, 2018; Liu and Qin, 2019; Wang et al., 2019; Wei and Huang, 2020; Yang, 2020; Deng and Chen, 2021) were compared with WM. The heterogeneity results of the pairwise meta-analysis are shown in Table 7. As seen in Table 7, three CHIs combined with WM (XBJ + WM: SMD = −1.72, 95% CI: 2.26, −1.17; TRQ + WM: SMD = −2.42, 95% CI: 3.56, −1.28; and XTP + WM: SMD = −3.41, 95% CI: 5.40, −1.43) had excellent performance in decreasing the time of mechanical ventilation compared to WM alone, and the results were statistically significant.

Results of ranking analysis manifested that XYP + WM was efficacious in decreasing the time of mechanical ventilation. Other beneficial treatments were TRQ + WM and XBJ + WM (Table 8 and Figure 5).

A cluster analysis was performed on the primary outcome and secondary outcomes for the seven treatments. As shown in Figure 6, in terms of the clinical effective rate and the level of WBC, the clinical effective rate and ICU length of stay, and the clinical effective rate and the time of mechanical ventilation, TRQ and XYP combined with WM were similarly superior. According to the clinical effective rate and the level of CRP, TRQ and RDN combined with WM were more beneficial. About the clinical effective rate and the level of PCT, TRQ and XBJ combined with WM were preferred.

Among 64 RCTs, a total of 18 RCTs reported the ADRs/ADEs of interventions. There were 14 RCTs (Gao et al., 2014; Zhang et al., 2014; Ma, 2016; Wu and Ren, 2016; Pan et al., 2017; Meng and Xie, 2018; Zhou, 2018; Shang and Zhang, 2019; Song et al., 2019; Wang, 2019; Wang et al., 2019; Xin, 2020; Deng and Chen, 2021; Qiu, 2021) involving 1,095 participants with XBJ group that reported ADRs, including headache and dizziness (32 cases in 10 RCTs), diarrhea (10 cases in 7 RCTs), nausea and vomiting (4 cases in 3 RCTs), chest discomfort or dyspnea (4 cases in 4 RCTs), itchy skin or rash (9 cases in 6 RCTs), and myelosuppression (4 cases in 4 RCTs). Three RCTs (Huang, 2015; Xi et al., 2016; Zhang, 2021) reported ADRs/ADEs of TRQ, one of the RCTs reported two cases of nausea and vomiting and one case of rash. Another two RCTs did not report ADRs/ADEs. In addition, only one RCT (Sun et al., 2012) mentioned the ADRs of RDN and one case reported thirst. The rest of the included studies did not provide information on any ADRs/ADEs. All of the symptoms were alleviated after corresponding treatment and did not influence the RCTs.

To assess the robustness and reliability of the primary outcome results, a sensitivity analysis was conducted based on the results of quality assessments of studies. Two studies were excluded for their high risk of bias (Tian et al., 2019; Deng and Chen, 2021), and the remaining 44 studies conducted a network meta-analysis again. Results did not show relevant deviations compared with the original network meta-analysis. The Bayesian ranking results from high to low for clinical effective rate were TRQ (66.3%), XYP (62.9%), RDN (61.4%), XBJ (59.1%), SF (58.7%), SM (39.5%), and WM (2.0%), respectively. The sensitivity analysis showed that the results of clinical effective rate were robust and reliable.

Comparison-adjusted funnel plots for different outcomes are displayed in Figure 7. There were three funnel plots of outcomes that were generally symmetrical visually, including the clinical effective rate, ICU length of stay, and the time of mechanical ventilation. Therefore, they had no publication bias. Among the remaining outcomes, the funnel plots were not symmetrically visual, which revealed that there were small sample size and publication bias.

This study also has some limitations. First, all RCTs were carried out in China, and the data of clinical studies in other languages were lacking. Second, the qualities of included RCTs in this study were not high. About 58% of RCTs described the method of generating random sequences. Only two RCTs described information of allocation concealment and one RCT accurately set blinding. Third, there was a lack of large sample directly comparing the two injections. The difference among the sample sizes of different injections would also reduce the strength of the evidence for the results. It is necessary to conduct a subgroup analysis based on background diseases, different types of pneumonia, duration of treatment, and Western medicine treatment measure. However, except XBJ injection, the number of included studies for other CHIs was not enough to conduct a subgroup analysis. In addition, it is difficult to conduct a subgroup analysis as the currently included studies could not be accurately categorized based on the above variables. We hope that more RCTs could include only one background disease, one type of pneumonia (CAP, HAP, or VAP), one kind of antibiotics, and a fixed treatment duration in developing the inclusion criteria.