Comparative efficacy of Chinese herbal injections for treating acute cerebral infarction: a network meta-analysis of randomized controlled trials

Background Chinese herbal injections (CHIs) are prepared by extracting and purifying effective substances from herbs (or decoction pieces) using modern scientific techniques and methods. CHIs combined with aspirin + anticoagulants + dehydrant + neuroprotectant (AADN) are believed to be effective for the treatment of acute cerebral infarction (ACI). However, no randomized controlled trial (RCT) has been performed to directly compare the efficacies of different regimens of CHIs. Therefore, we performed a systematic review and network meta-analysis (NMA) to compare the efficacies of different regimens of CHIs for ACI. Methods We conducted an overall and systematic retrieval from literature databases of RCTs focused on the use of CHIs to treat ACI up to June 2016. We used the Cochrane Handbook version 5.1.0 and CONSORT statement to assess the risk of bias. The data were analyzed using STATA 13.0 and WinBUGS 1.4.3 software. Results Overall, 64 studies with 6225 participants involving 15 CHIs were included in the NMA. In terms of the markedly effective rate, Danhong (DH) + AADN had the highest likelihood of being the best treatment. In terms of the improvement of neurological impairment, Shuxuening (SXN) + AADN had the highest likelihood of being the best treatment. Considering two outcomes, injections of SXN, Yinxingdamo (YXDM), DH, Shuxuetong (SXT), HongHuaHuangSeSu (HHHSS), DengZhanXiXin (DZXX) and Shenxiong glucose (SX) plus AADN were the optimum treatment regimens for ACI, especially SXN + AADN and YXDM + AADN. Conclusions Based on the NMA, SXN, YXDM, DH, SXT, HHHSS, DZXX and SX plus AADN showed the highest probability of being the best treatment regimens. Due to the limitations of the present study, our findings should be verified by well-designed RCTs. Electronic supplementary material The online version of this article (10.1186/s12906-018-2178-9) contains supplementary material, which is available to authorized users.


Background
Acute cerebral infarction (ACI) is one of the most common cerebral vascular diseases, also referred to as ischemic stroke, which is result from ischemia, hypoxia and cerebral blood circulation [1][2][3]. ACI has the characteristics of high disability and recurrence [4][5][6]. Besides, it is a major disease leading to serious damage of central nervous system or death [4,7]. It was estimated that ACI cause 6.2 million mortalities annually worldwide [8]. In traditional Chinese medicine (TCM) theories, ACI refers to "apoplexy", majorly due to blood stasis syndrome [3]. Therefore, promoting blood flow is of primary importance. It has been proven that TCM is an effective complementary intervention for stroke, especially in the treatment of ischemic stroke [9][10][11][12][13].
Currently, therapies for ACI include thrombolytics, antithrombotics, anticoagulants, and neuroprotectants [14]; this was the Grade-I recommendation in the guidance of diagnosis and treatment of acute ischemic stroke in China 2010, which fully considered the national conditions and clinical experience. Among them, thrombolytic has a short therapeutic time window. Thus, many patients easily miss the effective time window of thrombolysis. ACI patients who are not eligible for thrombolysis therapy should be given oral aspirin, which was approved by the US Food and Drug Administration (FDA), at a dose ranging from 150 to 300 mg/d as soon as possible (level of evidence: A) [15,16]. And patients with brain edema could use a mannitol intravenous drip. Although more high-quality clinical trials are needed to further demonstrate the efficacy and safety of neuroprotective agents, a number of RCTs have suggested that edaravone and cerebroprotein hydrolysate improve functional outcomes and the safety of patients with ACI [17][18][19][20]. Additionally, the therapeutic principle of invigorating blood circulation for removing blood stasis of TCM holds a significant position for ACI. Chinese herbal injections (CHIs) have the characteristics of rapid efficacy and high bioavailability [21][22][23].Presently, 37 injections are often used in the treatment of cerebral infarction, such as Xueshuantong injection, Shuxuening injection, Mailuoning injection, and Danshenchuanxiongqin injection [24][25][26][27]. Clinical data [28][29][30][31][32] from systematic reviews of RCTs have demonstrated the beneficial effects of CHIs for inhibiting platelet aggregation, improving blood microcirculation and nerve function, enhancing the tolerance of ischemic tissue to hypoxia, and protecting against ischemic reperfusion injury.
Hence, this study systematically evaluated the clinical effectiveness of CHIs combined with an aspirin + anticoagulants + dehydrant + neuroprotectant (AADN) regimen in ACI patients that conformed to the standardized treatment of ischemic cerebrovascular disease: integration, individualization and sequencing. However, there is no direct head-to-head evidence revealing the best CHIs for ACI treatment. Determining the superiority of a treatment based on a pairwise comparison meta-analysis is difficult. A network meta-analysis (NMA), which is an extension of a traditional meta-analysis, synthesizes the available evidence to enable simultaneous comparisons of different treatment options that lack direct head-tohead evaluations [33][34][35]. Therefore, the present study performed a NMA to compare the clinical efficacy of 37 CHIs combined with the AADN regimen to reveal the best CHIs for ACI, aiming to provide more sights for selection of ACI.

Eligibility criteria
Studies meeting the following criteria were included: (1) Clinical randomized controlled trials (RCTs) using CHIs + AADN to treat ACI regardless of blinding. (2) Cerebrovascular disease was diagnosed according to the standards revised by the Fourth National Conference on Cerebrovascular Disease by the Chinese Medical Association in 1995 [36]. The acute phase of ACI generally refers to 2 weeks after the onset of disease. Thus, this NMA enrolled patients with the course of disease within 2 weeks. No cerebral hemorrhage was detected using cranial computed tomography (CT) or magnetic resonance imaging (MRI). There were no limits on age, gender, race or disease severity. (3) Eligible comparisons were CHIs + AADN regimens versus the AADN regimen alone and CHIs + AADN regimens versus other CHIs + AADN regimens. There was no limitation on the dosages or treatment courses. (4) Outcome measures included the markedly effective rate, improvement of neurological impairment, activities of daily living function, and death from all causes within the treatment and during the entire follow-up period. The following formula was used: the markedly effective rate (%) = (number of recovered patients + number of patients with significant progress) / total number × 100%. The efficacy criteria were predominantly based on the reduction of the neurological deficit score and could be divided into four grades: recovery, significant progress, progress, and no change or deterioration. Recovery, significant progress, progress, and no change or deterioration were determined when the neurological deficit score decreased from 91% to 100%, between 46% and 90%, between 18% and 45%, and < 17%, respectively. The improvement of neurological impairment is expressed as the mean ± standard deviation.
The following studies were excluded: (1) studies that did not refer to the acute phase; (2) studies that did not meet the curative effect valuation standard; (3) studies involving patients who had a severe cognitive disorder, hemorrhagic tendency, or serious complications, such as atrial fibrillation, severe heart failure, severe liver and kidney diseases, undergoing surgery, acupuncture or other physical therapy; (4) data that were incorrect, incomplete or unavailable; (5) reviews or meta-analyses, experimental research, retrospective studies, case reports, and conference abstracts.

Data sources and search strategy
A systematic literature search was performed using the following databases from inception to June 2016: PubMed, Cochrane Library, Embase, China National Knowledge Infrastructure Database (CNKI), Wanfang Database, and Chinese Biomedical Literature Database (CBM). The medical subject headings (MeSH) and free text words were used. No language or other restrictions were imposed. Furthermore, we hand searched the reference lists of all retrieved studies. The specific Chinese and English search terms for each CHIs are shown in Additional file 1: Table S1 and the detailed search strategy is shown in Additional file 2.

Literature selection and data extraction
Two reviewers independently read the titles and abstracts of the literature to exclude literature that was obviously not relevant as well as reviews and pharmacological experiments. We retrieved the full text of the articles to determine whether they were eligible.
The data of interest from each included RCT were collected using a standard data abstraction form created in Microsoft Excel 2013 (Microsoft Corp, Redmond, WA, USA). The main components of the extracted data were as follows: (1) General information: author names and publication data; (2) Patient information: median age, number of patients, gender, and acute phase; (3) Intervention: names, dosages, and treatment; (4) Outcomes: the markedly effective rate, improvement of neurological deficit score, adverse drug reactions/adverse drug events (ADRs/ADEs), activities of daily living function, and death from all causes within the treatment and during the entire follow-up period.

Quality assessment
The methodological quality of each included study was evaluated using the Cochrane Risk of Bias tool [37] and the CONSORT statement [38]. The items included randomization, blinding, dropout, eligibility criteria for participants, adverse events, and statistical methods. The judgments for each entry involved were divided into 3 grades: "high", "unclear", and "low". A quality assessment was performed by two independent reviews, and disagreements were resolved by consensus.

Statistical analysis
We performed a pairwise meta-analysis using STATA 13.0 software (Stata Corporation, College Station, TX, USA). The pooled odds ratios (ORs) were calculated for dichotomous data and standardized mean differences (SMDs) were calculated for continuous variables, both with 95% confidence interval (95% CI). The Chi-squared test was used to evaluate the heterogeneity between studies, and I 2 was used to show the extent of heterogeneity. When P was ≥0.1 and I 2 was ≤50%, no statistical heterogeneity was suggested and the Mantel-Haenszel fixed-effects model was used for the metaanalysis. When P < 0.1 and I 2 was > 50%, we explored the sources of heterogeneity using a subgroup analysis and meta-regression. When there was no clinical heterogeneity, the Mantel-Haenszel random-effects model was used to perform the meta-analysis [37].
A Bayesian NMA was conducted using WinBUGS 1.4.3 software (MRC Biostatistics Unit, Cambridge, UK). The random-effects model with vague priors for multiarm trials was used [39]. The model parameters were estimated using a Markov chain Monte Carlo method called Gibbs sampling. Convergence was found to be adequate after running 1000 samples. These samples were discarded as "burn-in," and posterior summaries were based on 100,000 subsequent simulations. The results are reported as the OR and SMD with 95% CI. To evaluate the inconsistency between direct and indirect effect estimates for the same comparison, we evaluated each closed loop in the network. In a closed loop, we employed the inconsistency factor (IF) to evaluate heterogeneity among the included studies. If the 95% CIs of the IF values were truncated at zero, it indicated that the 2 sources were in agreement [39]. To rank the treatments, we used the surface under the cumulative ranking probabilities (SUCRA); a SUCRA value of 100% is assigned to the best treatment and 0% for the worst treatment [39]. A comparisonadjusted funnel plot was used to assess the presence of small-study effect [40]. Egger's test was used to assess the symmetry of the funnel plot [41].
To account for both the markedly effective rate and neurological deficits, we used multivariate methods to determine the dependency between outcomes. Clustering methods and 2-dimensional plots were used to produce clusters of treatments [42]. Using the clusterank command, clustered ranking plots can be obtained using the STATA program. The markedly effective rate and neurological deficits became the data variable containing the SUCRA scores for all treatments in this network. The different colors correspond to the estimated clusters and were utilized for grouping the treatments according to their similarity for both outcomes.

Results
Literature search and characteristics of the included studies Figure 1 shows the PRISMA flow diagram. A total of 13,764 articles were identified from electronic databases. After the exclusion of duplications, reviews, and obviously irrelevant studies by reading titles and abstracts, 3493 papers were downloaded for additional review. A total of 3429 RCTs were excluded for the following reasons: non-RCTs, non-acute phase, not meeting intervention and the curative effect valuation standard, incorrect data, no treatment time, and no outcomes of interest. Hence, 64 studies and 15 CHIs were included in the NMA. All studies were published in Chinese from 2006 to 2015.

Characteristics of the included studies
The 64 RCTs [28, included 6225 participants, with sample sizes varying from 26 to 300 participants. All RCTs were conducted among Chinese populations in China. All participants were evaluated using cranial CT or NMRI. The rang of participants was approximately 35 to 87 years. There were more male patients  Table 1. The acute phase was no more than 30 days, with 62.5% of the cases having an acute phase of less than 72 h. The duration of treatment for both the experimental and control groups was no more than 30 days. Figure 2 shows a network graph comparing fifteen CHIs. There were 120 pairwise comparisons including 40 direct comparisons. Table 2 provides a summary of the included studies. Additional file 3 showed the more details of included CHIs.

Quality of the included studies
We used the Cochrane Handbook version 5.1.0 and CONSORT statement to conduct a quality evaluation of the included studies. All studies mentioned the use of random distribution, while ten studies [44,64,68,73,75,82,[95][96][97]101] described a satisfactory method of randomization including random number tables or the envelope method. Two studies [60,85] reported information about blinding. All studies provided information on patients who were lost to follow-up or dropped out. All studies reported the eligibility criteria for participants and statistical methods. Approximately 74.6% of the studies provided  Figure S1.

Pairwise meta-analysis
Pairwise meta-analysis of the markedly effective rate Fifty-nine RCTs reported markedly effective rates; in these RCTs, 5864 patients were involved and 34 regimens were included. Table 3 summarizes the results of the pairwise meta-analysis regarding the markedly effective rates. There was no significant heterogeneity in the pooled analysis of all included studies (P > 0.1; I 2 < 50%); the results of the heterogeneity test are shown in Table 2. The direct comparison showed that DH and SXN were more beneficial in improving the markedly effective rate than AADN (AADN versus DH: OR = 0.61, 95% CI = 0.45-0.84; versus SXN: OR = 0.57, 95% CI = 0.44-0.73). SX and XST2 were more beneficial than FFDS (FFDS versus SX: OR = 0. 61, 95% CI = 0.38-0.98; versus XST2: OR = 0.54, 95% CI = 0.34-0.86). Other treatment comparisons failed to reach statistical significance (the 95% CI included 1). The detailed results are shown in Fig. 3.

Pairwise meta-analysis of the improvement of neurological impairment
Forty-one RCTs reported an improvement of neurological impairment; in these RCTs, 3828 patients were involved and 29 regimens were included. When P was ≥0.1 and I 2 was ≤50%, the Mantel-Haenszel fixed-effects model was used for the meta-analysis and vice versa. The results of the heterogeneity test are shown in Table 4. Table 4  Other treatment comparisons failed to reach statistical significance (the 95% CI included 0). The detailed results are shown in Fig. 4.

Pairwise meta-analysis of the death
Most studies did not report any deaths during the treatment period or during follow up after the treatment

Results of the Bayesian network meta-analysis
In the original analysis, most studies did not mention the activities of daily living function or death from all causes within the treatment period or during the entire follow-up period. Therefore, the present study did not compare death or the activities of daily living function among different treatments; we only performed a NMA to compare the markedly effective rate and the improvement of neurological impairment among the different regimens of CHIs for ACI.
Bayesian network meta-analysis of the markedly effective rate According to the network of comparisons (Table 3) Figure S2 shows the inconsistency plot used to identify heterogeneity among studies in the closed loop of this NMA. Eleven triangular loops and 25 quadratic loops were present in the NMA; 83% of IF values with 95% CIs were truncated at zero, suggesting no significant inconsistency.

Assessment of publication bias
The comparison-adjusted funnel plots (Additional file 6: Figure S3) for the markedly effective rate were asymmetric near the zero line. The result from Egger's test was P = 0.047. Therefore, this study may have a small sample effect and publication bias.

Bayesian network meta-analysis of the improvement of neurological impairment
According to the network of comparisons (Table 4)  As shown in Additional file 7: Figure S4, 9 triangular loops and seventeen quadratic loops were present in the NMA; 81% of IF values with 95% CIs were truncated at zero, suggesting no significant inconsistency.

Rank probability
The cumulative probability analysis (SUCRA results) showed that SXN + AADN had the highest likelihood of improving the neurological impairment scores (SUCRA-84.

Simultaneous ranking of the interventions for two outcomes
Clustered ranking plots of the network for the markedly effective rate and the improvement of neurological impairment score are shown in Fig. 7. Each color represents a group of treatments that belong to the same cluster. Treatments lying in the upper right corner are more effective than the other treatments. The upper right corner in Fig. 7 shows that SXN, YXDM, DH, SXT, HHHSS, DZXX and SX produce significantly better outcomes in ACI patients.

Discussion
Considerable evidence exists regarding the clinical effectiveness of CHIs in ACI patients. Some CHIs have been  widely used to strengthen clinical effectiveness, reduce neurological impairments and improve the patient's quality of life. However, the majority of these findings have not been analyzed in head-to-head comparisons. Clinicians must decide among several therapeutic options for ACI patients. To address the absence of comparative data, we conducted an NMA to comprehensively estimate the effectiveness of different CHIs combined with AADN for ACI.
This NMA consisted of 64 RCTs that included 6225 participants; fifteen CHIs were identified in the treatment of ACI, including injections of SXN, SXT, SX, MLN, HHHSS, FFDS, DZHS, DZXX, DSCXQ, DS, DH, YXDM, LI, XST1, and XST2. In terms of the improvement of the markedly effective rate, DH had the highest likelihood of being the best treatment in terms of the markedly effective rate. On account of improvement of neurological impairment, SXN had the highest probability of being the best treatment.
The clustered ranking according to two outcomes revealed that the markedly effective rate and the improvement of the neurological impairment cluster were best for SXN, YXDM, DH, SXT, HHHSS, DZXX and SX. SXN and YXDM are shown at the top right corner. Previous meta-analyses [106][107][108] found that SXN and YXDM as adjuvant treatments for ACI were beneficial compared to AADN alone. SXN and YXDM are Ginkgo biloba extracts (GBEs), both of which are extracted from Ginkgo biloba leaves. Ginkgo biloba leaves, the TCM for activating blood circulation, mainly contain ginkgo flavonoids, ginkgolides, and bilobalide and have been used as a therapeutic agent for managing cerebrovascular and neurological disorders [109,110]. GBE exhibits a wide variety of biological activities, including anti-inflammation and antioxidant effects [111,112]. ACI is the process whereby artery stenosis or blockage causes brain tissue hypoxic ischemia, resulting in brain dysfunction [6]. There is considerable evidence suggesting the active repair and recovery mechanisms following stroke, and neurogenesis is one of them [113]. GBE not only has antioxidant, anti-atherogenesis and angiogenic properties but can also strengthen repair and regeneration mechanisms and prevent cell death, protect the brain from further damage and improve neurological deficits following stroke [113][114][115]. The neuroprotective mechanism has been attributed to the heme oxygenase 1 (HO1)/Wnt canonical pathway as well as neuritogenic and angiogenic effects [113,116]. HO1, a key component of the EGb 761 neuroprotective signaling pathway, activates the signaling pathway mechanisms of angiogenesis, cell survival and neuroplasticity, and neurogenesis [113]. Thus, GBE could enhance the post-stroke regeneration process to improve treatments for stroke recovery. Further research is desirable to shed more light on the mechanism underlying the effects of GBE on ACI.
A NMA was used to compare the effectiveness of different CHIs to identify the best CHIs for ACI. This study is the first indirect comparison using a network approach to compare the effectiveness of CHIs, which is valuable for clinicians selecting CHIs for ACI treatment. However, some limitations existed in this NMA.
First, all trials reported random distribution, while ten studies described the randomization methods including random number tables or the envelope method. Information about allocation concealment and blinding was not clear in the majority of trials and may have therefore affected the reliability of the results. Second, the systematic review included only published studies in the database, with no relevant gray literature, which likely caused a selection bias in the literature. Third, the study aimed to use a NMA to evaluate the clinical effectiveness of 37 CHIs combined with an AADN regimen; however, only 15 CHIs were included in the NMA. Thus, more rigorously designed RCTs focused on the 22 additional CHIs are needed to confirm the effectiveness of CHIs for ACI. Fourth, due to the original research limitation, we failed to evaluate the long-term effect of CHIs. Additionally, with the limited data extracted from the original research, we failed to evaluate the ability of CHIs to improve the activities of daily living function and reduce mortality. Fifth, our results might have limited generalizability because all of the included RCTs were conducted in China among Chinese populations. Therefore, it is uncertain whether the effect may change when CHIs are used in populations of other ethnicities and in different geographical locations. In addition, though acute phases were limited, the severities of patient were various. This point may influence the results. Sixth, a NMA compares multiple treatments by incorporating direct and indirect evidence into a general statistical framework. One issue with the validity of a NMA is the inconsistency between direct and indirect evidence. Hence, to improve the reliability of our results, we used a random-effects model within a Bayesian framework. Although, headto-head trials provide the highest level of evidence when comparing interventions, the quantity of data for some CHI direct trials was small, such as DH versus FFDS, SXN, and XSTT. Large RCTs are needed to specifically compare CHIs with one another.

Conclusions
In summary, our evidence suggested that DH injection plus AADN was the optimum treatment regimen for patients with ACI to improve the markedly effective rate. SXN injection plus AADN was the optimum treatment