This study searched eight databases including PubMed, Cochrane Library, Embase, Web of Science, China Knowledge Infrastructure (CNKI), China Science Journal Database (VIP), Wanfang Database, and China Biomedical Literature Service System (CBM). Publications were limited to English and Chinese languages, with the search cutoff date set at July 21, 2025. Mesh terms were combined with free-text keywords for retrieval, with primary search terms including “Electroacupuncture,” “Stroke,” and “Depression.” The specific search strategy is detailed in Table 1 (using PubMed as an example). Search strategies for other databases and the PRISMA 2020 checklist are provided in the Supplementary material.

Inclusion criteria were established using the PICOS (Population, Intervention, Comparison, Outcome, Study Design) framework. Specific criteria were: (1) P: Participants with a confirmed diagnosis of PSD; (2) I: Treatment group receiving EA therapy; (3) C: Control group receiving conventional medication; (4) O: Using the HAMD as the primary indicator; (5) S: Study design being randomized controlled trials.

Exclusion criteria: (1) Studies with unclear diagnoses or comorbidities affecting outcome assessment; (2) Studies involving concurrent interventions; (3) Studies with non-representative outcome measures; (4) Not randomized controlled trials; (5) Studies lacking full-text access or incomplete data.

Two researchers independently screened and extracted data based on inclusion and exclusion criteria. Disagreements were resolved through consultation with a third researcher. The screening process involved: (1) Removing duplicate records using EndNote X9.3 literature management software. (2) Reviewing titles and abstracts to exclude reviews, theses, conference papers, scientific achievements, and other non-relevant literature. (3) Full-text review to determine eligibility.

After literature screening, the following information was extracted: (1) Basic literature information: author details, publication year, etc.; (2) Basic subject information; (3) Intervention method, intervention frequency, and treatment duration; (4) Outcome measures.

We used the Cochrane Risk of Bias tool (ROB 2.0) to assess the quality of included studies. This tool evaluates five domains: Bias in the randomization process, bias in deviation from the intended intervention, bias in missing outcome data, bias in outcome measurement, and bias in selective reporting of results. The bias in deviation from the intended intervention domain is further subdivided into two scenarios based on research objectives: evaluating the effectiveness of intervention assignment and evaluating the effectiveness of intervention adherence. Each domain contains multiple signal questions. When assessing the risk of bias in RCTs, researchers must make judgments and objectively answer these questions. Signal questions typically offer five response options: Yes (Y), Probably Yes (PY), Probably No (PN), No (N), and No Information (NI).

Following the recommendations of the Cochrane Collaboration and PRISMA guidelines, statistical analyses were conducted using Review Manager 5.4.1, reporting pooled risk ratios (RR) and mean differences (MD) with 95% confidence intervals (CI). Heterogeneity was quantified using the I² statistic. Heterogeneity was defined as low, moderate, or high based on I² values of 25, 50, and 75%, respectively. Publication bias for primary outcomes was visually assessed using funnel plots. The Egger regression test was applied to analyze publication bias.

To more accurately assess the robustness of the analysis results and the inclusion of bias in the studies, we shall employ Stata 18 to conduct sensitivity analyses and Egger’s test.

A total of 1,106 publications were retrieved from databases. After removing duplicates, 633 publications remained. Further screening based on title and abstract review yielded 27 publications. Full-text review of these 27 publications resulted in 22 eligible publications meeting inclusion criteria, all in Chinese. The specific screening process is illustrated in Figure 1.

A total of 20 studies were included (13–34). These comprised 22 individual research projects involving 1,761 patients. All studies were conducted in China across 12 provinces. All studies provided specific diagnostic criteria. The intervention group in each trial received EA. Among the control groups, 18 studies used fluoxetine as the control, two studies used sertraline tablets, one study used Flupentixol and Melitracen Tablets, and one study used sertraline hydrochloride. The intervention periods ranged from 4 weeks to 60 days across the 22 studies. Detailed information is presented in Table 2.

The Cochrane Risk of Bias tool (ROB 2.0) was used to assess the risk of bias in the included studies. Six studies used random number tables for allocation, three used computer software, 11 mentioned random allocation only, and two allocated participants based on order of arrival. One study employed an envelope method during randomization to ensure blinding. One study reported achieving double-blinding during the research process, one study reported achieving single-blinding, and the remaining studies mentioned blinding measures. This lack of consistent blinding protocols was the primary factor reducing study quality. Of the 22 studies, one was classified as “low risk,” one as “some concerns,” and 20 as “high risk.” Specific details are shown in Figures 2, 3.

We examined potential publication bias by plotting funnel plots, where symmetry indicates no publication bias. Figure 9A shows the funnel plot for HAMD, and Figure 9B shows the funnel plot for efficacy rate. As shown in Figure 9A, most studies cluster around effect sizes close to zero, suggesting consistent overall effect directions (no studies with markedly extreme effects). Observing the symmetry of the “funnel” reveals a degree of asymmetry in the distribution of study points (particularly in the region with larger standard errors, where points are skewed toward one side). This asymmetry suggests potential publication bias. As shown in Figure 9B, study points are primarily concentrated in the region where effect sizes are close to 1, indicating that the effect sizes of most studies are consistent in direction. Observing the symmetry of the “funnel” reveals a pronounced asymmetry in the distribution of research points—in regions with larger standard errors (i.e., small-sample studies), the points deviate from the symmetrical funnel shape, exhibiting a distinct “gap” at the bottom. This asymmetry suggests a higher likelihood of publication bias.

To accurately analyse publication bias, we employed the Egger test in Stata 18 to assess the combined effect size of the HAMD. The results indicate that, the Egger test p-value for HAMD scores was 0.15, and that for efficacy rate was 0.417, both exceeding the significance level of α = 0.05. This indicates no significant linear association between effect size and precision for either measure across the included studies. Furthermore, the scatter plot for the Egger test showed symmetrical distribution of individual studies around the regression line, with no evident skewing. These findings indicate that the included studies in this meta-analysis exhibit no statistically significant publication bias, with a low risk of publication bias. The pooled results for the improvement in HAMD scores and the difference in treatment response rates demonstrate good authenticity and reliability. This provides objective and robust evidence-based medical support for evaluating the clinical efficacy of electroacupuncture in treating post-stroke depression. See Figure 10 for details.

Observation of the meta-analysis results revealed considerable heterogeneity in the HAMD and SDS scores. We conducted sensitivity analyses on both datasets using Stata 18. The findings indicated varying degrees of stability in the outcomes across different endpoints within this study. The analysis using SDS as the outcome measure included six studies. Upon sequentially excluding individual studies, the pooled effect size estimates fluctuated between −1.87 and 0.35. In some instances, the effect size direction reversed and the confidence interval crossed the null line, indicating weak stability in the meta-analysis results for this measure and conclusions susceptible to interference from individual included studies. In contrast, the analysis using HAMD as the outcome measure included 18 studies. After excluding any single study, the pooled effect size consistently remained within the negative range of −0.76 to −0.43. Furthermore, all confidence intervals remained within the non-negligible threshold, with no reversal in effect size direction. This indicates robust stability and high reliability in the meta-analysis results for this measure. This further substantiates the robustness of the conclusion that electroacupuncture intervention demonstrates superior efficacy to conventional pharmacological interventions in ameliorating depressive symptoms (as measured by HAMD scores) in post-stroke patients, with minimal susceptibility to individual study data biases. See Figures 11A,B for details.

To further investigate the sources of heterogeneity, we conducted a subgroup analysis of HAMD-related studies according to treatment duration. The subgroup analysis revealed substantial heterogeneity in both studies with treatment durations not exceeding 4 weeks and those exceeding 4 weeks, suggesting that the sources of heterogeneity may be independent of treatment duration. See Figure 11C for details.

A review of relevant studies revealed inconsistencies between the MDs reported by Huang Shile, Long Haowen, Nie Bin, Xu Guoqing, and Zhu Genying and those from other studies (19, 23, 24, 28, 32). In Huang Shile’s study, the disease duration was significantly shorter than in other studies. In Long Haowen’s study, the control group received an uncertain dosage of 10–40 mg/day, which may have contributed to the heterogeneity. Regarding studies involving the Barthel Index, excluding Hong Zhenmei’s study yielded I² = 0%, p = 0.1. Reviewing this study revealed that the control group received sertraline tablets, whereas the other two trials used fluoxetine in their control groups, which may explain the heterogeneity.

We used the GRADE approach to assess the quality of evidence for this study. The meta-analysis results for HAMD and Barthel Index showed very low and low certainty, while meta-analyses for other indicators demonstrated high certainty. Detailed information is provided in Table 3.

Notably, the pooled results for HAMD scores and SDS in this study exhibited high heterogeneity (I² = 89%; I² = 97%), indicating significant clinical or methodological differences among studies. Sensitivity analysis revealed that inconsistencies in patient disease duration, drug dosage, and types were likely primary sources of heterogeneity. For example, the shorter disease duration in Huang et al.’s (19) study and the wider range of drug dosages in the control group of Long et al.’s (23) study may have impacted the consistency of results. Future studies should standardize intervention protocols and patient stratification to enhance the reliability of findings.

Regarding secondary outcomes, no significant differences were observed between EA and medication in improving the Barthel Index and SDS scores. This may stem from limited sample sizes or varying sensitivities of assessment tools across different dimensions of psychological and functional status. Furthermore, the Barthel Index primarily reflects activities of daily living, whereas the SDS is a self-report scale; its assessment of depression severity may differ from clinician-rated HAMD scores.

Furthermore, all included studies were conducted in China, and most exhibited suboptimal methodological quality, particularly with inadequate implementation of blinding and allocation concealment, potentially introducing implementation and measurement biases. The findings of this study are primarily applicable to the Chinese population and should be applied with caution to other regions or populations. Future research requires more high-quality, multicenter, large-sample randomized controlled trials to enhance the generalizability and strength of evidence.

Furthermore, this study applied more precise inclusion criteria for the experimental group. For the experimental group, the studies included in our analysis featured subjects receiving only EA treatment, while the control groups received only conventional drug therapy. This allows for a more precise analysis of the advantages and limitations of acupuncture compared to conventional therapies. Previous meta-analyses (44), however, appear to have applied broader criteria for intervention methods, incorporating studies using various acupuncture techniques such as EA, body acupuncture, auricular acupuncture, and scalp acupuncture. While this approach yielded more comprehensive results, it lacked the necessary precision.