We searched PubMed, Embase, Cochrane Central Register of Controlled Trials, Web of Science, China National Knowledge Infrastructure Database, Chinese Scientific Journal Database, Wanfang Database, and Chinese Biomedical Database to get relevant articles with no language restrictions published before May 18, 2021, using as the main search term (“Coronary Diseases”) or (“Heart failure”) and (“Injection”) within the restriction limit of (“randomized controlled trial”). A subset of Chinese and English journals that might publish studies relevant for our subject were also searched manually. The detailed search strategy is described in Supplementary Table S2.

Two review authors (PW and KY) independently reviewed the titles and abstracts of trials retrieved by the search for potential eligibility. Then, we acquired the full texts of trials considered potentially eligible for inclusion in the review. We sought further information from the authors of the trial, which was not sufficient to determine eligibility. Any differences were resolved through consensus and arbitration by a panel of adjudicators (PW, DL, YT, WX, and XL).

We included published RCTs that met the following criteria:

• Participants: all the enrolled participants were required to accord with the current or past definitions of CHD and HF (Hu et al., 2009; Zhu and Han, 2014; Xian et al., 2016; Xu and Cao, 2019; Pan et al., 2005; Yuan and Du, 2012; Feng, 2013; Wang et al., 2019; Li et al., 2016a; Shi et al., 2016; Shen et al., 2017; Ji, 2019; Zhou et al., 2005; Wu and Duan, 2009; Yang, 2009; Zhao et al., 2011a; Cao, 2012; Dong, 2012; Shen, 2012; Wu and Huang, 2012; Yang and Li, 2012; Zhou et al., 2013; Luo et al., 2015; Xiu and Chen, 2015; Li et al., 2016b; He, 2016; Mao, 2016; Wang et al., 2016; Wu, 2016; Li et al., 2018; Wang and Jang, 2018; Li, 2019a; Zhou and Luo, 2020; Lu, 2005; Zhu et al., 2008; Xing et al., 2009; Wang, 2012; Zhao et al., 2012; Wu, 2014; Yang et al., 2014; Zhang, 2015a; Teng, 2016; Xu, 2016; Tian et al., 2017; Zhan et al., 2017; Han, 2018; Ni et al., 2020; Zhang, 2020; Huang et al., 1999; Zhao et al., 2011b; Wang et al., 2011; Zhang, 2015b; Wei and Lu, 2020; Ren, 2021; Guo et al., 2012). Trials without a description of the detailed diagnostic criteria but which reported patients with definite CHD-HF were also included (Zhou et al., 2002; Kuang, 2004; Xin and Shan, 2012; Wu et al., 2017; Gong et al., 2018; Li, 2019b).

Studies not meeting all these inclusion criteria were excluded. In addition, the following exclusion criteria were applied:

• Interventions in the control group included other traditional treatments, such as other TCMI, acupuncture, or Chinese herbal medicine.

Data extraction and quality assessment were independently performed by two investigators (PW and HW). Data on trial details are as follows: (1) Basic information of the eligibility, including the content of study ID, first author, nationality, publication year, and study design; (2) Basic characteristics of included patients: sample size, sex composition, average age, course of treatment, and population distribution with the NYHA class; (3) Details of interventions; (4) Details of outcomes; (5) Information of quality assessment of RCTs. Two investigators (PW and KY) independently assessed risk of bias of individual studies. Discrepancies were resolved through consensus and arbitration by a panel of adjudicators (PW, DL, YT, WX, and XL). We also made attempts to contact the study authors by means of email, phone, or fax to obtain missing demographic information, such as the sample size, sex distribution, age, etc. When studies had multiple publications, we sorted all reports of the same study, so that each study, not each report, was the unit of interest in the review, and these studies were given a single study ID.

We assessed risk of bias of included RCTs using the Cochrane Risk of Bias Tool (Zhao et al., 2011b; Wang et al., 2011; Zhang, 2015b) based on the following items: random sequence generation, allocation concealment, blinding of participants and personnel, blinding of outcome assessment, incomplete outcome data, selective reporting, and other bias. Each item was scored as low, unclear, or high risk of bias. Any disagreements were resolved by a third researcher.

We synthesized all direct and indirect evidence to compare different treatments in terms of efficacy and safety, reported as odds ratios for binary outcomes (total effectiveness rate and adverse events) along with the corresponding 95% confidence intervals (CIs). Using Stata (version 16.0), we generated network diagrams for different outcomes to illustrate geometries, to clarify which treatments were directly or indirectly compared in the included studies (Chaimani et al., 2013). We analyzed frequency and random effects and conducted pairwise meta-analysis for head-to-head comparisons based on two or more trials. We assessed heterogeneity between the studies using Q test and the I² statistic within a visual forest plot. A p value less than 0.05 was regarded statistically significant. Heterogeneity was considered low, moderate, or high for estimated I² values under 25%, between 25% and 50%, and over 50%, respectively (Higgins et al., 2003).

Network meta-analyses were performed in a Bayesian framework using a Markov Chain Monte Carlo simulation technique using OpenBUGS (version 3.2.3). For all outcomes (total effectiveness rate, LVEF, NT-proBNP, BNP, and 6MWT), 150,000 sample iterations were generated with 100,000 burn-ins and a thinning interval of 1. We evaluated convergence of iterations by visual inspection of the three chains to establish homogenous parameter estimates in accordance with the Brooks-Gelman-Rubin diagnostic (Supplementary Figure S1) (Zhao et al., 2011b; Wang et al., 2011; Zhang, 2015b). Within the Bayesian framework, the network meta-analysis estimates the overall ranking of treatments by calculating the surface under the cumulative ranking curve for each; it is equal to 1 when the treatment is definitely best and 0 when the treatment is definitely the worst (Salanti et al., 2011). To assess the robustness and reliability of the results, we performed sensitivity analysis. We restricted the case number ≥100 to observe the effect of various treatments in patients with HF.

Our search strategy initially identified 1,284 records. After removal of duplicates, 981 remained for screening based on their titles and abstracts, of which 851 were excluded as irrelevant. We reviewed 130 full-text articles or, if these were not available, abstract publications or trial registry entries. Finally, we identified 61 RCTs (Hu et al., 2009; Zhu and Han, 2014; Xian et al., 2016; Xu and Cao, 2019; Hu et al., 2009; Zhu and Han, 2014; Xian et al., 2016; Xu and Cao, 2019; Hu et al., 2009; Zhu and Han, 2014; Xian et al., 2016; Xu and Cao, 2019) for inclusion, all of which had published. A study flow diagram is presented in Figure 1.

Among 61 RCTs (Hu et al., 2009; Zhu and Han, 2014; Xian et al., 2016; Xu and Cao, 2019; Hu et al., 2009; Zhu and Han, 2014; Xian et al., 2016; Xu and Cao, 2019; Hu et al., 2009; Zhu and Han, 2014; Xian et al., 2016; Xu and Cao, 2019) for inclusion, a total of 5,567 patients enrolled who received one of the 15 different treatments TCMIs, as listed in the Methods, in combination with conventional treatments (diuretics, ACEIs, ARBs, β-receptor blocker, aldosterone-receptor blocker, digoxin, or vasodilator substance). The main characteristics of all included studies are depicted in Table 1. The detailed information of TCMIs is described in Supplementary Table S3. Available data about absolute efficacy of various TCMIs are described in Supplementary Table S4.

The detailed risk of bias assessments of the included studies is summarized in Figure 2 and Supplementary Figure S2. (1) Selective bias (random sequence generation and allocation concealment): The randomization of 12 RCTs (Hu et al., 2009; Zhu and Han, 2014; Xian et al., 2016; Xu and Cao, 2019; Shi et al., 2016; Wu et al., 2017; Wang et al., 2019; Shi et al., 2016; Wu et al., 2017; Wang et al., 2019; Shi et al., 2016; Wu et al., 2017; Wang et al., 2019; Shi et al., 2016; Wu et al., 2017; Wang et al., 2019; Shi et al., 2016; Wu et al., 2017; Wang et al., 2019; Shi et al., 2016; Wu et al., 2017; Wang et al., 2019; Shi et al., 2016; Wu et al., 2017; Wang et al., 2019; Shi et al., 2016; Wu et al., 2017; Wang et al., 2019; Shi et al., 2016; Wu et al., 2017; Wang et al., 2019) was generated via random number table, and three studies (Huang et al., 1999; Hu et al., 2009; Xian et al., 2016) via computer randomization, and two studies (Li et al., 2016a; Tian et al., 2017) via random parallel grouping method, and one study (Shen, 2012) via dynamic random grouping; therefore, the risk of selection bias was considered low. The remaining RCTs referred to only random grouping, and the risk of selection bias was considered unclear. (2) Performance bias (blinding of the participants and personnel): Three studies (Pan et al., 2005; Wang et al., 2011; Xian et al., 2016) were double-blind, and two studies (Dong, 2012; Luo et al., 2015) were single-blind, which were considered low risk. Other studies did not provide information on blinding, so the performance bias was evaluated as unclear risk. (3) Detection bias: There was not enough information to evaluate its risk level; therefore, the risk is unclear. (4) Attrition bias: None of the included RCTs had incomplete data, so the risk of attrition bias was considered “low.” (5) Reporting bias: Taking into account the inability to acquire a complete implementation scheme, the risk of reporting bias was considered “unclear.” (6) Other bias: The risk of this bias was considered “low,” because no other obvious bias was observed in all studies.

Network meta-analysis included 13 treatments for the total effective rate, 12 treatments for LVEF, three treatments for NT-proBNP, nine treatments for BNP, and five treatments for 6MWT (Figure 3). In terms of the total effectiveness rate (Figure 4A), Shuxuetong injection was superior to all other therapies [vs. Huangqi injection (OR: 7.8, CI: 1.17–27.41), vs. Dengzhanxixin injection (OR: 7.34, CI: 1.39–23.76) and vs. CT (OR: 9.36, CI: 3.11–23.50)]. Compared to CT alone, CT combinations with the following TCMIs were significantly more effective: Gualoupi injection (OR: 13.42, CI: 1.29–59.68), Danshenduofenyansuan injection (OR: 6.24, CI: 1.66–17.60), Shenqi Fuzheng injection (OR: 4.69, CI: 1.94–10.12), Danhong injection (OR: 5.07, CI: 1.54–12.32), Xinmailong injection (OR: 4.30, CI: 2.18–7.85), Shenfu injection (OR: 3.83, CI: 2.56–5.59), Shenmai injection (OR: 3.52, CI: 1.91–6.13), and Shengmai injection (OR: 3.26, CI: 1.54–6.26).

For LVEF (Figure 4B), when compared to CT alone, Shenmai injection yielded best results (OR: 8.97, CI: 4.67–13.18); significant improvements were also found for Yiqi Fumai injection (OR: 8.74, CI: 2.67–4.85), Danshenduofenyansuan injection (OR: 5.34, CI: 0.60–10.08), Shenfu injection (OR: 5.12, CI: 2.35–7.89), and Xinmailong injection (OR: 4.37, CI: 0.55–8.17).

For NT-proBNP (Figure 4C), all the compared therapies significant differed from each other. Xinmailong injection, vs. Shenfu injection (OR: 41.48, CI: 22.50–60.49), and vs. CT (OR: −317.70, CI: −331.10–303.10). Shenfu injection with CT was superior to CT alone (OR: −275.60, CI: −288.50–262.70), whereas Shenmai injection with CT was not different from CT only.

For BNP (Figure 4D), Shenqi Fuzheng injection + CT was the best of all therapies [vs. Xinmailong injection (OR: 106.50, CI: −143.00–70.43), vs. Gualoupi injection (OR: 175.60, CI: 141.40–210.20); vs. Dazhuhongjingtian injection (OR: 191.00, CI: 127.70–255.20), vs. Yiqi Fumai injection (OR: −210.90, CI: −290.40–131.60), vs. CT (OR: −257.30,CI: −308.40–242.80), vs. Shenmai injection (OR: −252.40, CI: −286.80–218.60), vs. Danhong injection (OR: 428.70, CI: 283.70–517.70)]. The curative effect is in order for Xinmailong injection, Gualoupi injection, Dazhuhongjingtian injection, Yiqi Fumai injection, CT alone, Shenmai injection, and Danhong injection. Also, Shenmai injection and Danhong injection were not better than CT alone.

For 6MWT (Figure 4E), all TCMIs were superior to CT alone. Of those, Danhong injection was the best [vs. Shenmai injection (OR: 39.87, CI: 13.69–66.06); vs. Yiqi Fumai injection (OR: 38.79, CI: 4.18–73.67); vs. Dazhuhongjingtian injection (OR: 60.06, CI: 30.99–89.23); vs. Xinmailong (OR: 60.00, CI: 35.88–84.06); vs. CT (OR: 84.40, CI: 62.62–106.20)].

Figure 5 shows the Bayesian ranking profiles of the compared treatments with the detail ranking results summarized in Supplementary Tables S5–S9. Shuxuetong injection was most likely to be ranked first for the total effectiveness rate (cumulative probability 39.24%), Shenmai injection for LVEF (20.37%), Xinmailong injection for NT-proBNP (99.99%), Shenqi Fuzheng injection for BNP (42.55%), and Danhong injection for 6MWT (99.24%).

Forest plots for the four available pairwise comparisons that include heterogeneity estimates are shown in Supplementary Figure S3. Our assessment suggested minimal heterogeneity (I² = 0.00%) in half of all comparisons regarding different outcomes. However, moderate to high heterogeneity was detected in the following comparisons:

Shenfu injection vs. CT alone for LVEF (93.24%) and BNP (98.29%); Shenmai injection vs. CT alone for LVEF (98.89%), NT-proBNP (99.53%), and BNP (98.51%); Shengmai injection vs. CT alone for LVEF (97.04%) and 6MWT (39.21%); Yiqi Fumai injection vs. CT alone for LVEF (78.35%); Xinmailong injection vs. CT alone for LVEF (99.58%), NT-proBNP (99.24%), and 6MWT (58.09%); and Shenqi Fuzheng injection vs. CT alone for BNP (98.44%). The fit of the consistency model was similar or better than that of the inconsistency model (Supplementary Table S10).

Thirty-one reports (Shi et al., 2016; Wu et al., 2017; Wang et al., 2019; Shi et al., 2016; Wu et al., 2017; Wang et al., 2019; Shi et al., 2016; Wu et al., 2017; Wang et al., 2019; Shi et al., 2016; Wu et al., 2017; Wang et al., 2019; Shi et al., 2016; Wu et al., 2017; Wang et al., 2019; Shi et al., 2016; Wu et al., 2017; Wang et al., 2019; Gong et al., 2018; Ji, 2019; He, 2016; Wang et al., 2016; Zhou et al., 2013; Wu, 2016; Dong, 2012; Li et al., 2018; Wang and Jang, 2018; Li, 2019a; Zhou and Luo, 2020; Wu and Duan, 2009; Luo et al., 2015; Zhu et al., 2008; Zhao et al., 2011a; Zhao et al., 2012; Li, 2019b; Yang et al., 2014; Zhang, 2020; Teng, 2016; Ni et al., 2020; Huang et al., 1999; Zhou et al., 2002; Wang et al., 2011; Zhang, 2015b) considered the occurrence of ADRs; of those, 25 reports listed no ADRs, and 6 records described specific ADRs (Huang et al., 1999; Xin and Shan, 2012; Zhao et al., 2012; Shi et al., 2016; Wang and Jang, 2018; Ni et al., 2020), such as renal dysfunction, liver dysfunction, urinary system infection or urine protein, etc. (Table 2). The ADR rates of Shenfu injection, Shengmai injection, Xinmailong injection, Dazhuhongjingtian injection, and CT alone were 21.05, 5.56, 5.26, 6.67, and 7.28%, respectively.

A total of 2,959 patients in 22 trials with case numbers ≥100 trials (Zhu and Han, 2014; Xian et al., 2016; Xu and Cao, 2019; Zhu and Han, 2014; Xian et al., 2016; Xu and Cao, 2019; Zhu and Han, 2014; Xian et al., 2016; Xu and Cao, 2019; Zhu and Han, 2014; Xian et al., 2016; Xu and Cao, 2019; Zhu and Han, 2014; Xian et al., 2016; Xu and Cao, 2019; Zhu and Han, 2014; Xian et al., 2016; Xu and Cao, 2019; Zhu and Han, 2014; Xian et al., 2016; Xu and Cao, 2019; Zhu and Han, 2014; Xian et al., 2016; Xu and Cao, 2019; Zhu and Han, 2014; Xian et al., 2016; Xu and Cao, 2019; Zhu and Han, 2014; Xian et al., 2016; Xu and Cao, 2019; Zhu and Han, 2014; Xian et al., 2016; Xu and Cao, 2019; Yang et al., 2014; Xu, 2016; Zhang, 2020; Yang et al., 2014; Xu, 2016; Zhang, 2020; Yang et al., 2014; Xu, 2016; Zhang, 2020; Yang et al., 2014; Xu, 2016; Zhang, 2020; Yang et al., 2014; Xu, 2016; Zhang, 2020; Yang et al., 2014; Xu, 2016; Zhang, 2020) were included in sensitivity analysis. The results did not show any obvious deviations from the original network meta-analysis (Supplementary Figures S4–S5). Among the findings, Yiqi Fumai injection yielded the best therapeutic effects for LVEF and 6MWT, which were only slightly different from the original meta-analysis.

Our study had several limitations. First, our analysis could be complicated by various confounding factors beyond our control, because most treatments were indirectly compared and most direct evidence was derived from one trial in the present network. Second, despite our best efforts, the included RCTs were of relatively poor quality. For example, although all trials reported that patients were randomly assigned into different groups, only some of 61 RCTs described the specific methods of generating random sequences, such as a random number table or a random parallel grouping. Only five studies mentioned the blinding method, and most trials were low sample size tests with positive findings which are particularly prone to various biases. Third, most of the included studies have not been registered.

By synthesizing all evidence from multiple RCTs, this review helps to identify current problems and areas worthy of improvement. Due to the poor quality of the included studies, the evidence obtained from our network meta-analysis is not sufficient for a comprehensive comparison of different TCMI combinations with various CTs for treatment of CHD-HF. Based on our findings, we propose the following two recommendations for further studies on TCMI-CT in the treatment of CHD-HF: (1) clinical trials should be prospectively registered in recognized clinical trials registry platforms; and (2) the quality of study designs should be improved, including randomization, allocation concealment, and blinding.