Transcranial Magnetic Stimulation Combined with Speech-Language Therapy for Post-Stroke Aphasia Recovery: A Systematic Review and Network Meta-Analysis

Introduction

Stroke is the second leading cause of death and disability-adjusted life years (DALYs) lost worldwide and represents a major public health burden.^1,2 Aphasia, defined as partial or complete impairment of language function, affects approximately 30% of stroke survivors.³ This condition is associated with increased risks of anxiety and depression, reduced social participation, and diminished quality of life.^4–7 Effective therapeutic strategies are therefore urgently needed.

Current rehabilitation approaches for PSA include pharmacological treatments, music therapy, speech-language therapy (SLT), and non-invasive brain stimulation (NIBS).⁸ Pharmacological interventions, including conventional drugs and herbal formulations, have shown limited and inconsistent efficacy.^9,10 Their mechanisms remain unclear, and no agent has received regulatory approval specifically for the treatment of aphasia.^11,12 Music therapy has shown potential benefits, but further validation is needed before broad clinical implementation.^13,14 SLT remains the gold-standard behavioral intervention. However, its effectiveness often depends on intensive and sustained practice, which may not be feasible for all patients.^15,16 More effective and individualized strategies are therefore needed.

Non-invasive brain stimulation (NIBS) techniques, particularly transcranial magnetic stimulation (TMS) and transcranial direct current stimulation (tDCS), have emerged as neuromodulatory approaches that enhance cortical neuroplasticity and support language recovery.^8,17,18 TMS comprises repetitive TMS (rTMS) and theta-burst stimulation (TBS); TBS subdivides into excitatory intermittent (iTBS) and inhibitory continuous (cTBS) forms.¹⁹ Combining NIBS with SLT is neurophysiologically motivated: transient modulation of cortical excitability primes the learning and reorganisation triggered by subsequent language training.^20–22

Several meta-analyses^18,22–24 reported beneficial effects of rTMS on language recovery in PSA, with preliminary evidence that rTMS combined with SLT yields greater improvements than SLT alone, particularly in auditory comprehension, naming, repetition, and spontaneous speech.^25,26 The synthesis base is nonetheless limited. An AMSTAR 2²⁷ appraisal of existing rTMS reviews²⁸ found multiple methodological shortcomings: unregistered review protocols, missing justifications for excluded studies, inadequate description of participants and stimulation parameters, neglected risk-of-bias when interpreting results, incomplete reporting of stimulation sites and coil types, and inconsistent data extraction across reviews. Three further constraints persist: (i) some syntheses evaluated rTMS without requiring concurrent SLT,^29,30 however, given that SLT is the cornerstone of aphasia rehabilitation, such a requirement is essential for clinical applicability, and its absence limits the relevance of those syntheses; (ii) recent RCTs of newer protocols such as iTBS and cTBS^31–34 have not been incorporated into existing networks.^{18,22–24,35,36} (iii) some reviews included non-randomized studies,²⁹ provided focused syntheses of specific domains such as action naming³⁷ rather than a comprehensive multi-domain comparison,or on restricted outcome selection,³⁶ or omitted detailed stimulation parameters²⁴ or long-term effects.³⁸

This network meta-analysis addresses these gaps by restricting inclusion to randomized trials of TMS combined with concurrent SLT, incorporating recent TBS trials, evaluating 11 outcome measures across 7 language domains, and assessing both immediate and follow-up effects. The certainty of each ranking was rated under the GRADE framework.

Methods

Design and Registration

This network meta-analysis followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Network Meta-Analyses (PRISMA-NMA).³⁹ The protocol was prospectively registered in PROSPERO (CRD42024602404).

Search Strategy

Two independent reviewers (J.E.L. and Y.J.B.) searched Web of Science, EMBASE, PubMed, and the Cochrane Library from inception to January 24, 2026. No restrictions were applied to date, publication status, or document type. The search strategy combined Medical Subject Headings (MeSH) and free-text terms for keywords such as “stroke,” “cerebrovascular accident,” “transcranial magnetic stimulation,” “rTMS,” “cTBS,” and “aphasia,” along with related variants (see Table S1 for the full search strategies).

Inclusion and Exclusion Criteria

Eligibility followed the Population, Intervention, Comparison, Outcome, and Study Design (PICOS) framework, restricted to English-language randomized controlled trials (RCTs). Eligible participants were adults aged ≥18 years with PSA diagnosed by standardized clinical assessment. Eligible interventions were any TMS protocol (LF-rTMS, high-frequency rTMS (HF-rTMS), iTBS, cTBS, or Dual-rTMS) combined with SLT, provided that identical adjuvant treatments were used in both groups. Comparators were sham or placebo TMS combined with SLT, SLT alone, or other conventional therapies. Outcomes were immediate and long-term language improvement plus adverse events, assessed using validated instruments: the Montreal-Toulouse Protocol,⁴⁰ Western Aphasia Battery (WAB),⁴¹ Aachen Aphasia Test (AAT),⁴² Concise Chinese Aphasia Test (CCAT),⁴³ Boston Diagnostic Aphasia Examination (BDAE),^44,45 WAB-Revised,⁴⁶ Hemispheric Stroke Scale (HSS),⁴⁷ Standard Language Test of Aphasia (SLTA),⁴⁸ Boston Naming Test (BNT),⁴⁹ semantic verbal fluency tasks,^50–52 36-item Token Test (TT),⁵³ Test of Language Assessment in Aphasia (ADD),⁵⁴ Picture Naming Test-Turkish (T-RAT),⁵⁵ International Picture Naming Database,⁵⁶ Functional Independence Measure (FIM),⁵⁷ Amsterdam-Nijmegen Everyday Language Test (ANELT),⁵⁸ and the Snodgrass-Vanderwart picture inventory.⁵⁹ These instruments mapped to seven outcome domains: aphasia severity, naming, repetition, spontaneous speech, comprehension, functional communication, and writing. Immediate and long-term effects were analyzed separately within each domain. Where multidomain scales were used (eg., WAB, AAT, BDAE), subscale scores were assigned to prespecified domains following the original test structure and prior literature.

Because very small studies are susceptible to small-study effects and yield unstable estimates, we excluded those with fewer than 20 participants^60–62 (one study was excluded on this basis). Studies were also excluded for unclear diagnostic or efficacy criteria, non-extractable data, or TMS combined with medications or acupuncture. Non-original studies (reviews, meta-analyses, case reports, animal studies, conference abstracts, guidelines, letters) were also excluded.

Study Selection

After duplicates were removed in EndNote 21, two reviewers (J.E.L. and Y.J.B.) independently screened titles and abstracts. Full texts of potentially eligible studies were then assessed. Disagreements were resolved through discussion or consultation with a third reviewer (M.F). The selection process is shown in the PRISMA flow diagram (Figure 1).

Figure 1 PRISMA flow diagram for search and selection of eligible studies included in the NMA.

Data Extraction and Quality Assessment

Data were extracted using standardized Cochrane forms, capturing study characteristics (title, author, year, and country), participant characteristics (sample size, age, sex, handedness, stroke type, aphasia type, and time post-stroke), intervention details (TMS protocol and parameters, stimulation site, content and timing of SLT relative to TMS, treatment duration, and follow-up period), and outcome data (means, standard deviations, numbers of adverse events, and measurement scales). For multi-arm trials with both eligible and ineligible neuromodulation arms, only the arms relevant to the TMS-based network were extracted. One reviewer (J.E.L.) extracted the data, a second (Y.J.B.) verified them, and a third (M.F.) adjudicated discrepancies.

Risk of bias for each included RCT was assessed independently by two reviewers (J.E.L. and Y.J.B.) using the Cochrane Risk of Bias tool for randomized trials (RoB 2.0).⁶³ Each study was evaluated across five domains and assigned a judgment of “low risk,” “some concerns,” or “high risk.” An overall high risk of bias was assigned if any domain was rated as high risk, whereas an overall low risk required all domains to be rated as low risk. Disagreements were resolved through discussion or consultation with a third reviewer (M.F).

Certainty of evidence was assessed using the Grading of Recommendations Assessment, Development and Evaluation (GRADE) approach.^64,65 For each comparison and outcome, certainty was rated as high, moderate, low, or very low,^66,67 based on risk of bias, inconsistency, indirectness, imprecision, publication bias, intransitivity, and incoherence. Imprecision was rated following GRADE guidance.⁶⁸ Where incoherence occurred, the estimate with higher certainty between direct and indirect evidence was adopted.

Statistical Analysis

A Bayesian framework was applied using JAGS through the gemtc (version 0.8–2) and rjags (version 4–10) packages in R (version 4.4.2). Continuous outcomes were expressed as standardized mean differences (SMDs) with 95% credible intervals (CrIs). Random-effects models were used to account for clinical and methodological heterogeneity across studies. The surface under the cumulative ranking curve (SUCRA) and mean ranks were calculated to rank the interventions. Higher SUCRA values, ranging from 0% to 100%, indicated a greater probability that an intervention was the most effective or safest option. Consistency between direct and indirect comparisons was assessed using the deviance information criterion (DIC), with a difference of less than 5 indicating acceptable consistency. Publication bias was assessed visually using comparison-adjusted funnel plots in Stata 18.0.

Prior distributions and model implementation: Non-informative priors were used for all model parameters to let the data dominate the posterior. The relative effect parameters were assigned a normal distribution with mean 0 and SD = 15 x om.scale, where om.scale is a data-derived scale parameter defining the prior range for continuous outcomes. The between-study heterogeneity SD (τ) followed a uniform prior on [0, om.scale] (τ ~ Uniform(0, om.scale)). Markov chain Monte Carlo (MCMC) simulations used four independent chains with 5,000 burn-in and 20,000 sampling iterations. Heterogeneity, convergence, and visual diagnostics: Heterogeneity was assessed using tau values categorised by established guidelines: low (<0.04), low-to-moderate (0.04–0.16), moderate-to-high (0.16–0.36), and high (>0.36).^69,70 Uncertainty of the heterogeneity estimate was quantified by the 95% credible interval (CrI) for τ². Model convergence was assessed by the Brooks-Gelman-Rubin diagnostic, with PSRF < 1.05 indicating adequate convergence for all parameters.⁷¹ Diagnostic plots provided visual validation: trace plots confirmed adequate chain mixing and the absence of trends, and posterior density plots characterised the shape and variability of the posterior distribution.

Results

Search Results

The initial database search identified 1,900 records. After removal of 474 duplicates, 1,426 studies underwent title and abstract screening according to predefined criteria, and 1,341 were excluded. A total of 85 full-text articles were sought for retrieval; 28 could not be obtained, and the remaining 57 were assessed for eligibility. Of these, 31 were excluded because the total number of participants was less than 20 (n=1), the article was not in English (n=2), the intervention was not relevant (n=1), the study was a case report (n=4), the study was not an RCT (n=8), the target population was not eligible (n=1), the data were not extractable (n=9), or the intervention was not combined with SLT (n=5). Ultimately, 26 RCTs were included in the network meta-analysis. The PRISMA flow diagram is shown in Figure 1.

Table 1 Characteristics of the Included Studies

Study Characteristics

Twenty-six RCTs published between 2012 and 2026 were included (Table 1): 2 multinational (Canada/United States/Germany),^72,73 1 in the United States,⁸⁷ 3 in Germany,^78,84,91 1 in Turkey,⁷⁴ 1 in Canada,⁸⁶ 15 in China,^{25,31–34,75,76,79–82,85,88,89,92} 2 in Poland,^77,90 and 1 in Egypt.⁸³ The trials enrolled 1,136 patients with PSA; sample sizes ranged from 28 to 96 (mean 43.7, median 41), mean ages from 45.3 to 71.2 years, and follow-up from 1 to 180 days.

Across the included trials, the network comprised the following nodes: LF-rTMS alone;⁷⁴ LF-rTMS with SLT (n=19);^{34,72–79,81,82,84–88,90–92} LF-rTMS with SLT and music (n=1);⁸⁰ dual-hemisphere rTMS with SLT (n=2)^25,83 iTBS with SLT (n=5);^31–34,89 and cTBS with SLT (n=3).^31–33 The intensity of rTMS ranged from 80% to 110% of resting motor threshold (RMT), and the mean duration of illness from 15.9 days to 6.25 years. A detailed breakdown of stimulation parameters (intensity, frequency, and total pulses per session) is provided in Table 1.

Risk of Bias and GRADE Evidence Grading

Overall, 18 studies (69.2%) were low risk, 5 (19.2%) had some concerns, and 3 (11.5%) were high risk (Figure 2).

Figure 2 Summary of risk of bias for the included studies.

Most studies showed a low risk of bias across domains: outcome reporting (100%), missing data (88.5%), outcome measurement (96.2%), randomization (76.9%), and protocol deviations (88.5%). Unclear risk was noted for inadequate reporting of missing data (7.7%), protocol deviations (11.5%), and randomization methods (19.2%). High risk was identified for poorly described randomization (3.8%), missing outcome data (3.8%), and outcome measurement (3.8%). Studies failing to report relevant information were classified as unclear risk. Three trials were rated high risk for specific reasons: Yaşa et al⁷⁴ for insufficient randomization description allowing prediction of group assignment; Bai et al⁸⁵ for not reporting outcome-assessor blinding (possible detection bias); and Dresang et al⁸⁷ for missing outcome data with evidence that missingness depended on the true value (attrition bias).

The certainty of evidence varied substantially across comparisons. Conventional protocols, notably LF-rTMS+SLT versus SLT alone, had moderate-to-high certainty across several domains (eg., immediate naming, comprehension, and writing), reflecting narrow credible intervals and direct robust comparisons. However, network estimates involving novel or multi-component interventions (eg., LF-rTMS+Music+SLT) had predominantly low to very low certainty. Downgrades were driven by inconsistency (larger upper bound of tau-squared), indirectness (sparse network connections), and imprecision (wide credible intervals crossing the null). Risk of bias was generally not downgraded, as approximately 70% of studies showed low overall risk per the Cochrane RoB 2.0 assessment. A detailed GRADE summary for all network estimates is provided in Table S2.

Heterogeneity and Model Convergence Assessment

The between-study heterogeneity variance (τ²) varied across outcomes, with posterior medians ranging from 0.0033 (95% CrI: 0.00001 to 0.0404) for immediate comprehension to 0.1728 (95% CrI: 0.0057 to 0.6349) for immediate aphasia severity. Most τ² estimates were small, although long-term repetition showed greater uncertainty (τ² = 0.0673, 95% CrI: 0.00013 to 1.7868). Overall model convergence was satisfactory, with all PSRF values < 1.05 (Table S3). Trace plots showed adequate mixing without drift, and posterior density plots were generally unimodal and stable, supporting model convergence and stability (Figure S1). These findings were consistent with the Brooks-Gelman-Rubin diagnostic plots (Figure S2).

NMA

Immediate Improvement in Aphasia Severity

Twenty studies^{31–34,72,75,76,78–86,88,89,91,92} Involving seven interventions reported immediate aphasia severity outcomes (Figure 3A). LF-rTMS+Music+SLT was significantly more effective than several alternatives (Figure 3B): compared with LF-rTMS+SLT (SMD =1.48, 95% CrI [0.21, 2.73]); music+SLT (SMD = 1.77, 95% CrI [0.43, 3.14]); iTBS+SLT (SMD = 1.51, 95% CrI [0.18, 2.82]); and SLT alone (SMD = 2.05, 95% CrI [0.84, 3.29]). SUCRA ranked LF-rTMS+Music+SLT highest (96.86%) (Figure 3C).

Figure 3 NMA of immediate improvement in aphasia severity: (A) network plot; (B) comparative effectiveness; (C) SUCRA rankings. SMDs with 95% CrIs; significant results are shown in bold.

The certainty of evidence for this outcome is shown in Table S2; heterogeneity and convergence diagnostics are summarized in Table S3.

Immediate Improvement in Repetition

Twenty studies^{25,32–34,74,76–86,88,89,91,92} involving eight protocols assessed immediate repetition improvement (Figure 4A). LF-rTMS+Music+SLT was superior to cTBS+SLT (SMD = 1.07, 95% CrI [0.12, 1.98]), iTBS+SLT (SMD = 1.38, 95% CrI [0.53, 2.21]), LF-rTMS (SMD = 1.68, 95% CrI [0.49, 2.86]), LF-rTMS+SLT (SMD = 1.29, 95% CrI [0.49, 2.08]), music+SLT (SMD = 1.06, 95% CrI [0.05, 2.05]), and SLT (SMD = 1.63, 95% CrI [0.86, 2.39]) (Figure 4B). SUCRA ranked LF-rTMS+Music+SLT highest (97.25%) (Figure 4C).

Figure 4 NMA of immediate repetition: (A) Network plot; (B) Comparative effects; (C) SUCRA rankings. SMD with 95% CrI; significant results in bold.

The certainty of evidence for this outcome is shown in Table S2; heterogeneity and convergence diagnostics are summarized in Table S3.

Immediate Improvement in Comprehension

Twenty^{25,32–34,72–74,77–86,89,91,92} studies involving eight interventions evaluated immediate comprehension outcomes (Figure 5A). Dual-rTMS+SLT was superior to iTBS+SLT (SMD = 0.84, 95% CrI [0.35, 1.30]), LF-rTMS (SMD = 1.38, 95% CrI [0.39, 2.37]), LF-rTMS+SLT (SMD = 0.91, 95% CrI [0.48, 1.33]), music+SLT (SMD = 1.1, 95% CrI [0.34, 1.859]), cTBS+SLT (SMD = 0.72, 95% CrI [0.06, 1.37]), and SLT (SMD = 1.18, 95% CrI [0.77, 1.57]) (Figure 5B). Dual-rTMS+SLT had the highest SUCRA (97.57%) (Figure 5C).

Figure 5 NMA of immediate comprehension: (A) Network plot; (B) Comparative effectiveness; (C) SUCRA rankings. SMD with 95% CrI; significant results in bold.

The certainty of evidence for this outcome is shown in Table S2; heterogeneity and convergence diagnostics are summarized in Table S3.

Immediate Improvement in Naming

Twenty-two studies^{32,33,72–91} involving eight interventions evaluated immediate naming outcomes (Figure 6A). LF-rTMS+Music+SLT was superior to cTBS+SLT (SMD = 1.51, 95% CrI [0.53, 2.46]), Dual-rTMS+SLT (SMD = 1.21, 95% CrI [0.10, 2.41]), iTBS+SLT (SMD = 1.56, 95% CrI [0.65, 2.47]), LF-rTMS (SMD = 2.59, 95% CrI [1.53, 3.61]), LF-rTMS+SLT (SMD = 1.55, 95% CrI [0.72, 2.36]), music+SLT (SMD = 1.38, 95% CrI [0.36, 2.38]), and SLT (SMD = 1.96, 95% CrI [1.15, 2.76]) (Figure 6B). LF-rTMS+Music+SLT ranked first by SUCRA (99.69%) (Figure 6C).

Figure 6 NMA of immediate naming: (A) Network plot; (B) Comparative effects; (C) SUCRA rankings. SMD with 95% CrI; significant results in bold.

The certainty of evidence for this outcome is shown in Table S2; heterogeneity and convergence diagnostics are summarized in Table S3.

Immediate Improvement in Spontaneous Speech

Sixteen studies^{25,32–34,75,76,78–82,85,86,88,89,92} involving seven interventions reported immediate spontaneous speech outcomes (Figure 7A). LF-rTMS+Music+SLT significantly outperformed cTBS+SLT (SMD = 2.53, 95% CrI [1.4, 3.7]), iTBS+SLT (SMD = 1.23, 95% CrI [0.24, 2.21]), LF-rTMS+SLT (SMD = 1.05, 95% CrI [0.1, 1.99]), music+SLT (SMD = 1.27, 95% CrI [0.17, 2.36]), and SLT (SMD = 1.48, 95% CrI [0.57, 2.4]) (Figure 7B). SUCRA for LF-rTMS+Music+SLT was 95.67% (Figure 7C).

Figure 7 NMA of immediate spontaneous speech: (A) Network plot; (B) Comparative effectiveness; (C) SUCRA rankings. SMD with 95% CrI; significant results in bold.

The certainty of evidence for this outcome is shown in Table S2; heterogeneity and convergence diagnostics are summarized in Table S3.

Immediate Improvement in Functional Communication

Seven studies^{25,34,75,76,78,86,92} involving four interventions assessed immediate improvement in functional communication (Figure 8A). Dual-rTMS+SLT was significantly more effective than SLT alone (SMD = 0.62, 95% CrI [0.06, 1.17]) (Figure 8B). SUCRA ranked Dual-rTMS+SLT highest (91.49%) (Figure 8C).

Figure 8 NMA of immediate functional communication: (A) Network plot; (B) Comparative effectiveness; (C) SUCRA rankings. SMD with 95% CrI; significant results in bold.

The certainty of evidence for this outcome is shown in Table S2; heterogeneity and convergence diagnostics are summarized in Table S3.

Immediate Improvement in Writing

Seven studies^{34,78,81,82,84,91,92} involving three interventions evaluated immediate writing outcomes (Figure 9A). LF-rTMS+SLT was significantly more effective than SLT (SMD = 0.34, 95% CrI [0.10, 0.58]) (Figure 9B). SUCRA for LF-rTMS+SLT was 87.73% (Figure 9C).

Figure 9 NMA of immediate writing: (A) Network plot; (B) Comparative effectiveness; (C) SUCRA rankings. SMD with 95% CrI; significant results in bold.

The certainty of evidence for this outcome is shown in Table S2; heterogeneity and convergence diagnostics are summarized in Table S3.

Long-Term Improvement in Aphasia Severity

Five studies^{72,75,76,83,86} involving three interventions reported long-term aphasia severity outcomes (Figure 10A). Dual-rTMS+SLT showed a larger improvement than SLT (SMD = 1.23, 95% CrI [0.10, 2.36]) (Figure 10B). Although the credible interval did not cross the null, the lower bound was close to zero and the interval was wide. This imprecision, largely attributable to the limited number of studies informing this comparison, contributed to the evidence being rated low certainty in the GRADE assessment. SUCRA ranked Dual-rTMS+SLT highest (95.82%) (Figure 10C).

Figure 10 NMA of long-term severity of aphasia: (A) Network plot; (B) Comparative effectiveness; (C) SUCRA rankings. SMD with 95% CrI; significant results in bold.

The certainty of evidence for this outcome is shown in Table S2; heterogeneity and convergence diagnostics are summarized in Table S3.

Long-Term Improvement in Repetition

Five studies^{74,76,83,86,91} involving four interventions assessed long-term repetition improvement (Figure 11A). Dual-rTMS+SLT was superior to SLT (SMD = 1.82, 95% CrI [0.44, 3.17]) (Figure 11B). Dual-rTMS+SLT had the highest SUCRA (97.90%) (Figure 11C).

Figure 11 NMA of long-term repetition: (A) Network plot; (B) Comparative effects; (C) SUCRA rankings. SMD with 95% CrI; significant results in bold.

The certainty of evidence for this outcome is shown in Table S2; heterogeneity and convergence diagnostics are summarized in Table S3.

Long-Term Improvement in Comprehension

Six studies^{72–74,77,83,86} involving four interventions assessed long-term comprehension outcomes (Figure 12A). Pairwise comparisons showed no statistically significant differences among the interventions, as all 95% credible intervals crossed the null value (Figure 12B). Although Dual-rTMS+SLT had the highest SUCRA value (95.06%) (Figure 12C), this ranking reflects relative probability rather than statistically confirmed superiority. Therefore, these long-term comprehension results should be interpreted cautiously.

Figure 12 NMA of long-term comprehension: (A) Network plot; (B) Comparative effects; (C) SUCRA rankings. SMD with 95% CrI; significant results in bold.

The certainty of evidence for this outcome is shown in Table S2; heterogeneity and convergence diagnostics are summarized in Table S3.

Long-Term Improvement in Naming

Ten studies^{72–77,83,86,87,90} involving four interventions assessed long-term naming outcomes (Figure 13A). Dual-rTMS+SLT was significantly more effective than LF-rTMS (SMD = 1.99, 95% CrI [0.82, 3.07]), LF-rTMS+SLT (SMD = 1.23, 95% CrI [0.29, 2.15]), and SLT (SMD = 1.47, 95% CrI [0.55, 2.37]) (Figure 13B). SUCRA ranked Dual-rTMS+SLT highest (99.74%) (Figure 13C).

Figure 13 NMA of long-term naming: (A) Network plot; (B) Comparative effects; (C) SUCRA rankings. SMD with 95% CrI; significant results in bold.

The certainty of evidence for this outcome is shown in Table S2; heterogeneity and convergence diagnostics are summarized in Table S3.

Adverse Events

The reporting of adverse events was inconsistent across studies: some trials reported the number of affected participants, while others reported the total number of events or sessions during which events occurred. We therefore present the data as originally described, without pooling, and highlight relevant limitations. Among the 26 included studies, seven studies^{31,34,72,73,75,80,86} documented at least one treatment-related adverse event, all occurring during or shortly after the TMS procedure. Headache was reported in four studies: Ren et al³¹ (1 participant), Zumbansen et al⁷² (1 participant), Zumbansen et al⁷³ (1 participant), and Low et al⁸⁶ (10 episodes across sessions; affected participants not specified). Dysesthesia was reported in Zumbansen et al⁷² (1 participant) and Zumbansen et al⁷³ (1 participant). Dizziness was reported in Chou et al³⁴ (2 participants) and Liu et al⁸⁰ (2 participants). Ren et al³¹ also reported seizure (2 participants) and involuntary movement (1 participant), and Wang et al⁷⁵ reported dull pain (1 participant). No serious adverse events were reported. Because reporting formats were heterogeneous, the true number of participants experiencing adverse events may differ from the figures above.

Consistency and Publication Bias

Model consistency was assessed using DIC, with all closed-loop models showing differences <5, indicating adequate consistency. Node-splitting analysis was conducted for outcomes with available closed loops in the network. Only four outcomes (immediate aphasia severity, immediate repetition, immediate comprehension, and immediate spontaneous speech) had sufficient closed loops for this assessment; the remaining outcomes lacked closed loops and were therefore not amenable to node-splitting (Table 2). No significant inconsistency was detected in any of the assessable comparisons (all P > 0.05). Comparison-adjusted funnel plots detected potential publication bias (Figure S3).

Table 2 Local Inconsistency Assessment for Outcomes with Available Closed Loops

Discussion

Across 26 randomized trials, no single TMS protocol led every language domain. Three protocols partitioned the leadership: LF-rTMS+Music+SLT produced the largest immediate gains in aphasia severity, repetition, naming, and spontaneous speech; Dual-rTMS+SLT led short- and long-term comprehension and the long-term outcomes for aphasia severity, naming, and repetition plus immediate functional communication; LF-rTMS+SLT topped immediate writing. Whether these domain-specific advantages survive critical appraisal of the underlying evidence is the focus of the remainder of this Discussion.

The ranking hierarchy carries an interpretive ceiling. Trials varied in aphasia subtype, time since stroke, lesion characteristics, stimulation targets, and co-intervention details, so apparently direct protocol contrasts often draw on quite different study settings. The ceiling tightens further for nodes informed by a single trial or only by indirect comparisons; those rankings should be read as hypothesis-generating, not confirmatory.

LF-rTMS+Music+SLT ranked first for immediate aphasia severity, This finding is broadly consistent with prior studies supporting LF-rTMS-based approaches in post-stroke aphasia.^30,80,93 However, this estimate rests on a single trial and should be treated as hypothesis-generating. The mechanism is plausible, with neuromodulation likely facilitating transient cortical reorganization while structured language therapy plus music-based input reinforces engagement and language processing,^14,94–96 yet head-to-head replication is needed before this protocol can be positioned ahead of established LF-rTMS+SLT for severity recovery.

LF-rTMS+Music+SLT led immediate repetition, contradicting Li et al³⁸ who identified iTBS as the optimal protocol. Two non-exclusive explanations apply. Mechanistically, repetition draws primarily on frontal language regions including Broca’s, Wernicke’s, and Geschwind’s territories and their associated pathways,^97,98 and although iTBS spreads dynamic neurophysiological effects from the stimulation site to distributed fronto-limbic circuits,⁹⁹ pairing LF-rTMS+SLT with music input may amplify repetition recovery through synergistic neuromodulation and intensive language engagement. Methodologically, Li et al³⁸ did not require concurrent SLT as an inclusion criterion, whereas the present network restricted comparators to TMS+SLT, which can shift the relative ranking. Head-to-head trials are needed to clarify the comparative efficacy of LF-rTMS– and iTBS-based combinations.

Dual-rTMS combined with SLT was the most effective intervention, not only for short- and long-term improvement in comprehension, but also for long-term improvement in aphasia severity, naming, and repetition. This finding is consistent with Li et al³⁸ Khedr et al⁸³ Dang et al²⁵ and supports the relevance of bilateral neuromodulation for auditory–semantic processing. Comprehension is largely mediated by a ventral neural pathway.^100,101 In the dual-hemisphere protocol, low-frequency stimulation is applied to the right Broca’s homologue to reduce transcallosal inhibition, while high-frequency stimulation is delivered to the left Broca’s area to enhance connectivity in metabolically compromised regions. Sequential bilateral stimulation, first targeting the unaffected hemisphere and then the affected hemisphere, may reduce pathological right-hemisphere hyperactivity. This approach restores interhemispheric balance and creates favorable conditions for excitatory stimulation.⁸³ The dependence of comprehension on coordinated bilateral networks may explain the pronounced benefit of Dual-rTMS+SLT.

Immediate naming improved most with LF-rTMS, SLT, and music therapy. This contrasts with Ding et al³⁰ who identified Dual-rTMS as optimal and cTBS as ineffective, and with Li et al³⁸ who reported cTBS as superior. Variability in adjunctive interventions may partly explain these differences. Naming outcomes are also influenced by the duration of therapy and the first language of participants,³⁰ and environmental factors such as background noise may affect performance.¹⁰¹ The phase of stroke recovery represents a critical moderator. During the subacute phase, contralesional hyperexcitability may increase transcallosal inhibition of perilesional networks. Inhibitory LF-rTMS targeting the right inferior frontal gyrus may mitigate this maladaptive effect.¹⁰² In the chronic phase, when structural reserve in the left hemisphere is diminished, compensatory recruitment of right-hemisphere homologues may become the primary recovery pathway. In such cases, facilitatory stimulation, including HF-rTMS within dual-mode protocols or iTBS, may provide greater benefit.¹⁰³ Differences in cohort composition across trials likely influenced pooled estimates and may explain why studies with a higher proportion of chronic patients³⁰ reported stronger effects for bilateral stimulation.

For immediate spontaneous speech, LF-rTMS+Music+SLT demonstrated the most favorable outcomes, consistent with Ding et al³⁰ Beyond direct modulation at the stimulation site, LF-rTMS may suppress maladaptive activation in frontal and temporal regions of the right hemisphere while facilitating functional engagement of corresponding areas in the left dominant hemisphere. This coordinated modulation promotes restoration of interhemispheric balance and supports functional reorganization within preserved language networks.⁸¹

Dual-rTMS combined with SLT was most effective for long-term naming. This finding agrees with Ding et al³⁰ but differs from Li et al³⁸ who reported cTBS as optimal. Differences in adjunctive therapy and follow-up duration may account for the discrepancy. Importantly, SLT remains the essential behavioral component across all protocols. Structured and goal-directed language practice is necessary to guide and consolidate neuroplastic changes.¹⁰⁴ The failure of some studies³⁸ to distinguish between immediate and long-term outcomes may further contribute to inconsistent conclusions. Future trials should analyze short- and long-term effects separately across language domains.

Consistent with the findings of Lee et al⁹² LF-rTMS+SLT produced the greatest immediate improvement in writing. One possible explanation is that LF-rTMS promotes language recovery by modulating neural activity in the right hemisphere, particularly in the right frontotemporal cortex, caudate nucleus, and insular cortex.¹⁰² These neuroadaptive changes have been linked to enhanced language production.

Strengths and Limitations

This network’s principal strength is comparing multiple TMS protocols head-to-head while holding SLT as a concurrent co-intervention, which keeps the clinical question intact. Separating language domains and distinguishing immediate from follow-up outcomes further yielded a treatment-effect picture that conventional pairwise meta-analysis cannot produce.

Five limitations bound the inferences. First, included trials varied in patient characteristics, stroke chronicity, stimulation targets, treatment intensity, and SLT content, which limits cross-study comparability and likely shaped the rankings. Second, several nodes (particularly newer or combined protocols) rested on sparse or single-trial evidence, so any node-level claim is provisional. Third, the top-ranked combinations (notably LF-rTMS+Music+SLT) carry low to very low certainty driven by imprecision and indirectness, even though standard TMS-vs-SLT-alone comparisons reached moderate-to-high certainty. Fourth, long-term outcomes drew on fewer studies than immediate outcomes, making follow-up estimates less stable Fifth, inconsistent adverse-event reporting precluded reliable cross-protocol safety comparison. Together, these constraints place the protocol hierarchy on hypothesis-generating rather than confirmatory footing, with adequately powered head-to-head trials of the top three combinations as the most informative next step.

Conclusions

This NMA suggests that the optimal TMS protocol depends on the rehabilitation goal in post-stroke aphasia. LF-rTMS combined with music therapy and SLT showed the most favorable short-term pattern across several language outcomes, whereas Dual-rTMS+SLT ranked highest for most follow-up outcomes. LF-rTMS+SLT appeared to offer the clearest short-term benefit for writing. These findings may guide targeted rehabilitation strategies, but remain uncertain because most comparisons rest on low-certainty, largely indirect evidence. Direct comparative trials are needed to inform firm clinical recommendations.