Our systematic review was designed and implemented based on the Preferred Reporting Items for Systematic Reviews and Meta-analysis (PRISMA) guideline (26). The study has been registered with Prospero (CRD42022363899).

In the initial screening, two researchers (CG and YXD) independently searched RCTs related to the topic in seven databases: Web of Science, PubMed, Embase, Cochrane Library, CNKI, Wanfang Data, and PEDro. Search for studies published between the date of database creation and September 28, 2022. We searched for standardized disease names in the International Classification of Diseases, 11th edition (ICD − 11). Ultimately, we identified the keywords for this study as “Stroke,” “Aphasia,” “Language Expression Disorder,” “Listening Comprehension Disorder,” and “Repetitive Transcranial Magnetic Stimulation.” In addition, we manually searched other relevant literature, such as studies included in some systematic reviews and meta-analyses, to broaden the search for eligible articles. As an example, the search strategy for the PubMed database is as follows (Table 1).

Included studies were required to follow our pre-defined inclusion and exclusion criteria strictly. According to the PICOS principles, the inclusion criteria of our review were as follows: (1) participants: patients diagnosed with post-stroke aphasia; (2) interventions: rTMS; (3) comparison: experimental group (rTMS) versus control group (placebo or no treatment) condition; (4) outcomes: Western Aphasia Battery (WAB), Aphasia Quotient (AQ), Aphasia Battery in Chinese (ABC), Boston Diagnostic Aphasia Examination (BADE), Aachener Aphasie Test (AAT), Concise Chinese Aphasia Test (CCAT) and Computerized Picture Naming Test (CPNT); (5) type of studies: RCT; (6) studies published in English or Chinese. Exclusion criteria for the literature: (1) duplicate data; (2) full-text content not available; (3) data not extractable.

After completing the database search, we imported all retrieved studies into Endnote 20’s document management system (Endnote 20, United States) and removed duplicate studies using the software management function. Two researchers (CG and LX) then read the title and abstract of each study simultaneously and screened studies based on the inclusion and exclusion criteria we had previously developed. For initially screened studies, the two researchers would downloaded and read through the full text, removing articles that do not meet the inclusion criteria and discussed them to confirm their eligibility. If the two researchers disagree on the screening process of a study, the principal investigator (ZYD) was asked to provide advice and reach an agreement.

Two researchers (MH and LX) independently extracted the following data and items from the included literature: first author of the study, year of publication, the sample size of participating studies, age, gender, duration of disease, interventions tested, outcome indicators, and adverse effects. In addition, when the two researchers encountered difficulties in understanding or extracting the complete literature data during the data extraction process, the original authors of the literature would be contacted by sending an email to obtain the full trial data. When no response was received from the original author after three consecutive contacts, the study will be defined as missing data. Suppose two researchers disagree during the data extraction process. In that case, both would be placed in a research team with the Principal Investigator to discuss and resolve the issue. If the two researchers disagree on the screening process of a study, the principal investigator (ZYD) would be asked to provide advice and reach an agreement. The principal investigator will convene a meeting of the research team to discuss the reasons for any disagreements; once the sources of conflict are resolved, consensus will be reached.

The quality assessment of the literature studies was completed independently by two researchers (MH and JHH), then discussed to produce consistent results. Risk bias was assessed using the Cochrane Risk of Bias tool (Revman 5.40, Nordic Cochrane Center). A total of seven items were considered: random sequence generation, allocation concealment, blinding of participants and personnel, blinding of outcome assessment, incomplete outcome data, selective reporting, and other biases. The risk bias assessment was mapped, and different colors differentiated the results into three levels: high risk—red, unknown risk—yellow and low risk—green. Heterogeneity between studies was statistically analyzed by Revman 5.40. The magnitude of heterogeneity was expressed as I², with heterogeneity judged as high risk when I² ≥ 75%, moderate risk when 75% > I² ≥ 50%, low heterogeneity when 50% > I² ≥ 25%, and no heterogeneity if I² = 0% (27). I² quantifies the extent of heterogeneity between studies. On the one hand, we select the appropriate effect model for the forest plot according to the magnitude of I² to minimize the impact of high heterogeneity on the pooled results. On the other hand, during the assessment of evidence quality, we also use I² to grade the strength of the evidence. The quality of evidence for outcome indicators was assessed using the Grading of Recommendations Assessment, Development, and Evaluation (GRADE) system, which examines study limitations, intermittency, inconsistency, and imprecision of results (28). The results were assessed by grading the evidence for the outcome indicators as “high,” “moderate,” “low,” or “very low,” and the strength of the recommendations was divided into two levels: “strong” and “weak” (29).

The extracted study data were entered into Revman 5.40 software^a for statistical and analytical purposes. The decision to use a fixed or a random effects model for the meta-analysis was based on the magnitude of heterogeneity. A random effects model was used when I² ≥ 50%, and a fixed effects model was used when I² < 50%. Mean difference (MD) and 95% confidence interval (CI) were used to express the effect size for studies using the same measure. Vice versa, standardized mean difference (SMD) and 95%CI were used to express the effect size for studies using different measures. p < 0.05, statistically significant.

The number, type, and duration of adverse events that occurred during the rTMS intervention were counted in all the studies concerned. And statistically analyze the patients who experienced adverse events, as a percentage of the number of subjects. Record whether there were any intolerable or even life-threatening adverse reactions that caused the subjects to withdraw from the experiment. And, to track whether the adverse events in each study, still persisted during the follow-up time.

A total of seven databases were searched for literature, and the initial search resulted in 994 studies. Duplicate studies were screened out and removed by software, leaving 628 studies. Two researchers (YZL and CG) read the titles and abstracts of these studies and screened out 574 that were irrelevant to the topic. The remaining 54 studies were downloaded in full and read through, and 24 studies were still excluded (8 studies were screened for not using the internationally accepted aphasia ratings listed in the inclusion criteria or data on outcome indicators were not fully available; 12 studies were screened for not strictly using a randomized controlled trial design; four studies were screened for not using rTMS as the primary intervention in an experimental comparison study). Finally, 30 eligible studies were included (30–59) (Figure 1).

Across the included randomized controlled trials, rTMS was most commonly applied in conjunction with SLT, whereas a smaller number of trials compared rTMS plus SLT with SLT alone or with sham stimulation plus SLT. Only a few studies combined rTMS with pharmacological agents. This variability in study design illustrates that rTMS is more appropriately viewed as an adjunctive rather than independent therapy for PSA. To improve clarity, we have minimized use of acronyms in the Results section, spelling out the assessment tools (e.g., Western Aphasia Battery instead of WAB on first mention). Table 2 summarizes the basic data of the 30 RCTs. A total of 1,597 patients with PSA were included, with sample sizes ranging from 12 to 120, of which 819 patients with PSA were treated with rTMS. Subjects’ aphasia types had non-fluent aphasia, Broca aphasia, Motor aphasia, Global aphasia, and Various aphasia. Among the included RCTs, the outcome indicators for rating aphasia in post-stroke patients included Western Aphasia Battery (WAB), Aachener Aphasia Test (AAT), Aphasia Battery in Chinese (ABC), Aphasia Quotient (AQ); Concise Chinese Aphasia Test (CCAT); Computerized Picture Naming Test (CPNT) and Boston Diagnostic Aphasia Examination (BDAE).

In addition, Table 3 summarizes the intervention parameters of rTMS for post-stroke patients in each study, such as stimulation site, intensity, frequency, number of pulses, and stimulation time. Among the studies we included, 2 studies used rTMS at 0.5 Hz to treat PSA patients, 28 studies used rTMS at 1 Hz to treat PSA patients, and 3 studies used rTMS at 10 Hz to treat PSA patients. Overall, there are more clinical studies using low-frequency rTMS to treat PSA than high-frequency rTMS. And low-frequency rTMS treatment is mainly 1 Hz rTMS.

The risk of bias assessment showed that in all included RCTs, four RCTs (31, 40, 49, 55) did not use blinding for the assessment of outcome indicators and had a high risk of detection bias. And nine studies (31, 33, 34, 37, 41, 45, 50, 52, 58) did not explicitly report blinding for assessing outcome indicators, and the risk of detection bias was unclear. The risk of bias was low for all RCTs in the other items evaluated for risk of bias. Overall, the risk of bias was low in our included studies (Figures 2, 3).

We evaluated the level of evidence for the outcome indicators of the included studies by GRADE. Three outcome indicators were rated as intermediate, (AQ, ABC and WAB) due to high heterogeneity between studies (I ² > 80%) and were therefore downgraded in the inconsistency assessment. Four outcome indicators (AAT, BADE, CCAT and CPNT) were rated as intermediate because the sample sizes were too small (n < 100), which tended to influence the imprecision of the study results, and were downgraded in the imprecision assessment. The remaining outcome indicators were not found to be downgraded factors in each of the GRADE assessments. Overall, the GRADE recommended evidence level for the outcome indicator was “strong” (Table 4).

There are 12 studies (31, 32, 35, 36, 38, 39, 43–47, 49) rated the speech function of PSA patients by WAB with I² > 50% between studies and therefore used a random-effects model for data analysis. The results of the forest plot analysis showed that patients treated with rTMS had greater improvements in areas of verbal comprehension and expression. The specific improvement results were as follows: rTMS was more effective in improving auditory comprehension in PSA patients compared to control group (MD = 1.94, 95% CI = [1.16, 2.17], I² = 79%, p < 0. 001, Figure 4); rTMS was more effective in improving naming ability in PSA patients compared to control group (MD = 1.53, 95% CI = [0.82, 2.24], I² = 77%, p < 0. 001, Figure 5); rTMS was more effective in improving verbal repetition in PSA patients compared with the control group (MD = 1.79, 95% CI = [1.20, 2.38], I² = 50%, p < 0.001, Figure 6); and rTMS was more effective in improving PSA patients’ spontaneous speech (MD = 1.97, 95% CI = [1.65, 2.29], I² = 0%, p < 0.001, Figure 7).

The degree of impairment in PSA patients was assessed in 16 studies (30, 32, 33, 35–40, 43–49) using AQ scores with an I² > 50% between studies, so the data were analyzed using a random effects model. The results of data analysis showed that compared to the control group, the experimental group showed better improvement in AQ scores than the control group (MD = 13.82, 95% CI = [11.68, 15.97], I² = 52%, p < 0.001; Figure 8).

The degree of language loss in PSA patients was assessed by ABC in 7 studies (33, 34, 40–42, 48, 58) with I² > 50% between studies, and we analyzed the data using a random-effects model. The results of the forest plot analysis showed that compared to the control group, the experimental group had better outcomes in ABC scores (MD = 24.79, 95% CI = [17.80, 31.77], I² = 95%, p < 0.001; Figure 9).

In 4 studies (50, 53, 55, 56), the verbal function of PSA patients was assessed using the AAT scale, with an I² < 50% between studies, and we analyzed the data using a fixed effects model. The results of the forest plot analysis showed that compared to the control group, the experimental group showed a significant improvement in AAT scores compared with the control group (MD = 13.74, 95% CI = [9.43, 18.06], I² = 0%, p < 0.001; Figure 10).

The severity of aphasia in PSA patients was assessed in two studies (54, 59) using the BADE scale, with an I² < 50% between studies, and we analyzed the data using a fixed effects model. The results of the analysis of the forest plot showed that the experimental group had a better improvement than the control group in terms of BADE scores in patients with PSA (MD = 38.37, 95% CI = [6.32, 70.42], I² = 22%, p = 0.02; Figure 11).

Two studies (51, 52) used the CCAT scale to assess language function in PSA patients, and the I² value between studies was 0%, thus the data were analyzed using a fixed effects model. The results of the analysis of the forest plot showed that the experimental group had a more positive contribution in improving the CCAT scores of PSA patients compared to the control group (MD = 1.39, 95% CI = [0.25, 2.53], I² = 0%, p = 0.02; Figure 12).

In addition, 3 studies (51, 52, 56) tested the naming function of PSA patients by CPNT alone, with an I² < 50% between studies, so the data were analyzed using a fixed effects model. The results of the forest plot analysis showed that the experimental group was more able to improve the naming ability of PSA patients and promote the recovery of verbal function compared to the control group (MD = 3.95, 95% CI = [0.84, 7.06], I² = 8%, p = 0.01; Figure 13).

Of the 30 studies we included, only three studies reported the occurrence of adverse events. One of these studies (49) reported that two participants treated with rTMS experienced transient headache and nausea with a duration of <24 h, which accounted for 2/819 of the total number of participants in the experimental group of our study. Two other studies (36, 44) reported transient dizziness in two subjects treated with rTMS for <24 h, which accounted for 2/819 of the total number of participants in the experimental group of our study. In summary the number of patients with PSA treated with rTMS who developed adverse events as a proportion of the total number of participants in the experimental group was 4/819. In addition, based on the results reported in all studies, no patients withdrew from the experimental studies due to exhibited excessive adverse reactions. Moreover, only three of the 30 included studies reported adverse events (reporting rate 10%), and all RCTs had small sample sizes (n < 100). Therefore, we must consider the possibility of publication bias arising from unrecorded or unreported adverse events, which could underestimate the true risks of rTMS and thereby overstate its safety. We recommend that future studies continue to adhere strictly to established rTMS safety guidelines to ensure rigorous practice.

However, our meta-analysis also has several limitations that should be considered. First, the sample size of our included RCTs was small (n < 100), and too small a sample size tends to bias the assessment of treatment effects and overestimate the efficacy of rTMS. Second, although our studies all showed the positive impact of transcranial magnetic stimulation (TMS) in patients with post-stroke aphasia, most did not report patient follow-ups further to confirm the long-term effects of TMS. Third, the studies we included differed regarding the stimulation sites and the number of pulses. The number of available studies did not allow for more detailed subgroup analysis. Fourth, the absence of gray reports may lead to bias in comprehensive analysis results. Meaningful research is more likely to be accepted for publication, making us cautious about jumping to conclusions. Fifth, this study lack of patient-level data and possible cultural/geographical biases (majority of studies from China). Therefore, more multicenter follow-up, double-blind RCTs should be conducted to facilitate longitudinal and cross-sectional comparisons of different stimulation parameters of rTMS, to determine the optimal treatment protocol, and to improve the clinical efficacy of rTMS in patients with PSA. Moreover, heterogeneity in the site of stimulation across studies limits the generalizability of pooled results. Most studies targeted contralesional areas, but some used ipsilesional or bilateral protocols, which may lead to distinct therapeutic trajectories. The absence of detailed subgroup analyses also prevents us from assessing whether different aphasia phenotypes (e.g., Broca, global, motor) respond differently to rTMS. This remains an important future direction for tailoring interventions.