Fronto-cingulate gamma synchrony predicts deep brain stimulation response in treatment-resistant depression: preliminary evidence

Abstract

Background

Gamma-band activity has been linked to cognitive-affective processing in depression. However, its role in deep brain stimulation (DBS) response in treatment-resistant depression (TRD) remains unclear. This exploratory study investigated whether ventral capsule/ventral striatum DBS (VCVS-DBS) modulates gamma activity during cognitive-affective processing and whether DBS-related gamma changes are associated with improvement in depressive symptoms.

Methods

Seven participants with TRD completed an affective Multi-Source Interference Task during DBS ON and OFF. Source-localized electroencephalography estimated local gamma power, synchrony, and effective connectivity in fronto-cingulate regions. DBS- and task-related effects were identified with cluster-based permutation testing, and follow-up analyses used Bayesian modeling.

Results

During cognitive control under affective context, VCVS-DBS was associated with increased mid-gamma (41–55 Hz) power in left dorsal anterior cingulate cortex (L-dACC) and right ventrolateral prefrontal cortex (R-VLPFC) after affective stimuli. In the time window where power increases overlapped (0.76–0.86 s), DBS was also associated with increased mid-gamma synchrony between these regions. The DBS-related increase in synchrony—not local power—was associated with clinical improvement (Bayesian Spearman ρ = 0.52). Granger causality further indicated that effective connectivity shifted toward greater L-dACC influence over R-VLPFC during DBS. Exploratory cross-validation suggested that DBS-related synchrony change and baseline effective connectivity together accounted for individual differences in clinical improvement.

Limitations

The primary limitation was the necessarily small sample size, which was partially addressed using Bayesian analyses leveraging extensive within-participant data.

Conclusions

These preliminary findings suggest that VCVS-DBS may engage a fronto-cingulate gamma network involved in cognitive-affective control in TRD. They further suggest that clinical response may depend on both baseline dACC–VLPFC network organization and DBS-related modulation of gamma synchrony, highlighting potential relevance to cognitive-affective regulation across psychiatric disorders. Given the small sample size, these findings should be considered hypothesis-generating and require confirmation in larger cohorts.

Introduction

Patients with treatment-resistant depression (TRD), a severe form of major depressive disorder (MDD) characterized by inadequate response to antidepressant interventions, commonly experience cognitive impairments in domains such as attention, cognitive control, and executive functions, alongside persistent low mood.¹ Cognitive deficits are highly prevalent even after the remission of depressive episodes, and are associated with an increased risk of relapse and a more chronic illness course.² Understanding the mechanisms underlying these cognitive-affective symptoms is therefore essential for improving long-term recovery.

Emerging evidence implicates altered gamma-band activity in the pathophysiology of depressive disorders.^3,4 Gamma oscillations have been linked to cognitive-affective processes, including selective attention toward emotionally salient stimuli and cognitive reappraisal of affective information.^5,6 In MDD, magnetoencephalography studies have shown reduced frontal gamma responses to happy faces, which scale with depression severity.^7,8 In healthy adults, affective picture viewing was associated with increased frontal gamma, which varied with stimulus valence.^9,10 Together, these findings support the broader relevance of gamma-band activity to affective stimulus processing in depression.

Deep brain stimulation (DBS) has established itself as a promising approach to treat symptoms of TRD.¹¹ Electrophysiological data are often collected concurrently with DBS, which provides a compelling window to investigate disease mechanisms. However, DBS research in TRD has historically focused on lower-frequency dynamics. For instance, local field potential recordings from DBS leads have linked limbic alpha-band power to depressive severity,¹² while unilateral subgenual cingulate stimulation has been associated with reductions in ipsilateral theta activity.¹³ Electroencephalography (EEG) during ventral capsule/ventral striatum DBS (VCVS-DBS), a target implicated in mood and reward-related dysfunction, has shown increased prefrontal theta oscillations during cognitive control, which were correlated with clinical response.¹⁴ However, higher-frequency signals may also be modulated by VCVS-DBS: gamma activity within the VCVS is associated with reward expectation and motivational state dynamics.¹⁵ Therefore, examining gamma-band activity provides a complementary approach to the previously characterized lower-frequency findings.

This study aimed to examine the effects of VCVS-DBS on cognitive-affective processing in TRD. Considering the role of gamma oscillations in cognitive-affective processing along with their emerging implication in MDD,^3,6 we hypothesized that (i) VCVS-DBS would increase gamma-band activity in regions central to cognitive control under affective context, and (ii) this change would correspond to clinical outcomes. We then explored whether VCVS-DBS modulates gamma-band synchrony and directional influence between these regions.

Methods

Participants

This study analyzed experimental data from a VCVS-DBS trial (NEMAR—ds001784). Although the full dataset included 14 participants with treatment-resistant psychiatric conditions, EEG data were available for only eight participants with a primary diagnosis of obsessive-compulsive disorder (OCD) or MDD.¹⁴ For the present analysis, we restricted inclusion to participants with a primary diagnosis of MDD and baseline Montgomery–Åsberg Depression Rating Scale (MADRS) scores ≥ 33, resulting in seven eligible participants. Each participant first completed a session with DBS ON, followed by a 60 min washout period, and then a session with DBS OFF. Task-based EEG was recorded during both DBS ON and OFF conditions. MADRS scores were obtained at baseline (pre-implant) and again on the data-collection day, or within 1 month of that visit if the participant was unable to complete the assessment on the same day. The MADRS is a 10-item clinician-rated measure of depressive symptom severity, with each item scored from 0 to 6 and total scores ranging from 0 to 60; higher scores indicate greater symptom severity across affective, cognitive, and somatic symptoms.¹⁶

Experimental protocol

In the standard Multi-Source Interference Task (MSIT), recognized for eliciting robust, participant-specific effects on attention and cognitive control,¹⁴ participants identify the number differing from the other two numbers using three fingers of their dominant hand. Control trials align the target number with its corresponding response key, while interference trials present spatially misaligned targets surrounded by valid digits. In the affective MSIT (afMSIT), each trial began with a 400 ms image (positive, neutral, or negative) from the International Affective Picture System (IAPS) followed by a 1500 ms MSIT digit array and a fixation cross (3–5 s). Although the IAPS image was not required for the digit response, it introduced a response-irrelevant affective context in which participants performed conflict resolution and response selection. A total of 144 IAPS images—spanning the valence and arousal spectra—were counterbalanced across 72 control and 72 interference trials per block, ensuring each image appeared once per context (Fig. 1). Trial sequences were pseudo-randomized to prevent consecutive repetition of key attributes and reduce habituation. EEG was recorded at 1450 Hz with 60 scalp electrodes (10–20 system) using a Nexstim eXimia system. Electrode impedances were maintained <5 kΩ, and digitized scalp co-registration was used.

Fig. 1.

Affective Multi-Source Interference Task (afMSIT) Schematic. Each International Affective Picture System (IAPS) image (positive, neutral, or negative) was presented for 400 ms. Participants pressed the button matching the identity of the one digit that differed from the other two, while the IAPS image remained partially visible. Control trials used zeros and spatially congruent targets; interference trials introduced task-relevant distractor digits and incongruent target positions. Participants completed 20 feedback and 40 no-feedback practice trials, reaching ≥90% accuracy before starting. They were also instructed to minimize blinking during the task and to blink freely during the subsequent fixation period. The task was programmed using the Psychophysics Toolbox in MATLAB 2013a (http://psychtoolbox.org). Created in BioRender. (2025)—https://biorender.com/8p2of7tGW28ALSQZJ.

Reaction time was the primary behavioral outcome in the MSIT because participants were trained to perform the task with very low error rates. To determine whether the previously reported behavioral DBS effect was preserved in our participant subset, we fit the original mixed-effects model to reaction times.¹⁴

Data analysis

EEG pre-processing

Pre-processing was performed with concatenated DBS ON and OFF recordings to reduce bias. Continuous EEG was band-pass filtered between 1 and 60 Hz and trials with large residual artifacts (±150 µV) were rejected. Additional participant-specific notch filters were applied for DBS-related artifacts based on visual inspection of power spectra (Supplementary Fig. S1). Independent component analysis (ICA) was performed using the Extended Infomax algorithm¹⁷ after initial correction of large, stereotyped artifacts (e.g., eye blinks, muscle) with signal-space projection. ICA components were automatically labeled using MNE-ICALabel.^18,19 These labels were independently compared with a trained rater blinded to these automated labels to inform final component exclusion. Pre-processed data were segmented into epochs (from −1.5 to 3.4 s relative to IAPS-image onset) and decimated from 1450 to 483 Hz. Artifact-contaminated epochs were excluded (leaving 1917 of 2016 trial epochs for analysis), and time-domain baseline correction was applied to the retained epochs. The baseline window (−0.1 to 0 s) was selected from candidate windows by minimizing group-level root-mean-square power in frontal sensors during a pre-stimulus validation interval (−0.5 to − 0.1 s; Supplementary Fig. S2).

Source localization

Participant-specific structural MRI scans were processed using FreeSurfer (v7.4.1) to generate cortical meshes for EEG sensor co-registration and source localization. Digitized EEG electrodes (scalp locations) were used to align sensors with each participant's anatomical space. Regions-of-interest (ROI) were defined following the mappings from Widge et al.¹⁴ (Supplementary Table S1). Predefined ROIs included bilateral dorsolateral prefrontal cortex (DLPFC; middle frontal gyrus), ventrolateral prefrontal cortex (VLPFC; inferior frontal gyrus), and dorsal anterior cingulate cortex (dACC)—regions consistently implicated in conflict monitoring and affective regulation^20,21 (Fig. 2).

Fig. 2.

(A) Lateral and (B) medial views of the fsaverage left-hemisphere inflated surface, displaying the seven anatomically informed regions-of-interest (ROIs) used in this study. Labels were created by merging sub-parcels from the Lausanne 250-region atlas. Colors denote: dorsal anterior cingulate cortex (dACC, brown), dorsolateral prefrontal cortex sub-regions (DLPFC 1–4, navy, bright green, brick red, lime green), and ventrolateral prefrontal cortex sub-regions (VLPFC 1–2, magenta, yellow).

Source reconstruction employed a Linearly Constrained Minimum Variance (LCMV) beamformer²² to localize focal, high-frequency activity while suppressing interference from other sources. Epochs were cropped to stimulus- (−0.5 to 2.0 s relative to IAPS onset) or response-locked (−1.0 to 1.0 s relative to the button press) windows. To isolate induced activity, condition-specific evoked waveforms were subtracted from single trials before beamforming. Noise and data covariance matrices were estimated over pre- (−100 to 0 ms) and post-stimulus (0–1000 ms) windows, respectively. Adaptive diagonal-loading regularization was applied to ensure covariance-matrix stability (condition number < 300). Co-registration accuracy was confirmed by maintaining mean sensor-to-scalp distance < 5 mm (Supplementary Note 1 and Table S2).

Beamformer filters were computed with maximum-power orientation constraints and neural activity index normalization to enhance sensitivity to low-amplitude signals while reducing volume conduction bias. Per-trial source time courses were extracted via these spatial filters, using an eigenvalue-based (PCA flip) method. Finally, individual source estimates were morphed onto the fsaverage surface for across-participant comparison. All EEG processing and source localization were performed in MNE-Python (v1.9.0).

Time–frequency analyses

Power estimates were computed from source time courses in two gamma ranges: low-gamma (30–40 Hz) and mid-gamma (41–55 Hz). Analyses were restricted to frequencies below 60 Hz to reduce susceptibility to line noise and DBS-related high-frequency artifacts. These bands were selected to cover the gamma range previously reported to be sensitive to affective IAPS stimuli,⁹ while also overlapping with frequencies implicated in DBS-related gamma modulation.²³ Ten-cycle Morlet wavelets were used, ensuring frequency resolution and reliable band-averaged power estimates in source space. Power was normalized per subject and DBS condition using median baseline values (−0.1–0 s) and converted to dB for variance stabilization before ordinary least squares (OLS) cluster testing.

For power envelope correlation (PEC), five-cycle wavelets were utilized to improve temporal precision in a narrow interval of interest; all other preprocessing steps were held constant. For each trial (n = 1917), Spearman's correlation was computed across time between the average mid-gamma envelopes of the selected ROIs, then aggregated within each participant and DBS condition to compare DBS ON versus OFF. For the participant-level clinical association analysis, these changes were derived using the original 10-cycle envelope data to maintain consistency with the primary time–frequency pipeline and reduce trial-level variability before relating these changes to MADRS improvement. Frequentist p-values were corrected with the Benjamini–Hochberg procedure to control false discovery rate (FDR) across candidate predictors (n = 3), with significance defined as α = 0.05. Last, for local power analyses, baseline-ratio linear power was used to preserve sensitivity to transient increases (Supplementary Note 2).

Granger causality

To assess DBS-related shifts in directionality, each trial's signals were extracted and sorted by DBS condition for each participant. Following this step, DBS ON and OFF trials were each concatenated, yielding two continuous time-series per participant. A vector autoregressive (VAR) model (order = 8) was fit to each concatenated time-series and spectral Granger causality (GC) was computed for the band of interest. We calculated the logarithmic (natural-log) GC ratio for each participant under DBS ON and OFF, contrasting dACC → VLPFC vs. VLPFC → dACCW:

$$log\hspace{0.17em}GC=log(G{C}_{dACC\to VLPFC}+\epsilon )-log(G{C}_{VLPFC\to dACC}+\epsilon )$$

where log denotes the natural (base-e) logarithm and ε = 1 × 10^–5 prevents log (0). A Bayesian repeated-measures model with participant-specific random intercepts was fitted to these data to estimate DBS effect (Δlog GC ratio; ON–OFF). Robustness was assessed with jackknife resampling and posterior-predictive checks (Supplementary Note 3).

Statistical analysis

Primary—cluster permutation tests

Source-level data were analyzed using OLS regression. For each label and frequency band, a separate model was fitted at each time point, including the following predictors: “DBS” (DBS ON/OFF), “Interference” (Interference/Control task), “DBS × Interference” (interaction), “Arousal” (covariate; higher values = more arousing), “Valence” (higher values = more pleasant), “Trial” (trial number, as a nuisance regressor), “DBS × Valence” (interaction), and an intercept. Interference was modeled as cognitive-control demand, whereas valence and arousal were modeled as affective IAPS dimensions (normalized; 0–1). DBS interaction terms were included to test whether stimulation differentially modulated interference versus valence-related effects. For each predictor in the model, a time series of t-scores was calculated.

Permutation-based cluster correction was implemented to control for multiple comparisons across time points. For each label, frequency band, and predictor combination, 1000 permutations were performed through trial shuffling.¹⁴ The regression analysis was repeated to generate a null distribution of maximum cluster sizes (sum of t-values within a cluster) for controlling family-wise error rate. Observed clusters were identified using the t-score maps thresholded at z > 1.96 (p < 0.05, two-tailed), with contiguous suprathreshold time points grouped into clusters. Clusters shorter than 0.05 s were excluded, and FDR correction was applied across all ROIs and frequency bands.

Threshold-free cluster enhancement (TFCE) was also implemented to improve sensitivity to weaker signals without a fixed threshold.²⁴ TFCE clusters shorter than 0.03 s were excluded, reflecting our minimum resolvable interval at 55 Hz (Supplementary Table S3).

Secondary—Bayesian statistical models

Bayesian analyses were implemented in PyMC (v5.21.1) and summarized with ArviZ (v0.21.0), using weakly informative priors centered on 0. Inferences are reported as 95% highest-density interval (HDI) or the posterior probability that an effect is positive or negative.

For PEC, single-trial Spearman ρ values were averaged by participant and DBS condition for subject-level repeated-measures modeling and analyzed in trial-level hierarchical models with valence, interference, and their interaction as fixed effects. To evaluate time-on-task effects, PEC was regressed on trial number, DBS condition, and their interaction (DBS × Trial), yielding within-block coupling slopes for DBS ON and OFF. For power analyses, coefficients are expressed in baseline-ratio units (Δ-power relative to baseline). For clinical prediction, a Bayesian linear model regressed ΔMADRS on primary predictors—baseline GC ratio, ΔPEC, and their interaction—with covariates (stimulation frequency, baseline MADRS) evaluated for inclusion in subsequent models. Predictive accuracy was quantified with leave–one–out cross–validation (LOOCV): the model was re-fit after withholding each participant in turn and prospectively predicting the held-out participant. Posterior predictions were summarized with posterior-mean R² and mean absolute error (MAE). Robustness was evaluated with a 5000-shuffle label-permutation test, recomputing LOOCV R² for each shuffled outcome (Supplementary Note 4).

Ethics approval

This study involved a secondary analysis of an open-source dataset (NEMAR—ds001784). Written informed consent was obtained from each participant before any study procedures, and all procedures were approved by the Massachusetts General Hospital Institutional Review Board.

Findings

Seven participants (5 males and 2 females), aged 30–60 years with baseline MADRS scores ≥ 33, were included in the analyses. The average MADRS improvement was 21 ± 5.3 points from baseline to post-treatment; no outliers were observed in EEG or clinical data at baseline (Table 1). DBS ON reduced reaction time by 22.4 ms (Wald z = −4.98, p < 0.001), consistent with prior work.

Table 1.

Demographic and clinical characteristics.

Participant	Age	Sex	MADRS-BL	MADRS-FU	Stim. Freq. (Hz)
P01	50	M	33	10	130
P02	60	M	34	9	120
P03	60	M	33	10	100
P04	50	M	38	24	50
P05	60	M	36	10	130
P11	30	F	34	11	130
P13	50	F	42	29	90
Mean  ±  SD	51.43 ± 10.7	…	35.71 ± 3.3	14.71 ± 8.2	…

Note: All participants were right-handed and met criteria for moderate-to-severe depression (MADRS ≥ 33). The sample included 28.57% female participants. MADRS: Montgomery–Åsberg Depression Rating Scale; BL: baseline; FU: follow-up; Stim. Freq.: stimulation frequency.

Gamma power modulation

Four significant power clusters were identified, differentially modulated by stimulus valence and DBS (Table 2). Elevated post-stimulus gamma power was observed in L-dACC (0.75–0.90 s) and R-VLPFC (0.76–0.86 s), both in the mid-gamma range and modulated by the DBS predictor. Additionally, two pre-response clusters were identified for the valence contrast: L-dACC in the mid-gamma range, and R-DLPFC in the low-gamma range. These clusters were consistent across both traditional cluster-based permutation and threshold-free approaches.

Table 2.

Gamma-band activity clusters significantly modulated by stimulus valence and deep brain stimulation (DBS) predictors during an affective Multi-Source Interference Task (afMSIT).

Analysis	Band	ROI	Contrast	T window	p FDR
Response-locked	Low-gamma	R-DLPFC (1)	Valence (↑)	−0.98 to −0.70	0.008
	Mid-gamma	L-dACC	Valence (↑)	−0.82 to −0.66	0.008
Stimulus-locked	Mid-gamma	L-dACC	DBS (↑)	0.75–0.90	<0.001
	Mid-gamma	R-VLPFC (1)	DBS (↑)	0.76–0.86	0.033

Note: L-dACC: left dorsal anterior cingulate cortex; R-VLPFC: right ventrolateral prefrontal cortex; R-DLPFC: right dorsolateral prefrontal cortex. VLPFC (1) refers to anterior inferior frontal gyrus (aIFG) parcels (see Supplementary Table S1).

Envelope correlation

Motivated by the prominent temporal overlap between stimulus-locked power clusters under DBS (0.76–0.86 s), we analyzed functional connectivity in the mid-gamma band between L-dACC and R-VLPFC, alongside local power (Supplementary Note 5). In this overlapping window, a Wilcoxon's signed-rank test revealed that DBS ON significantly increased power envelope correlation (ΔPEC) between L-dACC and R-VLPFC (W = 0.00, p = 0.02), with a large effect size (paired Hedge's g = 1.54) (Fig. 3). A Bayesian repeated-measures model corroborated this observation, showing a 99.6% posterior probability (mean: 0.03, 95% HDI: [0.01, 0.05]) that DBS enhanced mid-gamma PEC between these regions. Sensitivity analyses confirmed that the effect remained robust (Wilcoxon p < 0.05) when ΔPEC was recalculated with 3–6 cycle envelopes but became non-significant at ≥7 cycles, consistent with a DBS-related increase in burst-like, phasic synchrony. To evaluate whether this coupling reflected stable neuromodulation or time-on-task effects, we compared within-block trial slopes. A statistically credible DBS × trial interaction (β = −0.05, [−0.10, −0.01]) was observed, where the within-block coupling slope for DBS ON was significantly less positive than for DBS OFF (Supplementary Table S4).

Fig. 3.

Baseline-normalized mid-gamma power and deep brain stimulation (DBS) effect on envelope correlation. Grand-average power envelope (41–55 Hz, 5 cycles) from International Affective Picture System (IAPS) stimulus onset to 1.5 s, for (A) DBS OFF and (B) ON conditions. Values are plotted as baseline-normalized power ratios (baseline = 1; mean ± standard error). (C) Per-participant change in mid-gamma envelope correlation measured in the 0.76–0.86 s overlap window, showing increases in 7/7 participants.

Trial-level analyses provided moderate evidence that PEC increased with stimulus valence (β = 0.02, 95% HDI: [−0.01, 0.06], P(β > 0) = 91.8%); coupling tended to rise for more pleasant images and fall for unpleasant ones. In this model, the positive effect of DBS alone on envelope correlation remained likely (P(β > 0) = 92.3%).

Granger causality

Spectral GC analysis within the same window revealed that DBS shifted effective connectivity towards greater dACC influence (Fig. 4). Fitting a repeated-measures model to the ΔGC ratio yielded a credible DBS effect of 0.45 log-units (95% HDI: [0.01, 0.90]), corresponding to an approximately 58% increase in dACC → VLPFC/VLPFC → dACC ratio during stimulation. The effect was robust to single-participant omissions (jackknife mean = 0.56, p = 0.01), with six of seven participants having shifted toward stronger dACC → VLPFC influence during DBS. Complementary phase-lag analyses likewise showed a DBS-related reversal from a VLPFC-leading to a dACC-leading gamma rhythm (Supplementary Note 6 and Fig. S3).

Fig. 4.

Deep brain stimulation (DBS)-related change (ΔGC) between left dorsal anterior cingulate cortex (L-dACC) and right ventrolateral prefrontal cortex (R-VLPFC) via Granger causality (GC). (A) Per-participant effects show individual log ΔGC ratio values (DBS ON–OFF); the violin-plot outline depicts the empirical distribution. (B and C) Scatterplot of baseline Granger-directionality ( x-axis) versus DBS Δ for each directionality component (y-axis). Near-unit negative slopes (Spearman ρ = −0.82 and −0.96) indicate that the direction and size of each subject's DBS shift chiefly reflect baseline bias. The negative slopes also indicate that DBS tends to move each component towards zero.

Testing for baseline dependency revealed that each individual component's Δ was strongly anti-correlated with their own baseline value (Spearman ρ ≈ −0.9; Figs. 4B and 4C). Accordingly, DBS tended to shift each ratio towards zero: the correlation between the baseline log-GC ratio and its DBS-related change showed a negative trend (ρ = −0.57, p = 0.18). This effect persisted in order-averaged VAR estimates (ρ = −0.82, p = 0.02), a lower-variance measure.

Four of seven patients began with a negative baseline ratio (greater VLPFC → dACC); their larger positive shifts outweighed the smaller negative shifts of dACC-dominant participants. As a result, a net group-level increase was observed for GC ratio—an apparent rise in dACC influence under DBS. Complete model-order statistics and posterior predictive checks are provided in Supplementary Table S5.

Clinical outcomes

Among task predictors, stimulus valence showed the strongest association with PEC (see above), supporting the affective relevance of connectivity during this time window. We next examined whether DBS-related change in PEC was associated with clinical improvement (ΔMADRS). A Bayesian rank-correlation model estimated a posterior mean ρ = 0.52 (95% HDI: [0.04, 0.91], P(ρ > 0) = 96.4%), indicating a positive relationship (Fig. 5). Frequentist analyses aligned with these findings (ρ = 0.82, p = 0.03), though FDR adjustment attenuated significance (p_FDR = 0.08). Neither dACC (ρ = −0.37, p_FDR = 0.41) nor VLPFC (ρ = −0.41, p_FDR = 0.55) power changes alone correlated with improvement. Sensitivity analyses are presented in Supplementary Note 7 and Fig. S4.

Fig. 5.

Correlation between clinical improvement and deep brain stimulation (DBS)-related change in dorsal anterior cingulate cortex (dACC)–ventrolateral prefrontal cortex (VLPFC) gamma power-envelope correlation (Δρ; 0.76–0.86 s). (A) Posterior distribution of Bayesian Spearman correlation coefficient (ρ) with 95% highest-density interval (HDI); posterior density (y-axis) and ρ (x-axis). (B) Scatterplot of DBS-related change in power-envelope correlation (ΔPEC) estimated with the original 10-cycle Morlet wavelet (x-axis) versus change in Montgomery–Åsberg Depression Rating Scale (MADRS) scores (y-axis). Red lines (posterior regression draws; 95% HDI) illustrate uncertainty. Longer wavelet cycles quantify sustained coupling; low responders show negative ΔPEC, whereas high responders remain positive.

The baseline log-GC ratio alone showed only a modest, non-significant relation to clinical improvement (ρ = 0.52, p = 0.23). Given its strong coupling to DBS-related directionality change, we tested whether this trait-like baseline bias moderates the effect of ΔPEC on clinical outcomes. Baseline log-GC ratio and ΔPEC were essentially orthogonal (ρ ≈ 0, p ≈ 1.00), showing minimal multicollinearity.

LOOCV in a ΔPEC-only model demonstrated weak explanatory power (LOOCV R² = 0.21, MAE = 4.05 MADRS points). Adding the baseline log-GC ratio and its interaction with ΔPEC raised LOOCV performance to R² = 0.96 and reduced MAE to 0.8.

When accounting for plausible covariates, introducing DBS stimulation frequency preserved high accuracy (LOOCV R² = 0.95 ± 0.04, jack-knife standard error (SE); MAE = 0.73 ± 0.92, SD) while significantly improving expected out-of-sample log-likelihood—supporting its inclusion (expected log point-wise predictive density (ΔELPD_LOO) = +6.9 ± 1.4; Fig. 6). In contrast, baseline MADRS did not improve model fit (Supplementary Table S6). All Pareto-k values were <0.7 (0.20–0.54), indicating no single participant unduly influenced predictive estimates. Last, permutation testing confirmed that explained variance was not due to chance (p = 0.003).

Fig. 6.

Clinical interaction model, 41–55 Hz gamma (0.76–0.86 s post-stimulus). (A) Red circles show predictions of Montgomery–Åsberg Depression Rating Scale (MADRS) improvement, defined as baseline MADRS minus follow-up MADRS, from the ΔPEC-only model (DBS-related Δ in power envelope correlation). Blue circles show prediction from the full moderator model, which includes baseline Granger causality (GC) ratio (log ratio of dACC → VLPFC to VLPFC → dACC directionality), its interaction with ΔPEC, and stimulation frequency. The full model nearly overlaps the line of identity (leave–one–out cross–validation (LOOCV) R² = 0.95, mean absolute error (MAE) = 0.7 points), whereas the ΔPEC-only model diverges widely (R² = 0.21, MAE = 4.1 MADRS points). (B) Prediction lines (blue = −1 SD low-PEC, yellow = mean, red = +1 SD high-PEC) illustrate posterior mean ± 95% Bayesian credible intervals. The x-axis shows standardized baseline GC ratio, and the y-axis represents standardized MADRS improvement (higher values indicate greater clinical improvement). Scatter symbols are colored on a gradient scale corresponding to DBS-related PEC change (blue = negative ΔPEC, yellow = values near zero, red = positive ΔPEC), with dot color intensity reflecting the magnitude of change. DBS: deep brain stimulation; dACC: dorsal anterior cingulate cortex; VLPFC: ventrolateral prefrontal cortex.

Discussion

This study provides preliminary evidence that VCVS-DBS increases mid-gamma synchrony (power envelope correlation) between L-dACC and R-VLPFC during cognitive-affective processing, with the magnitude of change correlating with clinical improvement in TRD. Stimulation was also associated with elevated local mid-gamma power in both regions. GC and phase-based metrics additionally converged on a temporally specific lead of dACC over VLPFC under DBS.

Gamma oscillations facilitate feed-forward signaling and rapid integration of distributed inputs within cortical circuits,^25,26 and their synchrony is associated with long-range communication from lower- to higher-order regions in early visual and somatosensory pathways.²⁷ In contrast, feedback influences, such as those related to top-down modulation, contextual updating, or expectation, are typically conveyed through alpha and beta band synchronization,²⁸ reflecting a frequency-based division of labor in cortical communication.

In preclinical work, low-frequency DBS of the rat ventral tegmental area (VTA) increased high-gamma phase coherence along the anterior–posterior VTA axis, which correlated with reduced immobility (despair-like behavior).²⁹ In our human cohort, VCVS-DBS was associated with increased dACC–VLPFC mid-gamma synchrony, and greater synchrony increases were associated with greater clinical improvement. These findings offer complementary evidence that mesolimbic stimulation may modulate inter-areal gamma synchronization. The frontal aslant tract—an association pathway linking the inferior frontal gyrus and the anterior cingulate³⁰—provides a plausible anatomical substrate for this coupling in our paradigm. Since greater synchrony was most strongly associated with higher valence (more pleasant images) and DBS ON, this pattern may reflect shared engagement of the fronto-cingulate network in TRD.

DBS was further associated with increased post-stimulus mid-gamma power in dACC and VLPFC. With the concomitant increase in synchrony, these changes support greater recruitment of task-evoked processes under DBS. Given the well-established role of the dACC in detecting emotional salience and initiating control, mid-gamma changes in this region may reflect integration of cognitive and affective conflict, while the VLPFC may act as a downstream implementer of this control, in line with its role in emotional reappraisal and inhibitory regulation of negative affect.^21,31,32

Beyond local power and synchrony, phase-related dynamics revealed a shift in inter-regional coordination during stimulation. DBS was associated with reversal of an early VLPFC-led (DBS OFF) state into dACC leadership, and then reinforced a pre-existing dACC-led alignment. This biphasic pattern supports the hypothesis that VCVS-DBS can induce “phase resetting”, where effective neuromodulation may first disrupt maladaptive patterns to allow for the emergence of more adaptive oscillatory structure.^33,34 Although the averaged cross-correlation lead was small (~1.6° of the gamma cycle), its consistency across trials suggests a reliable re-timing of coordination rather than a conduction delay. Overall, these results contribute to a growing body of literature hypothesizing phase resetting as a mechanism of action for acute electrical stimulation, particularly in disorders characterized by frontostriatal and limbic dysrhythmia.^35,36

Our findings also suggest a preliminary framework for understanding individual variability in treatment response. In an exploratory model, baseline dACC–VLPFC GC ratio interacted with DBS-related change in mid-gamma synchrony (ΔPEC) to predict clinical improvement. These predictors were statistically independent, indicating that they probe distinct network features. This supports a “trait × state” model of DBS, where baseline circuit characteristics may moderate how stimulation-driven dynamics translate into clinical outcomes.³⁷ In leave-one-out cross-validation, this model explained 95% of the variance in ΔMADRS, despite three participants having identical score reductions—an outcome that typically challenges predictive differentiation. Similar features emerged in the single participant with comorbid OCD, aligning with recent connectomic identification of a dACC–VLPFC “response tract” in DBS targeting a VC/VS “sweet spot” for OCD.³⁸ Therefore, gamma-synchronized communication between these regions may represent a transdiagnostic mechanism for cognitive-affective regulation.

Collectively, these findings suggest the possibility that gamma-mediated communication extends beyond sensory systems and may contribute to large-scale integration within prefrontal circuits. Looking ahead, integrating tractography-based targeting with real-time electrophysiological biomarkers may enable more individualized DBS protocols. Since inter–regional power coupling is direction-agnostic, can be estimated from standard scalp EEG, and is computationally inexpensive, it can be tracked online with minimal overhead. This presents a potential signal for future adaptive neuromodulation (e.g., closed-loop DBS), without reliance on intracranial sensors.

Limitations

These findings, while encouraging, should be viewed alongside several caveats. A primary limitation of our study is the small sample size. To mitigate this, we fitted Bayesian statistical models that draw on the large within-participant sample (n = 1917 trials). Permutation testing and leave-one-out methods were implemented to minimize the risk of over-fitting. While these safeguards strengthen internal robustness, they do not eliminate inherent constraints on external generalizability. Another limitation is our reliance on normative IAPS valence ratings; without participant-specific ratings, we could not fully account for idiosyncratic emotional reactions to certain IAPS themes that may reflect prior experiences.¹⁴ Accordingly, valence relationships were interpreted primarily at the group level.

Additionally, the spatial precision of source localization is limited by 60-channel scalp EEG. Intracranial recordings from the implanted VC/VS target or adjacent regions were not available. We therefore used participant-specific structural MRI processing, digitized scalp co-registration, optimized LCMV beamforming, and source-localization quality checks, including assessment of leakage between the dACC and VLPFC. Nevertheless, these findings should be interpreted as source-reconstructed cortical estimates rather than direct regional measurements.

Several potentially relevant electrophysiological dimensions were outside the scope of the present analysis, which focused on post-stimulus gamma activity. In particular, pre-stimulus dynamics were not formally modeled, although they may provide clinically relevant information about arousal, attentional readiness, preparatory control, and intrinsic rest–stimulus interactions. This is especially important in depression, where resting-state activity may shape subsequent stimulus-evoked responses rather than serving only as a neutral baseline.^39–42 Therefore, baseline-corrected post-stimulus estimates should not be interpreted as capturing the full DBS-related electrophysiological state. Future work should directly model pre-stimulus activity and rest–task interactions to clarify how DBS modulates both ongoing neural state and task-evoked cognitive-affective processing.

Finally, since DBS ON blocks were always presented first, order effects cannot be entirely ruled out. However, the overall mean coupling of the OFF block tended to remain lower than that of the ON block, suggesting that the observed difference in PEC is more likely driven by stimulation than practice effects.

Conclusion

To our knowledge, this study provides the first preliminary evidence that VCVS-DBS may modulate cortico-cortical gamma communication during cognitive-affective processing in TRD. These findings suggest that symptom improvements following VCVS-DBS may depend not only on local neural activation, but also on coordinated communication within fronto-cingulate networks involved in cognitive control and affective regulation. They further suggest that individual differences in baseline network architecture may influence treatment response to stimulation. If replicated, dACC–VLPFC gamma activity could help clarify mechanisms of DBS response and provide candidate electrophysiological markers of target-network engagement. Because inter-regional synchrony can be estimated from standard scalp EEG and is relatively easy to track, it may provide a practical signal for future adaptive DBS approaches. Similarly, pre-treatment gamma directionality within the dACC–VLPFC circuit may help inform stimulation optimization and more individualized neuromodulation strategies for patients with TRD. The trial-wise resolution of our sample and convergence of Bayesian and frequentist results indicate robust within-participant effects that merit further study; however, replication in larger cohorts is needed to establish generalizability and clinical utility.

Data availability

This study is a secondary analysis of an open-source dataset hosted on OpenNeuro (https://openneuro.org/datasets/ds001784).

Author contributions

Conceptualization: FGN, IT, SC

Data curation: IT

Formal analysis: IT

Investigation: FGN, IT, SC

Methodology: FGN, IT, SC

Supervision: VB

Visualization: FGN, IT, SC

Writing – original draft: FGN, IT, SC

Writing – review & editing: FGN, IT, SC, ID, AR, SK, MAK, WL, ML, RC, AML, VB

Funding information

The authors declare no specific funding for this work.

Supplementary material

Supplementary data are available with the article at https://doi.org/10.1139/jpn-2026-0016.