According to the PICOS model, we established the following criteria for all eligible studies to meet.

(i). Only optimum trials as randomized controlled trials (RCTs) without restrictions on follow-up, institutions, language, and publication time.

All ineligible studies must meet the following criteria: studies about patients with ascites or pericardial effusion; all patients receiving systemic chemotherapy, local hyperthermia or oral traditional Chinese medicine (TCM); both groups receiving kushen preparation; studies with unclear objectives; without any data about clinical responses, survivals, QOL, or adverse events.

The primary outcomes are clinical response and survival. Referring to previous studies (Paladine et al., 1976; Kessinger and Wigton, 1987; Keeratichananont et al., 2015; Jie Wang et al., 2018; Dipper et al., 2020; Xiao et al., 2020a), we integrated both Millar and Ostrowskimj criteria to measure the clinical responses as: (i) complete response (CR) is the disappearance of pleural effusion for more than 30 days, or the lack of accumulation of fluid; (ii) partial response (PR) is less than 50% reduction of pleural effusion for more than 30 days; (iii) no response (NR)/stable disease (SD) is less than 50% reduction of pleural effusion or less than 25% increase or the recurrence of fluid accumulation without further therapy; (iv) pleural progression (PP) is more than 25% increase of pleural effusion or symptomatic fluid accumulation again requiring further therapy. We set the pleurodesis failure as no response or stable disease plus pleural progression and assessed the clinical responses using complete response, pleurodesis failure, and pleural progression (Supplementary Material S2). Long-term survival was assessed by using overall survival (OS), progression-free survival (PFS), OS, and PFS rates. According to the Karnofsky performance status (KPS) scale, when a KPS score increased ≥10 after treatment, QOL was improved.

Adverse events (AEs) were assessed by using ADRs and thoracentesis-related adverse events (TRAEs). According to World Health Organization (WHO) or Common Terminology Criteria for Adverse Events (CTCAEs) criteria (Miller et al., 1981; Trotti et al., 2003), ADRs were measured by using the indicators myelosuppression, neutropenia, thrombocytopenia, anemia, hepatorenal toxicity, gastrointestinal reactions, thoracodynia, and fever. TRAEs were measured by using indicators including treatment-related death, respiratory failure, dyspnea, pneumothorax, chest infection, drainage tube detachment, tumor metastasis along the indwelling duct, catheter-related infection, or subcutaneous emphysema.

Adhering to a retrieval logic of patient plus intervention, we customized the retrieval strategies for each database using MeSH and free words (Supplementary Material S3). Yan Zhang and Hui Liu independently searched all related studies about “Kushen preparations in controlling MPE” from Chinese and English electronic databases (to February 2025) including the Guizhou Digital Library, SinoMed, China National Knowledge Infrastructure Database, WanFang Database, Chinese Scientific Journals Full-text Database, PubMed, Embase, Web of Science, and Cochrane Central Register of Controlled Trials (Issue 2, February 2025). We collected ongoing trials from the Chinese Clinical Trial Registry (http://www.chictr.org.cn), WHO International Clinical Trials Registry Platform (http://apps.who.int/trialsearch/), and US clinical trials (https://clinicaltrials.gov). Finally, we also identified eligible studies from the references of relevant SRs or network meta-analysis. Hui Liu and Yan Zhang independently selected eligibles and excluded ineligible studies following a predesigned inclusion and exclusion criteria.

For clinical responses, survivals, QOL, or adverse events, Da-chun Cai and Jiao Xu independently applied a revised Cochrane tool (RoB 2) to assess methodological quality arising from five domains: randomization process (D1), intended interventions (D2), missing outcome data (D3), outcomes measurement (D4), and selective reporting of results (D5) (Sterne et al., 2019; Higgins et al., 2021). We judged each quality based on the domain algorithm and made an overall judgment.

Yao-Qin Luo and Da-chun Cai independently collected all data using a predesigned data extraction form. The data were first author, time of publication, methodological features, demographic characteristics and cases; characteristics of patients as tumor types, pleural fluid volume, anticipated survival time (AST), KPS score, treatment history, and recurrence; drainage methods as indwelling pleural catheters (IPCs) or thoracentesis; kushen preparations, treatment dose, frequency and times, and sclerosants and uses; follow-up protocol, research institutions, criterion and time of evaluation. The outcomes were: complete response, pleurodesis failure, pleural progression, PFS, OS, QOL, ADRs, and TRAEs. Additionally, the authors of papers were contacted about available survival data. If they were unavailable, the Kaplan–Meier survival curves were transformed into data using Engauge Digitizer 4.1 (Guyot et al., 2012; Xiao et al., 2018).

All eligible studies were clustered into multiple homogeneous treatment units, and we further analyzed their clinical effectiveness and safety. The odds ratios (ORs) and their 95% confidence interval (CI) were applied to measure the complete response, pleurodesis failure, pleural progression, OS rate, QOL, ADRs, and TRAEs, with p < 0.05 being identified as statistically significant. Cochran’s χ² test and I ² statistic were performed to identify statistical heterogeneity among each unit. If the results showed significant heterogeneity and inconsistent directions or involved a single trial, we used forest plots to describe the result. When p ≥ 0.1 and I ² ≤ 50%, a fixed-effects model (FEM) was applied to pool the OR and their 95% CI. When p < 0.1, I ² > 50%, and the results had consistent direction, a random-effects model (REM) was applied. Yan Zhang and Feng Luo independently applied Review Manager 5.4 to pool the data from each unit. If the outcomes involved more than ten trials, a funnel plot and Egger’s test (STATA V.15.0 software, 401506209499) were applied to identify potential publication bias.

Referring to previous experience (Xiao et al., 2020b; Wang et al., 2021; Wang et al., 2022; Wang C. Q. et al., 2023), a subgroup analysis was implemented to reveal the potential clinical heterogeneity among the main treatment plans with enough trials to analyze the effects of patient related factors, interventions, and evaluation criteria on clinical responses and to further identify the suitable population and optimum usage. We further implemented univariate random effects meta-regression analysis to reveal the correlation between each factor and clinical responses and post hoc multiple regression analysis to identify it.

Following underestimation of effectiveness/safety, we implemented sensitivity analysis to identify robustness (Xiao et al., 2020b; Wang et al., 2021; Wang et al., 2022; Wang C. Q. et al., 2023). The consistency of results before and after excluding both trials with high risk and overestimation were analyzed. If consistency was good, the result was robust; otherwise, it was poor. To identify the required information size (RIS) for the results of main treatment units (Thorlund et al., 2016), we further applied Trial Sequential Analysis (TSA) software (version 0.9.5.10 Beta) to implement the analysis. In the light of previous experience, we set the risk of type I error as 5% with a power of 80%, relative risk reduction (RRR) as 25% for clinical responses and QOL, and 20% for adverse events (AEs) (Wetterslev et al., 2008; Thorlund et al., 2009). We used control event rates from this analysis for these calculation, and adjusted the information size for diversity (Wetterslev et al., 2009).

We integrated the results of sensitivity analysis into the GRADE approach (Guyatt et al., 2008; Xiao et al., 2020b; Wang et al., 2021; Wang et al., 2022; Wang C. Q. et al., 2023) and developed a revised approach to summarize the evidence. Quality was identified as “high”, moderate”, “low”, and “very low” following five domains: risk-of-bias of results, heterogeneity, indirectness, imprecision, and publication bias (Supplementary Material S2). Jun Huang and Yan-Yan Jin independently applied the GRADE profiler to summarize the evidence quality and generated the absolute estimates of effect for outcomes.

After retrieval, 1,269 records were identified. Two reviewers read the titles, excluded duplicates, and identified 443 records. After screening abstracts and excluding irrelevant and non RCTs, 147 RCTs, six SRs/meta-analyses (Tian et al., 2010; Tang et al., 2014; Biaoxue et al., 2015; Xu et al., 2015; Yang et al., 2016; Wu et al., 2018) and four network meta-analyses (Yang et al., 2017; Li B. et al., 2019; Li, 2022; Xu et al., 2022) were selected. Further evaluating full-texts and excluding 64 ineligible studies (Supplementary Material S3), 83 were considered eligible. Additionally, 42 studies were selected from previous studies. Finally excluding duplicates, 83 eligible studies were selected for this analysis (Figure 2).

We clustered the 83 eligible studies from 2001 to 2023 into four themes: intrapleural perfusion with CKI alone, CKI and sclerosants, kang’ai or matrine, and platinum for controlling MPE. Eleven trials reported CKI alone (Table 1a). CKI and sclerosant developed three comparisons as CKI-versus-cisplatin (Yuan, 2007; Hu Q. et al., 2008; Chen, 2010; Liang et al., 2011; Chen, 2013; Xing, 2013; Wang and Zhou, 2016; Yan et al., 2016; Wang R. et al., 2023), mitomycin (Zhang, 2011), or interleukin-2 (Huang, 2013). All trials recruited 796 inpatients—426 male and 244 female patients aged 20–82 years. Receiving CKI were 396 patients, while another 400 received sclerosants alone. Perfusion with CKI and sclerosants was reported in 59 trials (Table 1b). The CKI and chemical drug or BRM developed ten treatment plans: perfusion with CKI and cisplatin, nedaplatin (Li, 2014; Zhang S. et al., 2015; Li et al., 2017), bleomycin (Chen and He, 2003; Liu and Wan, 2011; Sun, 2012), hydroxycamptothecin (He et al., 2009; Wu et al., 2014; Cai and Wang, 2019), lobaplatin (Liu and Xu, 2016; Huang, 2021), carboplatin (He and Xie, 2010), mitomycin (Zhang et al., 2013), interleukin-2 (Hao and Liang, 2007; Zhou et al., 2010), OK-432 (Wei et al., 2014; Zhong et al., 2015), and Corynebacterium parvum (Huang et al., 2012). There were 41 trials which evaluated perfusion with CKI and cisplatin, recruiting 2,823 inpatients aged 15–91, with 1,346 male and 909 female patients. Some 1,424 patients received perfusion with CKI and cisplatin, while another 1,399 received cisplatin alone. CKI was administrated 10–60 mL/time, once to thrice per week, lasting one to twelve times; the cisplatin was administrated with 20–80 mg/time. Kang’ai or matrine and platinum developed four plans. Six trials involving 334 inpatients aged 36–84 years (Zhang, 2006; Hu J. et al., 2008; Xu and Xiong, 2008; He, 2011; Qu et al., 2012; Wang, 2016) evaluated perfusion with kang’ai and cisplatin (Table 1c). Received kang’ai and cisplatin were 168 patients, while another 166 received only cisplatin. Kang’ai was administrated 40–60 mL/time, once or twice per week, lasting one to four times. Six trials recruiting 319 inpatients aged 30–85 (Du et al., 2009; Li and Yang, 2009; He, 2010; Wang et al., 2010; Ji, 2011; Ji et al., 2012) evaluated perfusion with matrine and cisplatin (Table 1d). A total of 167 patients received matrine and cisplatin, while another 152 only received cisplatin. Matrine was administrated 150–800 mg/time, once a week, lasting 2 to 6 weeks.

Of 83 eligible studies, 58 (69.88%, 58/83) involved inpatients with miscellaneous tumors, 24 (28.92%, 24/83) with lung cancer, and only one with hematologic malignancies (Huang, 2013) or breast cancer (Wang R. et al., 2023). Most studies described demographic characteristics, but only 16 to 50 (19.28%, 16/83% to 60.24%, 50/83) reported the pleural fluid volume, KPS, AST, and treatment history. All studies reported the drainage methods and characteristics of interventions and assessed the clinical responses 5–10 weeks after treatment began using Ostrowskimj or Millar criteria. Only 36 studies (43.37%, 36/83) reported the QOL, and six reported overall survival (Cui et al., 2008; Chen, 2010; He, 2011; Han, 2013; Zhang S. et al., 2015). Some 79 studies (95.18%, 79/83) reported the AEs, 38 (45.78%, 38/83) assessed ADRs using WHO or CTEC3.0 criteria, and only four assessed TRAEs (Wang et al., 2010; Yang, 2012; Wei et al., 2014; Liu and Li, 2015; Song and Jia, 2015). No study reported conflicts of interest.

Of 83 studies, 79 (95.18%, 79/83) expressed concerns at overall bias for clinical responses, and four showed high risk (Qu et al., 2012; Wu et al., 2014; Wang and Zhou, 2016; Lin et al., 2023). At domain-level, only one study had low risk at D1 (Liu and Li, 2015), one showed high risk at D1 (Lin et al., 2023) or D2 (Wang and Zhou, 2016), and others had some concerns. All had low risk at D3 and D4. Two studies showed high risk at D5 (Qu et al., 2012; Wu et al., 2014), and others had low risk (Figure 3A; Supplementary Figures S1, S2). For overall survival, five studies had concerns of overall bias (Cui et al., 2008; Chen, 2010; He, 2011; Han, 2013; Zhang S. et al., 2015). All had some concerns at D1 and D2, and low risk at D3, D4, and D5 (Supplementary Figure S4).

Since studies were limited, we only assessed the methodological quality of QOL and adverse events in CKI versus cisplatin, and perfusion with CKI, kang’ai, or matrine and cisplatin. QOL was reported by 29 studies and showed high risk at overall bias. Only one study (Liu and Li, 2015) had low risk, and one (Lin et al., 2023) had high risk at D1. All showed some concern at D2, low risk at D3 and D5, and high risk at D4 (Figure 3B and S3). A total of 57 studies reported AEs, 35 (61.40%, 35/57) showed high risk at overall bias, and 21 had some concerns. There were 55 studies (96.49%, 55/57) with some concerns at D1and D2, two with low risk at D1 (Liu and Li, 2015; Lin et al., 2023), and one with high risk at D2 (Wang and Zhou, 2016). All studies had low risk at D3. High risk was shown by 16 studies (28.07%, 16/57), and 39 had low risk at D4. A total of 34 studies (59.65%, 34/57) showed high risk, and 21 had low risk at D5 (Figure 3C and Figure. S5).

Nine trials reported clinical responses about CKI versus cisplatin (Table 2a; Supplementary Figures S6–S8). Cochran’s χ² test and I ² statistic revealed no heterogeneity (I ² = 0%). We pooled the OR using a FEM. The results of meta-analyses revealed that CKI perfusion displayed a complete response (1.10, 95% CI 0.76 to 1.60), pleurodesis failure (0.80, 95% CI 0.56 to 1.14), and pleural progression (0.63, 95% CI 0.33 to 1.21) similar to cisplatin alone. Only single trial reported that CKI achieved clinical response similar to mitomycin and better than interleukin-2.

The CKI and chemical drug or BRM developed ten treatment plans (Table 2b; Figure 4C; Figure 5). Perfusion with CKI and cisplatin was evaluated by 41 trials. With no statistical heterogeneity (I ² = 0%), an FEM was used to pool the OR. The results demonstrated it significantly improving the complete response (2.71, 95% CI 2.30 to 3.19) and displaying a low pleurodesis failure (0.26, 95% CI 0.22 to 0.32) and pleural progression (0.22, 95% CI 0.14–0.36) than cisplatin alone. One to three trials reported nine other treatment plans. Compared with sclerosants alone, the results revealed that nine treatment plans achieved a low pleurodesis failure, while only CKI and bleomycin, hydroxycamptothecin, or interleukin-2 significantly improved the complete response.

Kang’ai or matrine and platinum developed four treatment plans (Table 2c, 2d; Supplementary Figures S9–S11). With no statistical heterogeneity (I ² = 0%), an FEM was used. The results demonstrated that perfusion with kang’ai or matrine and cisplatin significantly improved the complete response (3.04, 95% CI 1.76 to 5.26 and 1.87, 95% CI 1.26–2.78) and achieved a low pleurodesis failure (0.23, 95% CI 0.14 to 0.41 and 0.27, 95% CI 0.17–0.44) than cisplatin alone. Additionally, matrine and cisplatin achieved a low pleural progression (0.29, 95% CI 0.09–0.95).

Of 83 studies, only six (Cui et al., 2008; Chen, 2010; Chen et al., 2011; He, 2011; Han, 2013; Zhang S. et al., 2015) reported the OS of perfusion with CKI alone, CKI and cisplatin or nedaplatin, kang’ai and cisplatin, or matrine and carboplatin (Figure 5). Compared with sclerosants alone, only one trial reported that perfusion with CKI and cisplatin might improve the 0.5-year OS rate (Han, 2013), and it might prolong median survival time and PFS (Chen et al., 2011). Perfusion with CKI and nedaplatin might improve the 1-year OS rate (Zhang S. et al., 2015), and matrine and carboplatin might improve the 0.5-year, 1-year, and 1.5-year OS rates (Cui et al., 2008).

Due to limited trials, we only assessed the QOL of perfusion with CKI alone, CKI, kang’ai, or matrine and cisplatin (Table 3; Supplementary Figures S12, S13). Six trials reported the QOL about CKI alone (Yuan, 2007; Hu Q. et al., 2008; Liang et al., 2011; Chen, 2013; Xing, 2013; Yan et al., 2016). Statistical heterogeneity (I ² = 67%) was found, and an REM was used. Compared with cisplatin alone, CKI perfusion acquired a similar QOL. There were 21 trials reporting QOL about perfusion with CKI, kang’ai, or matrine and cisplatin. No heterogeneity was found (I ² = 0%), and an FEM was used to pool the OR. Compared with cisplatin alone, the results demonstrated that perfusion with CKI, kang’ai or matrine and cisplatin significantly improved QOL (3.60, 95% CI 2.84 to 4.56; 3.95, 95% CI 1.78 to 8.74 and 2.95, 95% CI 1.25–6.97).

Nine trials reported eight AEs about CKI alone (Table 3a; Supplementary Figures S14, S15, S18–S23). Cochran’s χ² test and I ² statistic only identified statistical heterogeneity for myelosuppression (I² = 69%), gastrointestinal reaction (I² = 67%), thoracodynia (I² = 69%), and fever (I² = 81%), and an REM or FEM was used to synthesize the OR. Compared with cisplatin alone, meta-analysis revealed that perfusion with CKI alone showed a low myelosuppression (0.02, 95% CI 0.00 to 0.15), leukopenia (0.10, 95% CI 0.03–0.35), gastrointestinal reaction (0.03, 95% CI 0.01 to 0.12), hepatotoxicity (0.09, 95% 0.02–0.33), nephrotoxicity (0.09, 95% CI 0.03 to 0.29), and thoracodynia (0.15, 95% CI 0.04 to 0.48).

Ten AEs were reported by 38 trials about CKI and cisplatin (Table 3; Supplementary Figures S14–S23). We only identified minimal heterogeneity for gastrointestinal reaction (I² = 8%), and an FEM was used. The results demonstrated that perfusion with CKI and cisplatin showed a low myelosuppression (0.34, 95% CI 0.24–0.47), neutropenia (0.35, 95% CI 0.26 to 0.46), gastrointestinal reaction (0.36, 95% CI 0.29–0.44), and hepatorenal toxicity (0.42, 95% CI 0.28 to 0.63 and 0.32, 95% CI 0.24–0.44) and fever (0.50, 95% CI 0.30–0.82).

Five trials reported six AEs to kang’ai and cisplatin (Table 3c; Supplementary Figures S15–S24). We only identified minimal heterogeneity leukopenia (I ² = 25%), and an FEM was used. The results revealed that kang’ai and cisplatin showed low neutropenia (OR = 0.20, 95% CI 0.11–0.38) and gastrointestinal reaction (OR = 0.34, 95% CI 0.19–0.63).

Five trials reported ten AEs to matrine and cisplatin (Table 3d; Supplementary Figures S14–S23). We only identified statistical heterogeneity for neutropenia (I ² = 66%) and minimal heterogeneity for myelosuppression (I ² = 43%), and an REM or FEM was used. The results revealed that matrine and cisplatin showed low neutropenia (0.10, 95% CI 0.02–0.61), gastrointestinal reaction (0.35, 95% CI 0.19–0.66), and thoracodynia (0.21, 95% CI 0.10–0.48).

In targeting perfusion with CKI and cisplatin, subgroup analysis revealed that under different primary tumors, drainage, and evaluation criteria, this treatment plan obtained significant improvement in the complete response and low pleurodesis failure. For patients with moderate-to-large effusion, KPS ≥50 to ≥70 scores, AST ≥3 months, or primary treatment, it significantly improved clinical responses. Perfusion with CKI (20–50 mL/time, once per week, two to four times) and cisplatin (20–80 mg/time) could significantly improve the clinical responses. Moreover, perfusion with low-dosage cisplatin and CKI could obtain clinical responses like high-dosage. However, the univariate regression and multiple meta-regression analysis did not reveal any correlation between clinical response and each variable (Table 4; Supplementary Figures S25–S72).

Only perfusion with CKI and cisplatin was included in more than ten trials (Table 5; Supplementary Figures S73–S83). The funnel plot and Egger’s test did not identify publication bias for the complete response, pleurodesis failure, pleural progression, myelosuppression, neutropenia, hepatorenal toxicity, thoracodynia, and fever, which were objectively reported. Significant publication bias was identified for QOL (coefficient = –2.47, 95% CI –4.62 to –0.32) and gastrointestinal reaction (coefficient = –1.49, 5% CI –2.71 to –0.21); both results were under-estimated.

In CKI versus cisplatin, 11 outcomes were pooled using meta-analysis. Before and after excluding the trials with high risk and over-estimating efficacy/safety, the OR of QOL, myelosuppression, gastrointestinal reaction, and thoracodynia showed poor robustness, and the others had good robustness. In perfusion with CKI and cisplatin, 13 outcomes were pooled, and QOL, thrombocytopenia and anemia showed poor robustness. In CKI and nedaplatin, lobaplatin, bleomycin, hydroxycamptothecin, interleukin-2, or OK-432, 12 outcomes were pooled, and only the complete response of CKI and nedaplatin, lobaplatin, or OK-432 showed good robustness. In kang’ai and cisplatin, six outcomes were pooled, showing poor robustness. In matrine and cisplatin, 11 outcomes were pooled, and the QOL, myelosuppression, neutropenia, and gastrointestinal reaction showed poor robustness (Table 6).

Since the trials were limited, we only assessed the RIS for clinical responses in CKI versus cisplatin. The TSA identified firm information size for supporting a similar complete response and pleurodesis failure between CKI and cisplatin, and no reliable information for pleural progression. We further assessed the RIS for clinical responses, QOL, and AEs in perfusion with CKI and cisplatin. Further analysis identified sufficient and conclusive information sizes for complete response, pleurodesis failure, QOL, neutropenia, and gastrointestinal reaction, and firm information for pleural progression, myelosuppression, and hepatorenal toxicity. Finally, we only assessed the RIS for clinical responses in kang’ai or matrine and cisplatin. The analysis identified firm information sizes for pleurodesis failure in both treatments and no reliable information for complete response (Table.7; Figure 6; Supplementary Figures S84–S94).

We applied a revised GRADE approach to identify the evidence quality as “high”, “moderate”, “low”, and “very low”. In CKI versus cisplatin, 11 results were pooled. Clinical responses, hepatorenal toxicity, and fever were summarized as moderate quality, while other five results were low to very low (Table 8a). In perfusion with CKI and cisplatin, 13 results were pooled. Clinical responses, myelosuppression, neutropenia, gastrointestinal reaction, hepatorenal toxicity, and fever were summarized as moderate, while the other four were low to very low. In CKI and nedaplatin, lobaplatin, bleomycin, hydroxycamptothecin, interleukin-2, or OK-432, 12 results were pooled. The clinical responses were very low to low (Table 8b). In kang’ai and cisplatin, six results were pooled at low to very low (Table 8c). In matrine and cisplatin, 11 results were pooled. The complete response and pleurodesis failure were summarized as moderate, with the other nine results as low to very low (Table 8d).