To ensure comprehensive literature retrieval, this study systematically searched eight databases, including PubMed, Web of Science, Embase, the Cochrane Library, ClinicalTrials.gov, China National Knowledge Infrastructure (CNKI), VIP Database, Wanfang Data, the Chinese Biomedical Literature Database (CBM), and the Chinese Clinical Trial Registry. The search period spanned from each database’s inception to March 21, 2025. To ensure an effective search, a combination of Medical Subject Headings (MeSH) and free words will be used. Boolean operators “AND” and “OR” were strategically combined between search terms. The search strategy was adapted to each database’s specific requirements, with PubMed serving as the representative example. Detailed search strategies are provided in Table 1.

The search results from each database were imported into EndNote (version X21). First, duplicate records across databases were removed. Second, preliminary exclusion of irrelevant studies was performed by reviewing titles and abstracts. Finally, full-text articles were assessed to further exclude studies unrelated to this research. The extracted data included: (1) general publication information (author names, publication year, country, sample size of included studies); (2) patient characteristics (age, sex, days since ischemic stroke onset, etc.); (3) specific details of interventions in both treatment and control groups (treatment duration, therapeutic methods); (4) outcome measure details (assessment metrics). The screening and data extraction were conducted independently by two researchers. Any discrepancies were resolved through consultation with a third-party reviewer.

Two researchers independently assessed the risk of bias in included RCTs using the ROB 2.0 tool. The evaluation covered five domains: (1) randomization process, (2) deviations from intended interventions, (3) missing outcome data, (4) outcome measurement, (5) selection of reported results. The two researchers then cross-verified their assessments. Any discrepancies were resolved through consultation with a third researcher to reach consensus on the study's inclusion.

Considering that neural recovery capacity may vary significantly with the stage of stroke and the severity of upper limb motor impairment, this study further conducted subgroup analyses based on baseline characteristics. Participants were stratified according to the stage of stroke, defined as hyperacute/acute phase (within 1 week after onset), subacute phase (within 12 weeks after onset), and chronic phase (more than 12 weeks after onset) (Joy and Carmichael, 2021). They were also classified by severity of upper limb motor dysfunction using the FMA-UE, with mild impairment corresponding to FMA-UE scores of 43–66, moderate impairment to scores of 29–42, and severe impairment to scores of 0–28 (Woytowicz et al., 2017).

For each subgroup, a Bayesian network meta-analysis consistent with the overall analysis was performed to integrate direct and indirect evidence. For the primary outcome measure FMA-UE, the differences in optimal intervention duration across various subgroups were individually assessed, with results visualized through forest plots and cumulative ranking probability plots.

When I² ≥ 50%, sensitivity analyses were conducted using the one-study-removed method to identify potential sources of heterogeneity.

We generated a funnel plot to assess publication bias.

This study employed the GRADE approach to evaluate the quality of evidence. Given that all included studies were randomized controlled trials, the evidence quality was primarily assessed based on five downgrading factors: risk of bias, inconsistency, indirectness, imprecision, and publication bias. The certainty of evidence was classified four levels: high, moderate, low or very low. The summary of findings tables was prepared as a summary of the certainty of evidence using GRADEpro V.3.6.1.

Following initial database searches, we identified 7,444 potentially relevant studies. After removing 3,451 duplicate records, 3,993 articles remained for title/abstract screening. This yielded 122 articles for full-text evaluation, from which 55 were excluded. Ultimately, 67 RCTs met inclusion criteria for this systematic review (Li L. Y. et al., 2021, 2022; Cui et al., 2021; Xu et al., 2015; Wang J. M. et al., 2021; Wang et al., 2023; Zhang et al., 2024, 2010; Xu, 2020; Yan, 2022; He et al., 2023; Hsieh et al., 2007; Gong and Xie, 2022; Bi et al., 2022; Wei et al., 2023; Liu et al., 2005; Lv et al., 2003; Ji, 2023; Xu et al., 2019; Yan et al., 2017; Wu et al., 2022; Yang et al., 2007; Liu, 2024; Bai et al., 2013; Tang et al., 2016; Yang, 2019; Zhong, 2021; Hou and Xu, 2022; Chen and Guan, 2022; Yu and Tian, 2023; Lang et al., 2013; Sun and Tu, 2024; Cao, 2022; Huang and Ou, 2021; Sun, 2020; Yin, 2021; Chen et al., 2014; Ding, 2024; Shen and Zou, 2025; Han and Shi, 2023; Wang, 2021; Pan and Jin, 2021; Xian et al., 2022; Ban et al., 2019; Kong, 2023; Xu, 2017; Ye et al., 2019; Xiong et al., 2020; Song, 2019; Zhang and Wang, 2019; Sang et al., 2023; Guo and Liu, 2022; Wang and Gong, 2022; Hou and Gao, 2024; Wang, 2022, 2024; Lu et al., 2025; Sha and Ma, 2020; Wu and Zhang, 2016; Yan et al., 2024; Qian, 2023; Liu et al., 2019; Li C. Y. et al., 2022; Zhou et al., 2023; Wang F. Q. et al., 2021; Fan et al., 2023; Li and Yuan, 2024). The selection flowchart is presented in Figure 1.

The analysis ultimately incorporated 67 studies, comprising 63 RCTs published in Chinese and 4 in English. All studies were published between 2003 and 2025, involving 5,635 participants. Sample sizes ranged from 20 to 200 participants, with most trials employing 1:1 allocation ratio. Patient ages spanned 22–85 years (except one RCT without specified age range). The included studies investigated various interventions comprising manual acupuncture, electroacupuncture, warm acupuncture, sham acupuncture, and their combinations with conventional therapy. The evaluated treatment courses spanned 2, 3, 4, 6, 8, 9, and 12 weeks. For outcome assessment, the Fugl-Meyer Assessment Upper Extremity (FMA-UE) scale was employed in 67 studies, the National Institutes of Health Stroke Scale (NIHSS) in 15 studies, and the Modified Barthel Index (MBI) in 12 studies. Complete details and results of all included studies are presented in Supplementary file S2.

Among the 67 included studies, 11 studies (Cui et al., 2021; Xu, 2020; Yan et al., 2017; Zhong, 2021; Xian et al., 2022; Kong, 2023; Zhang and Wang, 2019; Wang and Gong, 2022; Sha and Ma, 2020; Liu et al., 2019; Wang F. Q. et al., 2021) only mentioned “randomization” without specifying the method and were therefore judged to have some risk of bias. The remaining studies described specific randomization methods such as random number tables or lottery and were rated as low risk of bias. Only one study (Xiong et al., 2020) implemented double-blinding, while the others did not report blinding of researchers. Considering the feasibility of blinding in acupuncture interventions, these studies were assessed as having some risk of bias due to deviations from intended interventions. Three studies (Li et al., 2021; Wang et al., 2023; Ban et al., 2019) had cases of participant dropout and were thus rated as high risk of bias, while five studies (Li et al., 2021; Zhang et al., 2010; Hsieh et al., 2007; Xiong et al., 2020; Wang et al. F. Q., 2021) used third-party outcome assessors and were judged as low risk of bias in outcome measurement. One study (Xiong et al., 2020) was rated as low risk of bias across all domains, three studies (Li et al., 2021; Wang et al., 2023; Ban et al., 2019) were assessed as high risk overall, and the remaining studies were judged to have some concerns of bias overall. The risk of bias assessments for included studies are presented in Figure 2, the detailed assessment results are shown in Supplementary file S3.

Sensitivity analysis was conducted in this study by sequentially excluding individual studies from each high-heterogeneity subgroup.

In FMA-UE, the pooled MD and I² for the 4 W and 8 W groups remained within a relatively stable range after sequential exclusion of individual studies, indicating robust results. After excluding the study by Wang (2021) from the 6-week course group, the pooled MD was 5.21 (95% CI: 4.34–6.07, p < 0.001), and the I² value decreased from 98 to 0%; after excluding the study by Yan et al. (2024) in the 12-week course group, the pooled MD was 5.38 (95% CI: 4.27–6.48, p < 0.001) with I² declining from 87 to 16%, with both groups showing significantly reduced heterogeneity and maintaining statistically significant differences compared to controls. In NIHSS, after excluding the study by Gong and Xie (2022) in the 4-week course group, the pooled MD was 2.05 (95% CI: 1.35–2.75, p < 0.001) with I² decreasing from 82 to 44%; after excluding Pan and Jin (2021) in the 6-week course group, MD was 2.42 (95% CI: 1.73–3.11, p < 0.001) with I² dropping from 96 to 0%; after excluding Sang et al. (2023) in the 8-week course group, MD was 4.50 (95% CI: 3.49–5.50, p < 0.001) with I² reducing from 94 to 1%; and after excluding Li C. Y. et al. (2022) in the 12-week course group, MD was 5.33 (95% CI: 4.20–6.46, p < 0.001) with I² declining from 86 to 0%, with all groups maintaining statistically significant differences compared to control. In MBI, after excluding the study by Huang and Ou (2021) in the 4-week course group, the pooled mean difference was 7.24 (95% CI: 3.36–11.13, p < 0.001), with I² decreasing from 92 to 83%, indicating reduced heterogeneity and a statistically significant difference compared to the control group (see Supplementary file S5).

Publication bias assessment was conducted for the outcome measures (FMA-UE, NIHSS, and MBI) across the 67 included studies. The funnel plot for FMA-UE demonstrated good symmetry, with most studies clustered in the upper portion, indicating minimal publication bias; however, some data points located at the bottom and outside of the funnel suggested possible small-study effects and potential heterogeneity. The funnel plots for NIHSS and MBI demonstrated poorer symmetry, suggesting potential publication bias. As shown in Supplementary file S8.

The evidence quality for different acupuncture courses in treating PIS-ULMD was as follows: low for upper limb motor function improvement (FMA-UE), very low for neurological impairment (NIHSS), and very low for quality of life (MBI), with the overall evidence quality ranging from low to very low. This was primarily due to limitations, inconsistency, and publication bias (detailed GRADE assessments are provided in Table 2 and Supplementary file S9).

In this study, five publications explicitly documented the monitoring of adverse events (AEs) associated with acupuncture (He et al., 2023; Shen and Zou, 2025; Xu, 2017; Wu and Zhang, 2016; Yan et al., 2024). Among these, three studies reported no occurrence of AEs, while two randomized controlled trials (RCTs) recorded a total of eight AEs in the acupuncture treatment groups. One trial reported two cases of gastrointestinal reactions, two cases of abnormal liver function, and two cases of palpitation (He et al., 2023). Another trial documented two cases of infection at the acupuncture site (Wu and Zhang, 2016). The remaining studies did not explicitly mention safety outcomes or monitoring.

This study included 67 RCTs involving 5,635 PIS-ULMD patients across 7 different acupuncture courses, employing network meta-analysis to compare the effects of various intervention periods on upper limb motor function recovery using three outcome measures: FMA-UE (reported in 67 studies), NIHSS (15 studies), and MBI (12 studies).

Our findings demonstrate that acupuncture significantly improves upper limb motor function in PIS-ULMD patients compared to conventional therapy, with varying clinical efficacy across different treatment courses. For the primary outcome FMA-UE is the most widely used clinical measure for post-stroke upper limb motor recovery (Li et al., 2025), the optimal improvements occurred with 8-week course, 4-week course, and 2-week course. Secondary outcomes showed: NIHSS improvements were best with 12-week course, 6-week course, and 4-week course, while MBI improvements peaked with 8-week course, 4-week course, and 9-week course regimens. The observed discrepancy between NIHSS and FMA-UE results may stem from greater baseline NIHSS variability and milder neurological impairment in 8-week course groups, resulting in smaller measurable improvements. The collective evidence suggests 8-week acupuncture course may represent the optimal duration for simultaneously enhancing upper limb function and quality of life in PIS-ULMD patients.

In FMA-UE, heterogeneity significantly decreased in both the 6-week course and 12-week course groups after excluding Wang (2021) and Yan et al.’s (2024) studies, respectively. This primarily stemmed from notable discrepancies in baseline characteristics between patients in these two studies and others within their respective groups. Additionally, Yan et al. (2024) employed a special “hand and foot acupuncture with twelve needles “acupuncture protocol, potentially amplifying intervention effect differences. In NIHSS, high heterogeneity in 4-week course, 8-week course, and 12-week course groups was also linked to inconsistent baseline neurological deficit severity, particularly significant baseline differences between patients in Gong and Xie (2022), Sang et al. (2023), and Li C. Y. et al.’s (2022) studies versus others in their groups. In the 6-week course group, Pan and Jin (2021) included patients with longer disease duration, who were in the chronic phase and had poorer neurological recovery potential, leading to greater divergence in outcomes from other studies in the same group. For MBI, high heterogeneity in the 4-week course group similarly stemmed from Huang and Ou (2021) study, where patients were also in the chronic phase, and significant differences in disease duration from other patients in the same group contributed to this variability.

High heterogeneity in this study may primarily stem from variations in patient baseline characteristics (e.g., severity of motor function impairment, disease stage) and intervention protocols. These findings suggest that results for 8 weeks, as one of the optimal intervention periods, may represent an average of heterogeneous responses, necessitating further validation through stratified analysis to yield more precise evidence. Consequently, to address the substantial heterogeneity observed in this study, we conducted additional subgroup analyses based on the phase after stroke and the severity of upper limb motor impairment. The results indicated that an 8-week acupuncture course intervention yielded the optimal outcomes for patients in the subacute phase and severe PIS-ULMD. We found that the optimal acupuncture treatment course varies depending on the patient’s baseline characteristics. An 8-week course of acupuncture is particularly effective for subacute, severe PIS-ULMD patients.

Ischemic stroke is a disease with high incidence, disability, and mortality rates in clinical practice, often causing varying degrees of limb dysfunction, where PIS-ULMD is particularly challenging to rehabilitate and significantly impacts patients’ quality of life. As a traditional and effective therapeutic approach, acupuncture can significantly improve functional impairments and enhance daily living abilities in PIS-ULMD patients. Our findings align with conventional meta-analyses by Feng et al. (2022), and Li H. P. et al. (2022), confirming acupuncture’s efficacy in post-stroke upper limb functional recovery. Huang et al. (2024) analyzed four treatment duration intervals (10-day, 2-week course, 3-week course, and 4-week course) for acute ischemic stroke, suggesting 4 weeks as potentially the most favorable duration based on NIHSS scores; however, the lack of comparisons beyond 4 weeks prevented determination of whether longer courses provide additional benefits or the optimal treatment period. This study is the first to elucidate the relationship between multiple treatment courses and clinical efficacy through network meta-analysis. Furthermore, existing meta-analyses on acupuncture for PIS-ULMD have primarily focused on different acupuncture methods (Zhang et al., 2024) and intervention timing (Zhuo et al., 2021). Currently, there remains a paucity of high-quality RCTs directly comparing the efficacy of different acupuncture courses for PIS-ULMD, indicating that standardized acupuncture protocols for PIS-ULMD still require further development.

Acupuncture therapy exhibits cumulative effects, with therapeutic effects accumulating as the duration and frequency of acupuncture increase. However, upon reaching a certain level, the therapeutic effects reach a plateau or even decline (Cheng et al., 2022). Evidence for a precise efficacy plateau in acupuncture for PIS-ULMD is currently lacking. However, a related study on post-stroke wrist-hand functional reconstruction showed that electroacupuncture benefits change dynamically over time, featuring distinct phases of improvement, plateau, and decline (Wang et al., 2019). In line with the findings of this study, although the 12-week course involved a higher total intervention dose, its efficacy on the primary outcome measure was not significantly superior to that of the 8-week course. This observation aligns with the previously mentioned possibility of “acupuncture tolerance” in the introduction. We believe that the results of this study provide preliminary clinical evidence supporting the hypothesis that “acupuncture efficacy reaches a plateau.” However, due to the limited number of studies involving longer treatment durations included in this meta-analysis, the current evidence remains insufficient. Therefore, more high-quality randomized controlled trials are needed in the future to validate this hypothesis. Our results mark initial progress in exploring the optimal treatment duration of acupuncture for patients with PIS-ULMD, addressing a previous evidence gap. Despite the low quality of the current evidence, the findings may serve as a preliminary reference for selecting treatment cycles in clinical practice. Furthermore, this study revealed that the optimal treatment courses differ for motor function and overall neurological deficits, suggesting that future research should consider comprehensive selection of treatment courses based on patients’ motor abilities and neurological impairments.

This study possesses several notable strengths. In contrast to previous research, it is the first to specifically investigate the effects of different acupuncture intervention courses on patients with PIS-ULMD, conducting subgroup analyses that address a critical gap in the literature regarding optimal acupuncture treatment duration for PIS-ULMD. Furthermore, the application of network meta-analysis (NMA) has significantly strengthened the robustness of the evidence by synthesizing both direct and indirect comparative data. However, this study has several limitations that should be acknowledged. Firstly, since our research primarily focused on acupuncture treatment duration, and sensitivity analysis revealed that specific acupuncture protocols had minimal impact on the overall study findings. This approach is also consistent with existing research on acupuncture dosage, which has similarly not emphasized technical variations (Wang et al., 2024). We consequently pooled these interventions for analysis. However, we acknowledge that variations in specific techniques remain a key limitation and potential source of heterogeneity. Secondly, Studies on acupuncture treatment durations beyond the 4-week and 8-week courses remain relatively limited. This evidence gap may affect the reliability of treatment ranking results, particularly for less-studied treatment courses. Thirdly, since all included studies were conducted in China, the generalizability of our conclusions is limited. Caution should be exercised when extrapolating these findings to other populations or clinical settings. Fourthly, safety analysis indicated that most studies did not record acupuncture-related adverse events. Especially since safety data for the 12-week treatment group were nearly absent (only one study reported such information). As a result, this meta-analysis cannot reliably assess whether extending treatment beyond 8 weeks increases the risk of adverse effects. Furthermore, factors such as lack of allocation concealment and unclear blinding methods in some studies may reduce the reliability of the conclusions in this review.