TMS-induced modulation of brain networks and its associations to rTMS treatment for depression: a concurrent fMRI-EEG-TMS study

Abstract

Introduction:

Transcranial magnetic stimulation (TMS) over the left dorsolateral prefrontal cortex (L-DLPFC) is an established intervention for treatment-resistant depression (TRD), yet the underlying therapeutic mechanisms remain not fully understood.

Methods:

This study employs an integrative approach that combines TMS with concurrent functional magnetic resonance imaging (fMRI) and electroencephalography (EEG), aimed at assessing the acute/immediate effects of TMS on brain network dynamics and their correlation with clinical outcomes. Furthermore, this study explored the brain-state dependent effects of TMS, as the brain-state indexed by the phase of EEG prefrontal alpha oscillation.

Results:

Our study demonstrates that TMS acutely modulates connectivity within vital brain circuits, particularly the cognitive control and default mode networks. We found that the baseline TMS-evoked responses in the cognitive control and limbic networks significantly predicted clinical improvement in patients receiving a novel EEG-synchronized repetitive TMS treatment. Clinical outcomes in this novel treatment were linked to state-specific TMS-modulated functional connectivity within a pivotal brain circuit of the L-DLPFC and the posterior subgenual anterior cingulate cortex within the limbic system.

Conclusions:

These findings contribute to our understanding of the therapeutic effects underlying TMS treatment in depression and support the potential of assessing state-dependent TMS effects. This study emphasizes the importance of personalized timing of TMS for optimizing target engagement of specific clinically relevant brain circuits. Our results are crucial for future research into the development of personalized neuromodulation therapies for TRD patients.

1. Introduction

Transcranial magnetic stimulation (TMS) over the left dorsolateral prefrontal cortex (L-DLPFC) is a Food and Drug Administration-approved treatment for depression. TMS therapy has demonstrated efficacy and safety in the treatment of patients with treatment-resistant depression (TRD) [1,2]. However, the mechanism of action underlying the therapeutic effects of TMS is still unclear. Substantial evidence has shown that TMS not only stimulates the superficial cortex site right underneath the coil, but also has transsynaptic effects on deep brain circuits associated with the stimulation site [3–7]. Assessing such TMS-induced effects on brain circuits allows the quantification of network perturbation and identifies measures relevant to clinical improvement. To investigate TMS-induced effects, the propagation pattern of induced activity from the L-DLPFC to various downstream regions has been established in many studies [8–10]. However, such effects appear to be highly heterogeneous both at the stimulation site [11] and at the associated networks distal to the stimulation site [12,13], with reports of such stimulation both increasing and decreasing neuronal activity, depending on the region, network, and stimulation parameters. To directly assess TMS-induced acute effects on the brain, recent studies combined TMS with concurrent functional magnetic resonance imaging (fMRI) acquisitions, which allows monitoring of the acute/immediate subsequent effects of the TMS on brain dynamics. For example, Vink et al. investigated the propagation pattern of TMS-induced activity from the L-DLPFC to the subgenual anterior cingulate cortex (sgACC) [13]. Oathes et al. also assessed TMS-evoked response in the sgACC and found its pre-treatment magnitude and the post-treatment changes are both associated with depression improvement [6]. Despite these encouraging results, research on this topic is relatively sparse, and more studies are warranted to examine and quantify the TMS-induced acute effects on brain networks and characterize the variability between patients, which might inform prognosis in depression treatment [14].

In addition to assessing TMS-evoked response at a particular brain region, it is also important to explore how the induced local activity can drive modulations throughout large-scale brain network systems [15]. Numerous studies have explored the brain circuits affected by TMS perturbation. These efforts include investigating the lasting after-effects on brain connectivity minutes after a TMS session in offline setups, as well as assessing the immediate, acute effects using concurrent TMS-fMRI in online setups [16]. For example, previous studies have shown that TMS applied to the L-DLPFC attenuates hyper-connectivity between sgACC and the default mode network (DMN) in depressed patients [17], and TMS might modulate the abnormal or symptom-related network connectivity in depression [18–20]. These studies provide evidence for the mechanisms underlying its therapeutic effects. However, these TMS-induced effects on brain connectivity were assessed in an offline fashion. It remains unclear whether these lasting after-effects of TMS-induced modulation reflect direct engagement of targeted brain circuits or reflect indirect compensatory effects, such as induced adaptive plasticity or short-term reorganization across brain networks [21,22]. Even though some studies have shown promising results on the consistency of these acute and offline-lasting effects of TMS [16], more studies are needed to confirm the relationship between online-acute and offline-lasting effects of TMS-induced modulation. Currently, assessing the TMS-induced online-acute modulation with concurrent TMS-fMRI still provides stronger evidence for the direct engagement of target brain circuits compared to the offline setups [22].

While great progress has been made to integrate TMS with structural and functional MRI on the optimization of TMS spatial targets for depression treatment [23–29], relatively little has been explored on the optimization of TMS timing for target engagement [30–35]. Based on the substantial evidence of state-dependent effects of TMS on the brain and behavior [22,36–40], it is reasonable to hypothesize that TMS timing relative to the state of the brain matters for the TMS treatment of depression. However, the definition of brain state varies across subfields and contexts, and brain state fluctuates at different timescales [40,41]. In this study, we derive a brain state index varying at a relatively short timeframe. Specifically, we proposed to use the phase of prefrontal alpha oscillation in the electroencephalography (EEG) signal as an index of brain state. Many studies have demonstrated that the alpha phase is associated with an active inhibitory mechanism, and the timing of sensory stimulation relative to the alpha phase influences perception [42–46]. These studies indicate a gating mechanism of prefrontal alpha oscillation, with distinct phases reflecting different neural excitability, thus gating the information flow across brain networks. Thus, we hypothesized that prefrontal alpha oscillation might gate the propagation of TMS-induced effects across brain networks, and by targeting the TMS pulses to the personalized phase in the alpha cycle, a stronger effect may be induced at the distal target. Here, we integrated EEG with TMS to track brain state, which is used for personalized TMS pulse timing optimization. In our previous studies, we have shown the benefit of EEG-synchronized repetitive TMS (rTMS) treatment, where we observed progressive entrainment effects over the sessions of rTMS treatment [47], which are related to better TMS antidepressant response [30]. In another study, our group demonstrated that TMS-induced effects in the circuit between DLPFC and sgACC depended on the EEG prefrontal alpha phase [34]. However, TMS-induced modulation of large-scale brain network systems across the whole brain was not assessed, and their longitudinal changes over the pre- and post-treatment scans have not been explored.

Here, we aim to investigate and quantify the TMS-induced acute modulation of brain networks and to test its associations with the clinical improvement in rTMS treatment for depression. As shown in Fig. 1, in this study, the treatment was designed as follows: 1) At baseline, an integrated fMRI-EEG-TMS (fET) instrument was developed and used as a pre-treatment scan to select the personalized TMS timing target (optimal phase), which was defined as the prefrontal alpha phase that produced the largest blood-oxygen-level-dependent (BOLD) signal increase in the dorsal anterior cingulate cortex (dACC); 2) During the rTMS treatment, patients were randomized into two groups with either rTMS pulses synchronized to the personalized optimal phase as the timing target (SYNC group) or delivered at a random phase (UNSYNC group). Patients were treated with 30 rTMS sessions over six weeks; 3) After the treatment, another fET scan was acquired. In this study, we aimed to explore how TMS-induced acute effects propagate through brain network systems. Specifically, we assessed TMS-evoked BOLD responses and connectivity modulations, hypothesizing that TMS over the L-DLPFC would modulate not only the local brain networks but also distal networks associated with depression, such as the cognitive control network (CCN) and DMN. Furthermore, we tested the state-dependent effects of these modulations using fET scans, with prefrontal alpha oscillation phase indexing brain state. Additionally, we investigated TMS-induced acute effects on brain networks before and after a six-week rTMS treatment. We hypothesized that these TMS-induced effects, both at baseline and post-treatment, would be associated with clinical response. By quantifying the TMS-induced acute effects and state-dependent effects, this study is important for future efforts to temporally optimize TMS targeting in the treatment of depression, with potential implications for personalized treatment strategies in depression [30,31,34,47].

Fig. 1.

Experimental procedure and data analyses. TMS-induced effects were assessed at both pre- and post-treatment fMRI-EEG-TMS (fET) scans. Both baseline TMS-induced effects and the longitudinal changes in TMS effects were related to the clinical outcome (HRSD percent change). HRSD, Hamilton Rating Scale for Depression.

2. Methods

2.1. Participants and procedure

This randomized and double-blinded clinical trial was conducted at the Medical University of South Carolina (MUSC) Institute of Psychiatry (ClinicalTrials.gov ID: NCT03421808 [48]). The experimental procedures of our study and the recruitment process were approved by the MUSC institutional review board. All patients provided written informed consent to participate in the study.

Thirty-four TRD patients were enrolled in the study, and six subjects were excluded due to voluntary withdrawal, claustrophobia, or inability to complete the pre-treatment MRI scan. Pre-treatment fET data from the remaining twenty-eight patients (mean age ± SD = 45 ± 13 years, female/male = 19/9; SD, standard deviation) were collected. These twenty-eight patients were randomized into the SYNC (n = 15) or UNSYNC (n = 13) group before the treatment. Ten patients in the SYNC group completed the study, with five patients excluded due to 1) life stress, 2) inability to adhere to the treatment schedule, 3) equipment unavailable, and 4) patients opting not to complete it. Ten patients in the UNSYNC group completed the study, with three patients excluded due to 1) hospitalization, 2) equipment unavailable, and 3) opting not to complete. All enrolled patients had a baseline Hamilton Rating Scale for Depression (HRSD) (Ham-D 28-item) score greater than or equal to twenty at the time of enrollment. Details of patient enrollment and inclusion/exclusion criteria were described elsewhere [30].

After enrollment, an integrated fET scan was performed to determine an individualized optimal phase ϕ_opt, which was defined as the phase that produced the largest BOLD signal increase in the dACC. During the closed-loop EEG-rTMS treatment, patients in the SYNC group received rTMS treatment with the target phase set as the individualized ϕ_opt, while for the patients in the UNSYNC group, the target phase was randomly selected from a uniform distribution between 0 to 2π. rTMS was delivered at each patients’ daily individualized alpha frequency (6–13 Hz), with the initial pulse delivered at the pre-defined target phase. Thus, SYNC patients received rTMS pulses synchronized to the individualized optimal phase, whereas UNSYNC patients received rTMS pulses that were unrelated to the patients’ optimal phase (initial pulse was fired randomly). Each patient received one treatment session each weekday for six weeks, with a total of thirty closed-loop EEG-rTMS treatment sessions. During the treatment sessions, the TMS coil was placed at the L-DLPFC (F3 electrode) with the dose of 120 % motor threshold, 40 pulses per train, and 75 pulse trains per session (3000 pulses in total per session). Details of the closed-loop EEG-rTMS treatment were described in Refs. [30,47]. After the treatment, post-treatment fET scans were acquired from the remaining patients. Each patient’s HRSD was assessed during the pre- and post-treatment fET scans and before the first treatment of each week. The experimental procedure and data analyses are illustrated in Fig. 1.

2.2. Data acquisition and preprocessing

An integrated fET instrument was developed and used in this study [49], where simultaneous EEG-fMRI data were acquired during single-pulse TMS delivered at the L-DLPFC (marked under the EEG F3 electrode, Beam F3 locator) at the beginning of the TR gap, with inter-trial intervals drawn from a uniform distribution (4–6 TRs). The EEG data were acquired with a custom MR-compatible bipolar EEG cap (36 electrodes, sampling rate = 488 Hz, Innovative Technologies, CA, USA). A Siemens 3T Prisma MRI Scanner was used to acquire fMRI data with a custom 12-channel head coil (Rapid MR International, LLC, Columbus, OH, USA). A modified MR-compatible TMS coil was used (MagStim Rapid2) and configured to 100 %–120 % intensity of each subject’s motor threshold.

Functional MRI data were preprocessed with AFNI [50]. Structural T1-weighted MR images were processed with the FreeSurfer pipeline [51]. The local-global Schaefer cortical parcellation atlas [52] was used to define cortical regions of interest (ROIs) and network systems. The Schaefer atlas classified these 400 ROIs into 34 brain networks (17 per hemisphere) with 106 network nodes. The L-DLPFC EEG F3 stimulation site ROI was defined in the MNI152 space with a 10 mm radius sphere centered at the coordinate MNI = [−37 26 49] [53]. We also assessed L-DLPFC ROIs near the EEG F3 stimulation site by selecting the ROIs from the Schaefer atlas at the nearest distance to the EEG F3 coordinates: 1) dorsal-PFC in the subnetwork-A of the DMN; 2) lateral-PFC in the subnetwork-B of the DMN; 3) lateral-PFC in the salience/ventral attention network (subnetwork-B). Details of data acquisition and data preprocessing are included in the supplementary text.

2.3. Whole-brain general linear model analysis

To investigate the propagation patterns of TMS-induced effects on whole-brain BOLD signal, an event-related general linear modeling (GLM) analysis was performed, where each TMS pulse was modeled as an instantaneous event, convolved with FMRIB’s linear optimal basis set (FLOBS) [54]. The use of FLOBS compared to the canonical hemodynamic response function (HRF) allows more freedom in the variability of hemodynamic response, as we expect the TMS-induced brain activity might be different from that evoked by conventional fMRI task stimuli, and FLOBS allows control of HRF variability across different brain regions. The combined FLOBS parameters were used for group-level analysis. Specifically, we computed the signed root mean square of the regression parameter estimates from the subject-level [55]. As the distribution of the combined statistic is non-Gaussian, in the group-level analysis, we used FSL’s Randomise tool [56] for a nonparametric permutation statistical testing (10000 permutations; p < 0.05 with multiple comparison correction across space using voxel-wise family-wise error (FWE)).

2.4. Quantification of TMS-evoked BOLD response

To assess subject-wise TMS-induced acute effects and quantify inter-subject variability, both the spatial extent and amplitude of TMS-evoked BOLD response were quantified across brain networks. The spatial extent of the induced response was defined as the percent coverage of each network by the propagation pattern of TMS-evoked BOLD response. Specifically, based on the subject-wise whole-brain GLM analysis results, this measure was computed as the number of significantly activated voxels divided by the total number of voxels within each network. To quantify the amplitude of TMS-evoked BOLD response for each brain network, ROI analysis was performed, where the mean BOLD signal of each ROI was modeled by the TMS pulse timing regressor (TMS pulse timing boxcar function convolved with the FLOBS basis set) with GLM. The signed root mean square of the regression parameter estimates was summed across ROIs within each network to compute the amplitude of the network’s induced response.

2.5. Functional connectivity and hub analysis

To examine the TMS-induced acute modulation of functional connectivity, a whole-brain psychophysiological interaction (PPI) analysis [57–59] was performed. BOLD signal from each parcellated cortical region was extracted with SPM [60]. Then, to compute the TMS-dependent functional connectivity, we used multiple regression, where the BOLD signal of ROI A was modeled as a linear combination of multiple independent variables including 1) BOLD signal of ROI B; 2) TMS-induced BOLD response (a boxcar function representing the timing of TMS pulses convolved with FLOBS); 3) PPI interaction between the BOLD signal at ROI B and the timing of TMS pulses; 4) nuisance variables (motion parameters, large motion volumes, BOLD signal in white matter, and BOLD signal in lateral ventricles). The interaction between the BOLD signal from one ROI and the TMS regressor was modeled by deconvolving the BOLD signal from the canonical HRF, multiplying by the TMS timing boxcar function, and re-convolving with the HRF. As the PPI analysis focuses on second-order stimulation modulation, controlling the first-order stimulation modulation (TMS-evoked mean activation) is necessary, as the stimulation-evoked mean activation could be a potential confounder and drives the false positive observed interaction between regions [61]. FLOBS-based stimulation regression was performed to account for evoked mean activation, which has been shown to achieve relatively low risk of false positive results in controlling the confounds [61]. The TMS-modulated functional connectivity matrix (beta estimates of the interaction term in the GLM model) between each pair of ROI A and B was extracted, symmetrized (averaging the upper and lower triangles), and summarized within/between parcellated brain networks. Positive and negative node strength were calculated as the sum of all positive or negative connection weights associated with each node. In addition to the TMS-modulated PPI connectivity, we also assessed correlation-based functional connectivity (details in supplementary text).

2.6. Brain-state dependency analysis

The aim of this analysis was to assess whether single-pulse TMS delivered at different timing relative to the prefrontal alpha oscillation modulates brain network systems differently. To examine the brain-state dependency effect of TMS delivery, the phase of prefrontal alpha oscillation was used to index the brain state. Prefrontal alpha oscillation was extracted from EEG signals at channels FP1, F3, and F7, and the alpha phase at TMS onset was estimated (details in Ref. [34]). Based on the alpha phase, TMS trials were grouped into four bins (bin 1: −π to −1/2 π, bin 2: −1/2 π to 0, bin 3: 0 to 1/2 π, and bin 4: 1/2 π to π), and we investigated the TMS effects on brain network systems for the trials in different phase bins.

To examine the brain-state dependency of the TMS-evoked response, a GLM was used to model the BOLD signal at the L-DLPFC (EEG F3 stimulation site), with the trials in each phase bin as a separate regressor (Fig. 2). Our previous investigations evaluated the graded relationship between prefrontal alpha phase and BOLD response across all four phase bins [34]. To simplify the current analysis, we focused on identifying two TMS trial conditions for each subject: 1) high-load-phase (HLP) condition was defined as the condition where a high TMS evoked response was introduced at the stimulation site (L-DLPFC); 2) low-load-phase (LLP) condition was defined with a low evoked response at L-DLPFC. Because the phase-bin groups differ in the number of trials and the temporal spacing, this might potentially lead to biased estimation of BOLD response and the trial conditions identification. To prevent such bias during HLP/LLP conditions identification, bootstrapping was applied where the same number of trials were randomly chosen for each phase bin to construct the regressors. Here, we set the selected trial number as eighty percent of the trial number in the phase bin with the smallest number of trials, and the bootstrapping process was repeated 500 times. For each subject, the phase bins that generated the highest and lowest evoked response at L-DLPFC were identified as the subject-wise HLP and LLP bins, respectively. The contrast between the conditions of HLP and LLP bins was assessed using a whole-brain GLM analysis, as well as brain-state dependency of TMS-modulated PPI connectivity (details in supplementary text).

Fig. 2.

Brain-state dependency analysis of TMS-induced acute/immediate effects using alpha phase from the prefrontal EEG signal. TMS trials were grouped into four categories with different prefrontal alpha phase bins (illustrated with different colors), where the prefrontal alpha oscillation was extracted from the concurrent EEG. Effects from the TMS trials in different phase bins were modeled with general linear modeling (GLM), where the BOLD signal at L-DLPFC was modeled with the TMS trials in each phase bin as a separate regressor. The beta-weight of each regressor estimated in the GLM represents the amplitude of the TMS-induced response for that particular phase bin. The phase bins that generated the highest and lowest BOLD response at L-DLPFC were identified as the subject-wise high-load-phase (HLP) and low-load-phase (LLP) conditions, respectively. The current analysis focuses on the TMS-induced effects in the HLP and LLP conditions.

2.7. Associations between TMS effects and clinical response

This analysis aimed to evaluate brain circuit engagement during single-pulse TMS, and identify potential neurophysiological markers associated with the clinical outcome of SYNC treatment. Pearson correlation coefficient was used to assess the relationship between neuroimaging measurements of TMS effects and the clinical response. Clinical response was quantified as the percent improvement of the HRSD score since the baseline, i.e. 100 × (pre-treatment score – post-treatment score)/pre-treatment score. Pre-treatment baseline neuroimaging measurements (TMS-induced response and TMS-modulated PPI functional connectivity) were evaluated as potential predictors of clinical response. To explore the potential mechanisms of action underlying EEG-synchronized rTMS treatment, longitudinal changes in neuroimaging measurements of TMS effects were examined, including the changes in TMS-induced response, modulated PPI functional connectivity by all TMS trials, and state-specific modulated PPI functional connectivity by TMS trials in the HLP bin (more details in supplementary text). Associations between these longitudinal changes and clinical response were also tested. For analyses of the TMS-induced response, correlations were computed for each of the 34 brain networks, with multiple comparisons corrected using Bonferroni correction. For analyses of TMS-modulated PPI functional connectivity, multiple comparisons were corrected using the false discovery rate (FDR) procedure, given the larger number of possible connections.

3. Results

Patients in the SYNC group (N = 15) had HRSD scores of 30.50 ± 4.35 (mean ± SD) at baseline and HRSD scores of 14.50 ± 7.68 (mean ± SD) post-treatment, with a percent improvement of 52.56 % ± 25.37 % (mean ± SD). Patients in the UNSYNC group (N = 13) had HRSD scores of 28.40 ± 7.29 (mean ± SD) at baseline and HRSD scores of 13.70 ± 8.22 (mean ± SD) post-treatment, with a percent improvement of 55.31 % ± 19.04 % (mean ± SD). There was no significant difference in the HRSD percent improvement between groups (p = 0.78).

3.1. TMS-induced acute/immediate effects on whole-brain BOLD signal at baseline fET scan

The group-level TMS-induced BOLD activation map is shown in Fig. 3A (permutation test FWE-corrected p < 0.05). TMS significantly elevated BOLD signals in various brain regions, including dACC, occipital areas, insula, and thalamus. Next, we sought to quantify these TMS-induced acute effects on brain networks. Specifically, we quantified both the amplitude (Fig. 3B) and spatial extent (Fig. S01) of TMS-evoked responses across brain networks. As shown in Fig. 3B, TMS evoked the strongest response in the right hemisphere (RH) salience/ventral attention subnetwork-A and the smallest response in the left hemisphere (LH) limbic subnetwork-B, with the highest inter-subject variability in the RH somatomotor subnetwork-A and the lowest inter-subject variability in the RH limbic subnetwork-B. We hypothesized that specific neural circuits engaged at the baseline acquisition could predict the following rTMS antidepression response, and any significant results might potentially provide evidence of their mechanistic contribution to the rTMS therapeutic effects [6].

Fig. 3.

Quantification of baseline TMS-evoked responses on whole-brain BOLD signal. (A) Group-level activation map (t-value; FWE-corrected p < 0.05); (B) Amplitude of TMS-induced BOLD response in brain networks (Schaefer atlas parcellation; see network abbreviation in Table S1). TMS evoked the strongest response in the right hemisphere salience/ventral attention subnetwork-A (RH-SalVentAttn-A), with the highest inter-subject variability in the right hemisphere somatomotor subnetwork-A (RH-SomMot-A).

3.2. TMS-induced acute/immediate modulation of brain connectivity at baseline fET scan

From the PPI analysis, significant TMS modulations between networks are shown in Fig. 4 (FDR-corrected p < 0.05), with the strongest negative effect on the connectivity between LH DMN subnetwork-A and RH DMN subnetwork-B. Of note is that these were all negative effects, and none of the connections showed significant positive effects after multiple comparison corrections.

Fig. 4.

Quantification of TMS-induced functional connectivity (whole-brain psychophysiological interaction analysis) at baseline fET scan. TMS induced significant negative effects of connectivity between the default mode, control, limbic, and salience/ventral attention networks (FDR-corrected p < 0.05). None of the connections showed significant positive effects.

3.3. State-dependency of TMS-induced acute/immediate effects at baseline fET scan

We employed whole-brain GLM analysis to identify brain regions with stronger evoked responses during TMS trials in the HLP bins versus LLP bins, which were defined based on the evoked response in the L-DLPFC. Significant activations were found in the lateral frontoparietal network (L-FPN), including the bilateral DLPFC and inferior parietal lobule. Using a correlation-based approach, L-DLPFC seed-based functional connectivity analysis further showed that this HLP-vs-LLP contrast map overlaps with the connectivity pattern of the L-DLPFC (Fig. S03), suggesting that TMS-induced effect propagation from L-DLPFC follows its intrinsic brain connectivity. Based on PPI analysis, we also examined whether TMS modulates the connectivity between brain networks differently when the pulses were delivered at different prefrontal alpha phases. Our results showed that only the TMS trials in the HLP bins induced significant (FDR-corrected p < 0.05) negative functional connectivity modulations between 1) RH DMN subnetwork-A and RH DMN subnetwork-B; 2) RH DMN subnetwork-A and the RH control subnetwork-A; 3) RH DMN subnetwork-A and the RH control subnetwork-B. Whereas, no significant functional connectivity modulations were observed for the TMS trials in the LLP bins. These findings illustrate the influence of TMS timing on acute brain activation.

3.4. Associations of TMS-induced acute/immediate effects with the clinical response in rTMS treatment for TRD patients

In this section, we first tested the associations between the quantification of TMS-induced acute/immediate effects at baseline and the clinical response during rTMS treatment. As for the TMS-evoked BOLD response, we observed a significant correlation between the HRSD improvement and the amplitude of evoked response in the bilateral control subnetwork-B (LH: r = 0.8591, p < 0.0007; RH: r = 0.8601, p < 0.0007), and RH limbic subnetwork-B (r = 0.8991, p < 0.0002) for the patients in the SYNC group. There was no significant association between the clinical improvement and the amplitude of evoked response in any network for the patients in the UNSYNC group (p < 0.05; Bonferroni correction; detailed results are provided in Table SII). No significant association was found between the spatial extents of evoked response at baseline pre-treatment scan and clinical improvement (detailed results are provided in Table SIII).

Next, we assessed the longitudinal changes in the TMS-evoked response and tested their associations with the clinical outcome during rTMS treatment. We did not observe any significant result for each group. However, by pooling the patients from both groups, we found that TMS-evoked BOLD responses in the bilateral salience/ventral attention subnetwork-A significantly decreased from the pre-treatment to the post-treatment scan (LH: t = −4.0129, p < 0.0074; RH: t = −4.0403, p < 0.0070). But these longitudinal changes were not significantly associated with the clinical outcome (LH: r = −0.2505, p = 0.4851; RH: r = −0.2662, p = 0.4571).

Finally, we asked if any state-specific TMS modulations of connectivity at baseline or in its longitudinal changes are associated with the clinical outcome during rTMS treatment. We assessed the state-specific effects with the HLP condition. Here, as shown in Fig. 5B, we examined four L-DLPFC ROIs (details of ROI selection in supplementary text). We did not observe any significant association between the clinical outcome and the baseline state-specific TMS modulations of connectivity. Then, we tested the associations between the clinical outcome and the longitudinal changes in the state-specific TMS modulations of connectivity. As for the state-specific effects associated with the L-DLPFC of the DMN subnetwork-A, we observed that the clinical outcome was significantly correlated to the changes in the connectivity between L-DLPFC in the DMN subnetwork-B and orbitofrontal cortex in the RH limbic subnetwork-B only in the SYNC group (SYNC group: r = 0.9916, p < 1.4971e-6, FDR-corrected p < 0.01; UNSYNC group: r = −0.3841, p = 0.45). This relationship remained significant after regressing out the baseline connectivity measurement from the longitudinal changes (r = 0.7269, p < 0.0411). However, when testing other L-DLPFC regions, we did not observe any significant relationship between the clinical outcome and the connectivity changes after multiple comparison correction. For the connectivity between L-DLPFC and RH orbitofrontal cortex in the SYNC group, we did observe a trend for the state-specific effects associated with the L-DLPFC in the DMN subnetwork-B (p < 0.0023), but not L-DLPFC of EEG F3 (p = 0.4730) and L-DLPFC in the salience/ventral attention subnetwork-B (p = 0.0903). Our results suggest that EEG-synchronized rTMS treatment induces functional connectivity changes in specific neural circuits that are associated with the clinical outcome, i.e., state-dependent response in the L-DLPFC (DMN subnetwork-A) and the connectivity between L-DLPFC (in the DMN subnetwork-B) and RH orbitofrontal cortex (in the limbic subnetwork-B).

Fig. 5.

Associations of TMS-induced acute/immediate effects with the clinical response in rTMS treatment for depressed patients. (A) We observed a significant correlation between the clinical improvement (percent change in the HRSD) and the pre-treatment TMS-evoked response in the bilateral ContB network and RH-LimbicB network only for the SYNC group. (B) State-specific TMS modulations of connectivity were assessed for four L-DLPFC regions (EEG F3, L-DLPFC in DefaultA, L-DLPFC in DefaultB, and L-DLPFC in SalVentAttnB). (C) As for the L-DLPFC in the DefaultA, the pre- and post-treatment state-specific TMS-modulated connectivity changes (between the L-DLPFC in the DefaultB network and the RH orbitofrontal cortex in the LimbicB network) are significantly associated with the clinical outcome only for the SYNC group. No significant result was found for the other three L-DLPFC regions after multiple comparison correction. Panel (A) includes the results of 23 patients (11 SYNC patients and 12 UNSYNC patients) with pre-treatment integrated fMRI-EEG-TMS and HRSD available. Panel (C) includes the results of 14 patients (8 SYNC patients and 6 UNSYNC patients) with complete pre- and post-treatment data.

4. Discussion

In this paper, we investigated the TMS-induced acute/immediate effects on brain networks. First, we quantified its evoked BOLD response and modulated PPI functional connectivity. Then, we tested associations between TMS-induced effects and the clinical outcome. Additionally, we examined whether these TMS-induced effects depend on the prefrontal alpha oscillation phase. Our results showed that the spatial spread of TMS-induced phase-dependent effects follows a similar pattern to the correlation-based functional connectivity of the L-DLPFC. When assessing the TMS-modulated PPI functional connectivity with trials in the HLP bins (i.e., TMS trials that evoked a higher response at the L-DLPFC stimulation site), we observed significant negative modulations of the control network and the DMN. No significant modulation was found for trials where TMS was delivered at the LLP timing, and no significant modulation was observed for this state-specific effect associated with other L-DLPFC ROIs near the stimulation site. Lastly, we tested the associations between the quantification of TMS-induced acute/immediate effects and the clinical outcome, where TRD patients received a randomized six-week rTMS treatment. We found that the baseline TMS-evoked responses in the bilateral control subnetwork-B and RH limbic subnetwork-B significantly predicted the following clinical improvement, but only in the SYNC group. Finally, our results showed that the longitudinal changes in the state-specific TMS modulations of connectivity between L-DLPFC (part of the DMN subnetwork-B) and RH orbitofrontal cortex (in the limbic subnetwork-B) were significantly associated with the clinical improvement only for the SYNC group. The present study assessed TMS-induced acute/immediate effects on brain networks with concurrent TMS-fMRI and investigated the brain-state dependency of the induced effects with simultaneous EEG recordings. The results of its association with clinical outcomes have potential implications for developing efficient and personalized treatments for TRD patients.

The concurrent acquisition of fMRI data during single-pulse TMS enables the investigation of the acute/immediate BOLD response to TMS. This TMS-evoked BOLD response, frequently characterized in recent concurrent TMS-fMRI studies, serves to demonstrate target engagement or its associations with the clinical outcome. For example, Oathes et al. explored the TMS-evoked response in subcortical brain areas and the associated networks, such as the amygdala and sgACC [62], demonstrating engagement of these circuits underlying neuromodulation. Similarly, Vink et al. assessed the spatial propagation of TMS-evoked responses in the sgACC and across the whole brain [13]. In support of the findings in other literature, these studies highlighted the substantial individual variability in TMS-evoked responses. The observed variability may be attributable to neural activity at the stimulation site [11] and/or its functional connectivity with specific brain networks, such as the salience network [12]. These observations underscore the importance of characterizing such individual variabilities. Thus, in this study, we proposed to quantify both the amplitude and the spatial propagation extent of the immediate acute TMS-evoked responses. Such quantifications potentially validate target engagement and could have important clinical implications. For instance, a recent study suggested that a stronger evoked response in the sgACC during the pre-treatment TMS-fMRI session demonstrated better target engagement and subsequently correlated with improved therapeutic outcomes in the treatment [6]. Our findings align with these studies, where we demonstrated that the evoked response in the right orbitofrontal cortex (posterior portion of sgACC) within the limbic subnetwork-B during the pre-treatment session is relevant to the clinical outcomes of rTMS treatment. Additionally, we observed that baseline-evoked responses in the bilateral control subnetwork-B were significantly associated with clinical outcomes. However, these associations are only significant for the patients in the SYNC group. This suggests that patients exhibiting the strongest evoked responses in these networks, representing better target engagement of these networks, benefit most from EEG-synchronized rTMS treatment. Conversely, the nonsignificant results for patients in the UNSYNC group might imply that the therapeutic effects of unsynchronized rTMS treatment might not be specific to certain brain networks. Moreover, with the longitudinal data, we showed a significant decrease in the evoked response within the salience network after rTMS treatment, which does not correlate significantly with the clinical outcomes. These findings might suggest the potential impact of rTMS on anxiety symptoms, given the established link between the salience network and anxiety [63,64]. However, further research is warranted to validate this with a larger sample size.

Numerous studies have illustrated that TMS not only influences the stimulation site and its directly connected network but also has effects that propagate to other brain networks [65]. Considering these distributed brain regions collectively within the framework of brain circuits and networks, especially in relation to the impacts of TMS and the characterization of depressed patients [66,67], could enhance our understanding of the therapeutic effects of TMS in depression, and may also inform target selection. In support of this literature, we sought to characterize and quantify such effects at the level of network connectivity. Recent studies have demonstrated that rTMS may down-regulate and normalize hyperconnectivity of certain brain networks, such as the DMN [17,68] and the salience network [69]. Consistent with these findings, our study also revealed that TMS induces significant negative modulations in the connectivity between brain networks, including the DMN, control, and salience networks. These results are consistent with prior findings by Chen et al., who observed significant negative TMS modulations between the CCN (lateral frontoparietal network [70] or central executive network) and the DMN when single-pulse excitatory TMS delivered to the node of CCN [7]. Our results are also consistent with the findings that TMS can normalize depression-related hyperconnectivity associated with DMN [17]. These findings potentially suggest the engagement of these brain networks in the therapeutic mechanisms of rTMS by normalizing the network hyperconnectivity in depressed patients. However, unlike these studies, our concurrent acquisition allowed us to assess the acute modulatory effects of TMS on network connectivity and potentially provides more causal insights into the interactions between brain networks. Furthermore, our findings are in line with concurrent TMS-fMRI research, showing that TMS application to the central executive network (CEN) node induces causal negative downstream effects on the DMN [7].

In this study, we utilized the phase of EEG prefrontal alpha oscillation as an indicator of brain state and investigated the brain-state-dependent effects of TMS-induced brain activity. Our results suggest that the propagation of TMS-induced BOLD activity from the L-DLPFC to regions within the L-FPN is dependent on the EEG prefrontal alpha phase. Specifically, in conditions where TMS trials elicited a stronger response at the L-DLPFC (HLP timing), areas within the L-FPN also exhibited a stronger response to these TMS pulses, suggesting that the state-dependent spread patterns of TMS-induced activity closely follow the functional connectivity of the L-DLPFC, with a high spatial similarity. However, when examining other ROIs proximate to the L-DLPFC F3 stimulation site, the results suggest distinct spread patterns, underscoring the spatial specificity inherent to the localization or characterization of the L-DLPFC ROIs. These findings align with existing literature, indicating that the spread of TMS-induced activity from the L-DLPFC stimulation site follows the functional connectivity pattern of the site, with studies suggesting it depends on functional connectivity to the salience network [12]. Our results, along with the literature, highlight the critical role of leveraging correlation-based functional connectivity to guide TMS targeting [23]. However, it should be noted that we assessed functional connectivity of the L-DLPFC after regressing out TMS-evoked responses, to control the confound of TMS-related activations. Future studies should explore the spread of TMS-induced response and functional connectivity based on separate resting-state fMRI sessions to completely rule out the possible interactions between spontaneous brain activity and TMS-evoked responses. In this study, we used the prefrontal alpha oscillation phase as an index of brain state and assessed state-dependent effects on both evoked response and modulated functional connectivity. This choice was motivated by literature evidencing the gating mechanism of the alpha oscillation phase. However, to further elucidate the state dependency of TMS-induced activity, future research should consider other brain state indices, such as the power of alpha or the phase of theta oscillation. In our previous investigations, we demonstrated that the TMS-evoked BOLD response in the cingulate gyrus is dependent on the EEG prefrontal alpha phase [34]. Additionally, we have reported that alpha phase-synchronized rTMS treatment facilitates phase entrainment in patients [47], with increased entrainment over time correlating with improved rTMS treatment outcome [30]. Altogether, these findings contribute to the development of biomarkers for tracking the efficacy of EEG-synchronized rTMS treatment [31]. Following our previous findings, the current study focused on quantifying the acute effects of TMS on fMRI signals and evaluating its clinical relevance. These results add new perspectives to our understanding of TMS-induced effects on brain networks and also shed light on the therapeutic effects of rTMS, with implications for the design of future interventions in depression. An important future direction will be to conduct cross-modal analyses linking TMS-fMRI measures with TMS-EEG neurophysiological markers, which may provide a more comprehensive understanding of the therapeutic mechanisms underlying rTMS treatment.

The observed association between changes in state-specific PPI connectivity modulation and clinical outcomes suggested that patients who achieved a greater increase in modulated connectivity between the L-DLPFC and the right orbitofrontal cortex benefited the most from EEG-synchronized rTMS treatment. This brain circuit, which showed relevance to the clinical outcome, is associated with the personalized prefrontal alpha phase (HLP condition) in which TMS evoked the strongest response in the L-DLPFC within the DMN subnetwork-A. Furthermore, to examine spatial specificity, we assessed this relationship with other L-DLPFC ROIs. Our results revealed that, when evaluating the phase-dependent response, the strongest association to the clinical outcome was observed for the L-DLPFC within the subnetwork-A of the DMN, compared to ROIs in the EEG F3 area, DMN subnetwork-B and salience/ventral attention subnetwork-B. However, caution should be taken when evaluating the correlation coefficient results with a small sample size in this study. No such association was observed when using patients in the UNSYNC group as a control group. These findings highlight the potential systematic therapeutic effect of EEG-synchronized rTMS treatment in engaging more specific brain circuits, which might be more directly related to the clinical therapeutic effects in depression, as opposed to the possible widespread effects in the unsynchronized rTMS treatment. The latter (UNSYNC group treatment) may induce effects across a wide range of brain circuits, with TMS pulses during various brain states affecting distinct circuits. Randomly fired TMS pulses at different prefrontal alpha phases could modulate different network connections, with some pulses occurring at the right time actually having beneficial effects. Our findings align with existing literature that the therapeutic mechanism of rTMS treatment is likely to involve the circuit of the right orbitofrontal cortex (posterior portion of the sgACC) and the L-DLPFC [23,71]. Future research with a larger cohort is necessary to validate these results.

This study is subject to several limitations that warrant consideration in the future. First, while our findings are promising, the current study is limited by a relatively small sample size. Our results demonstrated that quantified effects of TMS are correlated with clinical outcomes only in the SYNC group, and results from the UNSYNC group served as a control. However, a direct comparison of the UNSYNC group’s results with those of conventional standard clinical TMS treatments is challenging. This is attributed to the fact that, for the UNSYNC group, the patients still received TMS pulses adjusted to their personalized prefrontal alpha frequency but with a random phase. Another notable limitation is the absence of a sham intervention. The sensations and auditory aspects associated with TMS may have influenced the quantified acute effects on the fMRI signal. The study’s double-blind design ensured that both groups experienced these factors during the intervention similarly. However, future studies employing sham-controlled interventions are needed to rule out any potential confounding effects caused by the sensory stimulation during TMS. Furthermore, this study focused on optimizing the timing of TMS targeting, but we did not employ a strategy for precise and individualized spatial targeting. There is a need for future research to optimize spatial targeting of TMS based on individualized functional connectivity and possible electric field modeling [72].

It should be noted that the effects of single-pulse TMS reported here may differ from those produced by repetitive stimulation in clinical rTMS protocols. Single-pulse TMS, combined with fMRI, is primarily designed to evaluate acute stimulation responses of brain circuits. Whereas the effects of rTMS depend strongly on stimulation frequency and pattern, with low-frequency rTMS (<1 Hz) generally inducing inhibitory effects and higher-frequency protocols (>5 Hz) typically producing excitatory effects. In the present study, we focused on single-pulse TMS with concurrent fMRI to characterize acute brain responses to stimulation, which provides a critical approach to demonstrate target engagement and to develop pre-treatment markers that may guide and optimize rTMS interventions. Future studies are needed to directly investigate the acute effects of rTMS on brain networks. However, concurrent rTMS-fMRI has significant technical challenges related to heating, safety, and potential artifacts, and most existing work has relied on offline assessments of rTMS-induced effects. While our study employed single-pulse TMS to evaluate acute network responses, the development of safe and effective protocols for concurrent rTMS-fMRI would enable direct characterization of the therapeutic effects of rTMS.

While the current study demonstrates phase-specific engagement of brain circuits in response to single-pulse TMS, and identifies an fMRI-based neurophysiological marker associated with clinical outcomes in synchronized rTMS treatment, these findings should be interpreted with caution regarding their clinical relevance. The results may inform future development of EEG-synchronized rTMS protocols for depression. However, our double-blind clinical trial did not demonstrate clinical superiority of the synchronized treatment, and the therapeutic effects of rTMS were not directly evaluated in this study. Thus, the clinical relevance of the present findings remains limited. Future studies are needed to determine whether an optimized synchronized rTMS design, guided by the neurophysiological marker identified here, can improve clinical outcomes, and an additional important future direction will be to test concurrent rTMS-fMRI approaches for directly assessing rTMS therapeutic effects.

In conclusion, the present study highlights the importance of quantifying TMS-induced acute effects on brain network systems, including evoked response and modulated functional connectivity. These quantifications demonstrate brain areas and circuits engaged by TMS perturbation and potentially their involvement in the therapeutic effects of rTMS treatment in depression. Additionally, the state-dependent analyses established that, by conditioning on different phases of the prefrontal alpha oscillation, the spread of TMS-induced activity follows the functional connectivity of the stimulation site. Finally, in an exploration analysis with a limited sample size, longitudinal changes in the state-specific TMS-modulated connectivity in the brain circuit of the stimulation site and the sgACC are associated with the clinical outcome. These results suggest that EEG-synchronized rTMS engages specific brain networks (i.e. right limbic and bilateral cognitive control networks) and brain circuits (i.e. L-DLPFC and right posterior sgACC) toward depression treatment. These results carry significant implications for the treatment of TRD and the development of more precise and personalized TMS treatment protocols, with a focus on both target selection and timing parameters.

Appendix A. Supplementary data

Supplementary data to this article can be found online at https://doi.org/10.1016/j.brs.2025.10.013.

Data and code availability

The datasets analyzed in the current study are available from the corresponding author on reasonable request. Custom codes used in this work are available at https://github.com/hehengda/fMRI_EEG_TMS.git.