Our systematic review was conducted following the PRISMA guidelines (Moher et al., 2009). We searched the following databases: PubMed, Embase, Cochrane Central Register of Controlled Trials, Cochrane Movement Disorders Group Trials Register. We also searched citing and cited articles in Google Scholar. The search was limited to human researches published in English. The following keywords were used to perform the search: “dystonia,” “torticollis,” “blepharospasm,” “Meige syndrome,” “deep brain stimulation,” “bilateral,” “globus Pallidus internus,” and “subthalamic nucleus.” The titles, abstracts, full texts, and references were independently screened and assessed by two investigators (FHY and ZZJ). We negotiated together to settle disagreements and reach a consensus.

The inclusion criteria for eligible studies were as follows: (1) the study used BFMDRS and/or TWSTRS scores, (2) the study reported the means and standard deviations (SD) of movement or disability BFMDRS/TWSTRS scores, SF-36 scores, and/or BDI scores, (3) the study used bilateral DBS, (4) the study was a randomized, controlled observational or experimental trial, (5) number of patients > 4, and (6) the follow-up duration was longer than 3 months and shorter than 4 years.

The exclusion criteria for eligible studies were: (1) indications for surgery other than dystonia, (2) target other than GPi or STN, (3) staged bilateral or unilateral DBS, (4) DBS with peripheral denervation surgery, (5) studies without outcomes of BFMDRS and/or TWSTRS scores (6) articles that included data that could not be extracted, (7) conference articles, (8) editorials, (9) reviews, (10) case reports, (11) duplicate publications (12) non-English articles.

A standardized form was used to extract the data. The following information was collected: (1) baseline characteristics of the patients (gender, age at surgery, age at onset, disease duration); (2) operation items (stimulation targets, programming parameters); (3) clinical outcomes (movement and disability BFMDRS/TWSTRS scores, follow-up duration, adverse effects, and other scoring scales at baseline/the last follow-up); and (4) information on body distribution/etiology (determined cases and undefined cases). Discrepancies were resolved by consultations between the authors (FHY, ZZJ).

The Meta-analysis Of Observational Studies in Epidemiology (MOOSE) was used to assess the quality of the studies included in this analysis (Stroup et al., 2000; Phan et al., 2015). Each of the following items was equal to one point, with a maximum of six points: (1) clear study population definition and enough patients (n > 10); (2) clear definition and assessment of outcomes; (3) independent evaluation of outcome parameters; (4) clear description of follow-up; (5) no selective loss during follow-up (<10%); and (6) identification of prognostic factors and important confounders. Only studies with a score >5, which were considered methodologically sound, were included in the main analysis. The sensitivity analysis included methodologically unsatisfactory studies. Using this strategy, the main analysis was not affected by unclear and small-sample studies since the sensitivity analysis included all data.

Regression analyses were performed to determine the potential predictors of the efficacy, quality of life, mood, and adverse effects, including age at onset, age at surgery, sex ratio, disease duration, dystonia type, target, and preoperative movement and disability scores.

All studies were included in the sensitivity analysis, although the methodologies of some were less clear (score ≤ 5 in the MOOSE assessment).

Each study’s effect size was determined by calculating the standardized mean absolute differences (SMD) in movement and disability BFMDRS/TWSTRS scores and 95% confidence intervals (CIs). Adverse events (AEs), including surgery-related, hardware-related, and stimulation-related AEs were recorded to evaluate the safety of DBS for dystonia. The Standard Cochrane Q and I² statistics were used to assess heterogeneity. If p < 0.10 or I² > 50%, the data were pooled by a random effect analysis model using a generic-inverse variance. Otherwise, a fixed-effect model was used. The means ± standard error was used as the form of pooled data. Comparisons of the patients’ baseline characteristics between the GPi and STN groups were detected by Student’s t-tests. Comparisons of the main outcomes of the two groups, including the surgical effects, quality of life, and adverse effect rates, were also performed using Student’s t-tests. P < 0.05 indicated a statistically significant difference.

Subgroup analyses of the body distribution and etiology classifications were also performed with total patients’ data (Steeves et al., 2012; Jinnah and Albanese, 2014; Volkmann et al., 2014). Dystonia was categorized as focal, segmental, multifocal, generalized, or hemidystonia based on classification by body distribution. For etiology, the primary hereditary, idiopathic, primary unspecified, secondary hereditary, and secondary non-hereditary groups were used to classify dystonia. Since some patients with primary dystonia lack information about family history or untested primary familial dystonia, another category of “primary unspecified” dystonia was added. Patients with dystonia gene (including DYT1, DYT6, and so on) positive patients were classified as the primary hereditary category. Patients with a negative family history of dystonia and no definite cause of dystonia, as well as primary familial dystonia with dystonia gene negative, were designated as “idiopathic.” Patients with pantothenate kinase-associated neurodegeneration (PKAN), Huntington’s disease (HD), familial myoclonic dystonia (FMD), and Wilson’s disease were classified as the secondary hereditary category. Patients classified as secondary non-hereditary dystonia include patients with cerebral palsy, patients with birth injuries, non-neonatal hypoxia, poststroke/trauma, patients with tardive dystonia, and patients who had other various causes. To detect any differences in the main outcomes, straight pairwise comparisons of the five groups were conducted. The p-values were calculated using Student’s t-tests and the Bonferroni multiple comparisons correction (Hsu, 1996); P < 0.05/N was considered statistically significant, where N was the final number of pairwise comparisons. To estimate the study variance, a simple linear meta-regression based on the unrestricted maximum likelihood model was performed, and P < 0.05/N was considered a statistically significant correlation. Comprehensive Meta-Analysis 2.2 (Biostat, Englewood, NJ, United States) and Stata 12.0 (Stata Corp, College Station, TX, United States) were both used to perform the statistical analyses. The data were managed using the Meta-analysis of Observational Studies in Epidemiology (MOOSE) Group and the Cochrane Handbook for Systematic Reviews of Interventions (Stroup et al., 2000; Higgins and Green, 2011). This analysis has been registered in PROSPERO under the number CRD42020146145.

According to the keyword search, 6,109 articles were identified. After duplicate articles were removed and titles and abstracts were filtered, 3,971 articles were excluded. The reasons for exclusion were that they were studies unrelated to dystonia, non-clinical studies, and low-quality articles (conference articles, letters, editorials, professional opinions, and case reports). According to the inclusion and exclusion criteria, the remaining 1,083 articles were secondarily screened by reading the full texts. The references of these studies were also screened. Finally, 103 studies that met all the criteria were filtered out for the MOOSE quality assessment. The specific screening process is illustrated in Figure 1.

Based on the MOOSE quality assessment, 70% of the studies lost points due to insufficient patient numbers or insufficient outcome parameters (Supplementary Table 1). Overall, the methodology of 68 articles, which included 140 patients, was considered not clear enough. Therefore, these articles were eliminated from the main analysis. The specific data and other information of all 103 articles are shown in Supplementary Table 1.

The baseline characteristics, including age at onset, age at surgery, and length of follow-up, were not significantly different between the groups. However, the disease duration in the GPi group was higher than that in the STN group (Table 1).

Through a simple linear regression analysis, we found that age at onset (p = 0.191), disease duration (p = 0.553), age at surgery (p = 0.154) and preoperative movement scores (p = 0.105) were not significant predictors of movement symptom improvement. They were also not significant predictors of disability symptom improvement, SF-36 score improvement, BDI enhancement, or AERs. A significant correlation was observed between disability symptom improvement and movement symptom improvement (Figure 2).

The 17 excluded articles were added, and all 103 cohorts were used to re-pool the data. Although the statistical values were slightly changed (Tables 2–4), the statistical significance of the data did not change for most outcomes with three exceptions. The focal group demonstrated significantly better movement symptom improvement than the segmental and generalized groups. The generalized group exhibited significantly less disability symptom improvement than the focal group (Table 3). The correlation between the preoperative movement scores and disability symptom improvement was also significant (Figure 2B-2).

Overall, both GPi-DBS and STN-DBS patients showed statistically significant improvements in movement symptoms, disability symptoms, SF-36 scores, and BDI scores in our analysis. Though the mean movement scores (SMD) and disability scores (SMD) in the STN group were numerically higher than those in the GPi group, they were not significantly different. To our knowledge, the clinical outcomes of GPi-DBS and STN-DBS in patients with dystonia have been evaluated in two other studies (Schjerling et al., 2013; Lin et al., 2019). The authors proposed that both GPi and STN targets are effective in treating dystonia; however, the extent of the movement and disability improvements was substantially larger in the STN group, which is comparable to the findings of this study. Schjerling et al. (2013) reported that: after stimulation of the STN, the mean 6-month improvement in BFMDRS movement score was 13.8 points; after stimulation of the GPi, this improvement was 9.1 points. Lin et al. (2019) also reported that the percentage improvement in the BFMDRS total movement score was significantly larger after STN DBS (64%) than after GPi DBS (48%) after the 12-month follow-up.

In contrast, the mean percentage of postoperative improvements (PIs) in the SF-36 and BDI scores in the GPi group was numerically higher than that of the STN-DBS group, but the differences were not statistically significant. According to multiple researchers, GPi-DBS improves the quality of life in patients with dystonia (Schjerling et al., 2013; Girach et al., 2019; Liu et al., 2019). However, few studies have investigated the quality of life in patients with dystonia who undergo STN-DBS (Lin et al., 2019). According to a follow-up assessment by Lin and Elkaim et al., no significant difference was observed between the groups in terms of the percentage improvement in quality of life (Elkaim et al., 2019; Lin et al., 2019).

Depression could be regarded as one type of stimulation-related AEs (Liu et al., 2019). Both STN-DBS and GPi-DBS could cause transient depression. In addition, both can result in surgery-related adverse effects, including hemorrhages and infections. Although GPi-DBS is known to have a direct influence on dyskinesia, STN-DBS is thought to be effective in these patients because of its effect on Parkinson’s symptoms, which leads to a substantial reduction in medication use, thus avoiding hyperkinesia (Dinkelbach et al., 2015). In our present study, the pooled AER after GPi-DBS was numerically lower than that after STN-DBS, but the difference was not significant. This outcome might partly be due to the different stimulation parameters used in GPi-DBS and STN-DBS considering the stimulation-related AEs (Tsuboi et al., 2020). More studies are needed to explore the relationship between programming parameters and safety.

Here, we focused on the focal, segmental, and generalized groups. The movement scores, disability scores, and the SF-36 and BDI scores after DBS all showed significant improvement in the three groups. The movement improvement for focal dystonia was slightly higher than segmental and generalized dystonia with no statistical difference. The segmental and generalized groups demonstrated significantly less disability improvement than the focal group, while no differences were observed between the segmental and generalized groups. A 42.9% improvement in dystonia, as assessed by the dystonia movement score, and a 63.8% improvement, as assessed by the dystonia disability score, were demonstrated by the first prospective, multicenter, single-blind study that assessed the efficacy and safety of DBS in cervical dystonia (Ostrem et al., 2007). In generalized and segmental dystonia, two double-blind and multicenter studies demonstrated a benefit ratio, with mean improvements in the dystonia movement score of 51 and 42% (Steeves et al., 2012; Zhan et al., 2017). Although patients with focal, segmental and generalized dystonia all exhibited good responses to DBS, patients with focal dystonia might demonstrate the most obvious improvement in motor symptoms and disability symptoms. The physiological and pathological mechanisms of dystonia and DBS might contribute to this phenomenon (Krauss, 2002; Kiss et al., 2007; Levinson et al., 2021).

Based on etiology, dystonia is classified as either primary hereditary, idiopathic, primary unspecified, secondary hereditary, or secondary non-hereditary. Primary unspecified means the genetic tests of those patients were unclear. The movement and disability scores both showed significant improvement in the five groups. We observed that all the primary groups demonstrated significantly better movement and disability symptom improvements than the secondary non-hereditary group, which showed greater disability symptom improvement than the secondary hereditary group. However, the mean values of movement symptom improvements in the primary groups were all higher than those of the secondary hereditary group. In previous studies, many authors noted a significant benefit of DBS in patients with primary dystonia (Benabid et al., 1987; Koy et al., 2013; Filip et al., 2017). Specifically, the dramatic response to DBS was shown in dystonia musculorum deformans-1 (DYT1 +) patients, by Markun et al. and other researchers (Reese et al., 2011; Cao et al., 2013; Quartarone and Hallett, 2013). Previous studies also noted that patients with primary dystonia exhibited a better response to DBS than those with secondary dystonia (Markun et al., 2012), a conclusion that is comparable with the outcomes of this study.

The frequency of AEs in the primary hereditary and idiopathic groups was significantly lower than that in the primary unspecified and secondary non-hereditary groups, and there were also significant differences between the primary hereditary and secondary hereditary groups. However, these results should be interpreted with caution. Although we analyzed 16 clinical studies in this analysis, the idiopathic group only included 22 patients from two studies, while the secondary non-hereditary group only included 23 patients from one study. Therefore, the statistical calculations containing these two groups should be carefully interpreted. Secondary hereditary dystonia includes pantothenate kinase–associated neurodegeneration (PKAN), Wilson’s disease, and Huntington’s disease (HD). Panov et al. (2013) proposed that the AER in PKAN patients was 23.2%, and the AER in the secondary hereditary group was 21.2 ± 7.4% in the current study. Due to the dramatic response of primary hereditary dystonia (Cif et al., 2010; Cao et al., 2013; Quartarone and Hallett, 2013), its AER was also much lower compared with other types of dystonia.

In this study, we attempted to determine the prognostic factors for DBS as a dystonia treatment. Compared with the baseline characteristics, it was difficult to compare the efficacy and safety of DBS between the GPI-DBS and STN-DBS groups because the disease duration was significantly different between these groups. However, no statistically significant change was observed in the efficacy or safety of DBS across the time of disease duration. Moreover, age at surgery, age at onset, and preoperative scores were also not significant predictors of movement symptom improvement, disability symptom improvement, SF-36 enhancement, BDI enhancement, or AERs. Nevertheless, we found patients who demonstrated significant improvements in movement symptoms likely also demonstrated excellent improvement in disability symptoms.

We cannot directly deny that these factors may contribute to a difference in the outcomes of patients with dystonia, and many studies have attempted to determine the possible prognostic factors for DBS in the treatment of different types of dystonia. In a previous study, Isaias et al. (2011) found that a younger age at surgery (<21 years of age) and shorter disease duration (<15 years) are the main predictive factors of good postoperative outcomes for primary dystonia (Eltahawy et al., 2004; De Vloo et al., 2019). Brüggemann et al. (2015) observed that caudate atrophy was a predictor of a less beneficial outcome. Rodrigues et al. (2019) recently showed that cortical plasticity can be used as a biomarker to verify outcomes of DBS treatment, which indicates a positive effect of DBS. Furthermore, Isaias et al. (2011) found that worse baseline severity is the main predictive factor of higher efficacy of DBS for the treatment of Meige syndrome. Actually, in our sensitivity analysis, the correlation between the preoperative disability score and disability symptom improvement was significant, which might be related to the floor effect. A floor effect, also known as a basement effect, means that when there is a certain improvement level, patients with mild symptoms may not have as much room for improvement as patients with severe symptoms (Isaias et al., 2011). However, we need to be cautious to conclude that a lower preoperative disability score is the main predictive factor of higher disability symptom improvement for all patients with dystonia. More convincing clinical trials are therefore needed.

Our study had several limitations.

First, most of the included studies were not randomized controlled studies, and only two studies contained comparisons of the two types of DBS, which had some advantages and some disadvantages. The disadvantage was high heterogeneity, which introduced biases and reduced the evidence level. The advantage was that a large number of patient samples from multiple centers was included in this study, which increased the statistical validity and universality of the results. Actually, valuable clinical information can also be provided by the one-arm meta-analysis (Weaver et al., 2005; Yin et al., 2019; Giordano et al., 2020).

Second, the number of patients treated with STN-DBS was not sufficient, and additional studies of patients treated with STN-DBS are therefore needed.

Third, SF-36 was used as the Qol tool in our analysis. Though it’s widely used and can evaluate patients’ health as a whole, it contains very few problems with dystonia.

Fourth, the follow-up duration (mean of 12.5 months) was short. Therefore, the long-term impacts and safety data could not be pooled or calculated.

Finally, in terms of the methodology, included studies were limited to those published in English and excluded some older articles that could not be retrieved.