This NMA was conducted under the Preferred Reporting Items for Systematic Reviews and Meta-Analysis extension statement for network meta-analysis (PRISMA-NMA) (Hutton et al., 2015). This study was registered with PROSPERO under registration number CRD42025632143.

To enhance the accuracy, the CCPPs in this study were reported in accordance with the requirements of the Consensus statement on the Phytochemical Characterisation of Medicinal Plant extracts (ConPhyMP) (Heinrich et al., 2022). Accurate scientific nomenclature for botanical drugs referred to Rivera’s suggestion (Rivera et al., 2014) and was validated taxonomically in the databases of “Medicinal Plant Names Services” (https://mpns.science.kew.org/mpns-portal/). The composition and standardised name for each CCPP were presented in Table 1. In addition, we referred to the Chinese Pharmacopoeia 2025 regarding the names of non-botanical drugs. The relevant information about CCPPs referred to the original study, the Chinese Pharmacopoeia 2025, and the National Medical Products Administration. The details were shown in Supplementary Appendix S2, S3.

From the time the database was established to 12 December 2024, eight databases and two registry systems were searched, including Web of Science, Cochrane Library, PubMed, Embase, China National Knowledge Infrastructure (CNKI), Wanfang database, China Science and Technology Journal Database (VIP), Chinese Biomedical Literature database (CBM), Clinical Trials, and the China Clinical Trials Registry. Clinical RCTs of nine CCPPs in treating CMD, including SXBX, TXL, SXTXD, YDXNT, KDL, XB, XKS, DAXXK, and YXTL were retrieved. Additionally, we manually searched references of eligible studies to identify other relevant research. The search is elaborated further in the Supplemental appendix 4, encompassing additional search strategies and outcomes information.

The Population-Intervention-Comparators-Outcomes-Timing-Setting (PICOTS) framework was used as the criterion for this study. Inclusion criteria: (1) Population: All patients were diagnosed with CMD. (2) Intervention: Conventional therapy combined with SXBX or TXL or SXTXD or YDXNT or KDL or XB or XKS or DAXXK or YXTL. (3) Comparator: Conventional therapy. (4) Outcomes: The primary outcome indicators were the IMR and CFR; The Secondary outcome measures were Angina attack frequency, hs-CRP, ET-1, NO, and LDL-C. (5) Timing: Studies with any follow-up duration were considered. (6) Setting: Studies conducted in any clinical setting (e.g., inpatient, outpatient) were eligible (7) Study design: RCTs.

Exclusion criteria: (1) Non-RCT. (2) The intervention was a combination of multiple therapies or did not specify a therapeutic agent. (3) Duplicate publication. (4) Retracted. (5) Full text unavailable. (6) Lack of complete data.

Two independent reviewers (Yudou Li and Xinyue Wang) extracted the following details from the included studies: (1) first author’s name, year, and country of publication; (2) sample size and mean age; (3) specific interventions, duration of interventions; and (4) outcome data. Any disagreements were discussed or consulted with the third researcher (Wujiao Wang).

The quality of the included studies was evaluated using the Cochrane Collaboration’s risk of bias assessment tool, version 2.0 (RoB 2). The quality assessment items were as follows: (1) Randomization process; (2) Deviations from intended interventions; (3) Missing outcome data; (4) Measurement of the outcome; (5) Selection of the reported result; and (6) Overall risk of bias. Bias in each aspect was evaluated as “low risk,” “some concerns,” and “high risk.” Any disagreements were discussed or consulted with the third researcher (Peifen Chang).

First, Risk ratios (RRs) and 95% CI were performed for dichotomous variables, and mean difference (MD) and 95% CI were performed for continuous variables. In certain multi-arm trials (such as three-arm studies), if two control arms both qualify as active controls, they are merged into a single active control group using established formulas (Supplementary Appendix S5), in order to prevent duplication of experimental group data and inflated contribution to the overall analysis. Data were analysed using a random-effects model under the frequentist framework with Stata version 15. Network diagrams were constructed to visualise the geometry of the treatment network. In these diagrams, node size is proportional to the total sample size for each treatment, and the thickness of the connecting lines represents the number of studies that directly compared the connected interventions. Provided that the closed loop of interventions was available, Global consistency was evaluated using the design-by-treatment interaction model, while local inconsistency was examined using node-splitting analysis, which compares direct and indirect evidence for specific treatment comparisons. Between-study heterogeneity was quantified by estimating the variance (τ²) of the underlying effect sizes, with parameters estimated using the restricted maximum likelihood method. Meta-regression analysis is employed to investigate the potential influence of covariates on intervention effect estimates. Sensitivity analysis is utilised to assess the robustness. Then, interventions were ranked using the Surface Under the Cumulative Ranking Curve (SUCRA) probability values, with higher SUCRA values indicating a greater likelihood of a treatment being ranked highly. Finally, the Funnel plots and Egger’s test were used to explore publication bias if >10 studies were included.

The quality of evidence was assessed using the GRADE method (Guyatt et al., 2011). It was categorized as high, moderate, low, or very low. RCTs received a high initial grade by default and were downgraded according to pre-specified criteria: risk of bias, inconsistency, indirectness, imprecision, and other considerations.

The preliminary search obtained 3,596 relevant papers, 2,952 papers were obtained after excluding duplicates, and 820 papers were obtained after excluding Non-RCTs, Reviews, Non-CMD, and duplicates. Thirty-nine papers (Bai et al., 2022; Chen et al., 2022; Chen et al., 2019; Chu, 2021; Fang et al., 2018; Feng et al., 2005; Fu, 2021; Fu et al., 2020; Gong et al., 2021; Jiang et al., 2024; Li, 2021; Li and Tang, 2009; Liang et al., 2019; Lin et al., 2023; Liu et al., 2003; Liu et al., 2018; Lv et al., 2014; Ma and Xu, 2006; Meng, 2018; 2019; Peng, 2011; Qi et al., 2023; Qin et al., 2017; Qin et al., 2023; Ren et al., 2023; Shen et al., 2021; Sun et al., 2022; Sun, 2021; Wang B. et al., 2024; Wang, 2022; Wang, 2015; Wang and Long, 2022; Wang et al., 2019; Wei, 2018; Wu et al., 2019; Xie et al., 2019; Zhang, 2022; Zhang et al., 2013; Zhao D. H. et al., 2021) were finally included after further reading the full text to exclude studies with Inadequate study design and unavailable research data. Tianli Li resolved any disagreements during the selection process. The Flow diagram of the literature search is shown in Figure 1.

A total of 39 studies were included, all of which were published in Chinese. There were 3,240 participants, including 1,622 in the treatment group and 1,618 in the control group. Detailed characteristics of the included studies are shown in Table 2.

The compared connections among interventions for each outcome are shown in Figure 2. Each node represents a different intervention and the size of nodes is positively correlated with the number of patients. The thickness of the line segment corresponds to the number of included studies for that intervention. The thicker the line segment, the larger the number of included studies for that intervention is. There were no closed loops formed between the studies.

A total of 39 papers were included in this study, in which seventeen studies (Bai et al., 2022; Chen et al., 2019; Fang et al., 2018; Fu, 2021; Fu et al., 2020; Gong et al., 2021; Lin et al., 2023; Meng, 2018; 2019; Qin et al., 2017; Ren et al., 2023; Shen et al., 2021; Wang B. et al., 2024; Wang, 2022; Wang and Long, 2022; Wang et al., 2019; Zhao D. H. et al., 2021) used the random number table method, 22 studies (Chen et al., 2022; Chu, 2021; Feng et al., 2005; Jiang et al., 2024; Li, 2021; Li and Tang, 2009; Liang et al., 2019; Liu et al., 2003; Liu et al., 2018; Lv et al., 2014; Ma and Xu, 2006; Peng, 2011; Qi et al., 2023; Qin et al., 2023; Sun et al., 2022; Sun, 2021; Wang, 2015; Wei, 2018; Wu et al., 2019; Xie et al., 2019; Zhang, 2022; Zhang et al., 2013) only mentioned randomization without detailing the randomization scheme. No study reported the use of opaque envelopes to conceal the randomization program. Only two studies (Chen et al., 2019; Xie et al., 2019) reported the blinding of participants and researchers; two studies (Liu et al., 2003; Peng, 2011) reported the blinding of participants; seven studies (Feng et al., 2005; Liu et al., 2003; Lv et al., 2014; Ma and Xu, 2006; Peng, 2011; Wang, 2015; Zhang et al., 2013) with small sample sizes. The quality assessment of the included RCTs is shown in Figures 3, 4.

As this network meta-analysis did not form a closed loop, node splitting could not be employed for inconsistency testing. First, we conducted a global consistency test. The results revealed that only the p-values for angina attack frequency, ET-1, and NO were below 0.05, indicating significant inconsistency, which may stem from diagnostic heterogeneity, dosing, follow-up length, and study quality. Secondly, we employed NMA within a frequency-based framework to fit a consistency model. Restricted maximum likelihood (REML) was used to estimate the global heterogeneity variance (τ²). Results indicated significant heterogeneity among the included studies. We conducted further meta-regression to identify sources of heterogeneity. Six characteristics were selected, including duration of intervention, CMD diagnosis methods, sample size, gender ratio, year of publication, and risk of bias. However, the regression analyses revealed no significant influence from these covariates, indicating that these characteristics were not sources of heterogeneity between studies. Subsequent sensitivity analyses demonstrated the stability of the results. Finally, we conducted sensitivity analyses excluding high-risk studies and non-validated CMD studies, further demonstrating the robustness of our findings (Supplementary Appendixs S9, S10)

A total of 12 studies reported adverse drug reactions (ADRs), with six studies indicating no ADRs occurred in either the intervention group or the control group. Two studies reported ADRs to SXBX, one study reported ADRs to TXL, one to YDXNT, one to XB, and one to DAXXK. All ADRs were mild, and no study reported withdrawal due to ADRs. The results of the forest plot revealed that there were no significant differences in the risk of adverse drug reactions across various CCPPs (Figure 12). Detailed information is provided in Supplementary Appendix S10.

The assessment of the level of evidence for inclusion of the outcomes was summarised using the GRADE methodology as shown in Table 3.

We assessed publication bias for indicators that included more than ten papers in the study, and the funnel plot results showed that angina attack frequency was roughly symmetrical (Figure 13). The ET-1 and NO funnel plots exhibited poor symmetry, and subsequent Egger’s tests revealed no significant evidence of publication bias (P-values of 0.145, 0.088, and 0.179, respectively). This inconsistency may stem from a potential small-sample effect, or may indicate that funnel plot asymmetry could be attributable to factors beyond publication bias (such as heterogeneity between studies). Nevertheless, we should exercise caution in interpreting the results, as the presence of publication bias cannot be entirely ruled out. It is worth noting that the included small-sample studies generally exhibited low methodological quality (such as more deficiencies in allocation concealment and blinding procedures), and the overestimation of effect sizes may partly stem from this. We have therefore interpreted these findings with caution. Although the possibility of publication bias cannot be ruled out, it is not the sole explanation for this result. It is undeniable that this bias is likely to have substantially impacted our SUCRA ranking results, which rely heavily on unbiased effect estimates. Given the current risk of potential bias, we should emphasize direct comparisons of point estimates and confidence intervals for clinical decision-making, rather than over-relying on specific ranking orders. Therefore, although XKS and YDXNT ranked highest in ET-1 and NO, respectively, further large-scale, high-quality studies are required to validate the relative efficacy of these interventions and provide more reliable evidence for ranking.

For other outcomes with fewer than ten included studies (CFR, IMR, hs-CRP, and LDL-C), formal statistical tests for small-study effects (e.g., Egger’s test) are underpowered. Therefore, the assessment of publication bias for these outcomes relies solely on qualitative interpretation of the funnel plots, which should be considered tentative. More primary studies are needed to allow for robust evaluation of publication bias for these endpoints.

A total of 39 RCTs involving 3,240 patients were included in the study. NMA results showed that the efficacy of CT combined with CCPPs was significantly better than CT alone. SXBX had the highest probability of being the best treatment on account of the reduction of IMR [MD = −5.93, 95% CI (−8.75, −3.11)] and LDL [MD = −0.56, 95% CI (−0.99, −0.14)]; XB showed better efficacy in CFR [MD = 0.71, 95% CI (0.53, 0.89)]; TXL showed better efficacy in angina attack frequency [MD = −5.30, 95% CI (−7.08, −3.53)]; YXTL showed better efficacy in hs-CRP [MD = −5.04, 95% CI (−8.38, −1.7)]; XKS showed better efficacy in ET-1 [MD = −43.3, 95% CI (−59.71, −26.89)]; YDXNT showed better efficacy in NO [MD = 17.69, 95% CI (6.07, 29.32)]. However, this finding must be interpreted with extreme caution, as the GRADE assessment indicates that the quality of evidence for all comparisons is low. This implies that our confidence in the accurate estimate of the effect size is limited, and future research is likely to alter or even reverse the current ranking and conclusions.

CMD is classified in TCM under “Xiong Bi” and “Xin Tong” (angina pectoris). Blood-activating and Qi-promoting CCPPs have been reported to improve coronary microcirculatory and vascular endothelial functions and alleviate pain. Our NMA is the first to compare various CCPPs in CMD systematically. The study highlighted differences in efficacy, but all shared the core TCM principle of “Blood Activation and Qi Promotion.” Furthermore, analysis of the composition of each CCPP revealed that the most frequently used herbs were Ginseng, Salvia miltiorrhiza, Panax notoginseng, Artificial musk, and Borneol. They benefit qi, improve blood circulation, and relieve pain. These herbs may offer potential therapeutic benefits for CMD. However, the exact mechanisms behind their effects require further investigation through modern pharmacological research.

Several CCPPs demonstrated significant efficacy in our NMA for CMD. Their benefits appear to stem from both the active herbal components they contain and the multi-target mechanisms these formulations employ. Below, we first discuss the major CCPPs and their pharmacological effects and then provide an overview of commonly used single herbal compounds.

SXBX is an aromatic and warming CCPP that benefits Qi and strengthens the heart. It was the most effective CCPP for decreasing IMR, which may be due to its effects in reducing lipid levels, plaque formation, and endothelial damage (Li D. et al., 2024), anti-inflammation, anti-atherosclerosis (Lu et al., 2019), and protection of endothelial function (Ning et al., 2011). Studies have shown (Wei et al., 2023) that it can promote angiogenesis via the GDF15-TRPV4 signaling pathway. It inhibits pyroptosis and improves I/R injury by promoting autophagosome generation and accelerating autophagic flux (Yu et al., 2022). Comprehensive metabolomics studies have shown it protects cardiac function by regulating amino acid, lipid, and energy metabolisms (Wu et al., 2020).

XB was the most effective CCPP in increasing CFR, possibly due to its improved energy metabolism, suppressed apoptosis, suppressed excessive autophagy, and endoplasmic reticulum (ER) stress effects. It has been shown to inhibit SGLT1 protein expression while upregulating the phosphorylation level of AMPK, promoting nuclear translocation of PPARα and enhancing its transcriptional activity, ultimately improving fatty acid energy metabolism in the heart (Pan et al., 2024). It also promotes mitochondrial homeostasis by inhibiting heme synthesis to increase succinyl-CoA (Chen et al., 2025). In addition, XB inhibits excessive autophagy by decreasing Beclin-1 and LC3II and increasing p62. It also inhibits ER stress by decreasing BIP expression and apoptosis by increasing Bcl2/Bax and decreasing caspase3 (Yang et al., 2022). DAXXK is second only to XB in increasing CFR. The main ingredient of DAXXK is total steroidal saponin, which is an Rhizome extract of Dioscorea nipponica Makino, and has been shown to reduce TC and TG levels, anti-inflammatory, anti-oxidative stress, and anti-atherosclerotic effects (Zhang et al., 2022).

TXL ranked highest for reducing angina attack frequency. It has the function of invigorating qi and promoting blood circulation, which can enhance myocardial contractility, inhibit platelet aggregation, and regulate the level of blood lipids (Chen et al., 2024). Studies have shown it could alleviate no-reflow by suppressing the interactions by modulating various leukocyte subtypes and inhibiting the expression of multiple inflammatory mediators (Liu S. et al., 2023). TXL also inhibited endothelial cell pyroptosis via the reactive oxygen species/nucleotide-binding oligomerization domain-like receptor family pyrin domain-containing 3/Caspase-1/GSDMD signalling pathway (Gu et al., 2023). YXTL ranked highest for decreasing hs-CRP levels. It has the effects of improving coronary microcirculation, anti-inflammation, and anti-platelet aggregation (Meng, 2020). KDL is a processed tablet made from Corydalis Rhizoma, which has anti-myocardial ischaemia, anti-thrombotic, and anti-arrhythmic effects (Sun et al., 2009).

XKS ranked highest for decreasing ET-1 levels, It can also elevate the nitric oxide content, improve the vascular endothelial function (Liu et al., 2022). Studies (Liu et al., 2016) have shown it protects cardiac function by inhibiting the myocardium Ca (2+) overloading and metabolic alterations. It also promotes angiogenesis through multiple signaling pathways, including metabolic pathways, the NOD-like receptor signaling pathway, the VEGF signaling pathway, the PPAR signaling pathway, and the PI3K/Akt signaling pathway (Liu Q. et al., 2023). SXTXD is second only to XKS in reducing ET-1. It can regulate the cellular autophagy process, promote smooth muscle cell proliferation and differentiation, exert anti-inflammatory effects, and optimize lipid metabolism (Chang et al., 2022).

YDXNT ranked highest for increasing NO levels. Studies have shown it has the effect of repairing damaged endothelial cells, reducing the release of endothelin, dilating blood vessels, and improving coronary microcirculation (Yang et al., 2023); it also can exert anti-inflammatory effects by decreasing IL-1β, IL-8, and IL-18 via the TLR4 pathway (Wang et al., 2014). In addition, it relieves atherosclerosis through regulating lipids, reducing lipid particle deposition in the endothelial layer of the artery, enhancing antioxidant power, and repressing inflammatory activity by inhibiting the nuclear factor-kappa B signal pathway (Cheng et al., 2015).

Overall, these commonly used herbs and their preparations appear to address key pathophysiological mechanisms of CMD, including oxidative stress, inflammatory responses, atherosclerosis, vascular endothelial dysfunction, and abnormal energy metabolism. By combining multiple active ingredients, CCPPs may offer synergistic effects; however, further well-designed studies are needed to determine optimal dosages, assess long-term safety, and elucidate their clinical utility.

Our study’s most frequently used herbs were Panax ginseng, Salvia miltiorrhiza, Panax notoginseng, Moschus, and Borneolum. A total of 45 experimental studies were identified to investigate the effects and mechanisms of the main active components of single-flavored Chinese medicine, which were frequently used in CMD. Table 4 lists the Mechanisms of the main active components of single-flavored Chinese Medicine on CMD. The structural formula of the main active components are showed in Figure 14. The possible mechanisms of them are summarized as follows:

This study compares the therapeutic effects of nine CCPPs and draws relevant conclusions. However, there are still some limitations here, including: (1) Interpretation and global relevance: All included trials were conducted in China, which limits generalisability to other populations, as genetic and environmental factors may influence drug efficacy; (2) Network geometry limitations: There is no closed loop between studies; NMA relies on indirect comparisons. Sparse connections, lack of closed loops, and absence of multi-arm trials may weaken the transitivity assumption and reduce the precision of indirect comparisons. (3) Diagnostic heterogeneity: Varying and sometimes non-validated definitions of CMD (e.g., symptom-based diagnosis, TTDE, variable CFR cut-offs) may introduce misclassification bias. (4) No adjustment for baseline covariates: Differences in patient characteristics, baseline CMD severity, and concomitant therapies were not accounted for in the NMA, potentially confounding results. (5) Potential publication bias: All included studies report positive effects; the absence of negative trials raises the possibility of reporting bias. (6) Evidence certainty: Given high/unclear risk of bias and lack of robust indirect evidence, the GRADE certainty for most outcomes is low; conclusions should be framed as hypothesis-generating rather than definitive. (7) Exclusion of international pharmacotherapy comparators: Standard CMD drugs (e.g., nicorandil, ranolazine, zibotentan) were not included in the network; therefore, the results cannot be directly compared to current international guideline-based treatments. (8) Research quality: Many trials have small sample sizes, increasing the risk of Type I/II errors and unstable SUCRA rankings. The absence of placebo-controlled and multicentre trials diminishes the robustness of the research findings.