Neuroimaging Mechanisms of Therapeutic Transcranial Magnetic Stimulation for Major Depressive Disorder

Abstract

Research into therapeutic transcranial magnetic stimulation (TMS) for major depression has dramatically increased in the last decade. Understanding the mechanism of action of TMS is crucial to improve efficacy and develop the next generation of therapeutic stimulation. Early imaging research provided initial data supportive of widely held assumptions about hypothesized inhibitory or excitatory consequences of stimulation. Early work also indicated that while TMS modulated brain activity under the stimulation site, effects at deeper regions, and in particular the subgenual anterior cingulate cortex, were associated with clinical improvement. Concordant with earlier findings, functional connectivity studies also demonstrated that clinical improvements were related to changes distal, rather than proximal, to the site of stimulation. Moreover, recent work suggests that TMS modulates and potentially normalizes functional relationships between neural networks. An important observation that emerged from this review is that similar patterns of connectivity changes are observed across studies regardless of TMS parameters. Though promising, we stress that these imaging findings must be evaluated cautiously given the widespread reliance on modest sample sizes and little implementation of statistical validation. Additional limitations include use of imaging before and after a course of TMS, which provides little insight into changes that might occur during the weeks of stimulation. Furthermore, as studies to date have focused on depression, it is unclear whether observations are related to mechanisms of action of TMS for depression, or represent broader patterns of functional brain changes associated with clinical improvement.

Introduction

Research into non-invasive brain stimulation is one of the fastest growing areas of psychiatric inquiry. Of these, repetitive transcranial magnetic stimulation (rTMS, hereafter simply TMS) is an important and relatively new treatment. TMS has been clinically available since 2008 when it was cleared by the U.S. Food and Drug Administration for pharmacoresistant major depressive disorder (MDD). TMS uses a pulsed magnetic field to induce neuronal depolarization in a targeted brain region. Since the initial multisite studies (1, 2), a number of groups have published on the efficacy of TMS in naturalistic samples (3), durability of effect (4), and efficacy across the lifespan (5, 6, 7). Furthermore, there is emerging literature supporting the use of TMS for other psychiatric conditions including schizophrenia (8, 9) and posttraumatic stress disorder (PTSD)(10, 11), and outside of psychiatry in areas such as tinnitus, migraine and pain syndromes (e.g., (12, 13, 14)).

While the putative therapeutic mechanism of action of TMS remains unknown, recent neuroimaging studies have set out to discover what is changing in the brain when a depressed patient receives multiple daily sessions of TMS delivered to the prefrontal cortex. This area of research is necessarily complex, requiring an interdisciplinary approach inclusive of expertise from neuroimaging, clinical research, engineering, etc. Given the myriad of approaches to data collection, processing, and analysis involved in human neuroimaging studies, and potential effects of the analytical decision-making process on study observations, a strong grounding in the fundamentals of neuroimaging methods and statistics is needed to appreciate the strength (or lack thereof) of evidence within the TMS/neuroimaging mechanisms literature.

In this review we synthesize findings from the key functional and resting state connectivity studies to identify potential mechanisms of action of TMS for MDD (see supplemental information for search details). To maintain a focus on clinically relevant mechanisms, all studies described below used therapeutic TMS. We considered performing a meta-analysis of these studies, but after reviewing the available literature we concluded that the vast heterogeneity of variables, including treatment parameters (stimulation site, number of sessions, etc.), imaging modalities (metabolic, resting state) and imaging analytic approaches (region of interest versus whole brain analyses), precluded the use of meta-analytic methods. To constrain the breadth of the review, we do not describe studies designed to test or manipulate neurotransmitter levels related to TMS. We acknowledge several studies reported TMS might be associated with increased dopamine release (reported in (15, 16), although not observed in (17, 18) and changes in gamma-aminobutryic acid (GABA) levels (e.g., (19, 20)). Our review also does not include diffusion or morphometry research, although there is a nascent literature suggesting TMS can impact these domains (e.g., (21, 22)).

The review begins with an overview of TMS to provide the reader context for the applications used in imaging studies. We then describe neuroimaging observations using metabolic approaches, followed by the more recent resting state functional connectivity studies. The review ends with an integrative summary of the current data, highlights important design limitations and conceptual assumptions, and suggests directions for future research.

TMS Overview

TMS for MDD starts with motor threshold (MT) determination, which calibrates the stimulator to an individual’s cortical excitability. During MT determination, a clinician delivers single pulse TMS to the motor cortex, and records the amount of stimulator output necessary to induce movement in the contralateral hand in 50% of delivered pulses. Following calibration, a course of TMS is delivered to the prefrontal cortex at 120% of MT on a daily basis for up to 30 (or more) sessions, often followed by a taper phase (for a review of clinical TMS see (23, 24)).

TMS parameters may vary, including stimulation intensity, location, frequency, and duration. These parameters have shifted over time, generally favoring higher stimulation intensity (i.e., increase from 100% (25) to 120% of MT (1), informed by research demonstrating that increased stimulation intensity was required to overcome coil-to-cortex variability associated with age and other factors (26)). Protocols have also evolved to incorporate more TMS sessions (i.e., increase from 10 (25) to >20 (1)), or multiple sessions given in a single day (termed accelerated TMS (e.g., (27))). Regarding location, in earlier studies TMS was delivered to the dorsolateral prefrontal cortex (DLPFC), usually the left DLPFC. This target was initially determined using a so-called “5cm rule,” where the TMS coil was moved 5cm anteriorly along the parasaggital line from the motor cortex. Follow up studies showed the 5cm rule could miss the DLPFC (28). Alternative targeting approaches utilize skull-based landmarks (29) or MRI-based neuronavigation (i.e., using MRI images co-localized with the TMS coil to enable placement over the DLPFC; e.g., (30)), and some evidence indicates that landmark-based techniques and neuronavigation approximate the same location (31). In recent years, various groups demonstrated the efficacy of different stimulation targets, including the dorsomedial prefrontal cortex (DMPFC) (32) and broader prefrontal cortex (33). When considering frequency, TMS pulses are typically considered to be either “high” (≥5Hz) or “low” (≤1Hz), where these frequencies are considered excitatory and inhibitory, respectively. These designations arose from corticospinal excitability studies measuring the size of motor evoked potentials following TMS to motor cortex (reviewed in (34)). This relationship was corroborated by metabolic neuroimaging (positron emission tomography (PET) or single-photon computed emission tomography (SPECT)) that suggested low frequency TMS reduced motor cortex activity, and higher frequency stimulation increased activity (35, 36, 37). Comparable results were observed in the DLPFC using near infrared spectroscopy (38). This apparently bimodal relationship between frequency and activity likely represents an oversimplification of stimulation-related brain changes, as connectivity studies in healthy controls suggest a more complicated relationship (39).

Resting State SPECT/PET Studies of TMS

SPECT/PET imaging studies are listed in Tables 1 & 2. TMS to the DLPFC was initially conceptualized as a way to reverse hypofrontality observed in depression (40, 41) and post-stroke depression (42). In the first TMS imaging study, Teneback et al. (43)(N=22) measured changes in regional cerebral blood flow (rCBF) in MDD patients scanned before and after two weeks of 20Hz or 5Hz TMS to the left DLPFC. They reported that increased inferior frontal lobe activity predicted subsequent TMS clinical response, and that active stimulation was associated with increased blood flow in the prefrontal cortex and limbic/paralimbic regions. Observed changes occurred under the TMS coil and distal to the stimulation site. Speer et al. (44)(N=10) then measured TMS-related rCBF changes before and after 10 sessions of 20Hz or 1Hz TMS to the left DLPFC. 20Hz TMS was associated with increased rCBF under the coil and in the amygdala, insula, hippocampus, parahippocampus, thalamus and cerebellum. 1Hz TMS was associated with distal reductions in rCBF in the right prefrontal cortex, left medial temporal cortex and amygdala. Changes in rCBF correlated with mood changes, and individuals whose mood improved with one frequency worsened with the other. Nahas et al. (45)(N=23; participants shared with (43)) delivered 5 sessions of 20Hz or 5Hz TMS to the DLPFC. They found higher frequency TMS caused greater prefrontal rCBF, relative to lower frequency stimulation, and that significant rCBF increases were observed both under the TMS coil and in distal regions. They also reported greater coil-to-cortex distance was associated with reduced brain activation, confirming observations by Kozel et al. (26). Loo et al. (46)(N=18) found similar results scanning during 15Hz and 1Hz TMS to the left DLPFC, with effects generally observed distal from the site of stimulation.

Table 1.

SPECT/PET TMS Studies: Experimental Design

Study	N per Group	Neuroimaging Methods (Imaging Interval)	Participants	Study Design	TMS Parameters
					Target	Frequency	Intensity (%MT)	Total Sessions a	Total Pulses a
Teneback et al., 1999 (43)	Active 13 Sham 9 Total 22	SPECT (Pre/Post)	MDD (n=14), Bipolar (n=8)	Sham-controlled (parallel)	L DLPFC	5Hz or 20Hz	100	10	16000
Speer et al., 2000 (44)	Total 10	PET (Pre/Post) b	TRD (n=8), Bipolar (n=2)	Sham-controlled (crossover)	L DLPFC	1Hz and 20Hz	100	20	16000 per frequency
Nahas et al., 2001 (45)	5Hz 5 20Hz 9 Sham 9 Total 23	SPECT (Pre/During 5th session)	MDD (n=16), Bipolar (n=7)	Sham-controlled (crossover)	L DLPFC	5Hz and 20Hzc	100	10	16000
Loo et al., 2003 (46)	Sham/1Hz 9 Sham/15Hz 9 Total 18	SPECT (During Sham/Active) d	MDD (n=12), melancholic subtype (n=4); Bipolar (n=2)	Sham-controlled (mixed)	L DLPFC	1Hz or 15Hz	90	1	1Hz: 360 15Hz: 675
Kito et al., 2008 (47)	Total 12	SPECT (Pre/Post)	TRD	Unblinded	R DLPFC	1Hz	100	12	3600
Baeken et al., 2009 (50)	Total 21	PET (Pre/Post)	TRD, melancholic subtype	Unblinded	L DLPFC	10Hz	110	10	15600
Kito et al., 2011 (48)	Total 26	SPECT (Pre/Post)	TRD	Unblinded	R DLPFC	1Hz	100	12	3600
Kito et al., 2012 (49)	Total 26	SPECT (Pre)	TRD	Unblinded	R DLPFC	1Hz	100	12	3600
Baeken et al., 2015 (51)	Active/sham 9 Sham/active 6 Total 15	PET (Pre/Post) b	TRD, melancholic subtype	Sham-controlled (crossover)	L DLPFC	20Hz	110	20 d	31200

Key: SPECT, single-photon emission computed tomography; PET, positron emission tomography; TMS, transcranial magnetic stimulation; MDD, major depressive disorder; MT, motor threshold; TRD, treatment-resistant depression; L, left; R, right; DLPFC, dorsolateral prefrontal cortex.

^aUpper limit of sessions/pulses

^bPatients completed 2 post-TMS scans, one at the completion of each treatment arm

^cParticipants switched to 10Hz at minute 18

^dFirst SPECT scan took place on day one during sham TMS, 2^nd scan took place on day two during active TMS. Scans occurred on consecutive days, 24 hours apart.

^eFive TMS sessions per day (separated by 15–20 minutes)

Table 2.

SPECT/PET TMS Studies: Regions of Interest and Results

Study	Region(s) of Interest	Pretreatment Predictor(s) of Improvement	Post-treatment TMS Effect(s)
Teneback et al., 1999 (43)	DLPFC, cingulate, caudate, anterior temporal poles, inferior frontal cortex, OFC, medial temporal cortex	Higher frontal lobe activity	Increased frontal lobe activity in responders vs. non-responders Reduced correlation between depression severity and limbic/prefrontal activity
Speer et al., 2000 (44)	Whole brain	-	Only 20Hz TMS increased global blood flow 20Hz: Increased prefrontal cortex (L>R), cingulate gyrus (L>R), L amygdala, bilateral insula, basal ganglia, uncus, hippocampus, parahippocampus, thalamus, cerebellum 1Hz: Reduced rCBF in R prefrontal cortex, L basal ganglia, L amygdala No correlation between clinical improvement and rCBF change
Nahas et al., 2001 (45)	L prefrontal cortex (stimulation site), paralimbic cortex	-	Increased rCBF at stimulation site and R medial frontal lobe Decreased ACC and anterior temporal activity 20 Hz with increased activity directly beneath TMS coil Blood flow under the TMS coil declined with increasing distance
Loo et al., 2003 (46)	Frontal cortex, then DLPFC, ACC, basal ganglia, thalamus.	-	15Hz: increased rCBF in inferior frontal cortex (L>R), R DMPFC, R posterior cingulate/parahippocampus; reduced rCBF to R OFC, uncus, R subcallosal gyrus 1Hz: increased rCBF in dACC, L insula, L somatosensory, precuneus, L cerebellum, R insula
Kito et al., 2008 (47)	Whole brain	Higher rCBF in L prefrontal, OFC, sgACC, anterior insula, parahippocampus and amygdala, inferior parietal regions	Bilateral prefrontal (including DLPFC), OFC, anterior insula, R sgACC, L parietal cortex
Baeken et al., 2009 (50)	DLPFC, dorsal and ventral ACC	Higher activity in DLPFC and ACC	Clinical response correlated with increased ACC activity
Kito et al., 2011 (48)	Whole brain	-	Decreased rCBF in prefrontal cortex, OFC, sgACC, basal ganglia, insula. Efficacy correlated with reduced rCBF in R prefrontal cortex, R putamen, R anterior insula.
Kito et al., 2012 (49)	DLPFC, VMPFC	Higher VMPFC rCBF	-
Baeken et al., 2015 (51)	sgACC (as a predictor), whole-brain (follow-up analysis)	Higher sgACC activity	Clinical efficacy associated with reduction in sgACC activity

Key: SPECT, single-photon emission computed tomography; PET, positron emission tomography; TMS, transcranial magnetic stimulation; L, left; R, right; DLPFC, dorsolateral prefrontal cortex; OFC, orbitofrontal cortex; rCBF, regional cerebral blood flow; ACC, anterior cingulate cortex; sgACC, subgenual anterior cingulate cortex; DMPFC, dorsomedial prefrontal cortex; VMPFC, ventromedial prefrontal cortex

Several SPECT/PET studies described predictors of response to TMS alongside effects of stimulation, potentially identifying requisite neural circuits for clinical improvement. Kito et al. (47, 48, 49)(total N=26) found that treatment response to 1Hz TMS was predicted by increased rCBF to the ventromedial prefrontal cortex (VMPFC). Efficacy was associated with reduced rCBF in the prefrontal cortex (including under the TMS coil), orbitofrontal cortex, subcallosal cingulate, putamen and anterior insula. Baeken et al. (50)(N=21) delivered 10 sessions of 10Hz TMS to the left DLPFC, and found that higher baseline activity in the DLPFC and anterior cingulate predicted superior clinical outcomes. Efficacy was associated with increased post-treatment activity in the anterior cingulate, bordering on the subgenual anterior cingulate cortex (sgACC). Recently, Baeken et al. (51)(N=15) delivered 20 sessions of sham-controlled accelerated TMS. They found that higher baseline sgACC activity predicted superior clinical outcomes, and clinical response was associated with reduced sgACC activity.

In summary, metabolic imaging studies found generally consistent effects of TMS. Higher frequency stimulation was associated with increased brain activity, and therapeutic efficacy was associated with changes in brain regions associated with emotion processing or mood regulation. Observed changes often occurred under the coil (DLPFC), in regions with direct anatomical connections (e.g., orbitofrontal cortex, ventromedial prefrontal cortex, basal ganglia), and regions with polysynaptic relationships to the DLPFC (e.g., sgACC and posterior cingulate cortex (52); reviewed in (53)). The sgACC has been implicated in a number of these studies, and most observed reduced sgACC activity following stimulation. The principle limitations of these studies are those associated with early TMS use, including lower stimulation intensity, fewer sessions, and modest sample sizes.

Resting State Functional Connectivity and Neural Networks in TMS

Resting state functional connectivity has been favored by recent studies examining TMS from a neural network perspective. The brain is organized into functional networks (54, 55), and capacity for dynamic network change in response to changing demands (environmental or cognitive) is a hallmark of healthy brain function (e.g., (55)). Network relationships are disrupted in MDD; patients consistently exhibit some degree of default mode network (DMN) dysfunction, typically hyperconnectivity (56), alongside disruptions in the frontoparietal executive control network (ECN) and the attention/limbic or salience network (SN). Taking the DMN as an example, pathological connectivity observed at the group level is typically ascribed to rumination since DMN regions are implicated in introspection in healthy controls (57). It is hypothesized that pathological sgACC activity in depression induces broader DMN dysfunction (57). While hypotheses at the population level may not extend to all patients with MDD, this model-driven approach provides a conceptual framework to examine the interplay between clinical phenotypes and imaging observations.

Two key studies by Fox et al. (58, 59) implicated sgACC-to-DLPFC connectivity in the mechanism of action of TMS, and laid the foundation for future network-related investigations. These studies built upon reports linking MDD treatment response (see Table 1 in (58)) to reduced sgACC hyperactivity (60, 61). In the first study, connectivity relationships between different DLPFC TMS targets (extracted from TMS efficacy studies) and sgACC were evaluated in data from 98 healthy controls and 13 MDD patients (58). Superior clinical outcomes were associated with targets exhibiting the greatest DLPFC-to-sgACC negative connectivity (described as “anticorrelation”). The importance of this result was underscored by their next report (59)(n=98 healthy controls used in (58); n=42 new healthy controls scanned 68±54 days apart; and n=2 MDD patients scanned before and after a course of TMS), where individual differences in DLPFC-to-sgACC connectivity were large and reproducible across imaging sessions. These papers suggested that remote suppression of the sgACC via DLPFC stimulation may be an antidepressant mechanism of TMS, and that connectivity could be utilized to optimize TMS therapy at the individual level.

Prospective Resting State Connectivity TMS Studies

TMS to the DLPFC

Resting state functional connectivity studies are described in Tables 3 & 4. Of these, most delivered TMS to the left DLPFC. Baeken et al. (62)(N=20) acquired resting state images before and after accelerated TMS to evaluate sgACC functional connectivity, though only a subset of imaging data was available (n=12; five responders and seven non-responders). At baseline, future TMS responders displayed greater negative connectivity between the sgACC and superior medial prefrontal cortex, including portions of the DMPFC. After TMS, responders demonstrated reduced negative connectivity between the sgACC and medial prefrontal cortex (MPFC). No sgACC changes were observed in non-responders.

Table 3.

Resting State TMS Studies: Experimental Design

Study	N per Group	Neuroimaging Methods (Imaging Interval)	Participants	Study Design	TMS Parameters
					Target	Frequency	Intensity (%MT)	Total Sessions a	Total Pulses a
Baeken et al., 2014 (62)	Active/sham 10 Sham/active 10 Total 20	RSFC (Pre/Post)b	TRD	Sham-controlled (crossover)	L DLPFC	20Hz	110	20c	31200
Liston et al., 2014 (63)	Total 17	RSFC (Pre/Post)	TRD (n=14), Bipolar II (n=3)	Unblinded	L DLPFC	10Hz	120 d	25	75000
Downar et al., 2014 (72)	Total 47	RSFC (Pre)	TRD (n=38), Bipolar I (n=2), Bipolar II (n=7)	Unblinded	DMPFC	10Hz	120	30	90000
Salomons et al., 2014 (74)	Total 25 h	RSFC (Pre/Post)	TRD (n=22), Bipolar I (n=1), Bipolar II (n=3)	Unblinded	DMPFC	10Hz	120	30	90000
Kang et al., 2016 (68)	Active 13 Sham 11 Total 24	RSFC (Pre/Post)	TRD	Sham-controlled (parallel)	L DLPFC	10Hz	110	10	10000
Baeken et al., 2017 (65)	Active/sham 21 Sham/active 23 Total 44	RSFC (Pre/Post)b	TRD	Sham-controlled (crossover)	L DLPFC	aiTBS	110	20c	32400
Ge et al., 2017 (70)	10Hz 9 iTBS 9 Total 18	RSFC (Pre)	TRD	Active-controlled (parallel)	L DLPFC	10Hz or iTBS	120	30	10Hz: 90000 iTBS: 18000
Avissar et al., 2017 (71)	Total 27	RSFC (Pre/Post)	TRD (n=24), Bipolar II (n=3)	Unblinded	L DLPFC	10Hz	120 e	25	75000
Philip et al., 2017 (67)	Total 33 h	RSFC (Pre/Post)	Comorbid PTSD/TRD	Unblinded	L DLPFC	5Hz	120	40	165000
Taylor et al i	Active 16 Sham 16 Total 32	RSFC (Pre/Post)	TRD	Sham-controlled (parallel)	L DLPFC	10Hz	120	25	75000

Key: TMS, transcranial magnetic stimulation; RSFC, resting state functional connectivity; MT, motor threshold; TRD, treatment-resistant depression; L, left; DLPFC, dorsolateral prefrontal cortex; MDE, major depressive episode; MDD, major depressive disorder; iTBS, intermittent theta-burst stimulation; aiTBS, accelerated iTBS; PTSD, posttraumatic stress disorder; DMPFC, dorsomedial prefrontal cortex

^aUpper limit of sessions/pulses

^bPatients completed 2 post-TMS scans, one at the completion of each treatment arm

^cFive TMS sessions per day (separated by 15–20 minutes)

^dProtocol for 120% of MT, average 86.5%, range 50–109

^eProtocol for 120% of MT, range 80–120 with 2 patients outside range (unreported MT)

^hn=25 with pre/post imaging

ⁱpersonal communication

Table 4.

Resting State TMS Studies: Regions of Interest and Results

Study	Seeds/Region(s) of Interest	Pretreatment Predictor(s) of Improvement	Posttreatment TMS Effect(s)
Baeken et al., 2014 (62)	sgACC	Greater negative connectivity between sgACC and superior medial prefrontal cortex/DMPFC	Reduced negative connectivity between sgACC-to-MPFC No sgACC changes observed in non-responders
Liston et al., 2014 (63)	sgACC DLPFC	Greater sgACC to-DMN and -ECN connectivity	Reduced sgACC-to-DMN connectivity Reduced DLPFC-to-DMN (MPFC) connectivity
Downar et al., 2014 (72)	Graph analysis; 516 atlas-based nodes	Greater reward circuit (ventral tegmental area, striatum, VMPFC), connectivity	-
Salomons et al., 2014 (74)	DMPFC (midcingulate) sgACC	Midcingulate: positive connectivity with sgACC, negative connectivity to thalamus, hippocampus, amygdala sgACC: Higher sgACC-to-DLPFC connectivity; lower sgACC-to-insula, putamen and parahippocampus/amygdala connectivity	Midcingulate: Reduced connectivity to insula and parahippocampus/amygdala sgACC: reduced midcingulate to ventral striatum connectivity; best responders had increased negative connectivity between midcingulate and sgACC
Kang et al., 2016 (68)	Bilateral DLPFC	Reduced DLFPC-to-caudate connectivity	Reduced DLFPC-to-caudate connectivity
Baeken et al., 2017 (65)	sgACC	Greater sgACC-to-OFC connectivity	Reduced connectivity with middle frontal gyrus and motor cortex, and increased with VMPFC
Ge et al., 2017 (70)	ICA-derived resting state networks	Stronger connectivity between DMN and SN regions	-
Avissar et al., 2017 (71)	Frontostriatal connectivity	Higher DLPFC to striatum connectivity	-
Philip et al., 2017 (67) a	sgACC, DLPFC, hippocampus, and amygdala; MVPA	Higher amygdala-to-MPFC connectivity MVPA: dorsal ACC, insula, inferior parietal lobule	Reduced sgACC to DMN connectivity; reduced hippocampus to insula connectivity MVPA: insula, dorsal ACC
Taylor et al b	Affective network (sgACC, amygdala) DMN (PCC) SN (dACC) ECN (DLPFC)	Higher connectivity between the posterior cingulate and insula predicted non-response	Reduced sgACC connectivity in all responders (sham and active), not observed in non-responders

Key: TMS, transcranial magnetic stimulation; sgACC, subgenual anterior cingulate cortex; DMPFC, dorsomedial prefrontal cortex; MPFC, medial prefrontal cortex; DLPFC, dorsolateral prefrontal cortex (left DLPFC unless specified otherwise); DMN, default mode network; ECN, executive control network; VMPFC, ventromedial prefrontal cortex; OFC, orbitofrontal cortex; SN, salience network; MVPA, multivoxel pattern activation; dACC, dorsal anterior cingulate cortex.

^aresults shown as related to MDD

^bpersonal communication

Liston et al. (63)(N=17 patients; 35 controls) delivered 10Hz TMS to the left DLPFC and imaged patients before and after TMS. Seeds were based on (58) and networks were small volume-corrected using the Shirer atlas (64). Compared to controls at baseline, MDD patients exhibited greater within-DMN connectivity (sgACC-to-DMN), reduced ECN connectivity (DLPFC-to-ECN), and disrupted between-network connectivity (reduced DLPFC-to-DMN connectivity and increased sgACC-to-ECN connectivity). These results were broadly consistent with other MDD imaging research (e.g., (56)). After MDD patients received TMS, sgACC-to-DMN connectivity was attenuated; sgACC-to-ECN connectivity changes were not observed. TMS also reduced connectivity between the DLPFC and the MPFC/VMPFC. From these results, the authors posited that TMS acts by reducing sgACC-to-DMN connectivity and inducing negative connectivity between the DLPFC and DMN. They also reported greater baseline sgACC connectivity predicted superior clinical outcomes, consistent with literature reviewed above.

Several important points arose from the study by Liston et al. (36). First, TMS selectively reduced pathological DMN connectivity and reduced ECN-to-DMN connectivity, without significant changes within the ECN. Another important observation was that “excitatory” 10Hz TMS was associated with reduced connectivity, emphasizing that simple assumptions regarding frequency and directionality of downstream effects may not be appropriate for connectivity studies. Importantly, no connectivity changes were associated with clinical improvement, which complicates the interpretation of their findings.

Baeken et al. (65)(N=44 patients, 44 controls) imaged before and after accelerated intermittent theta burst stimulation (aiTBS) as part of a larger inconclusive efficacy study (66). Compared to controls at baseline, patients had greater sgACC connectivity with the DLPFC and precuneus (consistent with (56)). Greater baseline connectivity between the sgACC and orbitofrontal cortex predicted clinical response. After TMS, sgACC connectivity with the middle frontal gyrus and motor cortex was reduced, and increased with the VMPFC. They did not observe TMS-induction of negative connectivity (i.e., increased anticorrelations) between the sgACC and prefrontal regions (e.g., (58, 63)).

We recently imaged participants with comorbid MDD+PTSD before and after 5Hz TMS (67)(N=33). Using a combination of seed-based and data-driven analyses followed by leave-one-out cross validation, we found baseline SN connectivity predicted subsequent clinical response, and MDD symptom reduction was associated with reduced sgACC-to-DMN connectivity. Interestingly, PTSD symptom improvement was associated with TMS-induced negative connectivity between the sgACC and DLPFC, similar to mechanisms proposed for TMS in MDD (58).

Several pilot or unpublished studies are relevant to potential mechanisms of TMS. Kang et al. (68)(N=24) delivered low-dose TMS and reported reduced DLFPC-to-caudate connectivity after treatment, using jackknife procedures (69) to validate their findings. Post-hoc analysis indicated DLPFC-to-caudate connectivity predicted improvement. Ge et al. (70)(N=20 patients, 21 controls) examined biomarkers of response, where patients received iTBS or 10Hz TMS. Responders (regardless of stimulation type) demonstrated stronger baseline DMN-to-SN connectivity. Avissar et al. (71)(N=27 patients, 27 controls; including a subset from (63)) reported higher left DLPFC to striatum connectivity predicted TMS response. Taylor et al. (personal communication)(N=32) delivered 10Hz TMS to the left DLPFC; while they observed no significant differences between active and sham TMS, post-treatment sgACC connectivity was reduced in all responders (i.e., sham and active) but not non-responders. Higher baseline connectivity between the posterior cingulate and insula predicted non-response.

TMS to the DMPFC

Downar et al. (72)(N=47) delivered 10Hz stimulation to the DMPFC, with treatment target based on prior work (32, 73). Non-responders demonstrated lower connectivity in reward pathways (ventral tegmental area, striatum, VMPFC), suggesting intact reward circuit function was necessary for response. Salomons et al. (74) (N=25, a subset of participants from (72)) delivered TMS to the DMPFC and compared connectivity before and after treatment. They seeded the DMPFC (composed of surface and midcingulate seeds from (75)) and a post-hoc defined sgACC; only midcingulate connectivity led to subsequent significant results. Several connectivity patterns predicted clinical response, including positive midcingulate-to-sgACC/MPFC connectivity, and negative midcingulate-to-thalamus, -hippocampus and -amygdala connectivity. Higher sgACC-to-DLPFC connectivity, and lower sgACC-to-insula, -putamen and -parahippocampus/amygdala connectivity also predicted clinical response. When evaluating TMS effects, they observed reduced DMPFC connectivity with the insula and parahippocampus/amygdala. When testing the sgACC seed, TMS reduced connectivity between the DMPFC and ventral striatum, and participants with the best clinical response developed more negative connectivity between the DMPFC and sgACC. These DMPFC studies use a less common treatment target but nevertheless reiterate the central role of the sgACC in mechanisms of action of TMS, and similar to DLPFC studies, they highlight how changes in regions and networks distal from the stimulation site are associated with clinical improvement.

New Approaches to Understanding Mechanisms of Action: Computational Psychiatry

Recently, novel computational approaches brought new insights into TMS mechanisms of action. In a recent study Drysdale et al. (76) used hierarchical clustering and machine learning to describe “biotypes” of depression using previously published neuroimaging data from multiple sites (N=1,188). While this was not formally a TMS treatment study, several findings are relevant to TMS mechanisms. Each of the 4 depression biotypes demonstrated distinct functional connectivity and clinical symptomatology profiles (see Figure 2d & 3c in (76)). Within this study, a subsample (n=124) had imaging prior to TMS to the DMPFC (other relevant findings described in (72) and (74)). A biotype characterized clinically by anhedonia, relatively severe anxiety, early insomnia, and middle insomnia (“biotype 4”) characterized patients who had the most robust clinical response to stimulation. Patients stimulated at other TMS targets were not included, so it is unknown whether this biotype is specifically responsive to DMPFC stimulation. Importantly, the biotypes defined by this analysis were stable over time (n=50, MRI scans 4–6 weeks apart), indicating that imaging changes observed after a course of TMS could be attributable to treatment and not biotype fluctuation. The authors also tested whether biotypes transcended categorical diagnostic boundaries by analyzing a cohort of patients with generalized anxiety disorder (GAD) (n=39) without comorbid MDD. Over two thirds of GAD patients were classified as belonging to a biotype, with nearly 60% assigned to the TMS-responsive biotype. This suggests functional circuits in mechanisms of TMS response might be independent of categorical diagnosis, and has implications for the nascent field of TMS for anxiety disorders (e.g., (77)). Whether conceptually similar results could be obtained when looking at more diagnostically heterogeneous groups (e.g., depressed patients with schizophrenia, substance use disorders, etc.) remains an important question.

Discussion and Future Directions

In the last ten years, research into mechanisms of TMS has rapidly expanded. The most commonly observation is that TMS applied near the cortical surface induces changes distal from the stimulation site and involves multiple neural networks. Interestingly, a similar thread emerges from studies using different stimulation sites, intensities, and frequencies: TMS is consistently associated with changes to the DMN (46, 62, 63, 67, 74) and induces some degree of change in multinetwork relationships. This unexpected consistency is poorly understood, though the common use of resting state scanning – designed to elicit the DMN – is a likely contributor. Distinguishing specific effects of TMS from non-specific changes related to symptom improvement will require more sham-controlled imaging studies. Despite direct stimulation of the ECN, the absence of significant ECN change across studies is also noteworthy. While statistical power considerations can lead to negative imaging results, it is also possible that local changes in metabolic demand induced by TMS complicate BOLD-based assessment of ECN connectivity. Regardless, the most parsimonious explanation for these convergent results is that changes near the cortical surface are less important for therapeutic response than those at distal locations and networks, inclusive of the subgenual cingulate.

Many important questions remain, and to this end, it is important to consider and challenge prior assumptions. Much emphasis has been placed on TMS stimulation site (DLPFC, DMPFC) and approach (high- versus low-frequency stimulation; repetitive, accelerated, and theta burst TMS)(Figure 1), yet there are strikingly few direct comparisons of these methods. Based on the general convergence of findings, it is possible that TMS parameters such as target site, intensity, and frequency either do not matter or are not relevant at the group level; without prospective data it will be impossible examine this issue. Assumptions about the nature and direction of connectivity effects from TMS, particularly as a function of pulse frequency, merit further testing. Summary findings from the metabolic imaging literature comparing higher and lower-frequency stimulation suggest opposing effects, whereas in connectivity studies higher frequency TMS is generally associated with reduced connectivity, an observation supported by a follow up analysis of participants (51) exhibiting increased GABA signal after 20Hz TMS (20).

Figure 1. Clinical Variables in TMS Neuroimaging Research.

Several important variables in TMS neuroimaging research include site of stimulation, stimulation frequency and treatment schedule. Each of these may impact imaging findings, the use of differing approaches complicates interpretation of the current literature.

Abbreviations: DMPFC, dorsomedial prefrontal cortex; DLFPC, dorsolateral prefrontal cortex; TMS, transcranial magnetic stimulation; TBS, theta burst stimulation

Like many emerging fields, sample sizes are modest (mean n=~24), which is likely related to the financial resources required for imaging and TMS. It is important to note that effects of limited power are more substantial when contrasting subgroups (i.e., responders versus non-responders). Beyond increased type II error, using small sample sizes in conjunction with stringent alpha correction negatively impacts predictive value and reproducibility (78, 79). Synthesis of this collective work is also limited by the heterogeneity of connectivity methods (Supplemental Table S1). For example, global signal regression, a preprocessing approach that complicates the interpretation of connectivity results (80, 81), has occasionally been utilized. Stringent motion correction methods (82, 83, 84) have often not been implemented, and many studies used multiple comparisons correction procedures recently associated with false positive inflation (85, 86). Only two reports utilized statistical cross validation (67, 68), a procedure that should be adopted to enhance the rigor of future TMS imaging studies.

Inconsistencies in reporting direction(s) of connectivity effects also complicate interpretation of this area. The use of “anticorrelation” can be confusing and it is unclear whether different groups utilize the term consistently. Furthermore, “positive connectivity” observed after treatment could reflect either an increase in Z scores, a reduction in positive Z scores that remains greater than zero, or a change from negative to positive Z scores. Greater negative connectivity could also be described as an increased correlation. Additionally, when papers describe directionality of effects it is not always clear whether they refer to raw BOLD timeseries relationships or whether covariates were incorporated.

Review of this literature highlights a number of recommendations for future work. One important consideration is selection of study designs that will generate data to fill important gaps in the current knowledge base. The majority of studies reviewed above imaged participants before and after TMS. This design provides information regarding overall changes, yet sheds little light on changes occurring over the course of stimulation (Figure 2). We have little information about how treatment-related changes in connectivity, neuronal activity, or regional metabolism might shift over time, or about the durability of such changes over the longitudinal course of MDD in remission, upon relapse, or in the context of persistent chronic depression. Serial imaging, interleaved MRI/TMS (e.g., (87)) or ambulatory (e.g., EEG) assessment methods capturing data at different treatment intervals in conjunction with behavioral data will be essential for elucidating answers to these questions. Also, because the field has relied heavily on resting state designs that target the DMN, future investigations should also include task-based imaging. For example, one research study suggested brain activation during a planning task might predict TMS response (88). Integration of multimodal neuroimaging measures to assess TMS-related structural changes is also an important and relatively understudied area. Preliminary evidence of increased fractional anisotropy after TMS has been observed (21), and exploratory morphometry studies observed small increases in DMN and salience regional volume after TMS (22), and pointed to baseline hippocampal volume as a possible predictor of TMS response (89).

Figure 2. Current and Future Approaches to TMS Neuroimaging.

The most common approach in TMS neuroimaging research is to scan participants prior to and following TMS procedures. While this approach captures change over time, it does not provide information on what happens during the stimulation itself. Future designs may include a) serial neuroimaging, where multiple scans or other imaging modalities are performed at multiple timepoints during a course of TMS, b) causal assessments of neural network function using interleaved MRI/TMS, and c) testing the durability of neuroimaging findings across the course of depressive illness (e.g., over time or in the context of clinical relapse, etc.).

Abbreviations: TMS, transcranial magnetic stimulation; EEG, electroencephalography; MRI, magnetic resonance imaging

As noninvasive neuromodulation expands to new disease indications and additional imaging data becomes available, it is critical to advance our understanding of the specificity of effects. It is possible the findings summarized above are not unique to MDD and instead represent broader epiphenomena of neural networks moving from a state of disease to a state of wellness. It is also possible we have incorrectly interpreted imaging findings as putative mechanisms of TMS antidepressant action simply because most TMS research to date has examined effects in depressed samples. Testing this hypothesis requires prospective comparisons in different disease groups and across various treatment conditions, e.g., TMS versus, medications, psychotherapy, or other neuromodulatory interventions (e.g., (90)).

In summary, this review reflects a body of work in its infancy, but one that has already revealed important findings on which to rationally build the next steps of research. Ongoing and future methodological innovations hold tremendous potential for shedding important new insights into the putative mechanisms of TMS and advancing our understanding of neural mechanisms of disease more broadly.

Supplementary Material

supplement

Table S1. Data Analyses in Prospective Resting State Imaging Studies of TMS