Two researchers (JRB, HWJ) independently conducted comprehensive searches of four English databases: MEDLINE (PubMed), EMBASE (Elsevier), CENTRAL (the Cochrane Central Register of Controlled Trials), and the Cumulative Index to Nursing and Allied Health Literature, via EBSCO. Additionally, searches were performed on three non-English databases: CNKI (Chinese National Knowledge Infrastructure Database) in Chinese, CiNii (Citation Information by NII) in Japanese, and KCI (Korea Citation Index) in Korean. The initial search was conducted on February 14, 2023, and updated on October 10, 2024, to include the most recent information and relevant evidence. Relevant studies were identified by reviewing the reference lists of previous studies (Zhang et al., 2012; Jin et al., 2018; Lai, 2018; Zhao J. et al., 2020b; Wang et al., 2022). No restrictions on language, publication date, or publication status were applied. For MEDLINE, the following search strategy was used: (“depressive disorder” [MeSH Terms] OR “depression” [MeSH Terms] OR depressive OR depression) AND (“stroke” [MeSH Terms] OR stroke) AND (jia-wei-xiao-yao-san OR xiao-yao-san OR xiao-yao powder OR xiao-yao wan OR kamishoyosan OR TJ-24 OR Soyosan OR Soyo-san OR Gami-soyosan OR Gamisoyo-san) (Supplementary Table S1).

After eliminating duplicate publications, two researchers (JRB, HWJ) independently reviewed the titles and abstracts to assess their relevance. The full texts of potentially eligible studies were then examined for the final selection. Any disagreements were resolved through discussion with another researcher (JTL). For this review, a formal inter-rater agreement statistic, such as Cohen’s kappa, was not calculated; instead, any discrepancies between reviewers were resolved through discussion and consensus with a third author. All retrieved studies were organized and managed using Zotero (Roy Rosenzweig Center for History and New Media, George Mason University).

Two researchers (JRB, HWJ) independently extracted data using standardized forms, followed by cross-verification. In case of conflict, a third researcher (JTL) was consulted to reach a consensus. For this review, a formal inter-rater agreement statistic, such as Cohen’s kappa, was not calculated; instead, any discrepancies between reviewers were resolved through discussion and consensus with a third author. The extracted information included the first author’s name, publication year, sample size, dropout number, participants, HM details, control interventions, intervention duration, outcome measures, and adverse events related to the intervention. If the data were insufficient or unclear, additional information was requested from the corresponding authors via email.

Two researchers (JRB, HWJ) independently analyzed the methodological quality and quality of evidence for each outcome of all included studies. Discrepancies were resolved through discussions with another researcher (JTL).

Methodological quality was assessed using the revised Cochrane Collaboration risk-of-bias tool for randomized trials (RoB2). The analysis involved the following domains: random sequence generation, allocation concealment, blind participants and personnel, blind outcome assessment, incomplete outcome data, selective reporting, and other potential biases. Each domain was categorized as “low risk,” “unclear,” or “high risk.” In the domain of random sequence generation, a study was considered to have “some concerns” or higher risk of bias if the term “randomization” was mentioned without an explanation of the randomization method or if there was no confirmation of double-blinding, even with an explanation.

The GRADE methodology was employed to assess the quality of evidence for each outcome. GRADEpro, an online software (https://gradepro.org/), was used to analyze risk of bias, inconsistency, indirectness, imprecision, and publication bias on a four-point scale (“very low,” “low,” “moderate,” or “high”).

RStudio (version 2023.06.1 + 524; Posit, Boston, MA, United States) and R (version 4.3.0; CRAN, WU Vienna), with the ‘meta’, ‘metafor’, and ‘dmetar’ packages for meta-analysis, and the ‘robvis’ package for visualizing the risk of bias assessments. Descriptive statistical analyses were performed for participant characteristics, interventions, and outcomes across all the included studies. Meta-analyses were performed on studies in which the treatment and control groups were within the same category, and separate analyses were conducted for each outcome variable. The main method for meta-analysis in the quantitative synthesis utilized a random-effects model, considering the inherent heterogeneity in clinical trials of traditional East Asian medicine (TEAM). The results of the common-effects model were presented for sensitivity analysis. Due to the differences in treatment components, study groups, and patient selection criteria among the included studies, potential heterogeneity in the actual treatment effects was anticipated, making the random-effects model more appropriate. Continuous outcomes were recorded as mean differences (MDs) with 95% confidence intervals (CIs). For dichotomous data, the relative risk (RR) was the primary analysis method, with odds ratios (ORs) provided as supplementary information. Point estimates and 95% CIs were used for representation. Heterogeneity among the effect sizes of the studies was assessed using the I ² statistic, with values > 75% indicating high heterogeneity.

Subgroup analyses were conducted based on the following criteria to account for heterogeneity or to analyze whether effect sizes were significantly different among subgroups, provided there were sufficient data: (1) type of antidepressant used in the control group, (2) treatment duration, (3) use of TEAM diagnostic tools, (4) use of HAMD criteria for participant selection, and (5) type of Soyo-san substance. Additionally, meta-regression analyses were conducted using treatment duration as a moderating variable to assess its impact on effect size.

A cumulative meta-analysis was conducted to analyze changes in effect size chronologically and by sample size. Additionally, leave-one-out analysis was performed to compare the effect size or heterogeneity when each study was included or excluded.

Publication bias was assessed for outcome measures in >10 included studies using funnel plots. Asymmetry in the funnel plots was evaluated, and if observed, Egger’s regression was performed using the metabias () function from the ‘meta’ package (considered present if P < 0.05). The trim-and-fill method was used to assess the degree of publication bias using the trimfill () function, and funnel plots before and after adjustment are presented.

After the exclusion of duplicates, the database searches identified 144 studies, and 33 studies were identified from reference lists of previous systematic reviews. After screening the titles and abstracts for relevance, 94 studies were excluded. Upon full-text review, three studies used HM other than Soyo-san, 13 studies compared HM alone with antidepressants, 17 studies employed treatments other than HM in the intervention group, or in combination with HM, or used treatments other than antidepressants in the control group. Twenty studies were excluded for various reasons. Finally, 41 RCTs involving 3,628 participants were included (Figure 1) (Supplementary Table S2).

The overall characteristics of the included studies are summarized in Table 2. All the RCTs were conducted in China and published in academic journals. The sample size ranged from 40 to 400 participants, with a median age of 88. The treatment duration was 2–12 weeks. Few studies applied TEAM pattern identification to categorize the signs and symptoms of each participant consistently to optimize treatment. Liver qi stagnation or phlegm in the heart in 3 studies (Xu, 2006; Ma, 2010; Gao, 2013), liver qi stagnation in 5 studies (Xu et al., 2011; Lu and Chen, 2013; Shao et al., 2017; Han et al., 2019; Hu et al., 2024), liver qi stagnation and spleen deficiency in 1 study (Jiang and Liu, 2019), liver depression and spleen deficiency in 1 study (Zhao D. et al., 2020a), and liver depression type of fire in 2 studies (Zhao S. et al., 2020c; Chen et al., 2021) were adopted as the TEAM pattern. Few studies limited participants’ conditions by severity of depression based on HAMD: mild or greater (Li et al., 2011; Lin et al., 2015; Cui et al., 2017; Xu et al., 2017; Yang et al., 2018; Zeng et al., 2018; Gong et al., 2020; Zhao D. et al., 2020a; Zhao S. et al., 2020c), moderate or greater (Li and Gao, 2006; Song and Yang, 2008; Wang, 2008; Zou et al., 2009; 2013; Zhang, 2010; Xu et al., 2011; Lu and Chen, 2013; Wang and Ni, 2014; Zhang et al., 2014; Yuan, 2015; Zhang and Zhang, 2016; Shao et al., 2017; Han et al., 2019; Wang et al., 2019), mild and moderate levels of depression (Jiang and Liu, 2019). Several studies specified the duration of PSD by indicating the time of onset before the intervention (Song and Yang, 2008; Zou et al., 2009; 2013; Ma, 2010; 2015; Zhang, 2010; Li et al., 2011; Xu et al., 2011; Zhou, 2017; Zhao D. et al., 2020a; Hu et al., 2024). In the control group, three types of antidepressants were used: SSRIs in 30 studies, flupentixol/melitracen in 9 studies (Zou et al., 2009; Li et al., 2011; Wang and Ni, 2014; Zhi et al., 2014; Lin et al., 2015; Ma, 2015; Zhang and Li, 2016; Han et al., 2019; Zhao D. et al., 2020a), and serotonin and norepinephrine reuptake inhibitors in 2 studies (Zou et al., 2013; Zhou, 2017). HAMD was the most frequently used outcome measure, followed by TER in 34 studies, MBI in 8 studies (Li and Gao, 2006; Song and Yang, 2008; Zhang and Li, 2016; Zhang and Zhang, 2016; Cui et al., 2017; Han et al., 2019; Zhao S. et al., 2020c; Zheng, 2021), 5-HT levels in 7 studies (Yang et al., 2018; Zeng et al., 2018; Han et al., 2019; Jiang and Liu, 2019; Wang et al., 2019; Zhao D. et al., 2020a; Zheng, 2021), SSS in 7 studies (Li and Gao, 2006; Song and Yang, 2008; Zou et al., 2009; 2013; Dang and Zhang, 2011; Zhang and Li, 2016; Zheng, 2021), NIHSS in 3 studies (Cui et al., 2017; Wang et al., 2019; Zhao D. et al., 2020a), PSQI in 3 studies (Yang et al., 2018; Zhao D. et al., 2020a; Zheng, 2021), FMA in 2 studies (Yang et al., 2018; Zheng, 2021), and MMSE in 2 studies (Song and Yang, 2008; Zhao S. et al., 2020c). TER was calculated based on the reduction of HAMD score only or together with symptom improvement in 28 studies, and solely on symptoms in 6 studies (Supplementary Table S3).

Based on the RoB2 analysis, 2 studies (Zhi et al., 2014; Zhou, 2017), used appropriate random sequence generation methods and explicitly conducted double-blinding, indicating a low risk of bias in the randomization process and outcome domains. In the remaining 39 studies, there was either no description of random sequence generation or insufficient information regarding allocation concealment and blinding of participants, personnel, or outcome assessors, resulting in some level of bias. No dropouts occurred in any of the included studies, indicating a low risk of bias due to deviation from the intended interventions. Regarding missing outcome data, almost all studies had complete data for the intended interventions, indicating a low risk of bias. The remaining 39 studies lacked protocols, indicating a high risk of bias. Lack of confirmation of the measurements before unblinding caused ‘some concerns’ about bias due to selection of the reported results. Overall, 2 studies were rated as having ‘some concerns’ about bias, whereas the rest had a high level of bias (Figure 2) (Supplementary Figure S1).

Twenty-seven studies used decoctions, 7 used pills (Wang, 2008; Zou et al., 2009; Wang and Ni, 2014; Zhang, 2014; Zhi et al., 2014; Zeng et al., 2018; Hu et al., 2024), and 7 used granules (Xu, 2006; Gao, 2013; Jiang and Liu, 2019; Xu et al., 2017; Zou et al., 2013; Zhou, 2017; Zhang and Li, 2016). Soyo-san was provided twice a day in 29 studies, three times a day in 7 studies (Zou et al., 2009; 2013; Wang and Ni, 2014; Zhi et al., 2014; Zeng et al., 2018; Chen et al., 2021), and 2 studies provided a combination of two and three times a day during different treatment periods (Li and Gao, 2006; Zheng, 2021). One study combined once a day and twice a day administration (Zhang et al., 2014), one study had once daily administration (Yuan, 2015), and one study did not specify dosing instructions (Zhang and Li, 2016). In addition to the original Soyo-san composition, 22 additional herbs were used. Except for 5 studies with unknown dosages of each herb (Wang, 2008; Zou et al., 2009; 2013; Gao, 2013; Zeng et al., 2018) and 6 other studies using medications with unknown composition (Wang and Ni, 2014; Zhang, 2014; Zhi et al., 2014; Zhang and Li, 2016; Jiang and Liu, 2019; Hu et al., 2024), the basic eight herbs of Soyo-san were used: Bupleurum chinense DC. [Apiaceae; Bupleuri Radix], Paeonia lactiflora Pall [Paeoniaceae; Paeoniae Radix], Angelica sinensis (Oliv.) Diels [Apiaceae; Angelicae Sinensis Radix], Wolfiporia cocos (Schw.) Ryvarden and Gilb. [Polyporaceae; Poria] (Fungus) in all studies (100%). Atractylodes macrocephala Koidz. [Asteraceae; Atractylodis Macrocephalae Rhizoma] in 33 (94.3%), Glycyrrhiza uralensis Fisch. [Fabaceae; Glycyrrhizae Radix et Rhizoma] in 33 (94.3%), Mentha arvensis L. [Lamiaceae; Menthae Haplocalycis Herba] in 24 (85.7%), and Zingiber officinale Roscoe [Zingiberaceae; Zingiberis Rhizoma Recens] in 24 (68.6%). Furthermore, 39 additional herbs were used as adjuvants based on pattern identification or symptoms (Supplementary Table S5).

With GRADE methodology, the quality of evidence was rated as “very low” or “low,” with no high-quality evidence available, mainly because of the high risk of bias in the RCTs included in the meta-analysis. Additionally, most outcome measures did not have an adequate number of eligible participants, leading to low precision, heterogeneity, and indirectness, which further downgraded the quality of evidence (Table 3).

In accordance with the guideline, an assessment of the herbal medicine interventions was performed using the ConPhyMP checklists for all included studies. Among 41 included studies, 20 were classified as Type A (Li and Gao, 2006; Wang, 2008; Zou et al., 2009; 2013; Dang and Zhang, 2011; Pan and Wan, 2012; Gao, 2013; Lu and Chen, 2013; Wang and Ni, 2014; Zhang et al., 2014; Yuan, 2015; Shao et al., 2017; Zhou, 2017; Yang et al., 2018; Zeng et al., 2018; Han et al., 2019; Jiang and Liu, 2019; Wang et al., 2019; Hu et al., 2024), 21 as Type B (Xu, 2006; Song and Yang, 2008; Ma, 2010; 2015; Zhang, 2010; 2014; Li et al., 2011; Xu et al., 2011; 2017; Li, 2013; 2015; Lin et al., 2015; Zhang and Li, 2016; Zhang and Zhang, 2016; Cui et al., 2017; Sun, 2017; Gong et al., 2020; Zhao S. et al., 2020c; Zhao D. et al., 2020a; Chen et al., 2021; Zheng, 2021). As shown in Tables 4–6, the response rates were very low in general. For Checklist 1, 41.5% (17/41) of studies provided a basic description of the ingredients (Item 3), 0% (0/41) of studies reported description of the botanical drug and taxonomic authentication (Item 2) and documentation of the legal basis for collection and processing (Item 4). Furthermore, almost every items in Checklist 2a and 2b were reported zero. The detailed study-by-study analysis documents are available in Supplementary data sheets.

When models of the effect sizes of HAMD were compared, the random-effects model (MD: −4.01; 95% CI: −4.72, −3.30; I ² = 94%) showed better improvement than the common-effects model (MD: −3.70; 95% CI: −3.85, −3.55; I ² = 94%), with both showing significant outcomes. Outcomes other than HAMD were also similar between the random- and common-effects models.

Leave-one-out analysis revealed no significant differences in either effect size or heterogeneity. For HAMD, heterogeneity was high (I ² = 94%, τ² = 4.2923). Leave-one-out analyses did not identify a single study that accounted for the dispersion. In the cumulative meta-analysis based on publication year and sample size, no significant differences were found, indicating robustness of the outcomes (Supplementary Figure S5).

A funnel plot was drawn for the primary outcome (HAMD) to analyze the publication bias of the included studies (Figure 5). The results of Egger’s regression, which was conducted for quantitative analysis, did not show a statistically significant or visually apparent publication bias. Similarly, no publication bias was observed for the TER (Supplementary Figure S4).

We systematically selected 41 RCTs that involved the combination of Soyo-san with conventional antidepressants to treat PSD and analyzed the effectiveness and safety using various methods. Combination treatment with Soyo-san showed significant effects on measures such as HAMD, TER, serotonin level, MBI, NIHSS, and SSS. The treatment group generally exhibited fewer adverse effects than the control group in all included studies. Subgroup analyses of HAMD, TER, MBI, and SSS outcomes did not reveal any significant differences. Meta-regression using treatment duration as the moderator variable for HAMD and TER did not yield significant findings. The risk of bias analysis indicated ‘some concerns’ of high risk of bias in the included studies. The quality of evidence for the outcome measures, assessed using GRADE, ranged from very low to moderate, with no high-level evidence. No publication bias was observed for HAMD and TER based on funnel plots and Egger’s regression.

PSD is typically diagnosed based on the DSM-5 as mentioned in the introduction. However, the RCTs included in this study employed various diagnostic criteria, such as the 2001 Chinese Classification and Diagnostic Criteria of Disorders, in addition to or instead of the DSM-5. Few studies diagnosed depression without describing the criteria used or based on arbitrarily set symptoms and signs, indicating an overall lack of clarity and consistency in diagnosis. Furthermore, the Soyo-san dosage and frequency for concomitant therapy varied across studies, and few studies lacked detailed explanations regarding its administration. Because the patients had PSD, the interventions applied to both the treatment and control groups were not consistently described, which compromised the consistency of the patient selection criteria across studies, leading to limitations in the interpretation of the results.

In a systematic review and meta-analysis of Soyo-san for PSD in 2022 (Wang et al., 2022), the diagnostic criteria were not described, and assessment of the diagnostic content in the selected studies was not conducted. The 2022 study reported significant improvements in HAMD, TER, SSS, 5-HT level, and BI as outcome measures that were similar to those of the present study: HAMD (MD: −4.56; 95%CI: −6.39, −2.74; I ² = 95% vs. MD −4.01, 95%CI −4.72, −3.30; I ² = 94%: random-effect), TER (RR: 1.21; 95%CI: 1.13, 1.29; I ² = 8% vs. RR: 1.21; 95%CI: 1.17, 1.25; I ² = 0%: common-effect), SSS (MD: −5.73; 95%CI: −9.86, −1.79; I ² = 92% vs. MD: −4.89; 95%CI: −7.47, −2.31; I ² = 97%: random-effect), BI (MD: 15.47; 95%CI: 12.89, 18.04; I ² = 13% vs. MD: 10.17; 95%CI: 9.58, 10.75; I ² = 97%: common-effect), and 5-HT (SMD: 5.11; 95%CI: 3.11, 7.12; I ² = 13% vs. SMD: 1.55; 95%CI: 1.33, 1.77; I ² = 97%: random-effect). However, the number of included studies for each outcome measure in the present study was significantly larger than that in the 2022 study: HAMD (36 vs. 11), TER (34 vs. 10), SSS (7 vs. 2), BI (8 vs. 2), and 5-HT (7 vs. 2), thus providing more precise effect estimates and helpful information for clinical judgment. Furthermore, the present study employed additional outcome measures such as NIHSS, FMA, MMSE, and PSQI, allowing for consideration of the complexity of PSD symptoms. Finally, subgroup analyses were conducted by applying various criteria to investigate whether the study outcomes differed among specific subgroups or remained robust.

The combination of Soyo-san and antidepressants demonstrated significant clinical effects beyond alleviating depression to include post-stroke functional recovery. In particular, the minimum clinically significant difference for the primary outcome, HAMD, has not been previously reported. The National Institute for Health and Care Excellence of the United Kingdom recommends a 3-point difference between the treatment and placebo groups to be considered clinically significant. Other studies suggest a 2-point difference indicates a clinical correlation (Kirsch et al., 2008; Montgomery and Möller, 2009). These reports compared treatment with placebo, which is not equivalent to this study. However, our result (MD: −4.01; 95%CI: −4.72, −3.30; I ² = 94%) can be considered clinically significant because it exceeds a 3-point difference. In TEAM, Soyo-san is prescribed to ‘Soothe the liver and relieve depression’ (疏肝解鬱), ‘fortify the spleen and nourish blood’ (健脾養血) on Liver qi stagnation with blood deficiency (肝鬱血虛), spleen failing in transportation (脾失健運) state provoked by emotional disorder (情志不暢). The antidepressant effects of Soyo-san have been observed in preclinical experiments demonstrating its ability to alleviate stress-related anxiety. Bupleurum chinense DC. [Apiaceae; Bupleuri Radix], a key ingredient in Soyo-san, has antidepressant properties, particularly when paired with Paeonia lactiflora Pall [Paeoniaceae; Paeoniae Radix] (Kwon et al., 2010; Zhang et al., 2020). Angelica sinensis (Oliv.) Diels [Apiaceae; Angelicae Sinensis Radix] helps regulate circulation and exhibits antidepressant effects. Wolfiporia cocos (Schw.) Ryvarden and Gilb. [Polyporaceae; Poria] (Fungus) enhances antidepressant function by regulating neurotransmission and decreasing inflammation in the brain (Gong et al., 2019; Huang et al., 2020). Glycyrrhiza uralensis Fisch. [Fabaceae; Glycyrrhizae Radix et Rhizoma] has neuroprotective effects, while Mentha arvensis L [Lamiaceae; Menthae Haplocalycis Herba], Zingiber officinale Roscoe [Zingiberaceae; Zingiberis Rhizoma Recens], and Atractylodes macrocephala Koidz. [Asteraceae; Atractylodis Macrocephalae Rhizoma] have antioxidant and anti-inflammatory properties (Koşar et al., 2004; Van Breemen et al., 2011; Kao et al., 2014; Hoang et al., 2016).

Patients with PSD experience various sequelae including motor impairment, cognitive impairment, and sleep disorders. Depression exacerbates their condition and reduces treatment compliance. Stroke survivors benefit from prompt and active rehabilitation to restore their ability to perform activities of daily living. PSD hinders recovery, worsens sequelae, and increases medication burden. Soyo-san may provide a comprehensive treatment that addresses this complex situation. Recent systematic reviews on SSRIs in post-stroke recovery reported improvements not only in preventing and treating depression but also in reducing anxiety and dependence in daily life and improving motor and cognitive function (Kalbouneh et al., 2022). However, SSRIs are associated with a higher risk of seizures than placebo. In this context, Soyo-san, when used in combination with SSRIs, demonstrated superior effects in improving various post-stroke sequelae, contributing to the recovery of daily life abilities. This suggests the potential of combination therapy to alleviate the burden of polypharmacy. Furthermore, the safety profile of combination therapy appears favorable, with fewer reported adverse effects than antidepressants alone. However, the effects of combined administration on seizures and increased risk of hemorrhage require further investigation.

The antidepressant effects observed with Soyo-san in this review are consistent with findings for other traditional herbal medicines. For instance, systematic reviews of Banxia-houpo-tang for depression (Kim et al., 2023) and Sihogayonggolmoryeo-tang for post stroke depression (Kwon et al., 2019) have also reported beneficial effects on depressive symptoms, although similarly limited by the methodological quality of primary studies. In particular, study of Sihogayonggolmoryeo-tang analysed effect on Barthel Index, same as our study. Situating Soyo-san within this broader context suggests a class of herbal formulas may target neuropsychiatric symptoms through shared anti-inflammatory and neuro-regulatory pathways, though direct comparative studies are lacking and urgently needed.

Regardless of the class of co-administered antidepressant, duration of administration, pattern identification, and dosage form, significant improvement in symptoms were observed based on the HAMD. Additionally, although not statistically significant, there was a trend toward better symptom improvement with administration >4 weeks and a preference for the order of dosage forms as pills, powders, and decoctions. Further research is needed to elucidate the potential differences in responses based on these factors.

Exploring the impact of severity classification, the study compared subgroups based on whether the HAMD criterion was applied during participant selection. Although not statistically significant, better outcomes were observed in the group for which the HAMD criterion was applied. Further subgroup analyses based on severity (mild, moderate) showed no significant differences; however, a tendency for better outcomes was observed in the moderate subgroup. This suggests that the co-administration of Soyo-san has a significant antidepressant effect, irrespective of severity, making it a potential treatment for various levels of PSD.

This study conducted a comprehensive search without language restrictions and employed various analysis methods, including subgroup analysis, meta-regression analysis, and GRADE, methods not implemented in the previous study, to evaluate the outcomes. Moreover, the study utilized a diverse set of outcome measures to confirm the positive effects of adjunctive therapy with Soyo-san, addressing not only PSD, but also overall post-stroke symptoms. The included studies exhibited unclear or inconsistent diagnostic criteria, limited geographic diversity, high risk of bias, and generally low levels of evidence. However, this study employed methods such as subgroup analysis, sensitivity analysis, and assessment of publication bias, which confirmed that the results remained stable and robust.

In geographic diversity, all included studies were conducted in China. As local healthcare practices and cultural frameworks may influence trial conduct and outcome selection, the findings may not generalize to other countries.

For heterogeneity shown from results, methodological quality (e.g., risk of bias in allocation concealment), variability in herbal medicine formulation (e.g., composition, dose), differences in stroke characteristics (e.g., severity, subtype, time since onset), diversity in co-interventions (e.g., type of antidepressants, rehabilitation, acupuncture) are considered to be associated. Exploratory analyses suggested few trends (e.g., longer treatment duration and decoction form were associated with greater reductions in HAMD), but residual heterogeneity remained high. Given the substantial heterogeneity, the pooled effect estimate should be interpreted with caution. We present it alongside sensitivity analyses and exploratory subgroup/meta-regression findings, but emphasise that the true effect may vary across contexts.

For The Total Effective Rate (TER), while widely used in Chinese clinical trials, lacks international validation and standardization. Its cultural specificity and composite nature limit comparability to globally accepted outcome measures, reducing generalizability of our findings. Functional outcomes such as Barthel Index and NIHSS were reported in ≤3 trials, providing insufficient evidence for firm conclusions. By contrast, HAMD was assessed in nearly all trials, albeit with high heterogeneity. While all included outcome measures showed statistically significant results, no validated Minimum Clinically Important Differences (MCIDs) were available to confirm their clinical relevance to PSD.

Adverse events were inconsistently reported, often without standardized severity grading. This underreporting prevents a balanced assessment of risks. Future trials should follow CONSORT-harms reporting to ensure adequate evaluation of safety.

The ConPhyMP assessment results highlight limitations regarding the heterogeneity and low reproducibility of herbal medicines in this review. Over half of the studies failed to report the basic description of their plant material, botanical authentication, and all of the included studies failed to provide chemical information. This deficiency reporting on plant origin, chemical identity, product quality information is associated to heterogeneity and low reproducibility in clinical outcomes. Our findings strongly underscore the need for future clinical trials of Soyosan to adhere to the ConPhyMP guidelines to ensure that systematic reviews and meta-analyses can draw reliable conclusions.

This study attempted subgroup analyses to explore the factors influencing the effects of Soyo-san combination therapy on PSD; however, no significant differences were observed. Additional research is needed to investigate potential differences based on covariates such as co-administered medication, treatment duration, baseline severity, the presence of pattern identification, and dosage form.

Limitations of the strictness of the diagnostic criteria and design of concomitant therapies for post-stroke management other than combined therapy were identified. This suggests participant selection may have been suboptimal, leading to potential distortions in outcome interpretation, including subgroup analyses. Hence, future RCTs should apply diagnostic criteria according to the DSM-5 and explicitly state stroke management for both patient and control groups.

Given the observed effects of Soyo-san beyond alleviating PSD, more proactive RCTs are needed to elucidate its effects on general post-stroke symptoms. Moreover, investigating whether Soyo-san positively addresses the issue of polypharmacy in patients with stroke is essential. Although this study confirmed that Soyo-san, when combined with antidepressants, shows superior effects in PSD, reduces the burden of adverse effects, and may resolve various symptoms related to polypharmacy, further research is required to define the specific conditions necessary to achieve these goals.