Brain Network Effects by Continuous Theta Burst Stimulation in Mal de Débarquement Syndrome: Simultaneous EEG and fMRI Study

Abstract

Objective:

Heterogeneous clinical responses to treatment with non-invasive brain stimulation are commonly observed, making it necessary to determine personally optimized stimulation parameters. We investigated neuroimaging markers of effective brain targets of treatment with continuous theta burst stimulation (cTBS) in Mal de Débarquement Syndrome (MdDS), a balance disorder of persistent oscillating vertigo previously shown to exhibit abnormal intrinsic functional connectivity.

Approach:

Twenty-four right-handed, cTBS-naive individuals with MdDS received single administrations of cTBS over one of three stimulation targets in randomized order. The optimal target was determined based on the assessment of acute changes after the administration of cTBS over each target. Repetitive cTBS sessions were delivered on three consecutive days with the optimal target chosen by the participant. EEG was recorded at single-administration test sessions of cTBS. Simultaneous EEG and fMRI data were acquired at baseline and after completion of 10-12 sessions. Network connectivity changes after single and repetitive stimulations of cTBS were analyzed.

Main results:

Using electrophysiological source imaging and a data-driven method, we identified network-level connectivity changes in EEG that correlated with symptom responses after completion of multiple sessions of cTBS. We further determined that connectivity changes demonstrated by EEG during test sessions of single administrations of cTBS were signatures that could predict optimal targets.

Significance:

Our findings demonstrate the effect of cTBS on resting state brain networks and suggest an imaging-based, closed-loop stimulation paradigm that can identify optimal targets during short-term test sessions of stimulation.

1. Introduction

Mal de Débarquement Syndrome (MdDS) is a neurological disorder of persistent oscillating vertigo caused by entrainment to background periodic motion such as occurs from sea, air, or land travel (Cha, 2009; Hain et al., 1999). The primary symptoms of MdDS are a rocking, bobbing, or swaying perception with associated symptoms of headache, fatigue, visual motion intolerance, and tinnitus suggesting that brain networks other than just the vestibular system are involved (Cha, 2009; Cha et al., 2008). Indeed, there are no peripheral vestibular deficits in MdDS (Cha, 2009); but functional imaging studies have shown central metabolic and connectivity abnormalities in MdDS (Cha et al., 2012). These studies have led to investigations using non-invasive brain stimulation in the treatment of MdDS (Cha et al., 2021; Cha et al., 2016a; Cha et al., 2019; Cha et al., 2016b).

Prior neuroimaging and non-invasive brain stimulation studies on MdDS have indicated that both long-range fronto-parieto-occipital connectivity as well as corticolimbic connectivity with the entorhinal cortex are related to symptom status (Cha et al., 2018; Cha et al., 2012; Ding et al., 2014; Yuan et al., 2017). The first neuroimaging study on MdDS using fluorodeoxyglucose positron emission tomography (FDG-PET) showed that there is hypermetabolism in the left entorhinal cortex and amygdala along with increased connectivity to posterior parietal and occipital cortex with these loci in individuals with MdDS (Cha et al., 2012). Using functional magnetic resonance imaging (fMRI), reduction in resting state functional connectivity measured between the entorhinal cortex and the inferior parietal lobule (IPL) and precuneus corresponding to the default mode network (DMN) was shown after repetitive transcranial magnetic stimulation (rTMS) of the dorsolateral prefrontal cortex (DLPFC) (Yuan et al., 2017). Resting state network connectivity measured by electroencephalography (EEG) showed reduction in connectivity in the medial frontal gyrus (MFG) which was correlated with fMRI connectivity changes in the default mode network (Chen et al., 2019). Morever, measurements of independent component phase coherence (ICPC), an EEG marker of synchronicity, have shown reductions in synchrony as a function of symptom status (Cha et al., 2018; Ding et al., 2014).

A critical challenge in the treatment of MdDS with neuromodulation has been in designing a more effective protocol that could achieve faster and higher response rates. A DLPFC protocol that administered 1 Hz right-side and 10 Hz left-side stimulation over a 5-day course of stimulation yielded a 30% treatment response rate (Cha et al., 2016b). Motivated by the DLPFC protocol, our current study considered a more intensive form of stimulation – theta burst stimulation (TBS) – in which trains of gamma frequency stimulation (50 Hz) are given at theta frequency (5 Hz) intervals (Cha et al., 2019). This form of stimulation more closely models the theta-gamma coupling of cortico-cortico communication. The more rapid stimulation parameters and generally longer stimulation effects compared to standard pulsed stimulation have led to increasing numbers of therapeutic trials of TBS for various clinical disorders (Di Lorenzo et al., 2020; Li et al., 2018; Schwippel et al., 2019; Zhao et al., 2020). Notably, intermittent TBS has been deemed as a non-inferior protocol to 10 Hz stimulation over DLPFC for depression (Dhami et al., 2019).

In our latest study of MdDS (Cha et al., 2019), a continuous TBS (cTBS) protocol led to a doubling of the response rate compared to the DLPFC protocol allowing a broader distribution of positive clinical effects on which to draw correlations to symptom improvement. On a similar base of 24 right-handed participants, 16 individuals, as opposed to 7 individuals reported at least a 10-point reduction in symptom severity of oscillating vertigo on a visual analogue scale of 0-100. A key remaining question in the cTBS study was whether any changes in neuroimaging data were related to the symptom changes in MdDS participants after completion of the stimulation protocol. Knowledge of imaging biomarkers that indicate treatment effects could be critical for establishing the neuromodulation mechanism of cTBS and for further individualized tailoring of treatment.

Each participant in our cTBS study received single adminitrations of stimulation at three different treatment targets with EEG recodings before and after each test stimulation. Since changes in the connectivity of the DMN and the visual network were closely related to symptoms in MdDS (Cha, 2009; Chen et al., 2019; Ding et al., 2014; Yuan et al., 2017), the target regions of stimulation were in the dorsal occipital cortex and cerebellar vermis, which are functionally connected to the fronto-parietal attention network (Halko et al., 2014). The lateral cerebellar hemisphere served as an active sham control. cTBS was employed in this study because it is an inhibitory protocol, while intermittent TBS is the physiologically opposite excitatory protocol that had been shown to increase DMN connectivity (Halko et al., 2014). Moreover, cTBS has been shown to decrease connectivity within visual areas – a theoretically favorable effect given the heightened connection between visual and entorhinal cortices previously noted in MdDS (Cha 2012; Chen et al., 2019). Since cTBS requires much shorter stimulation times than conventional TMS protocols (40 seconds vs 30-minutes) and yields more durable effects (Huang et al., 2005), we were able to provide a more tolerable therapy as well as to provide back-to-back sessions of treatment on the same day. The optimal target was determined based on the assessment of acute subjective symptom changes after each single test administrations, followed by 10-12 sessions of repetitive cTBS delivered over consecutive days with the chosen optimal target.

In our first analysis, we determined EEG and fMRI network markers that correlated with symtom change by obtaining both modalities of data before and after multiple treatment sessions. We were thus able to identify clinically relevant biomarkers of symptom reduction. MdDS participants have been shown to exhibit illness duration dependent changes in gray matter volume in visual-vestibular processing areas (Cha and Chakrapani, 2015) and our previous multi-modal study had revealed symptom associated changes in the visual network and default mode network, in which a greater decrease in connectivity was seen in individuals with a greater reduction of symptoms (Chen et al., 2019). Therefore, our analysis focused on the DMN and visual networks and derived functional connectivity metrics for regions of interests within these networks.

In our second analysis, we extended our investigation of symptom-correlated networks to EEG recordings obtained during test sessions of cTBS in order to identify early changes in brain network connectivity that predicted treatment response after multiple sessions. Our goal was to determine whether EEG changes that correlated with symptom improvement after multiple sessions could be identified after single test sessions of stimulation. If possible, transient effects measured by EEG could be used to determine optimal targets in rapid test sessions, thereby ultimately enabling an imaging-based, closed-loop stimulation paradigm that is independent of participants’ subjective reports. Early optimization of treatment targets could save time and reduce ambiguity, particularly when treatment targets may need to be individually tailored.

2. Methods

2.1. Participants

Study procedures adhered to Declaration of Helsinki guidelines and were approved through Western IRB (www.wirb.com). Participants gave written informed consent and were recruited under ClinicalTrials.gov study NCT02470377. Twenty-four right-handed individuals diagnosed with MdDS were recruited into this study; all but one completed the entire data collection process. The mean age for the 23 participants at the time of the study was 51.5 ± 12.0 years, with a median of 55 years and a range of 30 – 70 years. The duration of illness was a mean of 26.0 ± 22.9 months, a median of 18 months, and a range of 7 – 104 months.

All participants were women, which coincided with the high prevalence of female sex in MdDS (Cha, 2009). Participants were included if they had: 1) persistent oscillating vertigo after disembarking from sea, air, or land-based travel with symptoms that started within two days of disembarkation, 2) symptoms lasting at least 6 months, and 3) were evaluated by an experienced neurologist or otolaryngologist and were found to have no other cause for symptoms. In general, we recruited a medically refractory group that had failed at least one benzodiazepine, a selective serotonin reuptake inhibitor (SSRI) or selective norepinephrine reuptake inhibitor (SNRI), and physical therapy. Individuals were excluded if they: 1) had contraindications to undergoing cTBS, EEG, or fMRI, such as medications known to reduce seizure threshold or implanted metal, 2) were pregnant or planned to be during the course of the study, 3) had an unclear trigger for their symptoms, 4) were incapable of completing all study-related testing, or 5) had an unstable medical or psychiatric condition such as a history of bipolar disorder or psychosis.

2.2. Experimental procedures

The experimental protocol lasted 5 consecutive days, shown in Fig. 1. On Day 1, structural and functional MRI scans and 128-channel EEG were obtained (see details in the Simultaneous fMRI-EEG Acquisition section). On Day 2, participants received cTBS over three potential locations in a randomized sequence between participants. These targets included the midline dorsal occipital cortex, midline cerebellar vermis lobule VII, and right lateral cerebellar hemisphere over the horizontal fissure. cTBS is composed of three 50 Hz-pulses repeated every 200 ms (i.e., a burst of three pulses repeated at 5 Hz) given continuously in an uninterrupted manner for 600 pulses, which is about 40 seconds (Huang et al., 2005). The participants laid face down in a massage chair so that the back of the head could be easily accessed. Concurrent cTBS-EEG was set up for recording EEG signals immediately before, during, and after single administrations of cTBS (see details in the EEG Acquisition on Day 2 section). The Localite® TMS Navigator (Localite GmBH, Germany) frameless stereotaxy system was utilized for neuronavigation to localize each target using the participant’s own structural brain MRI obtained on Day 1. The optimal target was chosen by the participant based on the greatest reduction of scores within 60-minutes reported on a 0-100 visual analogue scale (VAS) (see Visual Analogue Scale section). The other two targets were considered non-optimal targets. The stimulation over each target was given in a randomized order between participants. If there was ambiguity in the optimal target, stimulation over two targets could be repeated on Day 3 in the reverse order in which they were administered on Day 2. In the event that the participant did not perceive any benefit from any target, they were randomly assigned to one of the three targets (undecided targets). Therefore, some participants received 10 “stacked” sessions while others received 12 “stacked” sessions. Stimulation over the selected optimal target was administered on Days 3 through 5, i.e., 3-4 sessions per day. Each treatment session consisted of 600 pulses given over 40 seconds, a 20-second break, then 600 pulses given over 40 seconds. On Day 5 after the stimulation protocol completed, post-treatment EEG and MRI were acquired.

Figure 1:

The experimental study protocol and analysis schemes.

2.3. Visual analogue scale

A self-reporting, visual analogue scale (VAS) of 0-100 was employed to evaluate the participants’ symptom changes from Day 1 to 5, in which 0 indicated no symptoms and 100 represented symptoms so severe that they could not remain standing. Prior studies showed that a decrease of 10 points was a meaningful detectable difference on this scale (Cha et al., 2013; Cha et al., 2016a; Cha et al., 2019; Cha et al., 2016b). The participants whose scores decreased by 10-points or more on Day 5 after the completion of stimulations relative to the score on Day 1 prior to any stimulation were considered to be “responders”, while those who showed a smaller decrease, no change, or increase were considered as “non-responders”.

On Day 2, the participants received 1 session over each of the 3 cTBS targets. Participants scored VAS prior to each target-specific session. Then following the end of every session, participants sat for 5-minutes before they were allowed to move (during which time the EEG was recorded). Each session was then followed by a 60-minute resting period, after which subjects scored symptom on VAS. After tying out all three targets, the stimulation target that most optimally lowered the participant’s symptoms on the VAS (i.e., post vs. pre scores per each target) was chosen for repetitive treatment sessions.

2.4. MRI and EEG acquisition on Day 1 and Day 5

On Day 1 and Day 5 of the protocol, structural and resting state functional MRI were collected using a General Electric Discovery MR750 whole body 3-Tesla MRI scanner (GE Healthcare, Milwaukee, WI). Whole brain resting-state fMRI data were acquired using the following parameters: FOV/slice = 240/2.9 mm, axial slices per volume = 34, acquisition matrix = 96×96, repetition time/echo time (TR/TE) = 2000/30 ms, SENSE acceleration factor R = 2 in the phase encoding (anterior-posterior) direction, flip angle = 90°, sampling bandwidth = 250 kHz and total number of volumes = 185. The EPI images were reconstructed into a 128×128 matrix, in which the resulting fMRI voxel volume was 1.875×1.875×2.9 mm³. Additionally, physiological pulse oximetry and respiration waveform recordings were collected (with 40 Hz sampling) for each fMRI run. A photo-plethysmograph with an infrared emitter placed under the pad of the patient’s left index finger was used for pulse oximetry while a pneumatic respiration belt was used for respiration measurements. The patient rested quietly with their eyes closed during the image acquisitions. A T1-weighted magnetization-prepared rapid gradient-echo (MPRAGE) sequence with sensitivity encoding was used to provide an anatomical reference for fMRI analysis, which had the following parameters: FOV = 240 mm, axial slices per slab = 190, slice thickness = 0.9 mm, image matrix = 256×256, TR/TE = 5/2.012 ms, acceleration factor R = 2, flip angle = 8°, inversion time TI = 725 ms, sampling band width = 31.2 kHz.

On Day 1 and Day 5 of the protocol, simultaneous fMRI-EEG were recorded through a 128-channel cap using BrainAmp MR Plus amplifiers (Brain Products GmbH, Munich, Germany) with sintered Ag/AgCI ring electrodes following the standard 10-5 system. Ten KΩ was the impedance upper limit for all EEG electrodes. Sampling rate was 5000 Hz. Participants lied down in the MRI machine and were instructed to keep still with eyes closed during the resting state recording of 6 minutes and 10 seconds. The SyncBox device (Brain Products GmbH) was used to synchronize the internal sampling clock of the EEG amplifier with the MRI scanner 10 MHz master clock signal.

2.5. EEG acquisition on Day 2

On Day 2, participants received 600 pulses of test cTBS over three potential locations in a randomized order between participants, including occipital, cerebellar vermis, and lateral cerebellar hemispheres. Concurrent cTBS-EEG was set up to record EEG signals immediately before, during, and after each test cTBS. 32-channel EEG data were recorded by a BrainAmp MR Plus amplifier (Brain Products GmbH, Munich, Germany) with sintered Ag/AgCl ring electrodes following the standard 10-20 system. Ten KΩ was the impedance upper limit EEG electrodes. Sampling rate was 5000 Hz. A 5-min EEG resting state recording was acquired immediately before and after the cTBS stimulation at each location. Only the EEG recordings before and after cTBS without TMS-induced artifacts were analyzed. In one participant, one of the post-cTBS EEG data was not available due to technical failure and therefore excluded from the Day 2 analyses.

2.6. Processing steps of EEG on Day 1 and Day 5

EEG data from Day 1 and Day 5 were cleaned of environmental and physiological artifacts following our previous analysis procedures (Chen et al., 2019; Yuan et al., 2016; Yuan et al., 2011; Yuan et al., 2018; Yuan et al., 2013; Yuan et al., 2012). Specifically, the gradient artifact and cardioballistic artifacts were removed in Brain Vision Analyzer 2.1 based on average artifact subtraction methods (Allen et al., 2000; Allen et al., 1998). Additionally, artifact residuals such as from muscle and ocular movements were removed through independent component analysis (ICA) in MATLAB 2017b, implemented in the EEGLAB toolbox (http://sccn.ucsd.edu/eeglab) (Delorme and Makeig, 2004). Bad channels with excessive artifacts were discarded via visual check and replaced by an average of its nearby four channels. Bad segments with various artifacts were also rejected by visual inspection. High-pass filtering at 0.5 Hz, notch filtering at 60 Hz, and low-pass filtering at 70 Hz was applied. All channels were re-referenced to the common average.

For EEG source reconstruction, individual MRI data were utilized to project scalp EEG measurements to cortical sources. Boundary element models for all participants were built based on individual structural MRI data segmented by FreeSurfer (https://surfer.nmr.mgh.harvard.edu/). The head model was constructed as a 3-layer boundary element model: scalp, skull, and brain based on triangular elements. Each layer was given its own conductivity value (Zhang et al., 2006). Individual digitization files of genuine electrode locations were co-registered to the scalp model. EEG source imaging was used to calculate the current dipoles on the cortex for each individual by a minimum norm method (Dale and Sereno, 1993; Hamalainen and llmoniemi, 1994).

To perform group-level analysis, individual head models were registered to the standard FreeSurfer “fsaverage” subject (Fischl et al., 1999). Continuous source data were down-sampled to the microstate based on corresponding scalp EEG data to retain data at a higher signal-to-noise ratio (Lehmann et al., 1987; Yuan et al., 2016). Absolute values of source data were then temporally concatenated across participants and a temporal ICA (Delorme and Makeig, 2004) was performed to reconstruct the EEG networks at the group level (Chen et al., 2019; Yuan et al., 2016). Because the participants in the current study turned out to be predominantly responders, we derived the projecting matrix based on responders only but applied the reconstruction of network connectivity to all participants. As a comparison, we also ran the ICA analysis on all participants; results were highly similar and are included in Supplemental Fig. 1. A total of 25 ICs was calculated. The number of ICs was chosen based on our prior investigations, which yielded cross-modal consistency between EEG and fMRI networks (Yuan et al., 2016).

Individual cortical connectivity of EEG networks was then calculated based on the group-level ICA analysis. Specifically, the projecting matrix was applied to individual participants’ continuous source matrices to calculate the time series of the EEG networks. Then, in every participant, Pearson correlation coefficients between the time course of each EEG network and the time course of each source dipole on the cortical surface were calculated, resulting in a map of correlation coefficients as an EEG connectivity map. For later statistical analyses, Pearson correlation coefficients underwent a Fisher’s z transform. Afterwards, spatial smoothing using a heat kernel (FWHM = 10 mm) (Chung et al., 2005) was employed to obtain a comparable spatial pattern with fMRI. Pre- and post-cTBS EEG connectivity maps were computed separately from data on Day 1 and Day 5. “Post-minus-pre-connectivity” maps represent the connectivity changes after cTBS, which were calculated by subtracting the pre-cTBS (Day 1) map from the post-cTBS (Day 5) map for each participant.

In the current study, only the visual network and the DMN were examined, as informed by our prior investigations of repetitive TMS over DLPFC (Chen et al., 2019). To identify the visual network, spatial correlation coefficients between the Yeo template of the visual network (Yeo et al., 2011) and all EEG network maps was calculated and the one with the highest spatial correlation was chosen as the EEG visual network. Similarly, a match for the Yeo template of the DMN was identified and designated as the EEG DMN. The significance of spatial matching was assessed by a bootstrap approach based on the null distribution of spatial correlation values established in (Yuan et al., 2016). Moreover, two additional EEG networks that appeared as parts of the DMN, i.e., the left IPL and right IPL, were manually selected. They were not identified as the highest match to any template but were chosen as additional nodes of the DMN that could be accessed from the cortical surface. Thus, a total of four EEG networks were selected, namely the visual network, default mode network focused on the medial frontal gyrus, the left IPL network and the right IPL network. Since multiple networks were selected, Benjamini-Hochberg method was used for multiple comparison correction (Benjamini and Hochberg, 1995). A cross-modal analysis of simultaneous EEG and fMRI data was performed to further confirm the correlation between these EEG networks and fMRI networks (see Supplemental Materials).

Regions of interests (ROI) were defined for each selected EEG network. An ROI with a diameter of 10 mm was set on the cortical surface centered at the point of maximum connectivity in an EEG network. Connectivity values, i.e., z scores, at all source dipoles within the ROI were extracted and averaged. Anatomical labels were identified using the Talairach-Tournoux Daemon atlas (Lancaster et al., 2000; Lancaster et al., 1997) provided in AFNI. The ROIs are identified as right inferior occipital gyrus (IOG), left MFG, left IPL and right IPL respectively for the visual network, DMN network, left IPL network and right IPL network.

2.7. Processing steps of EEG on Day 2

On Day 2 of the study protocol, a test session of cTBS was performed in which one stimulation session of cTBS was administered at three different sites for each participant (i.e., occipital, cerebellar vermis, and right lateral cerebellar hemisphere); only the site chosen as the optimal target based on the participant’s acute response was used for subsequent repetitive stimulation sessions. We employed EEG network analysis to examine the transient modulation of EEG connectivity immediately following a test stimulation at these three targets for the purpose of establishing a closed-loop stimulation paradigm based on the EEG network analysis. The EEG data were down-sampled to 250 Hz and independent component analysis was applied to remove artifacts related to eye movements, muscular, and cardiac activities. Bad channels with excessive artifacts were discarded via visual inspection and replaced by an average of its nearby four channels. Bad segments of various artifacts were rejected in the visual inspection. High-pass filtering at 0.5 Hz, notch filtering at 60 Hz and low-pass filtering at 70 Hz was applied. All channels were then re-referenced to the common average reference. After preprocessing, source images were calculated following the steps described above for the simultaneous fMRI-EEG signals. Because only 32 channels of EEG were acquired on the Day 2 EEG recordings, the lead field was selected from a subset of the model built for the 128-channel EEG.

Thereafter, we investigated the transient changes of EEG connectivity on the Day 2 associated with cTBS stimulations at each location. The projection matrix obtained from Day 1 and Day 5 data was adopted for calculating the EEG network time courses on Day 2. Pearson correlation coefficients between the time course of each network and the time course of source dipoles on the cortical surface area for each participant were calculated, resulting in a map of correlation coefficients as the EEG connectivity map. Similarly, Fisher’s transform and spatial smoothing were applied. Then “post-minus-pre-connectivity” maps representing the connectivity changes after cTBS, were calculated by subtracting the pre-cTBS map from the post-cTBS map for each participant and for each target location on Day 2. The ROIs that were defined based on Day 1 and Day 5 data were also adopted for calculating network connectivity values on Day 2. Connectivity changes within ROIs were extracted and averaged.

Once the optimal and non-optimal targets were determined for all participants, we compared each participant’s transient connectivity changes between optimal target and non-optimal targets by using a two-sided, paired t-test. In participants whose reported targets were undecided (self-reports indicated no favorable targets), we visualized their connectivity changes separately.

2.8. Analysis 1: Correlation of EEG connectivity changes with treatment response

We used an ROI analysis to examine the modulation on EEG networks with regard to symptom changes between Day 1 and Day 5. First, to test the normality of changes, we employed a corrected Kolmogorov–Smirnov test (Lilliefors, 1967; Öner and Deveci Kocakoç, 2017), with assumptions of unknown population mean and variance. We then calculated the correlation coefficients between connectivity changes measured by EEG and symptom changes measured by the VAS. A positive correlation coefficient indicated a decrease in EEG connectivity and a decrease of the VAS.

In addition, we grouped the participants as “responders” and “non-responders” to examine the changes within and between groups. Because our prior EEG and fMRI investigations have all suggested a greater reduction of connectivity associated with a decrease of symptoms and vice versa (Chen et al., 2019; Yuan et al., 2017), we compared the post-stimulation changes of connectivity between responders and non-responders using an unpaired, one-sided, two-sample t test. Finally, we assessed if any of the within-group connectivity differences were significant after cTBS using a paired, one-sided, two-sample t test separately in each group.

2.9. Analysis 2: Classification of optimal vs. non-optimal targets

In our current cTBS protocol, the optimal and non-optimal targets were determined by each participant’s self-report of symptoms on Day 2 test sessions. We determined whether connectivity metrics that became apparent after a full stimulation protocol by Day 5 could have some signatures that were detectable on the test sessions on Day 2. Initially, the optimal target was determined subjectively as the one leading to the greatest reduction in VAS within 60 minutes after stimulation. In our investigations, we explored linear discriminant analyses to classify the optimal vs. nonoptimal targets based on the network connectivity changes from EEG data immediately before and after stimulating test targets for the purpose of determining whether the EEG recordings by themselves could predict optimal target selection. We chose linear discriminant analysis because it is a robust classifier commonly reported to out-perform many nonlinear classifiers in brain-computer interface studies (He et al., 2020), especially useful given the relatively small sample size of our study.

Specifically, the EEG connectivity changes that were calculated in ROIs were used as features in the discriminant analysis. Only data that were labeled with optimal target or non-optimal target were used for training classifiers. We then adopted a leave-one-out-cross-validation strategy in the training: one target was selected as the test data while all others served as the training data, with the analysis repeated until all targets were exhausted as test data. The coefficients of a linear classifier were calculated according to the Fisher criteria (Fisher, 1936). The performance of the classifier was evaluated based on the leave-one-out data from all repetitions. Based on classification outcomes of test data, we calculated the percentage of true positive (optimal target classified as optimal target), true negative (non-optimal target as non-optimal target), false positive (non-optimal target as optimal target) and false negative (optimal target as non-optimal target).

Furthermore, in those participants who were not able to perceive any favorable target associated with reduction of symptoms (5 out of 23), their targets were considered to be “undecided” targets in our analyses. Therefore, we applied a trained discriminant classifier from all labeled data to the recordings with undecided targets in order to explore whether the neural signatures in those undecided targets were different from those of favorable targets.

3. Results

3.1. Clinical responses

A total of 24 right-handed participants were recruited for this cTBS study of which 23 had complete data. Sixteen out of the 23 participants were responders showing improvement by more than 10 points after completion of stimulation (VAS changes: mean −37.70 ± 16.30 points, range −88 to −15). The other 7 were non-responders (VAS changes: mean 0.71± 7.12 points, range −8 to 15), as shown in Fig. 2. Comparing the current study to our previous study of rTMS at the DLPFC site (Chen et al., 2019), more participants were responders (i.e. 69.6% in the current protocol vs. 30% in DLPFC protocol) with an overall greater magnitude of improvement (compared to VAS reduction of −32.50 ± 17.56 points in the responders of DLPFC protocol). The location of targets used for repetitive stimulations was not a significant factor in the participants’ response. Therefore, our analyses pooled all participants and focused on the relationship with symptom changes.

Figure 2: Symptom changes from Day 1 to Day 5.

The VAS changes of 16 responders are in blue and 7 non-responders are in red.

Among all participants’ data, the percentage of bad segments ranged from 0 - 5% and was not different between pre and post cTBS sessions (paired t-test, p > 0.1). The percentages of rejected segments did not correlate with symptom changes, either (r = −0.06, p > 0.1).

3.2. Analysis 1: Correlation of EEG connectivity changes with treatment response

Since our prior work had shown that rTMS modulation on EEG visual network and DMN were associated with symptom changes (Chen et al., 2019; Yuan et al., 2017), our current investigation focused on these networks. Through a group-level network analysis and a spatial matching approach with the Yeo template (Yeo et al., 2011), we selected the best matches for the visual network and DMN, respectively. An EEG visual network (Fig. 3) was identified with the highest spatial correlation of 0.61 (p < 0.01) among all EEG networks. The EEG visual network was represented by connectivity in bilateral primary visual cortices but showed the hot spot of highest connectivity value in the right inferior occipital gyrus (IOG), which was designated as the ROI for the EEG visual network. Meanwhile, an EEG DMN (Fig. 4A) was identified by matching to the Yeo template of DMN, with a highest spatial correlation of 0.31 (p < 0.05) among all EEG networks. The EEG DMN network primarily contains the medial frontal gyrus (MFG) and medial prefrontal cortex (MPC), although missing the posterior parts of the DMN template. The strongest connectivity value of the EEG DMN was found in the left MFG, which was designated as the ROI for the EEG DMN. Furthermore, two additional EEG networks were selected since they respectively constitute the left inferior parietal lobule (IPL) and right IPL areas as partial matches to the DMN template (shown in Figures 4B and 4C). Even though neither presented as the highest spatial match to the entire DMN template (left IPL: spatial correlation = 0.05, right IPL: spatial correlation = −0.04, both p > 0.05), they were selected to represent the posterior nodes of DMN. No other EEG networks presented as partial matches to the DMN template. Table 1 lists the coordinates of the center of ROIs in these four selected networks. Although the EEG networks were identified in a spatial matching approach with the Yeo template, the temporal correlation between EEG networks and their respective fMRI networks were confirmed in an analysis of simultaneous EEG and fMRI data (Supplemental Figures 2–3 and Supplemental Table 1).

Figure 3: EEG-derived visual network.

The network was identified by a spatial match approach with the visual network template. In the ROI (circled region in the insert), symptom changes after cTBS protocol were correlated with connectivity changes in the right inferior occipital gyrus (p_correced = 0.04). Post-vs-pre changes of the connectivity in the right inferior occipital gyrus did not differ between the responders and non-responders.

Figure 4: EEG networks associated with the default mode network (DMN).

The network A was identified by a spatial match approach with the DMN template. The networks B and C were identified for connectivity at the left inferior parietal lobule and right inferior parietal lobule. In the ROI (circled regions in inserts), symptom changes after cTBS protocol were correlated with connectivity changes in the right IPL (B2: r = 0.60, p_corrected = 0.007). Post-vs-pre changes of the connectivity in the left medial frontal gyrus is different between the responders and non-responders, but only marginally different in the right IPL. * indicates significance at p < 0.05. Δ Indicates marginal significance at p < 0.1.

Table 1.

Regions of interest in selected EEG resting-state networks

EEG Networks	Regions of Interest	MNI Coordinates (mm)
		X	Y	Z
Visual Network	Right inferior occipital gyrus	36.22	−87.53	−12.03
Default Mode Network	Left medial frontal gyrus	−15.06	42.13	40.53
Other Networks	Left inferior parietal lobule	−33.81	−73.56	41.48
	Right inferior parietal lobule	54.10	−41.85	39.71

We assessed whether the modulatory effects to these EEG networks were associated with individual participant symptom responses on Day 5. Considering that this cohort of participants were primarily positive responders, we first tested the normality of VAS and EEG connectivity changes using a Kolmogorov–Smirnov test, which indicated that the tested values were from a normal distribution. Regarding the visual network, a significant positive correlation was identified between the VAS changes and EEG connectivity changes in the ROI at the right IOG (Fig. 3: r = 0.43, P_corrected = 0.04; after controlling outliers, r = 0.49, p_corrected = 0.03). A positive correlation indicated that greater symptom reduction was associated with a greater decrease in connectivity. Similarly, a significant positive correlation was also found in the right IPL network (Fig 4C: r = 0.60, p_corrected = 0.007). However, in the ROI of the left MFG for the EEG DMN network, the connectivity changes were not significantly correlated to the symptom responses, although a positive trend was noted (Fig. 4A: r = 0.23, p_corrected > 0.1; after controlling outliers, r = 0.51, P_corrected = 0.03). In the left IPL network, the changes were not correlated with the responses (Fig 4B: r = 0.41, p_corrected = 0.075; after controlling outliers, r = −0.03, P_conected > 0.1). No other EEG networks yielded by this data-driven approach showed any significant correlation with symptom changes.

We also assessed within-group and between-group changes of EEG connectivity on Day 5. Regarding the EEG DMN (Fig. 4A), ROI analysis of connectivity changes at the left MFG region showed that the post-stimulation changes for the responders were lower than those for the non-responders (p = 0.02). Within-group assessment further revealed that responders’ decrease of EEG network connectivity in the ROI at the left MFG was marginally significant (p = 0.1), whereas non-responders’ increase was significant (p = 0.003). In addition, regarding the changes in the right IPL network, the responders showed marginal difference (Fig. 4C: p = 0.1). Interestingly, the effect appeared in the opposite direction among the two response types: stimulation decreased the connectivity in responders yet increased connectivity in non-responders, which was also observed in ROIs at left MFG and right IPL. However, the EEG visual network and left IPL networks did not show any significant within-group or between-group differences.

3.3. Analysis 2: Classification of optimal vs. non-optimal targets

We analyzed EEG data on Day 2, when single test sessions of cTBS were delivered at three different targets in a sequential and randomized order. Eighteen of the 23 participants were able to unambiguously choose an optimal target that decreased their symptoms more than the others; the other two targets were deemed as non-optimal targets. Five participants did not find any target effective and all targets in these participants were considered as undecided targets. Three of these participants were randomly assigned to each of the three targets; one worsened with all targets and was treated over the precuneus, and one did not complete the study but had Day 2 data available. On Day 2, after single-administration of cTBS, the VAS changes for the optimal targets were −6.23 ± 1.55 (range −20 to 10), the VAS changes for the non-optimal targets were 0.89 ± 1.83 (range −15 to 40), and VAS changes for the undecided targets were 2.63 ± 1.08 (range −5 to 10).

Our analysis examined whether transient changes could be detected in the EEG networks following single test sessions of cTBS. Specifically, we took a retrospective approach to quantifying the transient EEG network changes on Day 2 data by using ROIs defined in the Day 5 analyses, i.e., the regions that showed modulations that correlated with symptom changes after completion of the protocol. We analyzed each target category (optimal or non-optimal target) for EEG network connectivity immediately after the test stimulation sessions that could be detected relative to the baseline. Additionally, we also compared the transient changes between categories (i.e., optimal vs. non-optimal targets). Figure 5 shows the results in ROIs corresponding to four EEG networks – the right IOG, the left MFG, the left IPL and right IPL. In the ROI at the left MFG (Fig. 5B), the transient EEG changes after single administration of cTBS were significantly different between the optimal and non-optimal targets (p = 0.02), while the decrease for the optimal targets and the increase for the non-optimal targets respectively were marginally significant (p = 0.1). In addition, in the ROI of the left IPL network (Fig. 5C), the transient EEG changes after single session stimulation showed a marginal difference between the optimal and non-optimal targets (p = 0.06). Meanwhile, no significant transient changes were observed in the ROIs at the right IOG or the right IPL.

Figure 5: Transient changes in EEG network connectivity at test sessions.

Connectivity changes immediately after single administration of cTBS were derived from the ROIs at the right inferior occipital (A), gyrus left medial frontal gyrus (B), left inferior parietal lobule (C), and right inferior parietal lobule (D). The activities are grouped by whether the targets were selected as optimal targets, non-optimal targets, or undecided targets. * Indicates significance at p < 0.05. Δ Indicates marginal significance at p < 0.1.

In the other 5 participants, no optimal targets with favorable reduction of symptom were chosen, therefore all their targets were considered as undecided targets. The transient EEG connectivity changes are also plotted in Fig. 5. In the ROI at the left MFG, the undecided targets exhibited a strong increase in connectivity (p = 0.03), that appeared in the same direction as nonoptimal targets and opposite in direction to the optimal target in the participants that had decided targets. Likewise, in the ROI at left IPL, an increase in EEG connectivity was observed after stimulating undecided targets (p = 0.03), which also appeared in the opposite direction as optimal targets.

We explored whether the EEG transient connectivity changes could be used to distinguish optimal targets from non-optimal targets. By using an LDA classifier and a leave-one-out cross validation strategy, the overall accuracy of prediction was 79.3%. As shown in Fig. 6, the classifier yielded 75% accuracy in predicting optimal targets (i.e., TP = 75%), whereas a slightly higher accuracy of 80% was obtained in predicting non-optimal targets (i.e., TN = 80%). The false negative rate (FN = 25%) was slightly higher than the false positive rate (FP = 20%). Importantly, when applying the classifier that was trained by all optimal and non-optimal targets to the data in undecided targets, all targets of the undecided category were automatically classified as non-optimal targets.

Figure 6: Transient changes in EEG networks can predict optimal vs. non-optimal targets.

True positive (TP), false negative (FN), true negative (TN) and false negative (FN) are displayed for the classification results of a linear discriminate classifier with leave-one-out cross validation. Based on the connectivity features, all undecided targets were automatically classified as non-optimal targets.

4. Discussion

In a new cTBS protocol for treating a balance disorder MdDS, we have reconstructed resting state brain networks based on EEG source data and replicated changes in primary visual network and DMN correlating with treatment responses. This dataset of simultaneous EEG and fMRI acquisition allowed us to demonstrate the association between EEG and fMRI networks. The current study builds upon the analysis pathway reported in Chen et al. (2019), in which we investigated resting state brain network connectivity changes after rTMS of DLPFC. Our current analysis has shown that after a complete cTBS protocol, there is a positive relationship between symptom responses and connectivity changes. We further determined that on the Day 2 test sessions of cTBS targets, transient changes of EEG network connectivity induced by single administration of cTBS could be used to classify optimal targets from non-optimal targets in order to predict responses to repetitive stimulation.

Conventional rTMS for major depression disorder began with targeting left DLPFC using 10 Hz stimulation (O’Reardon et al., 2007). In the process of optimizing efficacy, a variety of approaches have been developed to optimize stimulation target identification, with the focus on identifying anatomical landmarks (Fitzgerald et al., 2009; Herbsman et al., 2009; Rusjan et al., 2010). Recently, a new targeting strategy was recommended based on fMRI connectivity with the subgenual cingulate cortex – a key region in antidepressant studies (Fox et al., 2012; Weigand et al., 2018). Importantly, fMRI studies in depression indicate that functional connectivity that resides in large-scale neural circuits composed of several distant cortical, and sometimes subcortical, structures are critical to the effect of rTMS (Drysdale et al., 2017; Fox et al., 2012; Liston et al., 2014; Philip et al., 2018; Weigand et al., 2018). Similarly, results from this current study and our previous work in rTMS study for MdDS (Cha et al., 2012; Chen et al., 2019; Ding et al., 2014; Yuan et al., 2017), indicate that a decrease in the functional connectivity of visual networks and DMN are related to symptom reduction. Importantly, a unique design of our work is that we employed an iterative targeting strategy: first we determined what functional connectivity measurements correlated with symptom improvement after 10-Hz/1-Hz rTMS over DLFPC (Cha et al., 2012; Cha et al., 2016a; Chen et al., 2019; Yuan et al., 2017) and then we adjusted the targets to posterior occipital/cerebellar regions in a subsequent cTBS protocol (Cha et al., 2019). Compared with rTMS over DLPFC, cTBS over the dorsal occipital cortex and cerebellar vermis led to higher response rates. These effects may also be due to better targeting (occipital/cerebellar vs. prefrontal), a higher frequency of stimulation (cTBS vs. 10-Hz/1-Hz), or more intensive treatment (10-12 treatments vs. 5 treatments).

Our current study of multimodal imaging data before and after cTBS revealed signatures of the neuromodulatory effect on functional connectivity. One of the prominent symptoms in MdDS is the development of visually induced dizziness (e.g., watching action movies, playing video games) (Cha, 2009). In the current analysis, one of the areas of functional connectivity correlated with symptom change was within the visual network (right IOG), which is consistent with a previous discovery in a DLPFC neuromodulation protocol (Chen et al., 2019). Comparing these two studies, despite employing two different stimulation protocols and two non-overlapping cohorts of participants, a common anatomical region in the visual network was uncovered. This outcome emphasized the robust role of visual network as a response biomarker in non-invasive brain stimulation treatments for MdDS patients.

In addition, three other EEG networks showed connectivity changes associated with symptom reductions – the EEG networks focused on the left MFG, the left IPL and right IPL. These key regions that have been identified from EEG networks are commonly recognized as the key nodes of the DMN (Buckner et al., 2008). The DMN has been considered to be a fundamental intrinsic network for normal brain function. Its dysfunction has been implicated in a variety of neurological disorders (Broyd et al., 2009). The DMN network is composed of several key nodes in the medial prefrontal cortex, posterior cingulate cortex/precuneus, the intraparietal sulcus, and the hippocampal formation, which includes the entorhinal cortex (Buckner et al., 2008; Greicius et al., 2003). The DMN connection to the hippocampal formation flows through the entorhinal cortex (Ward et al., 2014), which all play critical roles in memory retrieval, navigation, internal mental monitoring and taking on the perspective of others (Buckner et al., 2008; Menon, 2015). In MdDS, abnormal metabolic activity and functional connectivity have been documented in comparison to healthy individuals. The left entorhinal cortex, a pivotal part of the DMN, has been shown to be the single cluster of hypermetabolism in a PET study of MdDS (Cha et al., 2012). Later, our multimodal imaging data of EEG and fMRI showed modulation of the connectivity in the DMN in response to rTMS at DLPFC (Chen et al., 2019; Yuan et al., 2017). In particular, when DLPFC was stimulated, the stimulation did not lead to increased connectivity between the stimulation site and the entorhinal cortex (Yuan et al., 2017), even though resting state functional connectivity between left prefrontal cortex and entorhinal cortex was shown to be lower in MdDS than in healthy controls (Cha et al., 2012). Instead, our data showed that stimulation at DLPFC played a modulatory role, resulting in decreases of functional connectivity between the left entorhinal cortex and multiples nodes of the DMN – right inferior parietal lobule, precuneus and the right entorhinal cortex (Yuan et al., 2017). By using EEG network analysis, the current cTBS study again showed correlations between improvement in symptom severity and connectivity changes in multiple nodes of the DMN – left MFG, left and right IPLs, which are consistent with previous findings despite employing a different stimulation protocol. In both the DLPFC and cTBS protocols, EEG DMN with a hot-spot ROI at the left MFG was identified across non-overlapping cohorts. Connectivity patterns showed significant network-level decrease in positive responders as compared to non-responders.

There are a couple of differences when comparing the findings between the DLPFC 1-Hz/10-Hz and occipital/cerebellar cTBS protocols. Regarding the left MFG, even though a prominent modulatory effect was evident, the cTBS protocol did not reveal a linear correlation with symptom responses, which was otherwise observed in the DLPFC protocol. In addition, only in the cTBS protocol were the left IPL and right IPL networks found to show connectivity to symptom correlations (Figures 4B and 4C: left and right IPL, respectively). Therefore, our findings seem to indicate that cTBS to the occipital/cerebellar regions may exert stronger effects in posterior regions than frontal regions within the DMN.

In addition to the above connectivity modulations observed after the completion of stimulation, our current study for the first time has revealed transient changes of network connectivity after single administration of cTBS, in which multiple targets were tried before an optimal target was selected for repetitive stimulations. Our analysis using an ROI approach demonstrated that regions of symptom-associated modulation at completion of the protocol also exhibited early transient changes at single administration of stimulation. In particular, the ROI at the left MFG showed a significantly greater decrease of connectivity when optimal targets were stimulated compared with non-optimal targets. Likewise, the left IPL exhibited larger decreases with marginal significance for optimal targets, whereas the right IPL showed some but non-significant modulation. By combining the changes in three EEG networks, we have developed an LDA classifier that was built on transient changes as features that classify optimal vs. non-optimal targets. Leave-one-out-cross-validation has demonstrated that the classifier achieved an overall accuracy of 79.3%. More interestingly, when we applied the trained classifier to those undecided targets when self-reports indicated no favorable targets, all of the undecided targets were classified as non-optimal targets. Our results suggest a possible closed-loop paradigm that is based on early transient network changes, such that an optimal target can be determined by a “try-and-choose” approach. In particular, considering that the VAS score was completely a self-report and subjective biases might exist, an imaging-based biomarker could be an objective read-out independent of self-reports, which can facilitate administrators of non-invasive brain stimulation paradigms in deciding optimal targets.

Overall, our investigations have lent support for a feasible closed-loop strategy to optimize the target for cTBS in MdDS. Multimodal imaging data have indicated that connectivity within the DMN, measured by fMRI as well as EEG, can be an imaging-based, objective symptom biomarker for MdDS. Treatments guided by the modulation of this biomarker may be informative in trials of brain stimulation at different targets or other parameters, potentially not limited to transcranial magnetic stimulation. More importantly, although fMRI is capable of revealing symptom-related connectivity, an EEG-based network biomarker is compatible with magnetic stimulation and may provide instantaneous feedback in test sessions of stimulation protocols as in our current study. Furthermore, biomarkers from the two modalities can be integrated to guide treatment targets. For example, fMRI can be used to capture disease-modifying networks involving deep cortical or subcortical structures whereas connectivity involving superficial regions of the network can be identified with EEG, potentially through a matching procedure with the fMRI-measured network in spatial and temporal domains as described in our approach. New stimulation protocols could therefore be designed to promote network modulation in desired directions. Such a “try-and-choose” approach suggested by our data is different from existing closed-loop designs which primarily utilize imaging outcomes, mostly from fMRI (Assaf et al., 2018; Balderston et al., 2020; Li et al., 2017; Luber et al., 2017), as prior knowledge for targeting. Nevertheless, our EEG-network-based, closed-loop strategy could be integrated with other imaging-guided interventions. In individual patients, more than one target may be designed based on pre-stimulation imaging and then through a “try-and-choose” approach, an EEG-network-based biomarker could facilitate the decision in choosing an optimal target for repetitive stimulation.

Several factors do limit the interpretation of this study. First, we did not arrange a true nonstimulation sham arm. Thus, placebo effects were not excluded in the impact on clinical improvement. We used an active sham in the form of the right cerebellar hemisphere stimulation to create the same discomfort level as the other stimulation targets. We used the response as a continuous variable in determining correlation with functional connectivity. Symptom fluctuations are typical in most neurological disorders but with a disorder such as MdDS, which is not related to any structural injury, alterations in functional connectivity itself could be the basis of symptom severity. The treatment response rate was much higher than from the DLFPC stimulation, though both protocols had the same recruitment requirements and duration of study involvement. In both studies, individuals were medically refractory, having failed two medication trials and physical therapy. Placebo effects are generally minor in medically refractory groups (Cha et al., 2016a; Cha et al., 2016b; Li et al., 2018; Liston et al., 2014), but cannot be completely avoided. Second, the sample size was relatively small secondary to MdDS being a relatively rare disorder and recruitment requiring travel to participate. Despite the small sample size, a robust clinical response was noted. However, a larger study may have allowed some analyses to reach statistical significance.

5. Conclusion

Our results confirm the important role of connectivity within the visual network and DMN in MdDS, which may be potential imaging biomarkers of symptom status. Simultaneous measurements have validated the cross-modal association between EEG and fMRI networks. More importantly, the symptom-related networks exhibited an early biomarker effect after single-administration of stimulation at different anatomical targets. Since EEG is of economic cost and easily set up to be recorded at TMS sessions, a closed-loop design with EEG-based functional connectivity measurements may potentially be used to augment treatment protocol choices in concert with subjective assessments of treatment efficacy. These multimodal approaches could potentially guide decisions on the direction of network modulation by non-invasive brain stimulation and eventually enhance clinical outcomes.