2.1 Search strategy and inclusion/exclusion criteria

We conducted a systematic search of the literature in the PubMed, EMBASE, Web of Science, and Cochrane Library databases from the earliest available record until December 31, 2021, without restriction to regions, publication types, or languages. In addition, a list of references of relevant studies was reviewed to identify potential studies. The search process is described in Figure 1. The following search terms were used: (“Parkinson Disease” OR “Parkinson's Disease” OR “Parkinsonism” OR “Paralysis Agitans” OR “PD”) AND (“Gait Disorders, Neurologic” OR “Gait Disorder” OR “Locomotion Disorders” OR “Neurologic Ambulation Disorders” OR “Gait Dysfunction, Neurologic” OR “motor function” OR “freezing of gait” OR “gait impairments” OR “FOG”) AND (“Cognition” OR “Cognitive Function” OR “Cognitive Dysfunction” OR “Cognition Disorders” OR “Cognitive Impairment” OR “Neurocognitive Disorder” OR “Cognitive Decline” OR “Dementia” OR “MCI” OR “PDD”) AND (“transcranial magnetic stimulation” OR “repetitive transcranial magnetic stimulation” OR “TMS” OR “rTMS”) AND (“randomized controlled trial” OR “randomized” OR “placebo” OR “RCT”). The complete search strategies of the three electronic databases are provided in Table S1.

Trials that met all the following criteria (PICOS) were included: (1) “P”: The patients in the study were diagnosed with idiopathic PD according to the British Parkinson's Society Brain Bank or the Movement Disorders Society Clinical Diagnostic Criteria for Parkinson's Disease (Litvan et al., 2012); (2) “I”: Intervention using rTMS; the study provided a detailed description of the intervention, such as stimulation site, intensity, frequency, and time; (3) “C”: rTMS sham stimulation or dual-mode noninvasive brain stimulation (NIBS) as control; (4) “O”: The outcome of the study included gait, cognitive assessments, such as freezing of gait questionnaire (FOG-Q), 10-m walking time, timed up-and-go test (TUG), Montreal cognitive assessment (MOCA), and frontal assessment battery (FAB); (5) “S”: The study used randomized parallel or crossover design with a controlled group; (6) Data to calculate effect sizes (pre-post means, standard deviations, and sample sizes) were provided.

The exclusion criteria included: (1) case reports, animal experiments, review, and system review; (2) no control group; (3) no data available to determine effect sizes; (4) studies using rTMS combined with another therapy; (5) duplicated studies published by the same research group were excluded. Studies were excluded if they were a conference abstract or lacked the data required for estimating the effect size. If the included studies lacked key information, we contacted the corresponding authors for further information about their trials.

2.2 Data collection and extraction

Two authors (Shan Deng and Zhimei Dong) independently searched the titles and abstracts of the articles to assess eligibility for inclusion. The full texts of the eligible articles were reviewed and subsequently evaluated. The two authors independently extracted the following data from the included studies: (1) characteristics of the studies: author, year of publication, sample size, and study design; (2) participant demographics (country, age, gender, on or off state, duration of disease, and severity of disease), stimulation protocols (site, frequency, number of pulses, intensity, and duration), and the test used to quantify FOG and cognitive function outcome measures. Tables 1 and 2 show the extracted data in detail. Disagreements were resolved by discussion between the two authors. If further clarification was required, Professor Chao Qin was consulted. For each included study, quantitative data were extracted from the outcome text, tables, and graphs for the pre- and postintervention, stimulus, and control (sham) conditions. If the studies reported more than one “post” assessment, the time points after the first and last intervention were extracted and used for analysis. In two-phase crossover studies, data from the end of the first phase (treatment vs. placebo control) prior to the crossover studies were used.

2.3 Quality assessment

Two authors (Shan Deng and Zhimei Dong) independently assessed the methodological quality of the included trials using the risk of bias tool in the Cochrane Handbook for the Systematic Evaluation of Interventions, version 5.1.0. The tool reviews six potential sources of bias, including sequence allocation, allocation concealment, participant and personnel blinding, incomplete outcome data, selective outcome reporting, and other sources of bias. The risk of bias and quality of evidence were assessed as low, unclear (insufficient detail or not reported), or high, according to the Cochrane Handbook.

2.4 Statistical analysis

Stata version 14.0 (Stata Corp LP, USA) and RevMan5.3 software (The Nordic Cochrane Centre, The Cochrane Collaboration, Copenhagen, Denmark) were used to conduct the meta-analysis. The weighted mean difference (WMD) with a 95% confidence interval (CI) was used to show the combined results. The Cochran's Q statistic and I² test were used to test for heterogeneity. If the value of I² was less than 50%, the fixed effects model was used for the analysis. Otherwise, a random model was used. Where necessary, a sensitivity analysis was carried out to check the robustness of the results. Funnel plots were assessed for publication bias using Egger's regression test (where non-significant asymmetry indicated no bias). Outcome variables were compared using a value of p < .05 as statistically significant.

3.1 Results of the search

A total of 2469 studies were screened using the titles and abstracts, and another five references were obtained through other methods. The literature search identified 2453 references after the removal of duplicates. Of these, 2403 were excluded because they were not clinical trials or were outside the inclusion criteria. After full-text screening, 34 studies were excluded because not all participants were diagnosed with PD, they involved multiple interventions, had insufficient data, had no standard gait or cognitive outcomes, or the full-text was unavailable. Finally, 16 randomized controlled trials (Benninger et al., 2011, 2012; Chang et al., 2017; Cheng et al., 2022; Cohen et al., 2018; Dagan et al., 2017; W. He et al., 2021; Khedr et al., 2020; Kim et al., 2015; Ma et al., 2019; Maruo et al., 2013; Mi et al., 2019, 2020; Pal et al., 2010; Srovnalova et al., 2011; Zhuang et al., 2020) were included in the meta-analysis. The reasons for noninclusion and exclusion at the full-text screening stage are summarized in the PRISMA flow chart shown in Figure 1.

3.2 Characteristics of the included studies

The general characteristics of the included studies are shown in Tables 1 and 2. Studies were conducted in several countries. One was a multicenter study from Hong Kong and Taiwan (W. He et al., 2021).

Of the remaining studies, five were from China (Cheng et al., 2022; Ma et al., 2019; Mi et al., 2019, 2020; Zhuang et al., 2020), two were from the United States (Benninger et al., 2011, 2012), Korea (Chang et al., 2017; Kim et al., 2015) and Israel (Cohen et al., 2018; Dagan et al., 2017), and one was from Japan (Maruo et al., 2013), Hungary (Pal et al., 2010), Egypt (Khedr et al., 2020), and the Czech Republic (Srovnalova et al., 2011). The sample size ranged from seven to 42 (total number of participants was 419). The mean age of the study participants ranged from 59.94 ± 9.16 years to 74.57 ± 7.09 years, and 62.052% of the subjects were male. Among the 16 studies, 12 used a parallel-group design, while the rest adopted a crossover design. The average disease duration for all 16 studies was between 2.5 ± 1.1 years and 12.0 ± 6.3 years, and the Hoehn and Yahr scale ranged from 2 to 3.1 ± 0.5.

Eleven studies administered high-frequency TMS (> 1 Hz; range 5–50 Hz) (Benninger et al., 2012; Chang et al., 2017; Dagan et al., 2017; Khedr et al., 2020; Kim et al., 2015; Ma et al., 2019; Maruo et al., 2013; Mi et al., 2019, 2020; Pal et al., 2010; Srovnalova et al., 2011), one low (1 Hz) (Zhuang et al., 2020), and one study used both (Cohen et al., 2018). In addition, three studies used an intermittent theta-burst stimulation (ITBS: 50 Hz → 5 Hz) protocol (Benninger et al., 2011; Cheng et al., 2022; W. He et al., 2021). The most common site of rTMS stimulation was the primary motor cortex (M1) (Benninger et al., 2011, 2012; Chang et al., 2017; Cohen et al., 2018; Khedr et al., 2020; Kim et al., 2015; Maruo et al., 2013) and the prefrontal cortex (PFC) (Benninger et al., 2011; Cheng et al., 2022; Cohen et al., 2018; Dagan et al., 2017; W. He et al., 2021; Pal et al., 2010; Zhuang et al., 2020). Three studies stimulated the supplementary motor area (SMA) (Ma et al., 2019; Mi et al., 2019, 2020) and one study stimulated the inferior frontal gyri (IFG) (Srovnalova et al., 2011). Of these, seven studies (Benninger et al., 2011, 2012; Cohen et al., 2018; Dagan et al., 2017; Khedr et al., 2020; Srovnalova et al., 2011) used rTMS multisite stimulation and nine studies (Chang et al., 2017; Cheng et al., 2022; W. He et al., 2021; Kim et al., 2015; Ma et al., 2019; Mi et al., 2019, 2020; Pal et al., 2010; Zhuang et al., 2020) used rTMS single-site stimulation. Only four studies (Benninger et al., 2012; Cheng et al., 2022; Pal et al., 2010; Srovnalova et al, 2011) used an rTMS protocol with less than 1000 pulses per session, while the other 12 studies (Benninger et al., 2011; Chang et al., 2017; Cohen et al., 2018; Dagan et al., 2017; W. He et al., 2021; Khedr et al., 2020; Kim et al., 2015; Ma et al., 2019; Maruo et al., 2013; Mi et al., 2019, 2020; Zhuang et al., 2020) stimulated equal to or more than 1000 pulses per session. The total sessions of rTMS were between 1 and 24. Only one of the 16 studies used dual-mode NIBS as a control measure (Chang et al., 2017), the rest of the studies used sham stimulation as a control. The follow-up period ranged from immediately after rTMS to 90 days. The results from assessments conducted more than ten days following rTMS treatment were considered long-term outcomes, otherwise they were considered short-term results. Nearly all of the 16 studies conducted tests for outcome variables during the “on” state in PD. Only three studies reported 10-m walking time (Benninger et al., 2011, 2012; Maruo et al., 2013), five studies (Chang et al., 2017; Cohen et al., 2018; Kim et al., 2015; Mi et al., 2020; Pal et al., 2010) performed timed up-and-go test (TUG), and seven studies (Benninger et al., 2012; Chang et al., 2017; Dagan et al., 2017; Kim et al., 2015; Ma et al., 2019; Mi et al., 2019, 2020) assessed freezing of gait questionnaire (FOG-Q). In addition, cognitive function was assessed using Montreal cognitive assessment (MOCA) (Chang et al., 2017; Cheng et al., 2022; W. He et al., 2021; Khedr et al., 2020; Zhuang et al., 2020) and frontal assessment battery (FAB) (Benninger et al., 2011, 2012; Srovnalova et al., 2011).

3.3 Methodological quality of the included studies

The methodological quality of the included studies was assessed using the Cochrane Risk of Bias Assessment Tool. All 16 studies included in the meta-analysis used random allocation, with 12 studies (Benninger et al., 2011, 2012; Chang et al., 2017; Cheng et al., 2022; Cohen et al., 2018; W. He et al., 2021; Khedr et al., 2020; Ma et al., 2019; Mi et al., 2019, 2020; Pal et al., 2010; Zhuang et al., 2020) using a randomized parallel design and four studies (Dagan et al., 2017; Kim et al., 2015; Maruo et al., 2013; Srovnalova et al., 2011) using a randomized crossover design. Sixteen studies reported information related to blinding, of which 14 (Benninger et al., 2011, 2012; Chang et al., 2017; Cohen et al., 2018; Dagan et al., 2017; W. He et al., 2021; Khedr et al., 2020; Kim et al., 2015; Ma et al., 2019; Maruo et al., 2013; Mi et al., 2019, 2020; Pal et al., 2010; Srovnalova et al., 2011) were double-blind, and two (Cheng et al., 2022; Zhuang et al., 2020) were single-blind. Only two studies (Cheng et al., 2022; Zhuang et al., 2020) reported that the outcome assessors were not blinded, and two other studies (Dagan et al., 2017; Srovnalova et al., 2011) did not mention whether the outcome assessors were blinded.

Eight studies (Benninger et al., 2011, 2012; Cheng et al., 2022; Cohen et al., 2018; Dagan et al., 2017; W. He et al., 2021; Khedr et al., 2020; Kim et al., 2015) documented the number of dropouts, while the remaining eight (Chang et al., 2017; Ma et al., 2019; Maruo et al., 2013; Mi et al., 2019, 2020; Pal et al., 2010; Srovnalova et al., 2011; Zhuang et al., 2020) did not report any dropouts. Selective reporting of findings was not found in all 16 studies. Furthermore, other sources of bias identified were the small sample sizes of the 16 studies included. A detailed evaluation of the methodological quality is provided in Figures S1 and S2.

3.4 Freezing of gait questionnaire

Compared to the control group, five studies (Benninger et al., 2012; Chang et al., 2017; Kim et al., 2015; Ma et al., 2019; Mi et al., 2020) with 150 participants (n = 84 in the treatment group and n = 66 in control) focused on how rTMS affected gait in patients with PD using FOG-Q showed a statistically significant effect on post-training assessment within ten days (WMD: −0.925,95% CI: −1.642 to −0.209, p = .011; Figure S3) with low heterogeneity (χ² = 3.220, P for Q statistic = .522, I² = 0.000%; Figure S3). The funnel plot did not indicate significant asymmetry (Figure S4). The formal statistical tests did not indicate publication bias in the five studies (Egger's test, p = .727).

Seven studies (Benninger et al., 2012; Chang et al., 2017; Dagan et al., 2017; Kim et al., 2015; Ma et al., 2019; Mi et al., 2019, 2020) with 194 participants (n = 111 in the treatment group and n = 83 in control) assessed the long-term outcomes (assessed more than 10 days after training) of rTMS on gait using FOG-Q. Long-term efficacy of rTMS on the FOG-Q was statistically significant (WMD: −2.120; 95% CI: −2.751 to −1.489, p = .000; χ² = 7.100, P for Q statistic = .312, I² = 15.400%; Figure S5). No substantial asymmetry was observed in the funnel plot (Figure S6). Egger's test did not show publication bias for these seven studies (p = .072).

3.5 10-m walking time

Using data from three studies (Benninger et al., 2011, 2012; Maruo et al., 2013) with 73 participants (n = 36 in the treatment group and n = 37 in the control group), we found that compared to sham rTMS, active rTMS improved the 10-m walking time in the short term (WMD: −0.456, 95% CI: −0.793 to −0.119, p = .008; Figure S7). Heterogeneity between the included studies could be acceptable (χ² = 0.370, P for Q statistic = .829, I² = 0.000%; Figure S7). There was no significant asymmetry in the funnel plot (Figure S8). Egger's test did not suggest publication bias for these three studies (p = .668).

The long-term effect of rTMS on the 10-m walking time was reported in two studies (Benninger et al., 2011, 2012) involving 52 participants [24, 26]. The long-term effect on the 10 m walking time was evident after real rTMS (WMD: −0.526, 95% CI: −0.885 to −0.167, p = .004; Figure S9), and with a low level of heterogeneity in the estimates (χ² = 0.300, P for Q statistic = .582, I² = 0.000%; Figure S9). The funnel plot showed low evidence of publication bias (Figure S10).

3.6 Timed up-and-go test

Four studies (Chang et al., 2017; Kim et al., 2015; Mi et al., 2020; Pal et al., 2010) with 118 participants (n = 65 in the treatment group and n = 53 in control) reported that the TUG was performed within 10 days after rTMS treatment. The findings indicated that patients with PD treated with real rTMS took significantly less time in the TUG trial than sham controls (WMD: −1.064, 95% CI: −1.555 to −0.572, p = .000; Figure S11). There was no significant heterogeneity in the included studies (χ² = 1.170, P for Q statistic = .761, I² = 0.000%; Figure S11). The funnel plot was visually symmetrical. Egger's test did not suggest publication bias for these three studies (p = .146; Figure S12).

Furthermore, the long-term outcomes of rTMS of 160 participants in five studies (Chang et al., 2017; Cohen et al., 2018; Kim et al., 2015; Mi et al., 2020; Pal et al., 2010) (86 in the treatment group and 74 in the control group) using the TUG test were assessed more than 10 days after training. The pooled data results showed that the time spent by patients with PD in the TUG trial was significantly shorter after the real rTMS therapy compared to the sham control (WMD: −1.097, 95% CI: −1.422 to −0.772, p = .000; Figure S13). Heterogeneity among the five included studies was acceptable (χ² = 1.160, P for Q statistic = .884, I² = 0.000%; Figure S13). There was no significant asymmetry in the funnel plot (Figure S14). Egger's test did not indicate publication bias in these five studies (p = .179).

3.7 Montreal cognitive assessment

Five trials (Chang et al., 2017; Cheng et al., 2022; W. He et al., 2021; Khedr et al., 2020; Zhuang et al., 2020) with 160 participants (n = 84 in the treatment group and n = 76 in the control group) evaluated the efficacy of rTMS on global cognition using MOCA. There was a significant difference between the active rTMS and control group (WMD: 2.670, 95% CI: 0.513 to 4.827, p = .015; Figure S15). The heterogeneity of the included studies exceed the expected contingency (I² = 75.900%, χ² = 16.590, P for Q statistic = .002; Figure S15), suggesting that the results of the included studies exhibited substantial heterogeneity. The sensitivity analysis results (Figure S16) of the trials indicated that after removing the highly weighted study (Chang et al., 2017), the I² statistics between the studies was 0.00% (χ² = 1.770, P for Q statistic = .622; Figure S17). Moreover, the differences between the two groups were still significant (WMD: 3.714, 95% CI: 2.567 to 4.861, p = .000; Figure S17). The funnel plot of the four included studies revealed symmetrical distribution (Figure S18). The Egger's test reported insignificant results for publication bias (p = .927).

3.8 Frontal assessment battery

Three studies (Benninger et al., 2011, 2012; Srovnalova et al., 2011), which included 72 subjects (36 in the treatment group and 36 in the control group), evaluated the efficacy of rTMS on the FAB. The pooled statistics revealed a positive effect in favor of rTMS (WMD: −0.584, 95% CI: −0.934 to −0.234, p = .001; Figure S19) on FAB without heterogeneity (χ² = 2.630, P for Q statistic = .2680, I² = 24.000%; Figure S19). The funnel plot of the included studies was symmetrically distributed (Figure S20), and Egger's test did not suggest publication bias (p = .427).