The protocol for this study has been registered on the PROSPERO website (https://www.crd.york.ac.uk/prospero/) with number CRD42024582055 in Additional File 1. Two independent investigators systematically searched six databases (PubMed, Embase, Cochrane Library, CNKI, Wanfang, and CBM) for literature published up to 7 October 2024. The search strategy combined Medical Subject Headings (MeSH) terms with free words, incorporating key concepts such as ‘Traditional Chinese Medicine’ (TCM), ‘immune checkpoint inhibitors’, ‘PD-1′, ‘PD-L1′, and ‘ICIs’. The detailed search strategy is provided in Additional File 2. Relevant references from the retrieved articles were also screened for inclusion. The entire process adhered strictly to the PRISMA guidelines (Moher et al., 2009) (see Additional File 3). Botanical drugs included in this study were verified using the Medicinal Plant Names Services (MPNS) portal (http://mpns.kew.org/mpns-portal/), while other traditional medicines were authenticated via the Zhong Hui Zhong Yao Wang database (https://www.zhzyw.com/).

Two reviewers independently screened the titles and abstracts of retrieved records, followed by full-text assessment for eligibility. A third reviewer resolved any discrepancies and made final judgments based on the pre-defined protocol. The specific inclusion criteria were as follows:

All included literature must meet the following: 1) Study participants were clinically diagnosed cancer patients aged 18 years or older, with both an experimental and a control group; 2) The intervention consisted of ICIs combined with TCM, which could include TCM monomers, single herbs, herb pairs, prescriptions, or moxibustion/fumigation therapies; 3) The control group received ICIs for cancer; 4) Reported outcomes included at least one of the following: Overall Response Rate (ORR), Disease Control Rate (DCR), CD4+/CD8+ T-cell ratio, Progression-Free Survival (PFS), Overall Survival (OS), Karnofsky Performance Status (KPS), levels of Alpha-Fetoprotein (AFP), Carcinoembryonic Antigen (CEA), Carbohydrate Antigen 125 (CA125), Carbohydrate Antigen 19-9 (CA199), incidence of adverse effects, or TCM therapeutic evaluation; 5) Study design was a randomized controlled trial (RCT) comparing ICI combination therapy with TCM versus ICI monotherapy. Concurrent conventional or basic supportive care was permitted in both groups. Common ICIs considered included Pembrolizumab, Nivolumab, Sintilimab, Tislelizumab, Camrelizumab, and so on.

Data extraction and quality assessment were performed independently by two reviewers. A third reviewer was consulted to reconcile disagreements and finalize the data synthesis. Extracted data encompassed: publication year, corresponding author’s country/region, cancer diagnostic criteria, baseline characteristics of participants (sample size, age, gender), details of interventions (types, dosage, administration of ICIs; types, formulation, administration of TCM; other concomitant treatments; treatment duration), and all pre-specified outcomes (ORR, DCR, CD4+/CD8+ ratio, PFS, OS, KPS, AFP, CEA, CA125, CA19-9, adverse effects, TCM evaluation). Corresponding authors were contacted for missing or unclear data. Furthermore, ORR was defined as the proportion of patients achieving a complete response (CR) or partial response (PR). DCR was defined as the proportion of patients achieving CR, PR, or stable disease (SD) (Upadhyay et al., 2025; Bae et al., 2022). Both ORR and DCR are key metrics for tumor response evaluation.

The methodological quality of the included RCTs was assessed using the Cochrane Risk of Bias tool (RoB 2.0) (Higgins et al., 2011). This tool evaluates six domains: i) randomization process, ii) allocation concealment, iii) blinding of participants and personnel, iv) blinding of outcome assessment, v) incomplete outcome data, and vi) selective reporting. Each domain was judged as ‘low risk of bias’, ‘some concerns’, or ‘high risk of bias’. To enhance the reporting transparency and reproducibility of the included TCM interventions, the ConPhyMP tool was referenced, following the approach of Heinrich et al.

Data analysis was conducted using RevMan (version 5.3) and Stata (version 12). For dichotomous outcomes, data were pooled and expressed as Risk Ratios (RR) with 95% confidence intervals (CI). Continuous outcomes were categorized based on their distribution. Data following a normal distribution were presented as mean ± standard deviation (SD). Non-normally distributed data were converted to mean ± SD using established methods (Shi et al., 2023; Luo et al., 2018; Wan et al., 2014) via https://www.math.hkbu.edu.hk/∼tongt/pages/median2mean.html. The Weighted Mean Difference (WMD) and 95% CI were applied when outcomes were measured on the same unit across studies; otherwise, the Standardized Mean Difference (SMD) and 95% CI were used. Heterogeneity among studies was quantitatively assessed using the I² statistic and Cochran’s Q test, with results visualized using forest plots. An I² value less than 50% indicated low heterogeneity, warranting the use of a fixed-effects model. An I² value greater than 50% suggested substantial heterogeneity, leading to the adoption of a random-effects model (Schmidt et al., 2009; Higgins et al., 2003). If sufficient studies were available (n > 10), subgroup analysis or meta-regression was planned to explore potential sources of heterogeneity. Sensitivity analysis was performed by sequentially excluding each study to evaluate the robustness of the pooled results. Publication bias was assessed using Egger’s test (Egger et al., 1997) when more than five studies were included in a meta-analysis. A p-value <0.05 indicated potential publication bias, in which case the trim-and-filling method was employed to assess their stability.

A total of 6,826 articles were retrieved from the six databases. After removal of duplicates, 4,948 articles remained for screening by two independent reviewers based on the inclusion and exclusion criteria. Ultimately, 41 studies (Chai, 2023; Chen, 2022; Ding et al., 2024; Du and Liu, 2022; Fang, 2023; Jiang and Fang, 2024; Lin et al., 2022; Lin, 2023; Liu et al., 2022; Luo, 2023; Qu, 2023; Wang L. et al., 2023; Wang Q., 2023; Wang et al., 2024; Wang XH., 2023; Wang, 2021; Xiao et al., 2022; Yang, 2023; Yao, 2023; Zhang J., 2023; Zhang, 2022; Zhao et al., 2023; Zhao, 2024; Zhou, 2023; Zhu, 2024; Cao et al., 2024; Cao, 2023; Huang et al., 2024; Ye and Fang, 2022; Yu D. et al., 2023; Zhang LL., 2023; Zhong, 2023; Dou, 2022; Liu and Wang, 2024; Liu and Xia, 2024; Ye, 2022; Huang et al., 2023; Wang, 2024; Xu, 2023; Zhang et al., 2023; Zhou and Xie, 2024) were included in the meta-analysis. All included studies were conducted in China, with only three published in English databases. This analysis encompassed 2,612 cancer patients, including cases of lung cancer (n = 1,631), liver cancer (n = 490), gastric cancer (n = 154), esophageal cancer (n = 227), colorectal cancer (n = 50), and ovarian cancer (n = 60). Various ICIs were used, such as Sintilimab, Nivolumab, Camrelizumab, and Tislelizumab. TCM interventions included prescriptions, moxibustion, injections, and other formulations. The publication years of the included studies ranged from 2020 to 2024, indicating a focus on recent research. The characteristics of the included studies are summarized in Table 1, and the study selection flow diagram is presented in Figure 1. The TCM included in each study are detailed in Additional File 4, with all drugs having been verified.

The methodological quality of the 41 included studies, assessed using the Cochrane RoB 2.0 tool, is summarized in Figure 2 and detailed in Additional File 5. Most studies reported random allocation of participants to the case or control group; however, only two studies specified the involvement of a third researcher in the randomization process. Blinding procedures were seldom mentioned, leading to potential performance and detection bias. Several studies reported participant dropouts, with reasons for attrition not always being consistent between groups. All studies adhered to their pre-specified outcome measures without selective reporting. Overall, the quality of the included studies was variable, with some demonstrating robust methodology and others having unclear reporting of design elements. The assessment of TCM preparation reporting quality using the ConPhyMP tool is provided in Additional Files 6 and 7.

The comparison of ORR between two groups of patients is shown in Figure 3. The heterogeneity among 31 studies was low (p = 0.79, I² = 0%), therefore a fixed-effect model was chosen. The results showed that the patients in TCM + ICIs group had higher ORR values (RR = 1.34 [1.20–1.49]) compared to those in ICIs group, with RR = 1 as the reference. In addition, exploration on different cancers were also conducted (see Table 2), and the results showed that the TCM + ICIs group had higher ORR values in lung cancer, gastric cancer, esophageal cancer, and other tumors (RR = 1.37 [1.19,1.58], RR = 2.07 [1.05,4.08], RR = 1.63 [1.10,2.41], RR = 2.43 [1.09,5.40], respectively).

There are a total of 29 studies involving DCR, and the results of meta-analysis and heterogeneity analysis are shown in Figure 3. The fixed-effect model was applied (p = 0.46, I² = 0%). The results showed that the DCR values of TCM + ICIs group were significantly higher than those of ICIs group, with RR 1.15 (1.10, 1.21). The subgroup analysis results of different cancers also demonstrated that regardless of the type of cancer, patients in TCM + ICIs group have higher DCR in Table 2.

The meta-analysis and heterogeneity analysis of 16 studies are shown in Figure 4. There is obvious heterogeneity (p < 0.05, I² = 96%), so the random-effect model is the best choice. The CD4+T/CD8+T ratio in TCM + ICIs group was significantly higher than that in ICIs group, with WMD 0.25 (0.15, 0.35). Due to the presence of high heterogeneity, subgroup analysis and meta-regression were performed to further explore its sources. The results in Table 3 showed that heterogeneity among subgroups decreased when grouped based on different cancers or ICIs, but it cannot be determined whether it is the source of heterogeneity. Therefore, the results of meta-regression are particularly important. The results in Table 4 suggest that whether it is lung cancer and whether both groups receive chemotherapy regimens may be sources of high heterogeneity (p < 0.05). Unfortunately, further multi-meta-regression was not performed.

Table 5 contains a comparison of PFS and OS between two groups of patients. The meta-analysis of PFS used a random-effect model (p < 0.05, I² = 91%), and the results showed that the PFS of TCM + ICIs group was significantly higher than that of ICIs group (WMD = 0.96 [0.29, 1.63]). Due to the small number of studies included (n = 4), heterogeneity exploration was not conducted. Meanwhile, only two studies involved OS, and the heterogeneity among them was not high (p = 0.48, I² = 0%). After applying the fixed-effect model, the results showed that TCM + ICIs group had a longer OS period, with WMD 1.46 (0.62, 2.30).

14 studies are related to KPS, and there is significant heterogeneity among them (p < 0.05, I² = 72%). As shown in Figure 4, the results of the random-effect model demonstrate that the KPS score of TCM + ICIs group is significantly higher than that of the ICIs group, with WMD 6.35 (4.99, 7.70). The Galbr plot also showed high heterogeneity among studies. Thus, subgroup analysis and meta-regression were also performed. Through subgroup analysis of different cancers and ICIs, heterogeneity of each group has decreased, shown in Table 6, suggesting that they may be the source of heterogeneity (Table 7). Univariate meta-regression suggests that different types of ICIs are indeed sources of high heterogeneity (p < 0.05), which deserves further research.

Only four studies involved AFP (Table 5), and there was significant heterogeneity (p < 0.05, I² = 89%). After the application of the random-effect model, the AFP level in TCM + ICIs group was significantly lower than that in ICIs group, with SMD -0.75 (- 1.49, − 0.01). The source of heterogeneity has not been explored.

Nine studies are included in this section (Table 5), including seven on CEA, five on CA125, and two on CA199. Due to the presence of high heterogeneity, they all adopted a random-effect model (p < 0.05, I² = 66%; p < 0.05, I² = 87%; p < 0.05, I² = 95%, respectively). All results demonstrated that the levels of CEA and CA125 in TCM + ICIs group were significantly lower than those in ICIs group (SMD = −0.72 [-1.08, −0.37]; SMD = −0.77 [-1.46, −0.08], respectively), and there was no significant difference in CA199 levels between the two groups (SMD = −0.95 [-2.68, 0.78]). However, the number of studies included in this section is limited, which also affects the reliability of the conclusion.

The results of the adverse events of the two groups are presented in Table 5. There was high heterogeneity among the 12 included studies (p = 0.0005, I² = 67%), therefore a random-effect model was selected. The incidence of adverse events in TCM + ICIs group was significantly lower than that in ICIs group (RR = 0.82 [0.69, 0.97]). Further analysis was conducted based on different adverse events, including gastrointestinal reactions, myelosuppression, hypertension, thyroid dysfunction, liver dysfunction and kidney dysfunction. The final results showed that the various adverse reactions in TCM + ICIs group were significantly lower than those in ICIs group shown in Table 8.

Figure 3 shows the TCM Therapeutic Evaluation of two groups, which includes 16 related studies. Due to insignificant heterogeneity (p = 0.65, I² = 0%), a fixed-effect model was chosen. The overall TCM therapeutic evaluation of TCM + ICIs group was higher than that of ICIs group, with RR 1.42 (1.30, 1.55). Regardless of the type of cancer, the TCM therapeutic evaluation of TCM + ICIs group was significantly increased shown in Table 2.

The results of all sensitivity analyses can be found in Additional File 8. The stability of each meta-analysis can be determined by whether the changes of the results are significant or whether there is a reversal after removing each study one by one. After comprehensive evaluation, it was found that all 11 meta-analyses in this study have good stability (with no significant changes in the results).

Only meta-analyses with more than 5 included studies were evaluated for publication bias. The publication bias results of these seven meta-analyses were based on Eggar’s Test (Additional File 9). When it comes to ORR, DCR, CD4+T/CD8+T, Adverse Event, and TCM therapeutic evaluation, significant publication bias cannot be ignored (p < 0.05), while the other two meta-analyses are not. Therefore, the trim-and-filling method was executed to further evaluate whether the presence of publication bias affects the reliability and robustness of the results. This adjustment did not substantially alter or reverse the original pooled results, supporting the robustness and representativeness of the findings despite the presence of publication bias.