Introduction

Major depressive disorder (MDD) is a chronic and debilitating psychiatric disorder characterized by persistent and profound mood disturbances [1]. Despite receiving adequate antidepressant treatments, approximately 30% of patients exhibit treatment-resistant depression (TRD), wherein clinical symptoms remain refractory to conventional interventions [2]. Deep brain stimulation (DBS) has been proposed as a salvage therapy for TRD, demonstrating promising clinical efficacy in trials [3]. Currently, the primary DBS targets for MDD include the subcallosal cingulate gyrus (SCG), anterior limb of the internal capsule (ALIC), ventral capsule/ventral striatum (VC/VS), nucleus accumbens (NAc), bed nucleus of the stria terminalis (BNST), inferior thalamic peduncle, and lateral habenula. Among these, VC/VS, NAc, BNST, and ALIC are anatomically adjacent and structurally interconnected, suggesting that their stimulation modulates overlapping yet distinct neural circuits [4]. We hypothesize that combined DBS targeting the BNST-NAc may produce synergistic effects in TRD treatment.

Although DBS has shown clinical efficacy in TRD, the precise mechanisms underlying its therapeutic effects remain unclear. Most studies have focused on localized brain changes or alterations in structural/functional connectivity [5,6,7,8,9], yet emerging evidence highlights the importance of large-scale networks. Electroencephalography (EEG) microstate analysis provides a functional representation of brain networks with high temporal resolution. Microstates are a small set of semi-stable, polarity-insensitive scalp voltage topographies that can account for the majority of variance in resting-state EEG signals [10]. Each microstate typically lasts approximately 100 milliseconds before rapidly transitioning to another state. Despite minor variations in classification across studies, four canonical microstate classes (A–D) are consistently observed and are believed to reflect the coordinated activity of large-scale brain networks [11]. Simultaneous fMRI-EEG studies have linked specific microstates to well-known resting-state networks, such as the auditory, visual and default mode network (DMN) [12,13,14]. Thus, EEG microstate analysis provides temporally precise window into the dynamics of large-scale brain networks.

In addition, EEG microstates exhibit pathophysiological relevance in MDD. Cross-sectional studies have demonstrated that specific microstate properties can distinguish patients from healthy controls [15]. Longitudinal studies further indicate that effective MDD treatments, including selective serotonin reuptake inhibitors (SSRIs), magnetic seizure therapy, electroconvulsive therapy, and transcranial magnetic stimulation (TMS), influence specific microstate parameters, with changes correlating with clinical symptom improvement [16,17,18,19]. Collectively, these findings suggest that microstates may serve as both state and trait markers of the disorder.

Converging evidence indicates that DBS modulates large-scale brain networks. Previous research suggests that SCG DBS significantly modulates DMN, salience network, central executive network, and somatomotor network, with regional cerebral blood flow alterations in the DMN correlating with depression severity [20, 21]. These findings imply that localized deep brain stimulation can exert widespread effects on large-scale neural networks, influencing clinical symptoms. Supporting this, our prior positron emission tomography study in TRD patients undergoing BNST-NAc DBS revealed that short-term stimulation reduced dopamine D2 receptor binding in the amygdala, caudate nucleus, and substantia nigra [22], indicating that BNST-NAc DBS exerts widespread neurobiological effects. Based on these findings, we hypothesize that BNST-NAc DBS modulates large-scale brain networks, with potential relevance to patients’ emotional symptoms. Given that EEG microstates represent the rapid temporal dynamics of large-scale brain networks, they offer a powerful tool to examine this hypothesis. However, research on DBS-induced microstate alterations remains limited to two studies in Parkinson’s disease [23, 24], with no prior investigations in psychiatric disorders.

In this study, we designed a randomized, double-blind, crossover trial with two primary objectives: (1) to investigate whether BNST-NAc DBS induces measurable alterations in EEG microstate dynamics, and (2) to explore whether stimulation-dependent microstate changes correlate with clinical symptom improvements while assessing their synchronization with local field potentials (LFP). The clinical trial consisted of two phases: an open-label optimization phase lasting at least six months, followed by a randomized, double-blind, crossover phase. During the open-label phase, stimulation parameters were biweekly adjusted based on clinical response. Eligible participants then entered the double-blind phase and were randomized to receive either active-then-sham (On-Off) or sham-then-active (Off-On) DBS, with each stimulation period lasting two weeks and separated by a 2-day washout. Clinical assessments and simultaneous EEG-LFP recordings were conducted at pre-randomization baseline, and immediately following completion of each stimulation period and washout interval. A total of ten participants who completed the double-blind phase and had both EEG-LFP and clinical data under active and sham conditions were included in the present analysis. Through this multimodal approach, we aim to identify potential biomarkers associated with treatment response, ultimately enhancing the mechanistic understanding of DBS effects in TRD.

Methods

Study participants

A total of 26 participants were recruited from Ruijin Hospital, Shanghai Jiao Tong University School of Medicine. The inclusion criteria were as follows: (1) men and women aged 18–65 years; (2) diagnosis of non-psychotic major depressive disorder (MDD) based on the International Classification of Diseases (ICD)-10; (3) current depressive episode lasting ≥2 years and/or more than 4 recurrent episodes, with the current episode lasting ≥1 year and at least 5 years since the onset of the first depressive episode; (4) failure to respond to a minimum of three adequate antidepressant treatments, including at least two medications from different pharmacological classes, as well as unsuccessful psychotherapy and electroconvulsive therapy (ECT) (due to poor response, intolerance, or refusal); (5) stable use of current antidepressant medications for at least one month prior to enrollment; and (6) a baseline 17-item Hamilton Depression Rating Scale (HAMD-17) score of ≥17 [25].

Exclusion criteria included: (1) schizophrenia or other psychotic disorders unrelated to MDD; (2) severe personality disorders, neurological conditions, or significant medical comorbidities; (3) history of prior brain surgery; and (4) contraindications for anesthesia or stereotactic surgery. All participants provided written informed consent before enrollment. DBS was offered to participants exclusively as part of this clinical trial and they were not scheduled to receive DBS outside the context of the study. The study was conducted at Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, and was registered at ClinicalTrials.gov (Identifier: NCT04530942). Ethical approval was obtained from the Ruijin Hospital Ethics Committee, Shanghai Jiao Tong University School of Medicine (Approval Number: 2021-52). The research use of the implanted components was authorized by the Shanghai Testing & Inspection Institute for Medical Devices and approved by the China Food and Drug Administration.

Study design

A detailed description of the clinical trial procedure and dataset will be reported separately [26, 27]. Here, we provide a brief summary of the study protocol (Fig. 1). All participants underwent DBS implantation targeting the BNST and NAc. The study comprised two phases: an initial open-label phase, followed by a randomized, double-blind, crossover phase. Following surgery, participants entered an open-label optimization phase lasting at least six months, during which symptom assessments were conducted monthly, and stimulation parameters were adjusted biweekly based on clinical feedback to optimize treatment efficacy. After the open-label phase, eligible participants entered the randomized, double-blind, crossover phase. Randomization was conducted by an independent statistician who was not involved in the clinical trial. Random numbers were generated using SAS, based on a pre-specified seed to ensure reproducibility. The seed number and all relevant parameters were documented. Participants were randomly assigned in a 1:1 ratio to either the On-Off or Off-On group. In the On-Off group, participants received two weeks of active DBS stimulation, followed by two weeks of sham stimulation. In the Off-On group, the sequence was reversed. Each stimulation period was preceded by a 2-day washout interval, during which DBS was turned off. Throughout the double-blind phase, participants maintained stable medication regimens and unchanged stimulation parameters. Neither participants nor evaluators were aware of the stimulation condition (DBS-On vs. DBS-Off). A single unblinded investigator, who was not involved in any clinical assessments or participant interactions during the double-blind phase, was responsible for managing the DBS programming. This included switching the device on or off at the beginning of each period according to the randomized schedule.

Fig. 1
figure 1

Study Pipeline through the Clinical Trial.

Full size image

Clinical assessments were conducted at pre-randomization, after each washout interval, and following both active and sham stimulation periods. Depression severity was evaluated using the HAMD-17, while anxiety levels were assessed using the Hamilton Anxiety Scale (HAMA) [28]. Although additional clinical measures were collected, they were not included in this study as the primary focus was on the relationship between emotional symptoms and EEG microstate dynamics in MDD patients. For participants who were able to attend in-person assessments, simultaneous EEG-LFP recordings were collected at each clinical evaluation time point. Among those who completed the double-blind crossover phase, a total of 10 patients with EEG-LFP recordings and clinical data from the active and sham stimulation periods were included in the present study to analyze DBS-induced microstate changes (Table 1).

Table 1 Demographic and clinical characteristics.
Full size table

EEG data acquisition and preprocessing

EEG recordings were conducted in a sound- and electromagnetically shielded room, where participants were seated in a relaxed state with their eyes fixated on a cross displayed on a screen. Prior to data collection at each time point, DBS stimulation was turned off to avoid stimulation artifacts and to enable LFP recording, and EEG recording commenced immediately. EEG signals were acquired using a 64-channel Ag/AgCl cap in combination with an RSC-64R amplifier (Quanlan Technology Co., Ltd, China). The sampling rate was set to 1000 Hz, with FCz as the online reference. Offline preprocessing was performed using MATLAB (R2023b, The MathWorks Inc.) and the EEGLAB toolbox (v2023.1) [29]. A 1–100 Hz bandpass filter was applied using the pop_eegfiltnew function, and a 50 Hz notch filter was implemented to remove power line interference. EEG signals were re-referenced to the common average reference, and the original online reference electrode (FCz) was interpolated. Bad segments contaminated by artifacts were manually inspected and removed, and channels with excessive noise were deleted and subsequently interpolated using spherical spline interpolation. Independent component analysis (ICA) was performed to remove artifacts arising from ocular movements, cardiac activity, and muscular activity. On average, 7.66 ± 2.92% (mean ± SD) of data segments were identified as artifactual and excluded from further analysis. In addition, 10.35 ± 3.30 ICA components were identified as artifacts and removed, and 3.25 ± 0.77 electrodes were interpolated. The final EEG dataset for each participant and each experimental condition contained approximately 5 min of artifact-free data. Prior to microstate analysis, EEG data were bandpass-filtered between 2–20 Hz, and the sampling rate was downsampled to 250 Hz.

Microstate analysis

Microstate analysis was performed using CARTOOL (brainmapping.unige.ch/cartool, Denis Brunet). For each EEG dataset, global field power (GFP) was computed at each time point. Since local maxima of GFP correspond to time points with a high signal-to-noise ratio and maximal whole-brain neural synchronization, only GFP peaks were included for subject-level clustering [30]. Next, polarity-insensitive modified k-means clustering was applied to all EEG data [31]. The number of clusters (k) was set from 1 to 12, with 100 random start repetitions for each k to enhance reproducibility. The optimal number of clusters for each EEG dataset was determined using a meta-criterion based on multiple standards [11, 32]. Group-level clustering was then conducted separately for the DBS-On and DBS-Off conditions, and the meta-criterion was applied again to determine the optimal microstate classes and obtain the template maps of each microstate. These template maps were subsequently backfitted to the corresponding EEG datasets for each condition. At each time point in the original EEG data, the microstate with the highest topographical similarity (regardless of polarity) was assigned, and segments shorter than 8 ms were rejected for temporal smoothing. To further ensure data consistency, CARTOOL’s built-in automatic rejection of bad segments was applied. For each EEG dataset, three microstate metrics were extracted: duration, coverage, and occurrence. Duration refers to the average duration of each microstate, coverage represents the proportion of time occupied by each microstate, and occurrence denotes the mean number of occurrences per second. Additionally, the backfitted microstate sequences were treated as a Markov chain, and the actual transition probabilities between microstates were computed [33].

Source localization

Source analysis was performed using the Brainstorm Toolbox in MATLAB [34]. Each EEG dataset was segmented according to microstate classes and then imported into Brainstorm. A standard set of 10–20 electrode locations was co-registered to the ICBM152 template head, and a boundary-element head model forward solution was computed using OPENMEEG [35]. Constrained, standardized low-resolution electromagnetic tomography (sLORETA) was applied to localize scalp signals to cortical source space [36]. For both DBS-On and DBS-Off conditions, the following steps were conducted: for each participant, absolute values of source activation maps for different microstates were computed and averaged over the time dimension; next, individual-level microstate source activation maps were normalized using the mean current across voxels and then averaged at the group level. The activation threshold for the group-level microstate source activation maps was set above the 95th percentile, consistent with previous studies and fMRI conjunction analysis [17, 37].

LFP data acquisition and preprocessing

Synchronously with scalp EEG, LFP signals were acquired from the implanted DBS electrodes targeting the NAc. Model SR1202-S electrodes (SceneRay), with a diameter of 1.27 mm, were employed in this study. These electrodes consist of eight cylindrical contacts, each measuring 1.5 mm in length and separated by a distance of 0.5 mm. The LFP was recorded at a sampling rate of 1000 Hz through the internal recording function of the DBS neurostimulator (IPG; model: SR1103, SceneRay) and wirelessly transmitted to a terminal amplifier. Synchronization of the LFP and EEG signals was achieved using a third-party marker independent of the two recording systems. This marker periodically sent a high-level signal (transistor-transistor logic, TTL) to both the LFP and EEG systems. In the post-processing stage, synchronized EEG-LFP data were obtained by aligning the signals from the marker. Offline LFP preprocessing was performed using custom MATLAB scripts. Similar to EEG, signals were bandpass filtered between 0.5 Hz and 100 Hz and a 50 Hz notch filter was applied to remove power line interference. The preprocessed LFP data from the NAc contacts were used for subsequent analyses time-locked to EEG events.

Time-locked EEG-LFP analysis

To investigate the relationship between local NAc activity and large-scale network dynamics reflected by EEG microstates, we performed a time-locked analysis focusing on transitions identified as clinically relevant (C-to-D and D-to-C). The precise onset time points for each C-to-D and D-to-C transition were extracted from the back-fitted microstate sequences derived from the scalp EEG data (see Microstate Analysis section). Preprocessed NAc LFP signals were then segmented into epochs centered around these transition onset times.

Spectral and aperiodic analysis of LFP epochs

For each LFP epoch, time-frequency representations (spectrograms) were computed to visualize spectral power changes around the transition points. Spectrograms covering the 1–100 Hz range were averaged across all identified C-to-D events (N = 54) and D-to-C events (N = 47) across the 10 patients to assess consistent peri-event spectral modulation. To quantify changes specifically in the gamma band, average gamma power (30–100 Hz) was calculated for the 50 ms interval immediately preceding the transition onset (−50 ms to 0 ms) and the 50 ms interval immediately succeeding the transition onset (0 ms to +50 ms) for each event. Paired t-tests were used to compare pre-transition versus post-transition gamma power for both C-to-D and D-to-C events.

Furthermore, to disentangle oscillatory changes from broadband shifts in the power spectrum, the ‘Fitting Oscillations & One Over F’ (FOOOF) algorithm was applied to the gamma band (30–100 Hz) power spectra derived from the same pre-transition and post-transition LFP segments [38]. This algorithm parameterizes the neural power spectrum, separating the aperiodic (1/f-like) component from periodic oscillatory peaks. The parameters representing the aperiodic component (specifically, exponent and/or offset, as relevant to the gamma increase) were extracted for each segment. Paired t-tests were used to compare the pre- versus post-transition aperiodic component parameters for C-to-D and D-to-C events, assessing whether observed gamma power changes were attributable to the aperiodic background activity.

Statistics

Statistical analyses were performed using MATLAB and IBM SPSS Statistics 27.0. Paired two-tailed t-tests were used to assess differences in microstate parameters including temporal metrics (duration, coverage, and occurrence) and transition probabilities between DBS-On and DBS-Off conditions. Cohen’s d was calculated to quantify the effect size. Normal distribution was assumed for all microstate parameters. For parameters showing significant group differences, Spearman correlation analyses were conducted to examine their associations with clinical scores. Linear regression was used to model the relationship between microstate parameters and clinical scores. Spatial correlation of microstate topographies and source-localized activation maps was measured using Pearson correlation. For all box plots, red line indicates the mean, red bar shows the standard error of the mean, and blue bar shows the standard deviation. Multiple comparisons for paired t-tests were corrected using the Benjamini–Hochberg false discovery rate (FDR), applied separately to temporal metrics and transition probabilities. Significance thresholds for microstate-clinical score correlations remained uncorrected due to their exploratory nature. A significance level of α < 0.05 was adopted for all statistical tests.

Results

Consistent microstate patterns in both DBS-On and DBS-Off phases

Based on the meta-criterion, microstate analysis identified five distinct microstate classes during both the DBS-On and DBS-Off phases (Fig. 2). Among these, four canonical microstate classes (A, B, C, and D) were consistently observed, aligning with previous studies. Additionally, a fifth microstate class (E) was identified, which has also been reported in prior research, including studies on MDD and healthy controls, due to its visually similar topography [15, 32]. The spatial correlation coefficients between the DBS-On and DBS-Off template maps ranged from 0.8994 (microstate E) to a near-perfect 0.9939 (microstate B), indicating a high degree of topographical consistency and stability of the underlying neural generators between the two conditions. Together, these five microstate classes accounted for a substantial portion of the brain’s electrical activity, explaining 67.0% and 69.6% of the global explained variance (GEV) in the DBS-On and DBS-Off phases, respectively. Specifically, the GEVs for microstates A, B, C, D, and E were 10.1%, 10.4%, 29.0%, 10.5%, and 6.9% in the DBS-On phase, and 6.2%, 7.6%, 31.9%, 17.3%, and 6.7% in the DBS-Off phase, suggesting microstate C was the most dominant state in both conditions, particularly in the DBS-Off phase.

Fig. 2: Consistency of Microstate Topographies Across Deep Brain Stimulation States.
figure 2

Analysis of EEG data obtained independently from both the DBS-On (top) and DBS-Off (bottom) phases revealed similar microstate topographies. The results consistently identified the four canonical microstates (A-D) along with an additional microstate E. These five microstates exhibited high topographical consistency between the two conditions (spatial correlation coefficients >0.89), and together accounted for the majority of the global explained variance (GEV) (67.0% in the DBS-On phase and 69.6% in the DBS-Off phase).

Full size image

BNST-NAc deep brain stimulation modulated microstate occurrence and coverage

Significant changes in microstate coverage (the percentage of time spent in a microstate) and occurrence (how often a microstate appears per second) were observed between the DBS-On and DBS-Off phases (Fig. 3). During the DBS-On phase, microstate A exhibited significantly higher mean coverage (mean ± SD, DBS-On: 17.89 ± 4.16%, DBS-Off: 12.21 ± 5.09%, p = 0.01, pFDR = 0.044, d = 1.02) and occurrence (DBS-On: 3.20 ± 0.53, DBS-Off: 2.23 ± 0.82, p = 0.006, pFDR = 0.035, d = 1.12). Microstate B similarly showed increased average coverage (DBS-On: 20.08 ± 5.09%, DBS-Off: 14.98 ± 4.27%, p = 0.013, pFDR = 0.046, d = 0.98) and occurrence (DBS-On: 3.50 ± 0.62, DBS-Off: 2.70 ± 0.62, p = 0.003, pFDR = 0.028, d = 1.27) in the DBS-On condition. In contrast, microstate C, D and E showed no significant differences in mean coverage (C: p = 0.139, pFDR = 0.243; D: p = 0.03, pFDR = 0.075; E: p = 0.987, pFDR = 0.987) or occurrence (C: p = 0.036, pFDR = 0.084; D: p = 0.024, pFDR = 0.065; E: p = 0.956, pFDR = 0.984) after FDR correction. No significant changes in mean duration were observed for any microstate (A: p = 0.424, pFDR = 0.55; B: p = 0.476, pFDR = 0.595; C: p = 0.311, pFDR = 0.429; D: p = 0.061, pFDR = 0.133; E: p = 0.829, pFDR = 0.92).

Fig. 3: BNST-NAc Deep Brain Stimulation Modulated Microstate Occurrence and Coverage.
figure 3

Comparison of the mean duration (A), coverage (B), and occurrence (C) of each microstate between the DBS-On and DBS-Off phases. Compared with the DBS-Off condition, TRD patients in the DBS-On state showed significantly increased average coverage (B) and occurrence (C) of microstates A and B. In all panels, the red line indicates the mean, the red bar shows the standard error of the mean, the blue bar shows the standard deviation. Uncorrected p-values and the corrected p*-values are shown. Red values indicate statistically significant results (pFDR < 0.05). Statistical details (paired t-test, n = 10, DBS-On versus DBS-Off, mean ± SD): Microstate A: duration, DBS-On: 95.83 ± 6.65, DBS-Off: 93.21 ± 8.63, p = 0.424, pFDR = 0.550; coverage, DBS-On: 17.89 ± 4.16%, DBS-Off: 12.21 ± 5.09%, p = 0.010, pFDR = 0.044, d = 1.02; occurrence, DBS-On: 3.20 ± 0.53, DBS-Off: 2.23 ± 0.82, p = 0.006, pFDR = 0.035, d = 1.12. Microstate B: duration, DBS-On: 97.17 ± 7.97, DBS-Off: 94.84 ± 7.94, p = 0.476, pFDR = 0.595; coverage, DBS-On: 20.08 ± 5.09%, DBS-Off: 14.98 ± 4.27%, p = 0.013, pFDR = 0.046, d = 0.98; occurrence, DBS-On: 3.50 ± 0.62, DBS-Off: 2.70 ± 0.62, p = 0.003, pFDR = 0.028, d = 1.27. Microstate C: duration, DBS-On: 117.33 ± 16.86, DBS-Off: 123.41 ± 13.39, p = 0.311, pFDR = 0.429; coverage, DBS-On: 29.36 ± 7.85%, DBS-Off: 34.12 ± 7.66%, p = 0.139, pFDR = 0.243; occurrence, DBS-On: 4.12 ± 0.40, DBS-Off: 4.50 ± 0.39, p = 0.036, pFDR = 0.084. Microstate D: duration, DBS-On: 97.13 ± 6.65, DBS-Off: 104.31 ± 6.27, p = 0.061, pFDR = 0.133; coverage, DBS-On: 18.29 ± 4.49%, DBS-Off: 24.26 ± 4.42%, p = 0.030, pFDR = 0.075; occurrence, DBS-On: 3.23 ± 0.56, DBS-Off: 3.93 ± 0.46, p = 0.024, pFDR = 0.065. Microstate E: duration, DBS-On: 90.11 ± 9.65, DBS-Off: 90.68 ± 9.32, p = 0.829, pFDR = 0.920; coverage, DBS-On: 14.39 ± 6.79, DBS-Off: 14.43 ± 6.32, p = 0.987, pFDR = 0.987; occurrence, DBS-On: 2.69 ± 1.02, DBS-Off: 2.71 ± 0.97, p = 0.956, pFDR = 0.984.

Full size image

Alteration of microstate transition dynamics following BNST-NAc deep brain stimulation

BNST-NAc DBS significantly influenced microstate transition probabilities (Fig. 4). In the DBS-On phase, transition probabilities from microstate A to C (DBS-On: 0.06 ± 0.02%, DBS-Off: 0.03 ± 0.01%, p = 0.002, pFDR = 0.028, d = 1.39) and from C to A (DBS-On: 0.06 ± 0.01%, DBS-Off: 0.03 ± 0.01%, p = 0.004, pFDR = 0.028, d = 1.21) were significantly increased. Similarly, transition probabilities from microstate B to A (DBS-On: 0.06 ± 0.02%, DBS-Off: 0.03 ± 0.02%, p = 0.010, pFDR = 0.044, d = 1.03) and from microstate E to B (DBS-On: 0.04 ± 0.01%, DBS-Off: 0.03 ± 0.01%, p = 0.013, pFDR = 0.046, d = 0.97) were significantly elevated under active stimulation. In contrast, transition probabilities from microstate C to D (DBS-On: 0.06 ± 0.01%, DBS-Off: 0.12 ± 0.04%, p = 0.004, pFDR = 0.028, d = 1.20) and from D to C (DBS-On: 0.06 ± 0.01, DBS-Off: 0.11 ± 0.04, p = 0.004, pFDR = 0.028, d = 1.21) were reduced in the DBS-On phase. Overall, the DBS-On phase was characterized by increased unidirectional transition probability from microstate B to A and from microstate E to B, enhanced bidirectional transition probabilities between microstate A and C, and decreased bidirectional transition probabilities between microstate C and D. No significant differences were observed in other transition pairs. These alterations suggest BNST-NAc DBS reshapes the temporal sequencing and interaction between these large-scale network states.

Fig. 4: Deep Brain Stimulation Reshapes Microstate Transition Dynamics.
figure 4

Chord diagram illustrating microstate transition probabilities during the DBS-On and DBS-Off phases. Arrows indicate the direction of transition. Red signifies a significantly higher transition probability in the DBS-On phase compared to DBS-Off (pFDR < 0.05), while blue indicates the opposite (significantly lower probability in DBS-On, pFDR < 0.05); gray denotes non-significant differences between phases. Key significant changes include increased transition probabilities for A ⇋ C, B → A and E → B, and a decreased probability for C ⇋ D during the DBS-On phase. Statistical details (Paired t-test, n = 10, DBS-On versus DBS-Off, mean ± SD): A to B: DBS-On: 0.05 ± 0.02%, DBS-Off: 0.03 ± 0.01%, p = 0.020, pFDR = 0.058; A to C: DBS-On: 0.06 ± 0.02%, DBS-Off: 0.03 ± 0.01%, p = 0.002, pFDR = 0.028, d = 1.39; A to D: DBS-On:0.04 ± 0.01%, DBS-Off: 0.04 ± 0.02%, p = 0.899, pFDR = 0.954; A to E: DBS-On: 0.03 ± 0.01%, DBS-Off: 0.03 ± 0.01%, p = 0.319, pFDR = 0.429. B to A: DBS-On: 0.06 ± 0.02%, DBS-Off: 0.03 ± 0.02%, p = 0.010, pFDR = 0.044, d = 1.03; B to C: DBS-On: 0.07 ± 0.02%, DBS-Off: 0.07 ± 0.02%, p = 0.710, pFDR = 0.828; B to D: DBS-On: 0.05 ± 0.02%, DBS-Off: 0.04 ± 0.01%, p = 0.103, pFDR = 0.212; B to E: DBS-On: 0.04 ± 0.01%, DBS-Off: 0.04 ± 0.01%, p = 0.017, pFDR = 0.054. C to A: DBS-On: 0.06 ± 0.01%, DBS-Off: 0.03 ± 0.01%, p = 0.004, pFDR = 0.028, d = 1.21; C to B: DBS-On: 0.07 ± 0.02%, DBS-Off: 0.07 ± 0.02%, p = 0.632, pFDR = 0.763; C to D: DBS-On: 0.06 ± 0.01%, DBS-Off: 0.12 ± 0.04%, p = 0.004, pFDR = 0.028, d = 1.20; C to E: DBS-On: 0.05 ± 0.03%, DBS-Off: 0.06 ± 0.03%, p = 0.246, pFDR = 0.374. D to A: DBS-On: 0.04 ± 0.01%, DBS-Off: 0.04 ± 0.02%, p = 0.841, pFDR = 0.920; D to B: DBS-On: 0.05 ± 0.02%, DBS-Off: 0.04 ± 0.01%, p = 0.137, pFDR = 0.243; D to C: DBS-On: 0.06 ± 0.01%, DBS-Off: 0.11 ± 0.04%, p = 0.004, pFDR = 0.028, d = 1.21; D to E: DBS-On: 0.04 ± 0.02%, DBS-Off: 0.05 ± 0.02%, p = 0.211, pFDR = 0.336. E to A: DBS-On: 0.03 ± 0.01%, DBS-Off: 0.03 ± 0.01%, p = 0.204, pFDR = 0.336; E to B: DBS-On: 0.04 ± 0.01%, DBS-Off: 0.03 ± 0.01%, p = 0.013, pFDR = 0.046, d = 0.97; E to C: DBS-On: 0.05 ± 0.03%, DBS-Off: 0.06 ± 0.03%, p = 0.294, pFDR = 0.429; E to D: DBS-On: 0.04 ± 0.01%, DBS-Off: 0.05 ± 0.02%, p = 0.137, pFDR = 0.243.

Full size image

Correlation between microstate dynamics and depressive symptom

To explore the relationship between microstate parameters and clinical symptoms, we first examined correlations between significantly altered microstate metrics and HAMD or HAMA scores (Fig. 5A). Microstate B occurrence was negatively correlated with HAMD scores (r = −0.457, p = 0.043), whereas no significant correlation was observed between microstate A occurrence and HAMD scores (p = 0.061). For anxiety symptoms, the occurrence of microstate A (r = −0.497, p = 0.026) and B (r = −0.492, p = 0.027) were negatively correlated with HAMA scores. However, microstate A and B coverage showed no significant correlations with HAMA (A: p = 0.07; B: p = 0.087) or HAMD (A: p = 0.116; B: p = 0.115) scores. Linear regression results for these correlations are presented in S2.

Fig. 5: Microstate Changes Correlate with Clinical Symptom Severity.
figure 5

A Heatmap of Spearman correlation coefficient matrix between coverage and occurrence of microstate A and B, and HAMA (top) or HAMD (bottom) scores. (Spearman correlation, n = 10, for HAMA: coverage of microstate A, r = −0.413. p = 0.07; occurrence of microstate A, r = −0.497, p = 0.026; coverage of microstate B, r = −0.392, p = 0.087; occurrence of microstate B, r = −0.492, p = 0.027; for HAMD: coverage of microstate A, r = −0.363. p = 0.116; occurrence of microstate A, r = −0.426, p = 0.061; coverage of microstate B, r = −0.364, p = 0.115; occurrence of microstate B, r = −0.457, p = 0.043). B Heatmap of Spearman correlation matrix between transition probabilities of microstate A ⇋ C, C ⇋ D, B → A and E → B and HAMA (top) or HAMD (bottom) scores. (Spearman correlation, n = 10, for HAMA: A to C, r = −0.642, p = 0.002; B to A, r = −0.375, p = 0.104; C to A, r = −0.632, p = 0.003; C to D, r = 0.744, p < 0.001; D to C, r = 0.751, p < 0.001; E to B, r = −0.58, p = 0.007; for HAMD: A to C, r = −0.496, p = 0.026; B to A, r = −0.273, p = 0.244; C to A, r = −0.496, p = 0.026; C to D, r = 0.731, p < 0.001; D to C, r = 0.77, p < 0.001; E to B, r = −0.688, p < 0.001). Spearman’s rho of each correlation is displayed. * Uncorrected p < 0.05.

Full size image

We also evaluated the relationship between microstate transition probabilities and clinical scores (Fig. 5B). Transition probabilities involving A ⇋ C, and E → B were negatively correlated with both HAMA (A to C: r = −0.642, p = 0.002; C to A: r = −0.632, p = 0.003; E to B: r = −0.58, p = 0.007) and HAMD (A to C: r = 0.496, p = 0.026; C to A: r = −0.496, p = 0.026; E to B: r = −0.688, p < 0.001) scores. Conversely, transition probabilities involving C ⇋ D were positively correlated with both HAMA (C to D: r = 0.744, p < 0.001; D to C: r = 0.751, p < 0.001) and HAMD (C to D: r = 0.731, p < 0.001; D to C: r = 0.77, p < 0.001) scores. No significant correlations were observed between transition probability from B to A and HAMA (p = 0.104) or HAMD (p = 0.244) scores.

sLORETA-based source localization of microstate networks in DBS conditions

To investigate the neuroanatomical substrates of microstate activity and gain insights into the large-scale networks they represent, we applied standardized low-resolution brain electromagnetic tomography (sLORETA) to localize the five microstate classes. The source estimations for microstates A, B, C, D, and E were highly consistent between the DBS-On and DBS-Off phases, with spatial correlation coefficients of 0.9338, 0.8961, 0.9563, 0.8192, and 0.8754, respectively, further supporting the stability of these network representations across stimulation conditions. The cortical sources for the DBS-On phase are illustrated in Fig. 6 (see S3 for DBS-Off phase results). Key regions, including the posterior cingulate cortex and temporal lobes, were consistently activated across all microstates. Microstate A primarily involved the left temporal lobe, left insula, and occipital lobe; microstate B involved the right temporal lobe, right insula, and occipital lobe; microstate C involved the precuneus and anterior cingulate cortex (ACC); microstate D involved the parietal lobe and left insula; and microstate E involved the frontal, temporal, and occipital lobes. These findings align with previous studies, despite differences in source localization and microstate analysis methods [11].

Fig. 6: sLORETA-Based Source Localization of Microstate Networks in DBS Conditions.
figure 6

Group-averaged cortical source localization maps for microstates A–E during the DBS-On phase, estimated using constrained standardized low-resolution electromagnetic tomography (sLORETA) co-registered to the ICBM152 template head. The displayed maps represent activations exceeding the 95th percentile threshold across voxels. Primary cortical sources were identified for (A) Microstate A: left temporal lobe, left insula, occipital lobe; (B) Microstate B: right temporal lobe, right insula, occipital lobe; (C) Microstate C: precuneus, anterior cingulate cortex; (D) Microstate D: parietal lobe, left insula; and (E) Microstate E: frontal, temporal, occipital lobes. Common activation across all microstates was observed in the posterior cingulate cortex and temporal lobes. Notably, source estimations demonstrated high consistency between the DBS-On and DBS-Off phases (spatial correlations ranging from 0.8192 to 0.9563); results for the DBS-Off phase are available in supplementary materials.

Full size image

NAc gamma aperiodic activities driven transitions of microstates during deep brain stimulation

Given that transition probabilities between microstates C and D were identified as a significant correlate of the antidepressant effects of BNST-NAc DBS, we sought to explore the relationship between intrinsic activity within the targeted nuclei and these specific large-scale brain network state transitions, leveraging synchronously recorded LFP signals from the DBS electrodes and scalp EEG (Fig. 7A). Performing a time-locked analysis on the raw EEG-LFP data, aligning LFP signals to the onset of C-to-D transitions, revealed prominent gamma-band activity emerging in the Nucleus Accumbens (NAc) preceding the transition (Fig. 7B). This observation was further corroborated by spectrogram analysis (Fig. 7C). Consequently, averaging the LFP signals across all 54 C-to-D transition events identified in the 10 patients revealed a distinct peri-event increase in gamma power (Fig. 7D). This elevation in gamma power preceding the transition was found to be statistically significant (Fig. 7E). Analysis of the spectral components using the FOOOF algorithm indicated that this gamma power increase was primarily attributable to changes in the aperiodic component of the LFP signal, rather than changes in oscillatory power (Fig. 7F). Considering that D-to-C transitions also demonstrated significant correlations with clinical symptoms in TRD patients, we conducted a parallel time-locked analysis for the 47 identified D-to-C events across the 10 patients (Fig. 7G). In contrast to the C-to-D transitions, the spectrogram analysis revealed no discernible event-related spectral changes preceding the D-to-C transition. Statistical analyses confirmed no significant modulation of gamma power or the aperiodic component around these events (Fig. 7H, I). These findings suggest that fluctuations in NAc gamma-band aperiodic activity, potentially modulated in the context of ongoing DBS, may play a crucial role in driving the specific C-to-D whole-brain functional state transitions captured by microstate dynamics, while such a mechanism does not appear to underlie the D-to-C transitions.

Fig. 7: NAc gamma aperiodic activities driven transitions of microstates.
figure 7

A Schematic illustration depicting the whole-brain transition from microstate C to microstate D. B Raw trace of the 0.5–100 Hz filtered LFP signal recorded from the NAc, spanning 500 ms centered on the transition point from microstate C to microstate D. C Spectrogram of the NAc LFP signal corresponding to the data in (B), centered on the microstate C to D transition time point. An enhancement in gamma band activity is evident, initiating approximately 60 ms prior to the transition onset and persisting throughout the transition period. D Averaged peri-event spectrogram across 54 microstate C to D transition trials (recorded from 10 patients). Each trial’s spectrogram covers the 1–100 Hz frequency range. Pronounced high-frequency activity is evident surrounding the transition time point. E Comparison of gamma band power between the 50 ms interval preceding and the 50 ms interval succeeding the microstate C to D transition (n = 54 trials from 10 patients). Statistical analysis performed using a paired t-test; results are indicated within the figure panel. F Comparison of the gamma band aperiodic component between the 50 ms interval preceding and the 50 ms interval succeeding the microstate C to D transition (n = 54 trials from 10 patients). Statistical analysis performed using a paired t-test; results are indicated within the figure panel. G Averaged peri-event spectrogram across 47 microstate D to C transition trials (recorded from 10 patients). Each trial’s spectrogram covers the 1–100 Hz frequency range. H Comparison of gamma band power between the 50 ms interval preceding and the 50 ms interval succeeding the microstate D to C transition (n = 47 trials from 10 patients). Statistical analysis performed using a paired t-test; results are indicated within the figure panel. I Comparison of the gamma band aperiodic component between the 50 ms interval preceding and the 50 ms interval succeeding the microstate D to C transition (n = 47 trials from 10 patients). Statistical analysis performed using a paired t-test; results are indicated within the figure panel.

Full size image

Discussion

This study is the first to investigate the impact of BNST-NAc DBS on large-scale brain network dynamics in TRD patients using a randomized, double-blind, crossover design with EEG microstate analysis and time-locked NAc LFP recordings. 5 microstates, including 4 canonical microstates (A-D) and an additional microstate E, were consistently identified under active and sham conditions, indicating that the underlying cortical generators remain largely preserved under BNST-NAc DBS implantation and stimulation [15, 32]. Furthermore, several microstate parameters differentiated DBS-On from DBS-Off state, including coverage and occurrence of microstate A and B, as well as transition probabilities of C ⇌ D, A ⇌ C, E → B and B → A. Further correlation analysis confirmed the association between microstate parameters and clinical symptoms. Importantly, time-locked analyses of simultaneous EEG-LFP recordings demonstrated that elevations in gamma-band aperiodic activity within the NAc selectively preceded transitions from microstate C to D, but not the reverse direction. These findings suggest that combined microstate parameters and LFP signatures may serve as potential biomarkers for DBS treatment response.

Microstates are believed to reflect the dynamic temporal organization of large-scale brain networks. Our source localization results revealed that different microstates share partially overlapping but distinct cortical generators. Prior EEG-fMRI studies have shown that microstate A is associated with BOLD activation in the superior and middle temporal lobes, microstate B with the occipital cortex, and microstate C with the dorsal ACC, inferior frontal cortex and insula [11, 12]. These findings suggest that microstate A, B and C are related to auditory network, visual network and anterior DMN, respectively. To bridge neural activity with subjective experiences and behavior, a three-dimensional model comprising external, internal, and associative neural units has been proposed [39]. The external unit include sensory cortices (visual, auditory, and somatosensory) and motor cortex, with their reduced activity potentially contributing to psychomotor inhibition [40, 41]. DMN is classified as part of the associative unit, and its hyperactivation has been associated with repetitive thinking unrelated to the current environment [42, 43]. MDD is characterized by diminished interest in external stimuli, slowed motor response and excessive rumination. These clinical features are thought to be associated with disruption of these large-scale networks [40, 42, 44,45,46]. Thus, our findings, showing increased occurrence and coverage of microstate A and B, along with decreased coverage and occurrence of microstate C (although did not survive FDR correction), suggest that BNST-NAc DBS may alleviate depressive and anxious symptoms by reducing excessive internal focus linked to DMN and enhancing external focus through increased activation of visual and auditory networks.

Additionally, microstate D is associated with negative BOLD activation with dorsal and ventral prefrontal and parietal cortices, which are related to the cognitive control network (CCN). Previous studies have labeled the topography similar to our microstate E as microstate D [47, 48]. Because of identifying a topography resembling the classical microstate D, our results suggest that microstate E may represent a component of this canonical microstate class. Recent studies have also explored inter-network connectivity alterations in depression, revealing reduced connectivity between the anterior DMN and auditory network, as well as between the right frontoparietal network, also known as CCN, and lateral visual network in MDD patients compared to healthy controls [49]. In our study, DBS stimulation increased transition probabilities between microstate A (auditory network) and microstate C (anterior DMN), as well as from microstate E (a subcomponent of CCN) to B (visual network). These findings suggest that BNST-NAc DBS may improve the coordination between external perception and internal cognitive processing, thereby enhancing patients’ perception of the external world and their cognitive appraisal of internal thoughts.

Transition probabilities between microstate C and D were reduced under DBS-On condition and were significantly correlated with clinical symptoms. Importantly, time-locked analysis revealed a robust increase in NAc gamma-band (30–100 Hz) power immediately preceding C → D transitions, an effect primarily driven by changes in the aperiodic component of the LFP signal. Conversely, such modulation of gamma power or its aperiodic component was not found preceding D → C transitions. The observed transient rise in NAc gamma-band aperiodic activity may represent an excitatory drive signal that triggers the shift from microstate C (associated with anterior DMN) toward microstate D (associated with CCN). By comparison, the transition from microstate D to microstate C likely reflects a more passive process and thus does not require an excitatory trigger. Notably, previous study has shown that thought rumination in MDD was correlated with increased functional connectivity between DMN and CCN [50]. Therefore, BNST-NAc DBS may stabilize the local striatal circuitry, thereby attenuating the generation of these gamma aperiodic bursts. By suppressing this drive signal, DBS may effectively decouple the DMN and CCN, enabling the brain to escape from an inefficient ruminative cycle.

Based on the overall findings of our study, we propose a local-to-global hypothesis. This hypothesis posits that DBS may not directly induce a beneficial global brain state, but instead modulates local events to influence maladaptive large-scale network dynamics. Our simultaneous recording paradigm, with high temporal and spatial resolution, delineates an evidentiary chain from a local circuit event to large-scale network reconfiguration and, ultimately, to a clinical outcome. Although limited by the capabilities of our recording equipment, we were unable to directly capture how DBS modulates these local events in real time. Nevertheless, this cross-scale mechanism provides a novel, data-driven framework for understanding how DBS reshapes brain function to achieve therapeutic benefit.

Current research on microstate alterations in MDD remains limited and inconsistent. Zhao et al. reported that subclinical MDD patients exhibit increased mean duration of microstates B, C, and D, increased occurrence and coverage of microstate B, but decreased occurrence of microstates A, C, and D, along with coverage of microstates A and C [51]. Murphy et al. found that MDD patients had higher microstate A occurrence but lower mean duration, occurrence, and coverage of microstate D [15]. He et al. reported increased microstate B occurrence and coverage and decreased microstate D occurrence and coverage in MDD patients [52]. A meta-analysis of mood and anxiety disorders found that compared to healthy controls, mood disorder patients (including subclinical MDD, MDD, remitted MDD, and bipolar disorder) had significantly increased microstate B occurrence and decreased microstate D occurrence [53].

Several studies have investigated treatment-related microstate changes in MDD. Lei et al. found that after eight weeks of SSRIs treatment, MDD patients exhibited reduced microstate D duration and coverage and decreased microstate A occurrence, with microstate parameters resembling those of healthy controls [18]. This suggests that SSRIs may mediate improvement by restoring altered microstates to normal patterns. However, another study found that microstate A occurrence increased after two weeks of SSRIs treatment [19]. Other treatment modalities, such as TMS, have also been studied in relation to microstates. MDD patients undergoing TMS exhibited increased microstate C occurrence and coverage and decreased microstate D occurrence and coverage [17]. Similarly, seizure therapy, including electroconvulsive therapy and magnetic seizure therapy, significantly increased the mean duration of microstate A while reducing the frequency of microstates B, C, and D [16]. The reduction in microstate D occurrence observed in TMS, seizure therapy, and our study (though not statistically significant after FDR correction) suggests that microstate D occurrence may serve as a biomarker of neuromodulation therapy response in MDD.

In conclusion, this study demonstrates that BNST-NAc DBS modulates large-scale brain network dynamics in TRD, and these changes are correlated with clinical improvement. Our core contribution is the identification of a local NAc event that may drive microstate transitions. Although the current study is based on a limited sample size (n = 10), the microstate alterations and the driving role of NAc gamma activity were robustly observed, suggesting a strong effect size. These findings not only provide novel mechanistic evidence for DBS treatment but also suggest that combined EEG-LFP microstate signatures could serve as potential biomarkers for guiding and personalizing future DBS treatment.

Limitations

First, because of the highly selective nature of TRD and the invasive nature of DBS, the sample size was small (n = 10), which limits generalizability of the findings. Additionally, due to stimulation artifacts, EEG-LFP recordings for the DBS-On condition were acquired when the stimulation turned off. Therefore, these recordings more likely reflect BNST-NAc DBS aftereffects. Furthermore, microstates are influenced by brain developmental and behavioral states, as well as disease severity, chronicity, and medication status [10]. The broad age range and complex treatment histories of participants may have introduced additional variability in microstate features. Moreover, all participants were TRD patients, whose treatment resistance may represent a distinct MDD subtype with unique pathophysiological features. While the randomized double-blind crossover design helps minimize these confounding effects, direct comparisons with previous studies remain challenging. Inconsistencies in EEG-LFP acquisition and analysis also affect the reproducibility of findings. Individual variability and sample heterogeneity further contribute to inconsistencies in microstate results. Future research should consider these factors to enhance the reliability and generalizability of microstate analyses in MDD. Furthermore, although this study demonstrated a statistically significant effect size, the sample size remains one of the primary limitations. Future research should explore this issue with a larger population.