This systematic review was conducted in accordance with the Cochrane Handbook for Systematic Reviews of Interventions (26) and is reported in compliance with the Preferred Reporting Items for Systematic reviews and Meta-Analyses (27). The study protocol was prospectively registered with PROSPERO (https://www.crd.york.ac.uk/prospero/, CRD420251035172) prior to data extraction.

A comprehensive search was conducted across eight electronic databases: PubMed, Embase, Cochrane Central Register of Controlled Trials (CENTRAL), Web of Science, China National Knowledge Infrastructure (CNKI), Wan Fang, Chongqing VIP Chinese Science and Technology Periodical Database (CQVIP), and China Biomedical Literature Database (CBM), from inception to March 16, 2025. To minimize selection bias, supplementary sources were explored, including https://ClinicalTrials.gov, the Chinese Clinical Trial Registry (ChiCTR), the International Traditional Medicine Clinical Trial Registry (ITMCTR), and gray literature (e.g., dissertations). Reference lists of included studies were manually screened to identify additional eligible trials.

The search strategy combined Medical Subject Headings (MeSH) and free-text terms across four domains: (1) Population: “aged,” “elderly,” “geriatric,” “perioperative neurocognitive disorder,” “postoperative cognitive dysfunction,” “delirium,” “cognitive decline.” (2) Intervention: “electroacupuncture,” “acupuncture,” “EA.” (3) Context: “general anesthesia,” “surgery,” “postoperative period.” (4) Study design: “randomized controlled trial,” “RCT.” Boolean operators (AND/OR/NOT) and truncation symbols (*) were applied to optimize sensitivity and specificity (see Supplementary Table S1 for full search syntax).

The EndNote X20 (Clarivate Analytics) was used for deduplication, and then two independent reviewers (CW and XW) screened the titles, abstracts and full texts of studies based on eligibility criteria, in order to eliminate irrelevant publications and retain eligible studies. Discrepancies were resolved by a third reviewer (KW). A standardized extraction template (28) was developed, capturing: (1) Study characteristics (first author, publication year, nationality, etc.); (2) Participant demographics (age, sex ratio, sample size, type of surgery, etc.); (3) Intervention details [acupoints, stimulation parameters (waveform, frequency, intensity), timing (pre-/intra-/postoperative), session duration]; (4) Outcomes (evaluation indicators, evaluation time, assessment tools, diagnostic criteria for outcome events). Missing data were requested from corresponding authors via email; unresolved omissions led to exclusion from quantitative synthesis.

Two investigators (CW and XW) independently assessed the risk of bias of included literatures according to the Version 2 of the Cochrane risk-of-bias tool for randomized trials (RoB 2) (29). The following contents were assessed: random sequence generation, allocation concealment, blinding of participants and personnel, blinding of outcome assessment, incomplete outcome data, selective reporting, and other bias. Each domain was rated as “low,” “high,” or “unclear” risk by two reviewers (CW and XW), with arbitration by KW.

Two researchers (CW and XW) independently assessed the evidence certainty for each critical outcome using the Grading of Recommendations, Assessments, Developments and Evaluations (GRADE) approach (30), yielding classifications of “high,” “moderate,” “low,” or “very low” confidence. Disagreements were resolved by KW.

RevMan version 5.4 (Cochrane Collaboration, Oxford, UK) was used for meta-analysis. Mean differences (MDs) or standardized mean differences (SMDs) with 95% confidence intervals (CI) for continuous variables, and relative risk (RR) with 95% CI for dichotomous variables. Heterogeneity was quantified via I² statistic and Q test. I² ≤ 50% and p ≥ 0.1 indicated acceptable heterogeneity, and a fixed-effect model was selected for meta-analysis. Conversely (I² > 50% or p < 0.1), it indicated the presence of obvious heterogeneity, and a random-effects model was elected (31). Statistical significance was defined as two-tailed p < 0.05. The pre-specified subgroup analyses were conducted according to the EA intervention time, evaluation timepoints of outcome index, the number of EA intervention sessions, and surgical types to explore the reasons for heterogeneity. In addition, sensitivity analyses were conducted by excluding low-quality studies with high risk of bias per round to evaluate the outcome stability and investigate the sources of inconsistency among included literatures. When the number of included studies for an outcome indicator exceeded 10, funnel plot asymmetry test and Egger’s regression test were used to assess publication bias (32). p < 0.05 indicated a publication bias.

A total of 2,776 publications (2,525 Chinese and 251 English) were initially identified through database searches. After removing duplicates, the articles were screened by reading the titles, abstracts, and full texts, ultimately retaining 26 eligible studies (16, 18, 19, 33–55) for analysis. Supplementary searches of clinical trial registries (https://ClinicalTrials.gov, ChiCTR, ITMCTR) and reference lists of included studies retrieved no additional eligible publications. The selection process rigorously adhered to PRISMA guidelines, with a detailed flowchart outlining inclusion/exclusion decisions provided in Figure 1.

All 26 included trials were two-arm parallel studies conducted in China and published between 2011 and 2023. A total of 2,309 participants were included, with 1,150 in the EA group and 1,159 in the control group. Subjects spanned multiple specialties: 11 orthopedic surgery (18, 19, 35, 37, 40, 42, 43, 45, 47, 52, 53, 55), six gastroenterological surgery (16, 33, 36, 38, 39, 49), two hepatobiliary surgery (50, 51), two urological surgery (44, 48), four mixed types (34, 37, 41, 42, 46), and one unspecified study (54). Regarding the EA intervention timing, seven studies used EA before surgery, 11 during surgery, three after surgery, one both before and during surgery, three both before and after surgery, and one combined EA before, during, and after surgery. The acupoints used with a frequency of 3 or more times were Baihui (GV20, 21 studies), Neiguan (PC6, 18 studies), Zusanli (ST36, 12 studies), Hegu (LI4, 8 studies), Shenting (GV24, 5 studies), Sanyinjiao and Taichong (SP6, LR3, 4 studies each), and Sishencong and Shenmen (EX-HN1, HT7, three studies each). The waveforms of EA intervention were mostly chosen as dense-disperse wave (21 articles), with the frequency often set at 2/100 Hz (nine articles), and the stimulation intensity was generally based on patient tolerance (17 articles). The specific characteristics of the included literatures are shown in Supplementary Table S2.

Among the 26 included studies, 18 studies (16, 18, 19, 33–42, 44–47, 52, 54) (69.2%) used a random number table, one study (49) (3.9%) employed the coin toss, and seven studies (37, 42, 43, 48, 50, 51, 53, 55) (26.9%) stated “randomized” without method specification. Only two studies (18, 19) (7.7%) adequately described the methods of allocation concealment. Twenty-one studies (16, 33–45, 47–49, 52–55) (80.8%) were rated as high risk of bias due to lack of blinding of participants and personnel, whereas the remaining five studies (18, 19, 46, 50, 51) (19.2%) demonstrated low risk through implementation of sham EA controls. Only two studies (37, 42, 53) reported outcome blinding assessment (7.7%), others were unclear. All included studies (100%) demonstrated low risk of incomplete outcome data. Unavailable protocols (100%) precluded assessment of the selective reporting bias. No other sources of bias were identified across the studies (100%) (Figure 2).

Twenty-two studies (18, 19, 33, 34, 36–39, 41–53, 55); (n = 3,658) reported PND incidence after excluding one study (54) with unclear diagnostic criteria. Timepoints with single-study reporting (at 6, and 12 h postoperatively/on postoperative days 6 and 14) were excluded from pooled analysis to maintain statistical reliability. Perioperative EA significantly reduced overall PND incidence versus controls (RR = 0.47, 95% CI: 0.42 to 0.54, p < 0.00001; I² = 0%). Subgroup analysis by postoperative day (POD) revealed consistent reductions. POD 1 (16 studies): RR = 0.52, 95% CI: 0.43 to 0.62, p < 0.00001; I² = 0%. POD 2 (two studies): RR = 0.52, 95% CI: 0.36 to 0.74, p = 0.0003; I² = 50%. POD 3 (14 studies): RR = 0.42, 95% CI: 0.32 to 0.55, p < 0.00001; I² = 0%. POD 4 (two studies): RR = 0.40, 95% CI: 0.24 to 0.67, p = 0.0004; I² = 26%. POD 5 (two studies): RR = 0.43, 95% CI: 0.21 to 0.91, p = 0.03; I² = 0%. POD 7 (seven studies): RR = 0.49, 95% CI: 0.27 to 0.90, p = 0.02; I² = 0% (all analyses used fixed-effect model; Figure 3).

Given the limited number of available studies, we conducted subgroup analyses exclusively for the PND incidence on POD 1 and POD 3, MMSE scores on POD 1 and POD 3, and serum levels of IL-6, IL-1β, TNF-α, and S100β on POD 1. These analyses were stratified according to predefined characteristics: EA intervention timing, comparators, surgical categories, and EA sessions. The results are presented in Supplementary Table S4.

The EA-related reduction in PND incidence on POD 1 and POD 3 was independent of intervention timing. Preoperative (POD 1: RR = 0.55, 95% CI: 0.39 to 0.77, p = 0.0004, I² = 0%; POD 3: RR = 0.46, 95% CI: 0.26 to 0.83, p = 0.01, I² = 0%), intraoperative (POD 1: RR = 0.42, 95% CI: 0.28 to 0.63, p < 0.0001, I² = 0%; POD 3: RR = 0.52, 95% CI: 0.37 to 0.75, p = 0.0003, I² = 0%), postoperative (POD 1: RR = 0.61, 95% CI: 0.43 to 0.87, p = 0.006, I² = 0%; POD 3: RR = 0.33, 95% CI: 0.16 to 0.70, p = 0.004, I² = 0%), preoperative-postoperative combined EA (POD 1: RR = 0.49, 95% CI: 0.32 to 0.75, p = 0.001, I² = 0%; POD 3: RR = 0.22, 95% CI: 0.10 to 0.48, p = 0.0002, I² = 24%) consistently demonstrated significant lower PND incidence. Moreover, the reduction in PND incidence on POD 1 and POD 3 was consistent regardless of comparators, demonstrating significant effectiveness against both sham EA (POD 1: RR = 0.64, 95% CI: 0.47 to 0.86, p = 0.004, I² = 0%; POD 3: RR = 0.45, 95% CI: 0.27 to 0.74, p = 0.002, I² = 0%) and no-intervention controls (POD 1: RR = 0.46, 95%CI: 0.36 to 0.59, p < 0.00001, I² = 0%; POD 3: RR = 0.44, 95% CI: 0.32 to 0.60, p < 0.00001, I² = 0%). The operation-stratified analysis revealed that EA significantly reduced PND incidence on POD 1 for orthopedic surgery (RR = 0.50, 95% CI: 0.38 to 0.66, p < 0.00001; I² = 0%) and gastrointestinal surgery (RR = 0.57, 95% CI: 0.43 to 0.75, p < 0.0001; I² = 0%), whereas no significant superiority in urological procedure (RR = 0.50, 95% CI: 0.25 to 1.00, p = 0.05; I² = 0%), potentially attributable to limited study availability (n = 2). For PND incidence on POD 3, EA exhibited consistent reductions in orthopedic (RR = 0.36, 95% CI: 0.21 to 0.59, p < 0.0001; I² = 47%), urological (RR = 0.25, 95% CI: 0.07 to 0.84, p = 0.03; I² = 0%), and gastrointestinal surgery (RR = 0.47, 95% CI: 0.34 to 0.65, p < 0.00001; I² = 0%), suggesting neuroprotective benefit independent of surgical types.

EA intervention time is a key source of heterogeneity of MMSE scores on POD 1 and POD 3 (subgroup differences: p = 0.008; p = 0.002). The combination of preoperative and postoperative EA showed a significant improvement (MD = 3.51, 95% CI: 2.64 to 4.37, p < 0.00001; I² = 0%) in MMSE scores on POD 1. Preoperative (MD = 1.59, 95%CI: 0.48 to 2.69, p = 0.005), intraoperative (MD = 1.85, 95% CI: 1.38 to 2.33, p < 0.00001), or postoperative (MD = 4.20, 95% CI: 1.17 to 7.22, p = 0.007) EA revealed moderate improvement in MMSE scores on POD 1, but the high heterogeneity (I² = 83%; I² = 93%; I² = 97%) suggested potential confounding factors, such as surgical types, anesthesia drugs, patient demographics or EA parameters. Further homogeneous studies are needed to validate these findings and clarify the optimal timing of EA for cognitive protection. Preoperative (MD = 1.84, 95% CI: 1.05 to 2.64, p < 0.00001), postoperative (MD = 3.50, 95% CI: 2.94 to 4.06, p < 0.00001), preoperative-postoperative combined EA (MD = 3.07, 95% CI: 2.40 to 3.74, p < 0.00001) exhibited a significant improvement and consistent effects (I² = 0%) in MMSE scores on POD 3. Notably, postoperative EA intervention demonstrated superior effectiveness (MD = 3.50). Intraoperative EA showed mild improvement and more variable effects in MMSE scores on POD 3 (MD = 1.69, 95% CI: 0.45 to 2.92, p = 0.007; I² = 95%), warranting further investigation.

The number of EA sessions may be a significant and partial source of heterogeneity in serum IL-6 and IL-1β levels on POD 1 (between-subgroup difference: p = 0.05; p < 0.00001). Multiple EA sessions (≥ 2) produced a marked reduction in IL-6 (SMD = −1.65, 95% CI: −2.07 to −1.23, p < 0.00001; I² = 0%) and IL-1β (SMD = −5.37, 95%CI: −6.09 to −4.65, p < 0.00001; I² = 17%) levels on POD 1, while single EA session demonstrated no intergroup difference. These findings indicate that repeated EA interventions may exert cumulative anti-inflammatory effects.

EA intervention timing significantly contributed to heterogeneity in serum TNF-α levels on POD 1 (p < 0.00001). Intraoperative EA demonstrated a significant reduction in serum TNF-α levels on POD1 (SMD = −1.07, 95% CI: −1.43 to −0.72, p < 0.00001; I² = 22%). Surgical type substantially influenced outcome heterogeneity of serum S100β levels on POD 1 (p < 0.0001). While EA intervention did not significantly decrease S100β levels on POD 1 in orthopedic surgery patients (SMD = −0.21, 95%CI: −0.56 to 0.13, p = 0.22; I² = 33%), it demonstrated beneficial effects across multiple surgical types (SMD = −3.71, 95% CI: −5.91 to −1.51, p = 0.0009; I² = 96%). This variability likely reflects differences in the degree of cerebral injury associated with distinct surgical procedures.

To assess the robustness of the findings, sensitivity analyses were performed by sequentially excluding low-quality studies with high risk of implementation bias. For the incidence of PND on postoperative days 1 and 3, statistical significance and effect direction remained unchanged after exclusion of high-bias studies, indicating high result stability. After excluding the study by Gu et al. (35), the statistical significance shifted in MMSE scores on POD 1 (from p < 0.00001 to p = 0.05), suggesting result instability due to low-quality studies. Exclusion of Han et al.’ study (36) altered the significance level in MMSE scores on POD 3 (from p < 0.00001 to p = 0.07), further indicating methodological bias influenced the outcome (Supplementary Table S5).

Eight studies (16, 18, 19, 33–35, 40, 52) reported adverse events, including nausea and vomiting, minor acupuncture-induced hematoma, headache, dizziness, and venous thrombosis. No significant heterogeneity was observed across studies (I² = 0%, p = 0.42), and a fixed-effect model was applied for analysis. The pooled results demonstrated that the EA group had a significantly lower incidence of adverse events compared to the control group (RR = 0.52, 95% CI: 0.37 to 0.72, p < 0.0001; I² = 0%) (Supplementary Figure 1).

For postoperative day 1 PND incidence (16 studies), postoperative day 3 PND incidence (14 studies), and postoperative day 1 MMSE scores (18 studies), the funnel plots appeared symmetrical, and Egger’s test yielded no significant publication bias (p = 0.052, p = 0.063, and p = 0.087, respectively). However, for postoperative day 3 MMSE scores (13 studies), the funnel plot exhibited noticeable asymmetry, and Egger’s test indicated significant publication bias (p = 0.001) (Figure 5).

The overall certainty of evidence regarding the effectiveness of EA on relevant outcomes is summarized in Table 1. Moderate certainty evidence indicates that EA is associated with reduced incidence of PND on POD 1 and POD 3, and lower incidence of adverse events. Low or very low certainty evidence suggests that EA may be linked to decreased PND incidence on POD 2, POD 4, POD 5, and POD 7; improved postoperative cognitive function; and alleviated peripheral inflammatory responses (IL-6, IL-1β, and TNF-α) and cerebral injury severity (S100β).

Perioperative neurocognitive disorder (PND) imposes substantial clinical and socioeconomic burdens on elderly surgical patients, adversely affecting both immediate recovery and long-term outcomes (1, 56). This meta-analysis demonstrates that perioperative electroacupuncture (EA) significantly reduces PND incidence (POD 1, 2, 3, 4, 5, and 7) and enhances early cognitive recovery (MMSE at 1 h postoperatively, POD1, POD3, and POD 4) with a favorable safety profile. Low-to-moderate certainty evidence indicates EA’s neuroprotective potential through attenuation of systemic inflammation (reduced IL-1β, IL-6, TNF-α) and mitigation of neurological injury (lowered S100β) on POD 1. Crucially, PND reduction on POD 1 and POD 3 remained robust across EA timing, comparator types, and surgical categories, underscoring broad clinical applicability.

Subgroup analyses elucidated the critical influence of EA intervention timing and session frequency on efficacy. Combined preoperative-postoperative EA significantly improved MMSE scores on POD 1 (p < 0.00001), whereas postoperative-only EA enhanced MMSE scores on POD 3 (p < 0.00001). These findings support prioritizing either combined or postoperative-only EA as optimal timing strategies for cognitive protection in elderly patients under general anesthesia. A dose–response relationship was observed between EA sessions and anti-inflammatory effects, validating multi-session protocols (≥ 2 sessions) for maximal suppression of postoperative inflammation (serum IL-1β/IL-6 reduction; p < 0.00001).

Given inherent variability in anesthetic protocols and surgical trauma, procedure-specific subgroup analyses revealed differential neuroprotective effects of EA. Serum S100β reduction varied significantly by surgical category, suggesting procedure-related brain injury severity modulates EA efficacy. Sensitivity analysis sequentially excluding nonblinded trials demonstrated compromised robustness of MMSE improvements on PODs 1 and 3, indicating potential overestimation of EA effects in unblinded studies due to measurement bias or placebo effects. These results emphasize the necessity of rigorous blinding and standardized outcome assessments.

This work represents the first comprehensive evaluation of EA for PND prevention specifically in elderly patients under general anesthesia, overcoming limitations of prior reviews focused on transcutaneous electrical acupoint stimulation (TEAS) (20–25), mixed anesthesia techniques (15), or heterogeneous populations (57). Through predefined subgroup analyses and sensitivity testing, we identified previously overlooked heterogeneity sources and established a multidimensional evidence chain integrating clinical outcomes (PND incidence), cognitive metrics (MMSE), and mechanistic biomarkers (inflammatory/neurological markers), thereby elucidating EA’s “intervention-efficacy-mechanism” relationship.

Current evidence indicates that surgical trauma and stress-induced systemic inflammatory response syndrome can promote and exacerbate neuroinflammation through various mechanisms, including increased permeability of the blood–brain barrier and activation of glial cells. This neuroinflammatory cascade subsequently results in synaptic dysfunction, neuronal impairment and death, disrupted hippocampal neurogenesis, and compromised synaptic plasticity. Ultimately, these alterations contribute to the development of PND (58), as illustrated in Figure 6. EA’s neuroprotection aligns with established biological pathways. Preclinical evidence indicates acupuncture modulates neuroinflammation and systemic inflammatory responses (10, 59), enhances synaptic plasticity (10), inhibits microglial activation (13, 14), prevents neuronal apoptosis (14), and improves cognitive function (10, 13, 14, 60, 61). Neuroimaging studies further demonstrated acupuncture-mediated activation of cognition-related brain regions (e.g., prefrontal cortex) and strengthened functional neural connectivity (10, 60, 61). Our results suggest that perioperative EA may enhance cognitive function by potentially by mitigating the systemic inflammatory response and reducing neural damage, thereby decreasing the incidence of PND (Figure 6).

Beyond direct neuroprotection, EA mitigates preoperative anxiety, reduces intraoperative anesthetic requirements, and attenuates surgical stress responses—collectively contributing to PND prevention (1, 62, 63). Furthermore, EA harnesses the synergistic advantages of acupoint specificity, electrical stimulation, and cognitive protection. Unlike pharmacological interventions, which often exhibit limited efficacy and are associated with inherent side effects, EA offers a favorable safety profile, minimal invasiveness, and cost-effectiveness, significantly enhancing its appropriateness for perioperative application.

Despite promising results, several limitations merit consideration: Firstly, all included studies were conducted in China, limiting generalizability to diverse populations with varying cultural acceptance of EA. Secondly, high statistical heterogeneity (I² = 79–98%) arose from variability in EA parameters (waveforms, frequencies), surgical procedures, and outcome assessment protocols. Thirdly, inadequate blinding (80.8% of studies) and allocation concealment (92.3%) undermine internal validity. Lastly, EA-related adverse events (e.g., hematoma, syncope) were inconsistently documented.

To address these gaps, we advocate for: (1) Multicenter RCTs: Stratified by surgical type and patient demographics, adhering to Consolidated Standards of Reporting Trials (CONSORT) (64) and Standards for Reporting Interventions in Clinical Trials of Acupuncture (STRICTA) (65) guidelines. (2) Standardized Protocols: Explicit reporting of EA parameters (e.g., waveform: 2/100 Hz; intensity: patient tolerance-adjusted) and anesthesia regimens. (3) Extended Follow-up: Longitudinal assessments (≥30 days) to evaluate sustained cognitive benefits. (4) Mechanistic Studies: Integration of neuroimaging and biomarker analyses to clarify multimodal neuroprotective pathways. (5) Rigorous Controls: Implementation of sham acupuncture per Sham Acupuncture Reporting Guidelines and Checklist for Clinical Trials (SHARE) (66) to isolate EA-specific effects. Future trials should either focus on single surgical types to minimize heterogeneity or conduct stratified analyses to delineate surgery-specific EA efficacy. Protocol optimization accounting for these factors is critical for enhancing EA’s therapeutic potential.