Bed nucleus of the stria terminalis-nucleus accumbens stimulation for depression: A randomized, double-blind, crossover trial

Summary

Although deep brain stimulation (DBS) shows potential in treatment-resistant depression (TRD), the efficacy remains controversial. To assess the efficacy of DBS of the bed nucleus of the stria terminalis (BNST) and nucleus accumbens (NAc) for TRD, we conduct a randomized, double-blind, crossover trial after an open-label optimization stage. Stimulation leads to a mean decrease in the Hamilton Depression Scale-17 (HAMD) scores of 10.1 points (p < 0.001) with a 50% response rate at the end of the open-label stage. Anxiety, quality of life, and disability are also improved. The HAMD scores are significantly lower in patients during active DBS than during sham DBS (p < 0.001). Serious adverse events include suicide (1 patient) and seizure (1 patients). With individual stimulation analysis, we identify optimal stimulation sites, fiber tracts, and functional networks (false discovery rate [FDR] p < 0.05). This study demonstrates the efficacy and safety of BNST-NAc DBS for TRD and reveals optimal brain targets for stimulation. This study was registered at ClinicalTrials.gov (NCT04530942).

Graphical abstract

Highlights

• •BNST-NAc stimulation significantly alleviates treatment-resistant depression
• •The therapy also alleviates anxiety and enhances quality of life
• •Imaging revealed predictive sweetspots, fiber tracts, and functional networks

Lai et al. demonstrate in a randomized controlled trial that deep brain stimulation targeting BNST-NAc is an effective treatment for treatment-resistant depression. The findings pinpoint the optimal brain networks for stimulation, paving the way for more precise and personalized neuromodulation therapies.

Introduction

Depression is a leading cause of avoidable suffering in the world.^1,2 Approximately one-third of patients meet the criteria for treatment-resistant depression (TRD), for whom medication, cognitive behavioral therapy, and electroconvulsive therapy showed minimal effects.³ By directly modulating deep brain structures and neural circuits, deep brain stimulation (DBS) presents a promising potential to bridge this treatment gap.⁴

The striatal region is most commonly targeted in DBS for obsessive-compulsive disorder,⁵ and the structures around this area are increasingly valued as target for intervention in depression followed the subcallosal cingulate.⁶ Commonly targeted structures in this region are the anterior limb of the internal capsule (ALIC),^7,8,9 the ventral capsule/ventral striatum (VC/VS),^10,11 the nucleus accumbens (NAc),¹² and the bed nucleus of stria terminalis (BNST).^13,14 Although the initial open-label DBS study in 2009 targeting VC/VS reported significant improvements in depressive symptoms,¹⁰ subsequent randomized controlled trials (RCTs) failed to demonstrate a significant difference between the active and control groups.¹¹ Moreover, several RCTs have produced contrasting results.^7,14,15 Patient selection, trial design, outcome measures, anatomical targeting, and parameters of stimulation were argued to be responsible.³ Given that the published DBS RCTs targeting striatal region tested four targets in around 70 patients with contrasting results,^6,14 drawing definitive conclusions remains difficult without further evidence.

In this study, to evaluate the efficacy and tolerability of DBS targeting key nuclei within the striatal region in patients with TRD, we employed a targeting trajectory that traverses both the BNST and the NAc, distinguishing it from previous studies (see STAR Methods and Table S1). Importantly, we used eight-contact DBS electrodes, which enabled systematic exploration of effective stimulation sites across the striatal region, thereby maximizing the likelihood of identifying optimal targets during the open-label optimization phase. Patient-specific stimulation datasets were then analyzed alongside high-quality normative connectome to pinpoint optimal stimulation sites, fiber tracts, and networks for overall depression and depression dimension-related response.^16,17 The current study is designed with an open-label optimization stage, followed by a randomized, double-blind, crossover stage.

Results

Demographics and clinical characteristics

From March 29, 2021, to July 28, 2023, a total of 524 patients were screened for eligibility, of which 26 were included in the study to receive DBS at the BNST-NAc (Figures 1 and 2A). Patient demographic characteristics (85% male; mean age, 31.7 years [SD, 9.9]; mean age of depression onset, 17.8 years [SD, 4.6]; and mean hospitalization, 2.9 times [SD, 1.9]) are summarized in Table 1.

Figure 1.

Flowchart of the randomized controlled trial

Figure 2.

Active contact location and clinical results of open-label stage and randomized stage

(A) Intracranial electrodes (n = 26) and active contacts at the end of open-label phase.

(B) Depression (HAMD, MADRS), anxiety (HAMA), and anhedonia (DARS) scores at pre-versus postoperative 3 months (n = 26), 6 months (n = 26), and 1 year (n = 21). Masses in dark blue represent BNST; in light blue, NAc; and in pink, hypothalamus. Data are represented as median with interquartile range (IQR).

(C) Depression (HAMD, MADRS), anxiety (HAMA), and anhedonia (DARS) scores over time in each treatment group and their changes from post sham stimulation stage to post active stimulation stage.

HT, hypothalamus. Data are represented as mean ± standard deviation (SD). ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001.

Table 1.

Demographic and clinical characteristics

	All patients (n = 26)	Optimization stage			Randomization stage
		Responder (n = 13)	Non-responder (n = 13)	p	Sham-active (n = 10)	Active-sham (n = 8)	p
Age, years	31.7 (9.9)	32.7 (8.5)	30.6 (11.3)	0.610	34.4 (12.0)	29.3 (7.1)	0.309
Gender	4 F/22 M	4 F/9 M	0 F/13 M	0.096	0 F/10 M	2 F/6 M	0.094
Education, years	12.8 (3.1)	11.9 (2.7)	13.7 (3.4)	0.153	11.6 (2.8)	12.9 (3.0)	0.369
Occupation unemployeda	14 (54%)	7 (54%)	7 (54%)	1.000	5 (50%)	4 (50%)	1.000
Age of onset, years	17.8 (4.6)	18.2 (2.7)	17.4 (6.1)	0.682	19.6 (5.3)	16.5 (3.2)	0.165
Duration since onset, years	13.9 (3.6)	14.5 (8.4)	13.3 (8.0)	0.700	14.8 (3.0)	12.8 (6.1)	0.618
Previous psychiatric hospitalization, times	2.9 (1.9)	3.0 (2.2)	2.8 (1.5)	0.537	2.5 (0.8)	3.5 (2.9)	0.926
Follow-up in optimization stage, months	18.8 (5.6)	21.9 (8.7)	15.7 (6.2)	0.047∗	21.4 (2.5)	16.8 (9.2)	0.269
MoCA	26.5 (2.3)	26.8 (2.0)	26.3 (2.6)	0.617	26.4 (1.7)	26.1 (2.3)	0.774
Baseline HAMD	21.2 (3.8)	21.4 (3.9)	21.2 (3.7)	0.878	20.6 (3.6)	21.3 (4.0)	0.719
MINI risk of suicide score	17.2 (13.6)	13.9 (11.4)	20.5 (15.3)	0.229	17.7 (12.2)	16.8 (10.3)	0.863

F, female; M, male; MINI, Mini-International Neuropsychiatric Interview; MoCA, Montreal Cognitive Assessment.

^aUnemployed included schooling suspension (07, 09, 17), employed included schooling (24, 25).

During the open-label stage, three patients withdrew from the study at months 3, 9, and 10 of optimization; five patients were not randomized in the crossover trial (Figure 1). Consequently, 18 patients entered the crossover stage, with 10 randomized to sham-active and 8 to active-sham DBS. All 18 patients completed the RCT, although 8 had to be prematurely crossed over during the sham DBS phase: 4 responders (06, 10, 21, and 26) and 2 non-responders (24 and 25) with abrupt increase of symptoms within 1 day and 1 responder (08) and 1 non-responder (15) with gradual increase of symptoms of more than 2 days.

Open-label stage

At the last follow-up ([LFU], 18.8 ± 5.6 months) of the open-label optimization stage, stimulation led to a mean decrease in Hamilton Depression Scale-17 (HAMD) scores of 10.1 (SD, 7.1) points (p < 0.001; 95% confidence interval [CI], 7.2–13.0) and in Montgomery-Asberg Depression Rating Scale (MADRS) score of 13.5 (SD, 7.8) points (p < 0.001; 95% CI, 10.4–16.6) (Figure 2B; Table S2). Based on the HAMD scores at LFU, 13 of 26 patients (50%) were classified as responders, with 9 of them (35%) achieving remission. Through between-group comparisons, while most demographic and clinical characteristics did not differ significantly between groups, responders had a longer optimization stage duration than non-responders (p = 0.047, Table 1).

The self-rating Depression and Somatic Symptoms Scale (DSSS) score also decreased from 32.9 (SD, 13.2) at baseline to 14.9 (SD, 10.9) at LFU (p < 0.001). Additionally, compared to baseline, there were significant improvements in Hamilton Anxiety Scale (HAMA) (mean improvement of 10.4 points; SD, 7.2; p < 0.001; 95% CI, 7.5–13.3) and Dimensional Anhedonia Rating Scale (DARS) (mean improvement of 7.2 points; SD, 14.7; p = 0.020; 95% CI, 1.2–13.1) scores, although the significance of the DARS scores did not hold after multiple comparison correction. Quality-of-life measures, including 36-Item Short Form Survey Instrument (SF-36) (mean improvement of 116.4 points; SD, 162.7; p = 0.001; 95% CI, 50.6–182.1) and World Health Organization Quality of Life-BREF (WHOQOL-BREF) (mean improvement of 11.5 points; SD, 14.2; p < 0.001; 95% CI, 5.8–17.3), and Sheehan Disability Scale (SDS) disability scores (mean improvement of 8.6 points; SD, 9.1; p < 0.001; 95% CI, 4.9–12.3) were also significantly improved (Table S2).

Randomized, double-blind stage

At the beginning of the randomized double-blind crossover stage, the percentages of responders classified based on HAMD scores at T1 were 80% (8/10) and 38% (3/8) in the sham-active group and active-sham group, respectively. The ratio of responders to non-responders was not significantly different between sham-active and active-sham groups (Fisher’s exact test, p = 0.145), and no differences in any other demographical variables were found (Table 1). While the active-sham group showed slightly higher mean scores on depression rating scales (HAMD and MADRS) compared with the sham-active group (Table S3), no statistically significant differences were detected at T0, T1, and T2 before they switched to the crossover settings (all ps > 0.05). The Pittsburgh Sleep Quality Index (PSQI) scores (p = 0.046), 16-item Self-Report Quick Inventory of Depression Symptomatology (QIDS-SR16) scores (p = 0.015), and DSSS scores (p = 0.029) were higher in active-sham group at T0, but the significance did not hold after multiple comparison correction.

The study met the primary endpoint of BNST-NAc DBS efficacy for depression under blinded, controlled conditions with optimized stimulation parameters (Table S4). As shown in Table 2 and Figure 2C, the mean difference in HAMD scores between active and sham stimulation was 9.4 (SD, 7.1; p < 0.001; 95% CI, 5.9–13.0). No significant period (p = 0.589) or carryover (p = 0.069) effects were observed on HAMD scores. The active stimulation also induced significant treatment effects on MADRS scores (mean improvement of 13.8 points; SD, 8.4; p < 0.001; 95% CI, 9.6–18.0), HAMA scores (mean improvement of 10.2 points; SD, 9.2; p < 0.001; 95% CI, 5.6–14.7), QIDS-SR16 scores (mean improvement of 8.2 points; SD, 6.8; p < 0.001; 95% CI, 4.8–12.5), DSSS scores (mean improvement of 12.3 points; SD, 10.8; p < 0.001; 95% CI, 6.9–17.6), and DARS scores (mean improvement of 11.5 points; SD, 17.3; p < 0.001; 95% CI, 2.9–20.1). Significant treatment effects were also observed on quality-of-life measures (WHOQOL-BREF and SF-36) and disability scores (SDS) between active and sham stimulation; no significant period or carryover effects were observed on secondary outcome measures (Table 2).

Table 2.

Mean scores during the crossover trial (n = 18)

Variable	T0	T1	T2	T3	T4	T5	Sham	Active	Changes between sham and active	95% CI	Treatment effect	Carryover effect	Period effect
HAMD	20.9 (3.6)	9.7 (5.8)	16.8 (5.6)	15.8 (6.8)	19.4 (6.2)	15.7 (9.3)	20.5 (6.4)	11.1 (6.6)	9.4 (7.1)	5.9–13.0	30.795 (<0.001)	3.844 (0.069)	0.305 (0.589)
MADRS	31.3 (5.3)	15.3 (7.8)	25.4 (9.1)	24.2 (9.2)	29.4 (9.6)	24.3 (14.0)	31.2 (10.0)	17.4 (9.0)	13.8 (8.4)	9.6–18.0	47.641 (<0.001)	4.225 (0.058)	0.675 (0.423)
HAMA	24.6 (6.3)	12.0 (5.6)	20.4 (7.0)	17.7 (7.6)	21.9 (7.5)	18.1 (9,7)	23.0 (7.3)	12.8 (6.5)	10.2 (9.2)	5.6–14.7	23.723 (<0.001)	3.089 (0.089)	0.525 (0.474)
QIDS-SR16	18.6 (5.8)	8.3 (5.5)	14.8 (6.6)	13.2 (6.8)	16.1 (6.7)	12.7 (7.7)	17.0 (6.5)	8.8 (5.3)	8.2 (6.8)	4.8–12.5	25.244 (<0.001)	3.514 (0.070)	0.064 (0.802)
DSSS	34.8 (13.3)	12.7 (8.2)	22.8 (10.2)	19.7 (10.4)	23.4 (10.7)	18.4 (11.2)	25.2 (10.0)	12.9 (7.2)	12.3 (10.8)	6.9–17.6	24.974 (<0.001)	0.975 (0.331)	0.001 (0.972)
DARSa	23.9 (12.2)	34.4 (15.4)	24.6 (14.4)	26.8 (14.5)	23.5 (15.3)	28.7 (17.6)	22.0 (14.1)	33.5 (16.0)	11.5 (17.3)	2.9–20.1	7.329 (0.016)	0.381 (0.546)	0.018 (0.896)
WHOQOL-BREFa	29.4 (12.7)	43.9 (17.8)	33.9 (14.7)	36.2 (16.4)	31.4 (13.9)	36.7 (18.0)	30.9 (14.4)	42.0 (18.0)	11.1 (20.9)	0.7–21.5	5.694 (0.023)	2.181 (0.150)	0.022 (0.884)
SF-36a	252.9 (95.3)	410.1 (163.2)	289.6 (85.7)	312.4 (113.8)	280.4 (105.6)	333.9 (151.3)	269.7 (95.7)	376.6 (144.4)	106.9 (155.9)	29.4–184.5	7.741 (0.013)	1.237 (0.284)	0.066 (0.801)
SDS	25.4 (5.5)	14.3 (7.8)	20.7 (6.5)	17.6 (8.1)	19.6 (7.8)	16.3 (9.4)	20.8 (7.4)	13.1 (8.3)	7.7 (8.7)	3.3–12.0	12.746 (0.003)	0.587 (0.455)	0.031 (0.863)
PSQI	12.3 (3.2)	8.6 (3.8)	9.4 (2.5)	9.8 (3.2)	9.7 (2.9)	9.9 (3.9)	10.4 (3.1)	9.2 (3.9)	1.2 (4.5)	−1.0∼3.4	1.323 (0.267)	0.171 (0.685)	0.053 (0.821)

Data were expressed in mean (SD) for rating scales and F value (p value) for effects.

^aHigher scores in DARS, WHOQOL-BREF, and SF-36 means better state.

Adverse events

Adverse events related to surgery, device, and stimulation are listed in Table 3. The most commonly reported adverse event during open-label stage is sleep disturbances, with 9 patients and 20 reports. During sham DBS phase, increase in depressive symptoms was most commonly reported, with 8 patients and 8 reports. The number of reported adverse events markedly reduce during active DBS phase. As for severe or permanent adverse events, one patient had postoperative seizures 2 months after surgery, one patient died by suicide after surgery, and another patient reported continuous stiffness in the shoulders and neck. The patient who died by suicide 9 months after surgery had clinically responded to DBS at 7 months after surgery, with HAMD rated 28 at baseline to 6 at 7 months after surgery. Although it was not known whether there was an acute relapse before the suicide, as it had been 3 weeks since her last monthly assessment, the suicide note suggested it was not related to the treatment. No infection or intracranial hemorrhage was recorded.

Table 3.

Adverse events

Adverse event	Duration of effect	No. of patients/reports
Perioperative and open-label phase
Surgery related
Nausea	Transient	1/1
Fall	Transient	1/1
Total		2/2
Device related
Shutdown with impulse generator (IPG) due to battery depletion	Transient	6/9
Discomfort over switching parameter settings	Transient	3/4
Discomfort due to vibration around IPG	Transient	1/1
Pain around the IPG	Transient	2/2
Electric sensation	Transient	1/1
IPG replacement		2/2
Total		10/19
Stimulation related
Abdominal bloating	Transient	2/2
Agitation	Transient	7/9
Anxiety	Transient	5/7
Chest tightness	Transient	1/1
Constipation	Transient	1/1
Decreased weight	Transient	1/1
Depersonalization	Transient	1/1
Dizziness	Transient	2/4
Excessive salivation	Transient	1/1
Fatigue	Transient	3/3
Fear	Transient	3/3
Forgetfulness	Transient	1/1
Frequent urination	Transient	1/1
Headache	Transient	1/3
Hypersomnia	Transient	2/2
Hypomania	Transient	1/1
Increased appetite	Transient	4/4
Increased libido	Transient	3/3
Increased weight	Transient	3/4
Mania	Transient	1/1
Nausea	Transient	1/1
Obsession	Transient	2/5
Palpitation	Transient	4/5
Sleep disturbances	Transient	9/20
Tingling sensation in the back	Transient	1/1
Urinary incontinence	Transient	1/1
Total		19/86
Others
Stiffness in the shoulders and neck	Permanent	1/4
Suicide	Permanent	1/1
Seizure	Transient	1/1
Total		3/6
Open-label phase total		22/113
Sham DBS phase
Agitation	Transient	2/2
Anxiety	Transient	3/3
Decreased appetite	Transient	1/1
Dysuria	Transient	1/1
Fatigue	Transient	1/1
Increase in depressive symptoms	Transient	8/8
Muscle soreness	Transient	1/1
Palpitation	Transient	1/1
Self-harm	Transient	1/1
Suicidal ideation	Transient	4/4
Total		10/23
Active DBS phase
Anxiety	Transient	2/2
Pain in the left ear	Transient	1/1
Sleep disturbances	Transient	1/1
Total		4/4
Randomized phase total		11/27

Sweetspots, fiber filtering, and network mapping analysis

To identify optimal stimulation sites and related brain networks, we reviewed 567 post-operative programming datasets and selected 150 sessions for subsequent connectomic analysis (detailed information of these 150 sessions were provided in Tables S5, S6, and S7). These sessions were conducted between 1 month and 1 year after surgery, with each spaced at least 2 weeks apart from the prior parameter session. By assembling qualified stimulation data from the optimization stage, we found that the electric fields (E-fields) predominantly overlapped with the BNST and extended to encompass the NAc and internal capsule (Figure 3A). Sweetspot analysis identified the optimal stimulation site in the ventral portion of the internal capsule and lateral to the BNST, with a lateral shift observed in sweetspot at more anterior planes (Figure 3B). Spatial map consisting of sweetspots was cross-validated across the entire cohort in a leave-one-patient-out design (r_s = 0.57, p < 0.001). The fiber tracts positively and negatively associated with outcome metrics diverge markedly near the stimulation site (Figure 3B, r_s = 0.33 at p < 0.001). Compared to previously reported depression-related fiber tracts, the tract identified as positively related to outcome in the current study—while not fully overlapping in trajectory—shared cortical projection areas with the superolateral branch of the medial forebrain bundle (slMFB) and the anterior thalamic radiations (ATRs), including the dorsolateral prefrontal cortex (dlPFC), orbitofrontal cortex, and the medial prefrontal cortex (mPFC) (Figure 4). The two primary tracts positively related to depression improvement overlapped at the sweetspot location, highlighting a common pathway associated with optimal clinical improvement (Figure 3B right upper and right middle panels).

Figure 3.

Location of the E-fields and results summary including the depression models from DBS fiber filtering, sweetspot mapping, and network mapping

(A) Location of E-fields in coronal, axial, and sagittal T1-weighted MRI (MNI152 NLIN 2009b) slices; the hot map denoted the number of binarized E-fields (thresholded at 0.2 V/mm for visualization) overlapped in each voxel. Green mass represents BNST; blue mass represents NAc.

(B) The three levels of analysis were able to explain the amount of variance of clinical outcomes (percent changes in HAMD-17 scores) when analyzed in a circular nature and led to significant leave-one-out cross-predictions of clinical outcomes across leave-one-patient-out cross-validation; plots show fitting of a linear model that represents the degree to which stimulating voxels (sweetspots), tracts (fiber filtering), and functional regions (network mapping) explain variance in clinical outcomes using Spearman’s correlation across the whole cohort (N = 150). Gray shaded areas represent 95% confidence intervals. In sweetspots, voxels with positive correlation to outcome were shown in yellow-red and negative ones were shown in green-blue. For clarity, 40% of the fiber tracts identified positively/negatively correlated to depressive symptom improvement are shown in the image. Three level analysis results were superimposed on slices of a brain cytoarchitecture atlas in MNI152 space (up-left). The displayed sweetspots, fibers, and projections survived FDR correction at 0.05 level of significance. Involvement of different network nodes is displayed as bar plot (the mean ± SD of r_s values) and yellow line (percentage involvement of the corresponding brain region).

Figure 4.

Comparison of current depression tract with slMFB and ATR

The fiber tracts (red) positively related to depressive symptom improvement isolated from Figure 3 overlaid with fiber tracts from the open-source tracts including slMFB (gray) and ATRs (blue) by Middlebrooks et al.¹⁸ Pu, putamen; STN, subthalamic nucleus; RN, red nucleus; HTH, hypothalamus; SNc, substantia nigra pars compacta; SNr, substantia nigra pars reticulate; VTA, ventral tegmental area.

The results from network mapping analysis survived the false discovery rate (FDR) correction at 0.05 significance level and leave-one-patient-out cross-validation (r_s = 0.14, p < 0.05). Increases in functional connectivity to the majority of some nodes, including dlPFC (92.9%), dorsal anterior cingulate cortex (dACC) (92.8%), ventral lateral prefrontal cortex (86.9%), and putamen (89.7%), and decreases in functional connectivity to the majority of the others, including amygdala (85.9%), SCC (84.9%), medial frontal cortex (MedFC) (89.5%), ventral medial prefrontal cortex (vmPFC) (46.8%), and NAc (47.6%), were suggested to be associated with better improvement (Figure 3B). The comparatively large standard deviations in optimal connectivity within certain nodes indicate diversity across their subregions, e.g., insular and superior parietal lobule. We leveraged dimensional HAMD symptom ratings to identify not only symptom-specific brain networks but also the associated white matter fiber tracts, which could serve as templates for both invasive and non-invasive neuromodulation (Figure S1, p < 0.05 FDR corrected).

Discussion

In this study, we conducted an RCT with a crossover design, utilizing eight-contact electrodes to assess the efficacy and safety of BNST-NAc DBS in TRD. Our results demonstrated that BNST-NAc DBS significantly alleviated depressive symptoms, with marked improvements in anxiety, quality of life, and functional disability. Connectomic analyses further identified stimulation sweetspots, fiber tracts, and functional networks associated with clinical benefit. These findings were supported by rigorous multiple comparison corrections and validated through out-of-sample cross-validation, underscoring the robustness and generalizability.

DBS holds promise for alleviating symptoms in patients with treatment-resistant psychiatric disorders, but RCTs conducted thus far for depression have yielded mixed results.^6,19 Moreover, findings from open-label studies may differ substantially from those of RCTs. For example, subcallosal cingulate and VS stimulation have shown efficacy for depression in open-label studies,^4,10,12 while recent RCTs did not fully support these findings.^11,20 Study design, targeting precision, postoperative programming, and disease symptom dimension are key factors contributing to this heterogeneity.^7,16,21

Study design is a critical consideration, as negative RCTs (targeting the SCC, the VC/VS, and BNST) did not include an initial open-label parameter optimization phase prior to the RCT.^11,14,20 This may have been crucial, considering inter-individual differences in anatomy and the need to optimize stimulation contacts and parameter protocols. In contrast, a positive RCT targeting the ventral ALIC included a longer (51.6 weeks) initial open-label stimulation parameter optimization phase.⁷ One additional advantage of DBS is that, once implanted, the device can be turned on and off, permitting a crossover study of its effects within an individual patient.²² In this study, we used the crossover design to control for placebo effects with active and sham stimulation phases⁷ and ensured most patients underwent more than one monopolar stimulation from an 8-contact electrode during over-6-month optimization stage, offering a much wider range of spatial and amplitude variability for optimization. The optimization stage in this study may also reduce the placebo effect evident.²³ Additionally, due to the high rates of suicides in TRD patients, we adopted the premature crossover strategy to ensure safety.⁷ Prior to any premature crossover, all participants who underwent early transition received a comprehensive assessment by our team, and the overall premature crossover rate was comparable to previous study.⁷

To date, neither have evidence-based guidelines for the use of DBS in depression been established nor is there a consensus on the optimal target. Following the subcallosal cingulate,^3,4,20,23,24 the striatal region is the second most commonly reported target in DBS for depression.⁶ The commonly targeted structures in striatal region for depression are the ALIC,^7,8,9 VC/VS,^10,11 NAc,¹² and BNST.^13,14 Although these targets are nominally different, it has been increasingly recognized that stimulation of these areas often encompasses similar regions. For example, the NAc is the most ventral extension of the striatum and is located directly below the ALIC, so targeting this region is likely to involve stimulation of both anatomical structures.¹² Strikingly, the sweetspots identified in this study lied ventral and medial within the striatal region, with the cortical projection area largely overlapping the two main tracts identified through fiber filtering analysis.¹⁸ While the sweetspots are very close to, but not exactly in, the BNST or NAc, a possible explanation could be that the stimulation covering both the BNST and NAc provides optimal coverage of these sweetspots.

Although the pathophysiology of depression is not fully understood, the accumulated evidence suggests that depression is a network disorder resulting from dysfunctional networks.⁶ We incorporated four established networks (affective network [AN], cognitive control network [CCN], default mode network [DMN], and reward network [RN]) into our analysis to contextualize the depression-improvement network associated with BNST-NAc DBS in the current study.^6,21,25 The identified depression-improvement network connected the areas of main networks related to mood, reward, and cognitive control. The propensity of optimal functional connectivity seeded from stimulation sites generally differs at the inter-network level but is consistent at the intra-network level. The functional network identified in this BNST-NAc DBS study replicates previous findings, which is inverted relative to the repetitive transcranial magnetic stimulation (rTMS) network for depression improvement, supporting that high-frequency rTMS and high-frequency DBS may modulate brain circuits in opposite directions.²⁶ The functional networks from our analysis may serve as guidance for refining invasive and non-invasive stimulation protocols.^27,28,29 It should be noted that, the spatial overlap among these networks—such as the shared regions of the CCN and DMN in the dACC and inferior parietal lobule, and of the DMN and RN in the MedFC and vmPFC—implies statistical non-independence in their connectivity measures, which poses a methodological challenge for reliably quantifying their unique contributions. A full decomposition of these overlapping networks is, therefore, precluded by the scope and data constraints of the present work.

Compared to previous studies (summarized in Table S8), besides the randomized, double-blind, crossover design, the strength of our clinical trial lies in three aspects. (1) We incorporated a sufficiently long (≥6 months) open-label optimization stage (also used by Bergfeld et al.⁷) prior to entering the randomized, double-blind, crossover stage. This design addresses a critical limitation of previous RCTs with negative results, which may have failed to demonstrate treatment effect due to inadequate parameter optimization. (2) The use of 8-contact electrodes traversing both the BNST and NAc, combined with our optimization stage, allowed us to search through adjacent targets for optimal stimulation parameters. (3) This work implemented a washout design in an active-sham crossover DBS study on depression. A washout period was suggested to minimize possible carryover effects in previous studies, while the time required for washout in DBS studies for depression is poorly understood.^3,7 Unlike other psychiatric disease, TRD is associated with high rates of suicide; the 2-day washout in this study was strictly designed to be longer than traditional washout periods in designs for movement disorders (45 min∼1 day)^30,31 and also to ensure patient safety.³² The analysis revealed no significant carryover effects in either primary or secondary outcome measures, indicating that the washout period was adequate for most assessments with minimal concern, providing a crucial framework for future invasive neuromodulation studies.

Limitations of the study

This study has several limitations. First, although perhaps less pronounced than in DBS for movement disorders, the surgery itself can induce a microlesion effect, which may confound analyses aimed at isolating the effects of stimulation.^33,34 To mitigate this effect, the sweetspot, fiber filtering and network mapping analyses were restricted to visits conducted between 1 month and 1 year after surgery and the stimulation parameters in each included visit were kept constant for at least 2 weeks.³⁵ Second, our observation that responders tended to present a longer optimization stage during the open-label stage emphasizes the importance of implementing a dedicated optimization design before crossover. In the absence of a treatment response, prolonging this optimization to 1 year or longer may be warranted. Third, while our study identified beneficial fiber tracts and networks through normative connectome analysis, we did not use patient-specific connectivity (diffusion tensor imaging/fMRI) data and included multiple sessions in a single patient. Although in studies with limited sample sizes (e.g., DBS for neuropsychiatric disorders), it is sometimes necessary to include data from all available treatment sessions,^30,36 we caution that high-quality, patient-specific functional or structural imaging data and multi-center design would be valuable for further validation.²⁶ Finally, the assessment scales in the current study are fundamentally constrained, preventing real-time measurement of disease state changes. In the future, digital tools powered by artificial intelligence may offer a promising solution to dynamically track individual disease state transitions.^24,37

To conclude, this study demonstrates the efficacy and safety of BNST-NAc DBS in patients with TRD, identifying sweetspots, fiber tracts, and a multiple comparison-corrected, cross-validated functional network responsive to depression. The specific connectivity map could also serve as a guide for optimizing both invasive and non-invasive stimulation. Additionally, our experience in study design may offer valuable insights for improving future DBS trial workflows for psychiatric disease.

Resource availability

Lead contact

Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Bomin Sun (sbm11224@rjh.com.cn).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •All data reported in this paper will be shared by the lead contact upon request.
• •This paper does not report original code.
• •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Acknowledgments

This study is funded by the National Natural Science Foundation of China (grant no. 82271515 [to B.S.]), the Fundamental Research Funds for the Central Universities (grant no. YG2025QNB35 [to Y.L.]), the National Natural Science Foundation of China (grant no. 82401813 [to Y.W.]), the Science and Technology Commission of Shanghai Municipality (grant no. 20410712000 [to D.L.]), Medical Research Council Senior Clinical Fellowship (grant no. MR/W020408/1 [to V.V.]), the SJTU Trans-med Awards Research (grant no. 2019015 [to B.S.]), the Scientific and Technological Innovation Action Plan of Shanghai (grant no. KY20211478 [to B.S.]), and the Shanghai Municipal Science and Technology Major Project (grant no. 2021SHZDZX 835 [to B.S.]). This deep brain stimulation intervention was supported by an unrestricted, investigator-initiated research grant by SceneRay Inc. (B.S. and D.L.), which provided the deep brain stimulation devices used. The funders of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. We sincerely thank all participants for taking part.

Author contributions

Conceptualization, B.S., S.Z., V.V., and Y.L.; methodology, B.S., S.Z., V.V., Y.L., and Y.Z.; investigation, Y.L., Y.Z., Y.W., L.D., K.Y., X.L., F.W., K.H., P.H., X.Z., X.Q., Y.P., and W.L.; writing – original draft, Y.L. and Y.Z.; writing – review & editing, Y.L., Y.Z., Y.W., L.W., J.Z., Y.Z., Q.L., Y.F., Z.Y., T.-F.Y., N.L., A.M.L., V.V., S.Z., and B.S.; funding acquisition, Y.L., Y.W., V.V., and B.S.; resources, C.Z., D.L., V.V., S.Z., and B.S.; supervision, A.M.L., V.V., S.Z. and B.S.

Declaration of interests

The authors declare no competing interests.

Declaration of generative AI and AI-assisted technologies in the writing process

During the preparation of this work, the authors did not use generative AI or AI-assisted technologies.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Software and algorithms
MATLAB R2023a	The MathWorks	https://matlab.mathworks.com/
SPSS 26	IBM	https://www.ibm.com/analytics/spss-statistics-software
Lead-DBS 2.5.3	Open source	http://lead-dbs.org
SPM 12	Open source	http://www.fil.ion.ucl.ac.uk/spm
DPABI toolbox 7.0	Open source	http://rfmri.org/dpabi

Experimental model and study participant details

Between July 2021 and April 2024, we did this single center, double-blind, randomized controlled and crossover trial at Ruijin Hospital (Shanghai, China). The study was approved by the Ruijin Hospital Ethics Committee, Shanghai Jiao Tong University School of Medicine. The research use of the implanted components was authorized by the Shanghai Testing & Inspection Institute for Medical Devices and approved by China Food and Drug Administration. All included patients provided written informed consent.

Inclusion criteria were as previously described^35,38,39: 1) patient aged 18–65 years; 2) meeting the international Classification of diseases (ICD)-10 definition of non-psychotic major depressive disorder (MDD); 3) current episode ≥2 years duration and/or more than 4 repeated episodes with current episode ≥1 year duration and a minimum of 5 years since the onset of the first depressive episode; 4) lack of antidepressant response to a minimum of three antidepressant treatments of adequate dose and duration, including at least two medications from two different classes, failure of adequate psychotherapy and electroconvulsive therapy (ECT) (either poor response, intolerance or rejection); 5) remaining stable with the current anti-depressive medicine for the last month and 6) Hamilton Depression Scale-17 (HAMD) score ≥17. Exclusion criteria were: 1) schizophrenia or psychosis unrelated to MDD; 2) severe personality disorder, neurological disorders or medical disorders; 3) history of brain surgery; 4) contraindications for anesthesia or stereotactic surgery.

Method details

Study design

The study started with a longitudinal, open-label trial followed by a randomized, double-blind crossover design (Figure 1). A study patient advisory group advised on study design in study set-up, and during recruitment. Following electrode implantation, patients underwent an initial open-label stage with parameter optimization for at least six months, during which psychometric assessments will be conducted and documented at each programming session and at three fixed time points: 3 months, 6 months and 12 months post-surgery. Patients should undergo observation for at least two weeks after programming to evaluate the therapeutic efficacy of the selected stimulation parameters. Following the open-label phase, patients were randomly assigned to either receive active DBS for two weeks followed by sham DBS for two weeks (the active-sham group), or vice versa (the sham-active group). Prior to each active and sham phase, a 2-day washout period was implemented. If deemed necessary by the research team or at the request of the patients, the phase was terminated, with patients progressing to the next phase upon termination. Throughout the crossover phase, both medication and DBS parameters were kept constant, and patients were evaluated at five key time points: the pre-randomization phase (T1), after two washout phases (T2 and T4), and after each active or sham DBS phase (T3 and T5).

We used an envelope procedure for randomization, wherein each randomly assigned group was enclosed within a sealed, opaque envelope. An independent researcher sequentially opened these envelopes and conducted the randomization by switching DBS on or off accordingly. Throughout the trial, patients, two assessors (YW, LD, psychiatrists), and other relevant researchers were kept blinded to the stimulation condition (patient/assessors-blind).

Procedures and programming

All patients received bilateral implants of 8 contact electrodes (SR1202-S; SceneRay, Suzhou, China). The target was set at the NAc, with the distal 1 or 2 contacts below the NAc and the proximal 1 or 2 contacts close to the dorsal BNST.^35,38,39 The length of each contact is 1.5 mm and the spacing between contacts is 0.5 mm. Target coordinates for the electrode tip were approximately 4–8 mm lateral to the midline, 1–3 mm anterior to the anterior border of the anterior commissure and 5–8 mm inferior to the anterior commissure. Postoperative CT was obtained to assess for intracranial hemorrhage and lead localization. One participant had an additional surgery 5 days after first implantation but before randomization to reposition a lead due to the lead not being in the ideal target region.

Stimulation was initiated one-week post-surgery and parameters were systematically adjusted approximately every two weeks.³⁵ Monopolar stimulation was predominantly utilized across the patient cohort. In practice, we increased the amplitude upon observing diminished benefits or limited clinical improvement, potentially reaching 6 V. Frequency and pulse width were subsequently tailored if needed. If these adjustments were ineffective, we introduced the second and, if required, the third preselected contact for monopolar stimulation.

Outcomes

During the study, patients underwent assessments using standardized measures for depression, anxiety, anhedonia, quality-of-life and disability. The primary outcome measure of the trial was the difference in HAMD scores between the active and sham stimulation phases in the randomized phase.⁴⁰ Additionally, a patient was considered a responder if they exhibited a decrease of at least 50% in their HAMD score compared to baseline. Remission was defined as achieving a HAMD score of less than 8.³⁹

Secondary outcome measures were the Montgomery-Asberg Depression Rating Scale (MADRS) for depression,⁴¹ Hamilton Anxiety Scale (HAMA) for anxiety,⁴² and the Dimensional Anhedonia Rating Scale (DARS) for anhedonia.⁴³ Other secondary outcome measurements included the 16-item Self-Report Quick Inventory of Depression Symptomatology (QIDS-SR16), Depression and Somatic Symptoms Scale (DSSS), World Health Organization Quality of Life- BREF (WHOQOL-BREF), 36-Item Short Form Survey Instrument (SF-36), Sheehan Disability Scale (SDS) and Pittsburgh Sleep Quality Index (PSQI). The psychiatrists (YW and LD) rated adverse events at monthly visits. An adverse event was rated as serious if it resulted in death, permanent loss of biological function, or a life-threatening situation, hospitalization, or prolongation of hospitalization. Events were classified as being related to surgery, the device, or the stimulation.

Quantification and statistical analysis

Localization of DBS electrodes and calculation of electric fields

Localization of DBS electrodes and calculation of corresponding electric fields (E-fields) were implemented using Lead-DBS software, SPM12 and DPABI.^35,44,45 DBS electrodes were localized using Lead-DBS using the PaCER algorithm. In brief, postoperative CT images were first linearly coregistered to preoperative MRI and normalized into ICBM 2009b NLIN asymmetric space using the SyN approach implemented in Advanced Normalization Tools. DBS electrodes were then localized using Lead-DBS and warped into the Montreal Neurological Institute (MNI) space using the PaCER algorithm after visual review and refinement of the coregistrations and normalizations. Anatomical segmentations of local subcortical structures at the DBS site as defined by the CIT-168 atlas.⁴⁶ Electric Fields Estimation: E-fields were calculated applying a finite element method (FEM)-based model in each patient. Conductivities of 0.33 and 0.14S/m were assigned to gray and white matter, respectively. Based on the volume conductor model, the potential distribution was simulated using the integration of the FieldTrip-SimBio pipeline. The surface of the volume mesh served as the anode. Subsequently, the gradient of the potential distribution was calculated by derivation of the FEM solution.

Normative connectomes, connectivity estimation and fiber tracking

Functional and structural connectivity estimates between each E-field and all other brain voxels were computed using 2 normative connectomes: the GSP1000 connectome⁴⁷ and the HCP32 connectome,⁴⁸ respectively. These connectomes were preprocessed and available in Lead-DBS toolbox.

Functional connectome (the GSP1000 connectome) and connectivity estimation: the GSP1000 Processed Connectome is derived from data acquired by the Brain Genomics Superstruct Project (GSP), in which the functional connectome data was acquired on matched Siemens 3 T MAGNETOM Tim Trio MRI systems (Erlangen, Germany) using the vendor-supplied 12-channel phased-array head coil. Functional preprocessing was performed with a modification of Thomas Yeo’s Computational Brain Imaging Group (CBIG) fMRI preprocessing pipeline to generate the BOLD runs used in the connectome.⁴⁷ Briefly, preprocessing included skipping first four frames, slice timing correction, motion correction, 0.0001–0.08 band-pass filtering, and regressing out the average BOLD time series over cerebrospinal fluid (CSF) and white matter. Normalization of functional volumes to the MNI space was then performed using FreeSurfer. Voxel-wised correlations between time series of voxels inside each bilateral E-field model and of remaining whole-brain voxels were calculated, averaged across the normative sample, and Fisher z-transformed to construct the functional connectivity profile. Values in E-fields served as weights to generate the connectivity profile using the Lead-DBS tool.^44,49

Structural connectome (the HCP32 connectome) and fiber tracking: briefly, multishell diffusion-weighted and T2-weighted imaging data from 32 subjects of the Human Connectome Project at Massachusetts General Hospital (https://ida.loni.usc.edu/login.jsp) were used as normative structural connectome. Whole brain tractography fiber sets were calculated using a generalized q-sampling imaging algorithm (GQI, a model-free deterministic tractography method).⁵⁰ Tracking parameters included a step size of 1 mm and an angular threshold of 60°. Sampling was performed within a white matter mask that was defined using the unified segmentation approach on T2-weighted structural acquisitions and that was coregistered to the b0 volume using SPM12. In each subject, 200,000 streamlines were sampled within this mask. Fiber tracts were then transformed into MNI space using a diffeomorphic registration algorithm as previously described.⁴⁸ For each patient, fibers passing through a non-zero voxel of the E-field were selected from the normative connectome and projected onto a voxelized volume in standard space (2mm resolution) while keeping count of the fibers traversing each voxel. Each fiber received the weight of the maximal E-field magnitude of its passage and fiber densities were weighted by these values.

Optimal stimulation sites (sweetspots), fibertracts and functional networks associated with DBS improvements in depression

As the implantation surgery could induce a microlesion effect, potentially interfering with stimulation-related analyses, the sweetspot, fiber filtering (to identify fibertracts associated with improvements) and network mapping (to identify networks associated with improvements) analyses were restricted to visits conducted between one month and one year after surgery.³³ The stimulation parameters in each included visit were kept constant for at least 2 weeks. Before these analyses, the impact of subjects’ demographic and clinical characteristics (including age at surgery, gender, education years, occupation, disease duration, hospitalization times, follow-up intervals, baseline MOCA scores, baseline HAMD scores and MINI risk of suicide scores) on the outcome were assessed via general linear model. As the results showed baseline (T0) HAMD score was the only clinical variable predictive of DBS response (p = 0.008), with higher baseline scores predicting greater percentage improvement, it was then added as a covariate in the following analysis. The identified sets of sweetspots, fiber tracts and networks were considered significant at FDR correction at a 5% α-level and later validated by estimating clinical outcome with leave-one-out cross-validation.^28,51

Sweetspot analysis: For each voxel covered by the group of E-fields, E-field magnitudes across patients were correlated with clinical outcomes using Spearman’s correlation. The resulting sweetspot maps (should also covered by ≥ 20% of E-fields with a magnitude above 0.2 V/m) would peak at voxels in which stronger E-fields were associated with better treatment responses.^44,49

Fiber filtering analysis: For each set of the stimulation parameters, fibertracts traversing each E-field voxel were selected and projected to a voxelized volume in MNI space. Each of these fibers were weighted by the Spearman’s correlation coefficient calculated between accumulative E-field vector magnitudes and clinical improvements across patients,⁵¹ only fibers that traversed >20% of stimulation volumes with an E-field magnitude >0.2 V/mm were considered.⁵² These r_s values represent the association strength between fiber tracts and clinical outcomes, and were used to color-code them accordingly.^44,49

Network mapping analysis: Voxel-wised Spearman’s correlations between Fisher-z-scored connectivity and clinical improvements (HAMD and its subscales) were calculated to identify functional networks predictive of the clinical outcome, which yields R-map models of optimal connectivity.^48,53 These R-maps denote optimal relative distributions of the connectivity; the more closely a patient’s connectivity profile matches this optimal map, the greater the expected clinical improvement.^44,49,54 To better describe this depression network with prior knowledge on mood processing, we assessed key areas including amygdala (Amy), hippocampus (Hip), insula (Ins), orbitofrontal cortex (OFC), SCC, dorsolateral prefrontal cortex (dlPFC), dorsal anterior cingulate cortex (dACC), posterior cingulate cortex (PCC), superior parietal lobule (sPL), inferior parietal lobule (iPL), precuneus (PCu), medial frontal cortex (MedFC), ventral medial prefrontal cortex (vmPFC), ventral lateral prefrontal cortex (vlPFC), caudate (Cau), putamen (Put) and nucleus accumbens (NAcc) in affective network (AN), cognitive control network (CCN), default mode network (DMN) and reward network (RN) to identify their proportion (the percentage of certain brain area included) and propensity (how certain brain area is functionally connected to form this depression network, namely the mean and SD of correlation coefficient) participated in this network.⁶ In addition, the patient-specific functional connectivity within the four networks (AN, CCN, DMN and RN) seeding from E-fields were.

Statistical analysis

Based on reference literature and incorporating clinical experience from earlier studies, it is expected that the active stimulation group in this trial will achieve a further reduction of 9.5 points in HAMD scores compared to the sham stimulation group, with a conservative estimate of a pooled standard deviation of 8.5 points for both groups. Under the parameters mentioned above, when the significance level of the statistical test is set at 5% (two-sided), and the statistical power is set at 90%, the sample size calculated following statistical principles is 17 cases. We anticipated enrolling 34 patients and considered the objective met when at least 17 patients entered the randomized double-blind crossover phase, factoring in a dropout rate of 50%. Data are presented as mean (standard deviation, SD) in tables and as mean (standard error, SE) in figures, and statistical significance was set at a 2-tailed 5% level. After data normality was determined by Shapiro-Wilk test, correlational analyses were performed using Spearman rank correlation or Pearson correlation, group differences were inferred using Wilcoxon signed-rank sum test or paired student t-tests. Categorical variables were analyzed by Fisher’s exact test or a χ² test as appropriate. Linear regression was used to model the relationship between a dependent variable and one or more independent variables using the stepwise method.

Intention-to-treat analyses (ITT) was performed to results in this RCT. In the open-label phase, outcome measures were analyzed using paired t-tests with Bonferroni correction. Descriptive statistics included the percentage of responders, non-responders, and remission cases. In the crossover phase, analyses of the primary efficacy outcome involved testing three effects: carryover, period (T3 and T5), and treatment (active and sham) effects. This was accomplished using a general linear mixed model for each tested outcome measure, with outcome scales as the dependent variables, group (active-sham and sham-active group) and treatment as independent variables, and outcome scores at T2 as a covariate. Carryover effects were assessed examining the between-group main effect. Period effects were assessed through the interaction term of treatment × group effects. Treatment effects were assessed using the within-group main effects. Data analyses were performed using SPSS 26.0 and MATLAB R2023a.

Additional resources

This randomized controlled trial is registered with Clinicaltrials.gov (NCT04530942, https://clinicaltrials.gov/study/NCT04530942?term=bomin%20sun&rank=2).

Published: March 31, 2026