The inclusion criteria of this study, according to the PICOS checklist (Higgins, 2008), are as follows: (1) Population: patients diagnosed with TRD based on the diagnostic and statistical manual of mental disorders, 4th edition (DSM-IV), and treatment resistance was defined as a failure of antidepressant therapies at an adequate dosage and duration; (2) Intervention: DBS targeting various nerve nuclei; (3) Control: placebo or sham stimulation in the trials if the data was available; (4) Outcomes: response decline, remission and recurrence rates, referred to as “efficacy,” and adverse events (AEs) rate, referred to as “risk” and (5) Study design: RTCs and open-label studies were included.

Two independent authors (JM and WH) constructed the corresponding search strategies in the Medline, Google scholar, Cochrane library databases, Embase and World Health Organization International Clinical Trials Registry Platform up to January 2019. The search results from the databases were obtained using the terms “deep brain stimulation” OR “DBS” AND “depression” OR “depressive disorders.” We searched for eligible studies by scanning the abstracts from the original articles and screening the list of references for relevant publications. Additionally, other references were manually scanned from relevant reference reports and clinical trial websites. Only eligible studies that met the predefined criteria were input into a bibliography management system.

For accuracy and completeness, two reviewers (JM and KZ) independently checked the full articles and extracted the information corresponding to the characteristics of the study population, the details of the surgery, the outcome measures, and the adverse events. We recorded the data, including the Hamilton Rating Scale for Depression (HRSD), Montgomery-Asberg Depression Rating Scale (MADRS), adverse events, suicidal ideation and suicide, in each target location. Discrepancies were resolved by discussion and consensus. Cochrane tools were not used to assess the quality of the included studies, because most were open-label studies.

A meta-analysis was conducted to calculate the studies outcomes, including the response, remission, recurrence and adverse events rate, using Stata software (version 13.0). A 50% or greater improvement from the baseline score to end of treatment on a depression rating scale (for example, the HDRS or the MADRS) was defined as the clinical response. A clinical remission was defined as a score on a depression rating scale within the normal range (for example, HDRS of 7 or less or MADRS of 12 or less). The adverse effects were evaluated by examining the proportion of adverse events.

During the systematic searching based on the searching strategies, 374 references from the electronic databases and eight additional records from other resources were identified and the last searching was performed on January, 30, 2019. After screening stage where the titles and abstracts of the retrieved records were reviewed, 360 records were excluded due to duplication or non-relevant content (Figure 1). Of the remaining 22 references, we excluded five (Jiménez et al., 2013; Mosley et al., 2015; Richardson et al., 2015; Narang et al., 2016; Bergfeld et al., 2017), because the primary outcomes presented by these trials did not meet the inclusion criteria. Thus, 17 studies (Mayberg et al., 2005; Lozano et al., 2008, 2012; Malone et al., 2009; Bewernick et al., 2010; Nahas et al., 2010; Kennedy et al., 2011; Holtzheimer et al., 2012; Puigdemont et al., 2012, 2015; Merkl et al., 2013; Schlaepfer et al., 2013; Dougherty et al., 2015; Accolla et al., 2016; Bergfeld et al., 2016; Fenoy et al., 2016; Williams et al., 2016), including eight studies target SCG (Mayberg et al., 2005; Lozano et al., 2008, 2012; Kennedy et al., 2011; Holtzheimer et al., 2012; Puigdemont et al., 2012, 2015; Merkl et al., 2013), 2 VC/VS (Malone et al., 2009; Dougherty et al., 2015), 2 EpC (Nahas et al., 2010; Williams et al., 2016), 1 Nac (Bewernick et al., 2010), 2 sMFB (Schlaepfer et al., 2013; Fenoy et al., 2016), 1 posterior gyrus rectus (Accolla et al., 2016), and 1 study targets vALIC (Bergfeld et al., 2016) were retrieved and included in this study. The study sample size was 233. All the studies reported demographic, clinical characteristics and surgical information, which are detailed in Table 1. The age of the sample population ranged from 42.0 to 50.7 years old. The follow-up period ranged from 3 months to 5 years. All the studies used the HDRS or MADRS as the primary outcome.

The response, remission and recurrence rate were used to evaluate the efficacy and the AEs rate was used to investigate the safety of the DBS treatment for TRD. Table 2 summarizes these finding from the open-label studies, and Table 3 represents the included RCTs. At total of 16 open-label studies (Mayberg et al., 2005; Lozano et al., 2008, 2012; Malone et al., 2009; Bewernick et al., 2010; Nahas et al., 2010; Kennedy et al., 2011; Holtzheimer et al., 2012; Puigdemont et al., 2012; Merkl et al., 2013; Schlaepfer et al., 2013; Dougherty et al., 2015; Accolla et al., 2016; Bergfeld et al., 2016; Fenoy et al., 2016; Williams et al., 2016) and 2 RCTs (Dougherty et al., 2015; Puigdemont et al., 2015) compared the response rate and showed a statistically significant heterogeneity (I² = 73.6%, p < 0.0001^*), using a random model (Figure 2). According to the result of the meta-analyses, the pooled estimate was 0.56 (95% CI = 0.43–0.69) in the non-RCT group (Figure 2) and the pooled response-rate value was 1.45 (95% CI = 0.50–4.21) in only two RCTs (Figure 3).

At total of 14 open-label studies (Mayberg et al., 2005; Lozano et al., 2008, 2012; Malone et al., 2009; Nahas et al., 2010; Kennedy et al., 2011; Holtzheimer et al., 2012; Puigdemont et al., 2012; Merkl et al., 2013; Schlaepfer et al., 2013; Dougherty et al., 2015; Bergfeld et al., 2016; Fenoy et al., 2016; Williams et al., 2016) compared the AEs rate and showed a statistically significant heterogeneity (I² = 76.8%, p < 0.0001^*), using a random model. According to the results of the meta-analyses, the pooled estimate was 0.67 (95% CI = 0.54–0.80) (Figure 2). Of the included articles, nine reported the rate of suicide and suicidal attempt. The median rate of suicidal attempt was 16.7% (ranged from 4 to 80%) and of suicide was 4.8% (ranged from 3.3 to 12.5%).

To assess the statistical heterogeneity for response, remission and AEs rate, funnel plots (Figures 4A–C) were generated and a sensitivity analysis (Figures 5A–C) was performed to evaluate the publication bias and sources of heterogeneity. In the final analysis, by omitting one study at a time, none of the studies attributed to the heterogeneity in the response rate, while three studies contributed to the heterogeneity in the AEs rate (Merkl et al., 2013; Fenoy et al., 2016; Williams et al., 2016). Moreover, as shown in the Table 4, the studied outcomes of the meta-analyses were reliable, since the effect of a larger or smaller sample size was excluded. When we exclude the larger or smaller sample size studies, they did not influence the pooled outcomes or the heterogeneity.

We reduced the bias in our study as much as possible. We used reasonable search strategies, and the contribution of three independent authors contributed to reducing the sampling bias. In addition, we used explicit and rigid inclusion criteria, independent searching and a sensitivity analysis to contribute to the reduction of selection bias. Moreover, no funding source influenced the reporting bias. However, almost all the included studies were open-label studies, which will inherently result in bias indicated by an asymmetry in funnel plots.

Currently, DBS therapy is still in the stage of exploration and cannot be used as a recommended intervention. Most patients with TRD, who undergo DBS treatment, also receive antidepressant treatment, psychotherapy and other neuromodulation therapies, according to the 2016 guidelines by the Canada Network for Mood and Anxiety Treatment (CANMAT) for management of major depressive disorder (MDD) in adults (Kennedy et al., 2016). Previous studies in this field are largely non-randomized controlled small-sample studies. In the scope of reviewing the published literature, the SCG was the main target region in the brain used for depression. Other regions were also published, including the EpC, the sMFB, the posterior gyrus rectus, the vALIC, the VC/VS, and the NAc. Case reports mention a potential region of the inferior thalamic peduncle conveying thalamo-cortical information (Jiménez et al., 2005) and habenula associating with monoaminergic neurotransmission (Sartorius et al., 2010). A preliminary trial of DBS described that targeting the nucleus accumbens was more promising than the caudate, as evaluated by positron emission computed tomography (PET) (Millet et al., 2014). An exploratory meta-analysis was published in 2014 (Smith, 2014), which concluded that the procedure may be 71% more effective than a sham treatment, and one out of three patients with depression were expected to benefit from DBS. However, the estimates of the sham response were by expert opinions plus a random number software rather than practice data. Mosley et al. (2015) conducted a systematic review to exam the impact of DBS on depression. They did not make a concrete conclusion and called for methodological refinements.

In our study, DBS, for all the discussed targets, seemed to be a potential treatment option for TRD. On the one hand, the response and remission rates were respectively 56 and 32%, while the recurrence rate was relatively low (14%), suggesting that this treatment might be considered as promising. On the other hand, the AEs rate was common (67%) and moderate, which can be resolved by the corresponding therapy. However, suicide and suicidal ideation still occurred in the patients who received this treatment. Two studies (Dougherty et al., 2015; Puigdemont et al., 2015) improved the design to perform RCTs and found that there were no significant differences in the active and sham groups. This controversial opinion should be studied more extensively, with more blinded, randomized controlled trials. Among the included studies, eight studies (Mayberg et al., 2005; Malone et al., 2009; Merkl et al., 2013; Schlaepfer et al., 2013; Dougherty et al., 2015; Accolla et al., 2016; Fenoy et al., 2016) referred the short-term (<1 year) effect in DBS for TRD, and the respond rates ranged from 20 to 91.7% and the remission rates from 20 to 66.7%. In addition, eight studies (Lozano et al., 2008, 2012; Bewernick et al., 2010; Kennedy et al., 2011; Holtzheimer et al., 2012; Puigdemont et al., 2012; Bergfeld et al., 2016; Williams et al., 2016) assessed the long-term (more than 1 year) effect, and the respond rates ranged from 40 to 75% and the remission rates were from 20 to 60%. Specifically, in a 5-year follow-up study (Williams et al., 2016), the mean improvements from the pre-implant baseline for the HRSD were 41.2, 53.8, and 45% in 1 year, 2 years, and 5 years, respectively. In addition, 3 of 5 (60%) patients continued to be in remission at the end of the follow up. The evaluation of the side effects of DBS for TRD requires a consideration of various factors, including the surgical procedures, infection and the stimulation in unrelated brain regions. No evidence currently reveals that the approach damages neurocognitive performance. However, DBS treatment could be conducive to psychiatric disorders, manic episodes and other psychiatric disorders, but these conditions are transient and can be reversed by adjusting the parameters (Fitzgerald, 2008).

The exact mechanism of DBS is unknown and studies are ongoing on this regard. Some studies have demonstrated that the gamma oscillations inhibition and the facilitation of theta-gamma coupling by DBS is probably mediated by activation of inhibitory circuits in SCG and the enhancement of plasticity in the frontal cortex (Sun et al., 2015). Depression is generally considered to involve three compartments of the neurocircuitry including the dorsal, ventral, and modulatory. The dorsal compartment mediates the cognitive and motor aspects, and the ventral compartment is associated with the somatic and vegetative aspects. The modulatory system mediates the mutual interactions of these two compartments through an inhibitory pathway consisting of the amygdala, hippocampus, rostral cingulate cortex, and the hypothalamic-pituitary-adrenal axis (Morishita et al., 2014).

During the last decade, several studies have utilized anatomical and functional neuroimaging modalities to investigate the neurobiology of depression (Holtzheimer and Mayberg, 2010). Structural neuroimaging studies consistently report smaller medial volumes, primarily the hippocampus, amygdala, and the entorhinal cortex, in depressive patients (Sheline et al., 1996, 1998; Hastings et al., 2004; Videbech and Ravnkilde, 2004). Zhang et al. (2009) suggested that the subtle structural/functional dysfunction in the right anterior cingulate, insula, caudate tail and amygdala-parahippocampal regions was associated with depression. Moreover, it is reported that non-responsive patients have a more organized white matter (Taylor et al., 2008) and a smaller right medial frontal and striatal volumes compared to responsive patients and comparison subjects (Shah et al., 2002). Functional imaging studies show that depressive patients have a higher ratio of amygdalar-hippocampal to cortical blood flow (Hornig et al., 1997) and lower concentrations of occipital γ-aminobutyric acid (Price et al., 2009). The aim of the future studies should involve defining the best localization within that target area. Moreover, imaging studies are important to elucidate the mode of action.

There are several limitations in this study. The specificity of the DBS surgery and the relevant ethical issue attributes to the difficulty of performing high quality and large sample RCTs. We, therefore, performed this meta-analysis using the rates to obtain a more reliable conclusion by pooling rates from multiple studies. In addition, the study implementation, patient characteristics, and different sample sizes also contribute to the bias. However, we illustrated the clinical heterogeneity by performing a sensitivity analysis. Further studies should refine the methodology, investigate the optimal targets structures and stimulation parameters, and maximize the consistency of outcomes measurement to make comparisons feasible. Moreover, it is of extreme importance to analyze the exact network effects to understand the neurobiological mechanisms.

We concluded that the DBS treatment had good prospects. However, the problem of the reported adverse events is worth paying attention to and resolving, because it could lower the quality of life. Further basic scientific research is needed to search for more optimal brain structures to modulate the neural circuits and to investigate the mechanisms of underlying the DBS treatment.