The Cochrane Handbook for Systematic Reviews of Interventions and the PRISMA extension statement (Hutton et al., 2015) were followed in this NMA study. We mainly searched for literature that focused on the application of TMS in PD patients. The online databases including Medline, Embase, PubMed Central, and Web of Science were searched for articles published date to December 2021. The search specified Participants and Interventions. The searching strategies were built as follows: (Parkinson Disease OR Idiopathic Parkinson’s Disease OR “other MeSH entry terms”) AND (Transcranial Magnetic Stimulation OR Magnetic Stimulation, Transcranial OR “other MeSH entry terms”) AND (max sensitivity filters for controlled trials). See Supplementary Appendix for actual searching terms. The PROSPERO registration number is: CRD42022329110.

Two independent authors screened the retrieved results by reviewing the titles, abstracts, and full texts of literature. Any disagreements were resolved with consensus after group discussions. Studies that investigated the effectiveness of TMS for improving motor and non-motor symptoms of PD were eligible for our NMA. The inclusion criteria were: (1) Study types: randomized controlled trials (RCTs) or crossover RCTs. (2) Participants: patients diagnosed with PD. (3) Interventions: the rTMS protocols with different parameters. (4) Comparison/Control: the comparison among different frequencies or targets or the control were sham stimulation or active control. (5) Outcomes: the measurement of motor symptoms for PD.

The exclusion criteria were: (1) patients with secondary Parkinsonism or Parkinson’s plus syndrome. (2) the interventions included other types of neuromodulation techniques, such as deep brain stimulation (DBS) or transcranial direct current stimulation (tDCS). The other modalities of TMS, including theta-burst stimulation (TBS) and quadripulse stimulation (QPS), were also excluded. (3) conference abstracts, reviews, and other no-RCTs trials. (4) studies did not offer sufficient data for calculation. (5) studies with low quality.

We performed the quality assessment of studies with the recommended Cochrane Risk of Bias Tool mean values (Higgins and Green, 2011), which rated studies as three levels (high risk, low risk, unclear) in six domains: (1) Selective bias, (2) Performance bias, (3) Detection bias, (4) Attrition bias, (5) Reporting bias, (6) Other bias. The CINeMA (Confidence In Network-Analysis) online Web¹ was applied to evaluate the quality of evidence considering six domains: Within-study bias, Reporting bias, Indirectness, Imprecision, Heterogeneity, and Incoherence. Any disagreements were resolved with consensus after group discussions.

The characteristics of included studies were extracted and summarized as follows: authors, publication year, study design, sample size, rTMS protocols (including stimulation intensity, frequency, target areas, and sessions), the group’s assignment, and the outcome measurements.

The primary outcome for analysis was the Unified Parkinson’s Disease Rating Scale part III (UPDRS-III), with higher scores indicating severe motor symptoms. For continuous data, we used the mean and standard deviation (m ± sd) of the changes in measurements after TMS administration for data synthesis. When studies provided these values (m ± sd) at baseline and post-intervention respectively, we used the following formula to calculate the changes:

The correlation coefficient (Corr) was calculated from the literature providing all the above indicators by the same formula. We used WebPlotDigitizer² to extract numerical data when results were presented with figures. Standard errors and interquartile ranges were also used to obtain target data. We used the data at the end of all rTMS sessions for the results reported at multiple time points. The follow-up period without rTMS was not considered. The data extraction process was conducted under the guidance of the Cochrane handbook.

The stimulation frequency was extracted as high-frequency (hf) when it was greater than 1 Hz, or low-frequency (lf) when it was below or equal to 1 Hz. The unilateral target was extracted regardless of stimulating the left or right hemisphere. For some multiple stimulation targets studies, we described the targets as bilateral stimulation or one target area plus another area, including simultaneously or sequentially stimulated multiple targets. Any disagreements were resolved with consensus after group discussions.

We performed this network meta-analysis based on the Bayesian hierarchical models (Lu and Ades, 2004). The network graph was plotted using STATA software version 16.0 (Stata Corp, College Station, Texas, USA) to present the evidence of direct and indirect comparisons between interventions. We used the gemtc (v.1.0-1) (Valkenhoef and Kuiper, 2021) and rjags (v.4-12) (Plummer, 2022) software packages to establish a consistency model and conduct subsequent analysis. These two packages were based on the Markov Chain Monte Carlo method (MCMC) and included in R software (v.4.1.2, R Foundation for Statistical Computing, Vienna, Austria) (R Core Team, 2022). The model parameters were set as a random effect model with MCMC number of chains: 4, tuning iterations: 30,000, simulation iterations: 70,000, thinning interval: 1, variance scaling factor: 2.5. The model convergence was assessed with the Brooks-Gelman-Rubin diagnosis plot and the Potential Scale Reduction Factor (PSRF) (Brooks and Gelman, 1998; Tunaru, 2016).

The node-splitting approach was used to assess inconsistency by comparing estimates from both direct and indirect evidence (van Valkenhoef et al., 2016). The heterogeneity among studies was evaluated with the I² statistics. The random-effect model was adopted with I² > 50% in the global heterogeneity. The Sensitivity analysis was conducted to explore the sources of heterogeneity by excluding one literature at a time. The meta-regression analysis was performed to investigate the effect of different covariates on pooled effects. Publication bias was assessed using funnel plots and Egger’s test.

Statistical effects were set to be significant at p < 0.05. Pooled effect sizes were reported with mean difference (MD) and 95% credible intervals (CrIs) or confidence interval (CI). Both pair-wise meta-analysis and network meta-analysis were analyzed and presented. The rankogram was used to report the probability ranking of different stimulation targets. We applied the surface under the cumulative ranking curve (SUCRA) values to assess the likelihood that the interventions rank the best (Salanti et al., 2011). The SUCRA value is between 0 and 1, SUCRA equals 1 indicates that the intervention is absolutely effective, and SUCRA equals 0 indicates that the intervention is absolutely ineffective (Cope and Jansen, 2013). The interventions with a higher SUCRA value mean having better treatment efficacy.

After removing duplicates, a total of 787 records were selected by titles and abstracts. Then, the full text of 144 articles was evaluated for eligibility. Finally, 36 RCTs on TMS improving motor symptoms of PD patients were included in this NMA. The exclusion reasons and the screening process are listed in Figure 1.

The descriptive data of 36 studies were presented in Table 1. A total of 36 English articles were included in our study, which contained 14 crossover trials (Börnke et al., 2004; Lefaucheur et al., 2004; Koch et al., 2005; Brusa et al., 2006; Rektorova et al., 2008; Filipović et al., 2009, 2010; Sedlácková et al., 2009; Maruo et al., 2013; Kim et al., 2015; Flamez et al., 2016; Dagan et al., 2017; Yokoe et al., 2018; Fricke et al., 2019) and 22 RCTs (Khedr et al., 2003; del Olmo et al., 2007; Hamada et al., 2008; Pal et al., 2010; Benninger et al., 2012; Shirota et al., 2013; Brys et al., 2016; Makkos et al., 2016; Shin et al., 2016; Aftanas et al., 2018; Cohen et al., 2018; Khedr et al., 2019; Mi et al., 2019; Randver et al., 2019; Chung et al., 2020; Hanoglu et al., 2020; Khedr et al., 2020; Li et al., 2020; Zhuang et al., 2020; Aftanas et al., 2021; Lench et al., 2021; Spagnolo et al., 2021). Of all publications, 24 articles (Khedr et al., 2003, 2019, 2020; Börnke et al., 2004; del Olmo et al., 2007; Hamada et al., 2008; Rektorova et al., 2008; Sedlácková et al., 2009; Pal et al., 2010; Benninger et al., 2012; Maruo et al., 2013; Kim et al., 2015; Brys et al., 2016; Makkos et al., 2016; Shin et al., 2016; Dagan et al., 2017; Aftanas et al., 2018, 2021; Yokoe et al., 2018; Mi et al., 2019; Randver et al., 2019; Hanoglu et al., 2020; Li et al., 2020; Spagnolo et al., 2021) used high-frequency stimulation (5–20 Hz), 7 articles (Brusa et al., 2006; Filipović et al., 2009, 2010; Flamez et al., 2016; Fricke et al., 2019; Zhuang et al., 2020; Lench et al., 2021) used low-frequency stimulation (0.1–1 Hz), 5 articles (Lefaucheur et al., 2004; Koch et al., 2005; Shirota et al., 2013; Cohen et al., 2018; Chung et al., 2020) compared the efficacy of both two frequencies. The stimulation targets involved unilateral or bilateral stimulation of M1, DLPFC, or both, and SMA, premotor cortex (PM), central parietal cortex (Pz), etc. All 36 studies set up comparisons with sham stimulation. The network plot depicted the direct and indirect comparisons among different rTMS targets (Figure 2). Node size and edge width were weighted by the involved sample size and number of studies.

Figure 3 shows the methodological quality assessment of the included studies. All studies used randomization, but 77.78% of studies did not report the method of generating random sequences, and 69.44% of studies had an unclear risk of allocation concealment. Two studies had participants lost to follow-up, making data incomplete (Brys et al., 2016; Cohen et al., 2018). Three studies only achieved single blindness in participants (Filipović et al., 2009, 2010) or investigators (Zhuang et al., 2020). All studies have no Reporting bias and Other bias. Overall, the quality of the included studies is moderate. The risk of bias in each study was summarized in Supplementary Figure 1. The CINeMA evidence grading results exhibited the confidence rating of all interventions relative to sham were available in Supplementary Figure 2. After accounting for all biases, the comparisons between hf M1+DLPFC, hf bi DLPFC, hf bi M1, and sham had high evidence grades. The comparisons with hf PM, hf SMA, lf DLPFC, lf PM+M1, lf Pz showed very low evidence grades. The comparisons with hf DLPFC, hf M1, lf M1, lf Pz, lf SMA, lf bi M1 showed low to moderate evidence grades.

The pair-wise meta-analysis investigated the efficacy of TMS with different parameters compared to control on PD (the inverse variance method with the random-effect model was used). As shown in Figure 4, TMS could significantly improve motor symptoms in PD patients compared with the control group (the total pooled effect sizes, MD: −3.72, 95% confidence intervals (CI): −5.01 to −2.43). Notably, the direct comparison results between active stimulation and sham condition indicated frequency- and target-dependent effects. The high-frequency stimulation targeting different areas showed a significant improvement in motor symptoms (MD: −4.63, 95% CI: −6.25 to −3.01), while low-frequency stimulation had not the same effects (MD: −1.21, 95% CI: −2.45 to 0.03). rTMS with hf M1+DLPFC induced the most significant improvement in UPDRS-III than other targets (MD: −7.39, 95% CI: −13.25 to −1.53).

Network meta-analysis estimates the efficacy of different rTMS targets in PD by incorporating evidence from direct and indirect comparisons. As exhibited in Table 2, the high-frequency stimulation over bilateral M1 (MD: −6.50, 95% CrI: −9.03 to −4.02), bilateral DLPFC (MD: −5.68, 95% CrI: −10.74 to −0.69), and M1+DLPFC (MD: −7.60, 95% CrI: −11.10 to −4.02) could significantly reduce the UPDRS-III scores when compared with sham condition. All low-frequency parameters were not effective in improving motor symptoms in PD patients. Concerning the comparison between active stimulations, the high-frequency stimulation over bilateral M1 had a better effect than targeting unilateral M1 (MD: −5.17, 95% CrI: −9.35 to −1.04). And, hf M1+DLPFC TMS showed more significant effects when compared to hf M1 (MD: −6.27, 95% CrI: −11.07 to −1.40), lf M1 (MD: −6.27, 95% CrI: −12.22 to −0.21), lf SMA (MD: −5.08, 95% CrI: −10.06 to −0.01), and lf Pz (MD: −6.46, 95% CrI: −12.67 to −0.17).

For UPDRS-III, the probability of hf M1+DLPFC being best-ranked was 39.14% and had the highest SUCRA value (90.66%). The hf bi M1 ranked third with a probability of 30.93%, and a SUCRA value was 84.90%. The lf PM+M1 had the worst ranking probability of 46.70%, and a SCURA value was 19.38%. Figure 5 demonstrates the rankogram and cumulative ranking plot of different TMS targets.

Model convergence was assessed using the Brooks-Gelman-Rubin method by comparing within-chain and between-chain variance. The Brooks-Gelman-Rubin diagnostic plot revealed that the median value and 97.5% value of the reduction factor tend to be one after 70,000 iterations, and the PSRF value was also close to 1, which suggested a satisfactory convergence had been reached (Supplementary Figure 3). The node-splitting method revealed that there was no significant inconsistency between the direct and indirect comparisons (p > 0.05), which means the results of the used consistency model were reliable (Supplementary Figure 4). The heterogeneity analysis showed global heterogeneity with I² > 50%. Therefore, the random effect model was used in both pair-wise meta-analysis and NMA. Sensitivity analysis was also performed by omitting one study at a time. The results of sensitivity analysis showed the robustness of pooled effects in the outcome. A meta-regression analysis using the sessions as a covariate indicated the rTMS session was negatively related to the improvement in UPDRS-III (i.e., the more sessions, the larger improvement in motor symptoms. See Supplementary Figure 5). However, the effect was not significant (the mean shared regression coefficient was −3.33, 95% CI: −6.87 to 0.30). The bias of publication was not observed in included studies, with a symmetrical funnel plot and p = 0.6727 in Egger’s test (Supplementary Figure 6).