Magnetic resonance imaging connectivity features associated with response to transcranial magnetic stimulation in major depressive disorder

Transcranial magnetic stimulation (TMS) is an FDA-approved neuromodulation treatment for major depressive disorder (MDD), thought to work by altering dysfunctional brain connectivity pathways, or by indirectly modulating the activity of subcortical brain regions. Clinical response to TMS remains highly variable, high-lighting the need for baseline predictors of response and for understanding brain changes associated with response. This systematic review examined brain connectivity features


Introduction
Major depressive disorder (MDD) affects 13 % of people during their lifetime (Kessler et al., 2015).It is a leading cause of disability worldwide (World Health Organisation, 2017).Medications or psychotherapy are effective for many but at least a third of people do not respond to first-line treatments (National Health Service Digital, 2020;Rush et al., 2006).Minimally invasive neuromodulation therapies represent another treatment option for MDD (Conroy and Holtzheimer, 2021).An increasingly used neuromodulation therapy is repetitive transcranial magnetic stimulation (rTMS), which delivers magnetic pulses to temporarily modulate the excitability of a target cortical area, leading to longer-term neuroplastic changes manifested in altered brain connectivity and changes in the activity of connected deeper, brain areas (Duprat et al., 2022;Klomjai et al., 2015;Oathes et al., 2021;Tik et al., 2023;To et al., 2018).rTMS is performed whilst a patient is awake and alert, and is usually well-tolerated, with transient side effects including discomfort or pain at the stimulation site, headache, and light-headedness (Li et al., 2021).Treatment usually consists of daily stimulation sessions across 4-6 weeks, although "accelerated" approaches delivering up to ten sessions per day across a single week, are gaining in popularity (Cole et al., 2020).
There are two types of rTMS in common clinical use: standard rTMS involving trains of pulses at a specific frequency, and "theta burst stimulation" (TBS), involving bursts of pulse triplets repeating at a slower frequency (5 Hz).TBS can be intermittent ("iTBS", usually two seconds of bursts followed by eight seconds of rest) or continuous ("cTBS").iTBS and cTBS exert opposite effects on immediate electrophysiological measures of cortical excitability (Huang et al., 2005).Whilst iTBS sessions can be much shorter than standard rTMS sessions, they exhibit similar efficacy (Blumberger et al., 2018).The optimal stimulation target, stimulation pattern, and intensity are not known.Most studies aim to stimulate the left dorsolateral prefrontal cortex (DLPFC), a functionally defined area often approximated by the centre of middle frontal gyrus (Dosenbach et al., 2008).Targeting may use scalp-based co-ordinates (for example, the "F3" location of the international 10-20 system (Herwig et al., 2003), or approximations of this), anatomical magnetic resonance imaging (MRI), or potentially functional magnetic resonance imaging (fMRI) (Fitzgerald, 2021;Trapp et al., 2020).
Identifying brain connectivity relationships that are modified by effective TMS (i.e., connectivity relationships that change in association with treatment outcomes) could help suggest new treatment targets.Moreover, some people improve rapidly with TMS, whilst most improve gradually over a period of weeks and then reach a plateau, and some do not respond at all (Kaster et al., 2019).Baseline predictors of response to a given type of TMS could minimise the experience of multiple failed treatment trials, which can contribute to hopelessness (Papakostas et al., 2003), and minimise the period of untreated depression, which is associated with poorer clinical outcomes and greater disability (Ghio et al., 2015).
We conducted a systematic review of studies that have examined baseline connectivity features, or changes in such features, associated with improvement of depressive symptoms following a course of TMS in people with MDD.We chose to focus on brain connectivity measured with magnetic resonance imaging (MRI) techniques, as: (1) MRI is widely clinically available; (2) MRI techniques have been the focus of most of the research literature; and (3) MRI techniques are able to give spatially precise measurements of connectivity changes between specific brain areas.
MRI can measure "structural" (white matter) connectivity (SC) between brain regions.A recent meta-analysis found that MDD was associated with widespread lower SC, particularly evident in the corona radiata, and genu and body of the corpus callosum (Schmaal et al., 2020).MRI can also be used to quantify "functional connectivity" (FC) between brain regions.FC is defined by the correlation of the activity time courses of two regions (greater correlation implying greater communication or co-ordination between regions).Intrinsic FC can be assessed during the task-free (resting) state, or during periods of activity induced by tasks.This popular technique has yielded descriptions of networks of brain areas with separable functions, such as the executive control network ("ECN"), involved in working memory and decision making (the DLPFC is mostly considered part of this network, though it contains portions of other networks) (Dosenbach et al., 2008); the salience network ("SN"), involved in assigning importance to internal and external stimuli (Seeley et al., 2007); and the "default mode network" (DMN), involved in rumination and other internally directed mental activity (Raichle, 2015).MDD is associated with increased FC between the DMN and both the ECN and SN, as well as reduced FC between the SN or ECN and a limbic network (Brandl et al., 2022;Kaiser et al., 2015).It is conceivable that some of these relationships could represent optimal treatment targets for TMS.An increasingly popular "accelerated" protocol, the "SAINT" protocol, which delivers fifty, ten-minute, stimulation sessions within a week, intends to modulate a core component of the limbic network, the subgenual anterior cingulate cortex (sgACC), indirectly via its connectivity with the DLPFC (Cole et al., 2020) (it has not yet been compared to an equivalent protocol without connectivity targeting).Alternatively, our "BRIGhTMIND" trial, comparing standard rTMS to connectivity-guided iTBS, based its DLPFC target on connectivity with the anterior insula (mostly part of the SN) (Morriss et al., 2024).
To the authors' knowledge, there has been one review to date that has explored brain connectivity changes in MDD following TMS (Schiena et al., 2021).This narrative synthesis of thirteen studies indicated FC changes after TMS amongst regions of the sgACC, DMN, SN and ECN, and increases in SC within the frontal lobe.That review only searched one database, included only left DLPFC high frequency rTMS protocols, and predominantly focused on changes in connectivity, rather than baseline predictors of improvement.Our work will extend this, therefore, by examining: 1. Baseline connectivity features associated with clinical improvement following a course of TMS (to any brain area, with any stimulation pattern); 2. Changes in brain connectivity features that are associated with improvement.

Study identification, inclusion, and exclusion criteria
This systematic review and meta-analysis was completed according to the 2020 Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines (Page et al., 2021).The study was registered in the "International Prospective Register of Systematic Reviews" (PROSPERO) in July 2022 (CRD42022346262).
A comprehensive database search was completed using Embase, MEDLINE and PsychINFO, from inception up to 26th May 2022.The search was repeated on 11th April 2023 to identify additional studies published in the intervening period.The search strategy included the following: (functional connectivity/ OR functional connect* OR effective connect* OR fmri connect* OR connectom* OR structural connect* OR dynamic connect*) AND (depress* OR depression/) AND (transcranial magnetic stimulation/ OR repetitive transcranial magnetic stimulation/ OR theta burst stimulation.mpOR theta burst*).Reference lists from relevant studies and reviews were also examined to add further studies meeting the eligibility criteria.We included peerreviewed reports of original research and excluded reviews, metaanalyses, conference proceedings, unpublished theses, case series and non-peer-reviewed articles.Bibliographic data was managed in Zotero software.
Studies were included if they met all the following criteria: (1) Major Depressive Disorder (MDD) without comorbid physical illness diagnosis.As per our pre-specified protocol, we included studies that contained patients with a depressive episode in the context of bipolar disorder, if patients with bipolar disorder made up no more than 20 % of the study's overall sample size; (2) A minimum of 10 sessions of any TMS protocol, delivered with the intent of improving clinical symptoms in a current MDD episode; (3) Structural or functional connectivity measured using an MRI methodology, prior to the first TMS session and optionally following the completion of TMS treatments.No language restrictions were applied.

Study screening and data extraction
Step 1) Each title and abstract were assessed for inclusion by two reviewers (PB, LW, HO, and CB), with initial agreement between reviewers at 93 %.In doubt or if consensus was not agreed at this stage, then abstracts were included for full article review.
Step 2) Full-text articles were split between the four reviewers and assessed for inclusion, independently checked by a second reviewer.Agreement at this stage was 88 %, with disagreements resolved by discussion between the reviewers.
Step 3) All four reviewers were involved in the data extraction process, with one reviewer extracting data from each of the studies.20 % of studies also had data extracted by a second reviewer to assess for agreement, with no significant discrepancies noted.Studies from the rerun search on 11th April 2023 were assessed for inclusion by PB and LW.
Details for each included study were inputted into a data extraction sheet consisting of sample characteristics, study design, TMS protocol, location of TMS target, neuroimaging details, measures of clinical response, relationships between baseline connectivity or change in connectivity and clinical response, limitations of studies, other potentially relevant studies found in the reference list, funding source and potential conflicts of interest.Related, previously published, articles were referred to where necessary to obtain this information.We sought to identify sample overlap between studies and indicate such overlap in the Tables.

Assessment of study quality
Quality assessments were completed by PB or LW.Imaging quality for all studies was assessed with a modified version of the 13-item tool used by Xu et al. (2021).Xu et al.'s item 8 ("Parcellation template clearly reported, reproducible") was modified to: "Parcellation template or regions of interest clearly reported, reproducible".Xu et al.'s item 9 ("Calculation of edge weights are clearly reported and are reproducible") was modified to: "Method for calculating edge weights, functional connectivity, or structural connectivity clearly reported and reproducible".General study quality was assessed by the NIH National Heart, Lung, and Blood Institute (NHLBI) quality assessment for controlled intervention studies for those following a randomised controlled trial (RCT) protocol, with all other studies assessed by the NIH NHLBI quality assessment tool for Before-After (Pre-Post) studies with no control group (https://www.nhlbi.nih.gov/health-topics/study-quality-assessment-tools).Where a criterion was not relevant to a specific study, it was scored as "low risk".Where a criterion could not be determined from available information, it was scored as "high risk".Criteria that were only partially met were scored as "high risk".Initially, a study was considered "good" quality if at least 80 % of criteria were met, and "fair" quality if at least 50 % of criteria were marked as low risk.However, studies with a sample size insufficient to detect a moderate effect with 80 % power (for example, minimum N = 82 for correlation analyses, determined using G*Power 3.1.9.7) could at most be considered as "fair" quality in the general study/intervention quality assessment.

Strategy for data analysis and synthesis
For reporting in the Tables, we extracted baseline connectivity features, or changes in connectivity features, that showed significant correlations with improvement on a measure of depressive symptoms or were significantly different between a group of responders/nonresponders, remitters/non-remitters, or improvers/non-improvers.
We extracted information on network assignment of regions involved in each connectivity feature where this was given.Where standard MNI co-ordinates were given, we assigned regions to networks using the nearest node in the widely used Power et al. (2011) atlas.To guide our synthesis of findings, we then considered the extracted relationships in terms of relationships between these networks (sgACC, amygdala, hippocampus, and the striatum remained as separate regions).
For baseline connectivity features, we next considered connectivity relationships between networks (or sgACC etc.) reported in at least two studies (i.e., relationships that have been studied at least twice), regardless of the effect direction in those studies.From this we produced Fig. 2, which was used to guide our narrative synthesis in the Results.A similar approach for change in connectivity features was trialled but there were few relationships that appeared in at least two studies.Relationships regarding change in connectivity features are discussed alongside baseline relationships involving the same networks.We also include discussion of key structural connectivity studies, and all studies that used a target within the DMPFC, so that these important, underresearched, avenues are given attention.
As per our study protocol, we did attempt Co-ordinate Based Random Effects Size (CBRES) meta-analyses of the baseline FC findings of studies that targeted the left DLPFC, using the ClusterZ algorithm implemented in the NeuRoi software (Tench et al., 2017).For each seed region/network (i.e., sgACC, ECN, DMN, SN), a list of studies that found a significant relationship involving that region/network was given, alongside co-ordinates of the connected regions and the effect directions and sizes.Due to heterogeneity in reported relationships, these analyses did not identify significant clusters.

Overview of included studies
Following the literature search and review of abstracts, 117 full-text articles were assessed for eligibility.Forty-one studies met inclusion criteria (Fig. 1, exclusion reasons in Table S4).These included a total of 1097 patients with a current major depressive episode, excluding likely     sample overlap.Fifteen of the patients (1 %) had a history of bipolar disorder.Most studies quantified clinical improvement using the Hamilton Depression Rating Scale (HDRS-17, N = 27), followed by the Montgomery-Åsberg Depression Rating Scale (MADRS, N = 8).Twenty-eight studies delivered 10 Hz rTMS and fifteen delivered iTBS (these numbers include five studies that delivered either 10 Hz rTMS or iTBS, and one study that delivered iTBS and cTBS), two studies delivered 20 Hz, one 5 Hz rTMS.In thirty-seven studies, TMS targeted the left DLPFC.The remaining four studies targeted bilateral DMPFC.Of the thirty-seven DLPFC studies, 33 reported functional connectivity ("FC") measured using resting-state fMRI (rsfMRI) or, in one case, arterial spin labelling (Table 1), the remainder reported white-matter structural connectivity (SC) derived from diffusion tensor imaging MRI (N = 4, Table 2).The four DMPFC studies all reported FC measured with rsfMRI (Table 3).Due to the paucity of studies examining SC, the focus of the remainder of this report is on FC.

Quality of included studies
Using the imaging quality assessment tool, adapted from Xu et al. (2021), 33 studies were graded as good quality (80 % of criteria met) and 8 as fair quality (50 % of criteria met).Twenty-three studies did not meet the resolution and motion criterion (at least 3.5 mm 3 voxels / 12 DTI directions and detailed motion thresholds), fifteen studies did not explore the impact of any potential confounding variables, nine studies did not correct for multiple comparisons where required.Using the intervention quality assessment tools (adapted from the NHLBI tools), for the nineteen studies that used an RCT approach, all were graded as fair quality (five were downgraded from "good" due to insufficient sample size to detect a medium effect).None reported a pre-specified analysis protocol.For the twenty-three studies that did not use an RCT approach, one was regarded as good quality, the remainder were regarded as fair quality (eight were downgraded from good due to insufficient sample size).One study incorporated multiple follow-up time points; one had an adequate sample size.Further details are given in the Supplement.It should be noted that the choice of using the RCT or non-RCT quality assessment tool was based on the nature of the underlying dataset used by a given study.Many studies did not analyse this data in such a way as to distinguish sham versus active effects.This is indicated where the study is discussed below.

Functional connectivity studies targeting left DLPFC
Fig. 2 illustrates relationships between baseline functional connectivity and clinical improvement.It includes connectivities, within or between networks (or sgACC), that were reported in at least two studies.These connectivities are the focus of this section.An equivalent figure was not possible for relationships between change in connectivity and clinical improvement due to a lack of studies reporting the same connectivity relationships.Findings for that analysis are included below where relevant.2013), 1-4 previous antidepressant trials).The opposite pattern was found for rostral ACC.These relationships were present immediately, as well as three months after the end of treatment.Likewise, Rosen et al. ( 2021) (1.23) found that the mean target location for TMS responders (>50 % improvement) on the HDRS-17 was in left ECN and showed negative baseline FC with sgACC (N = 23, severe depression, 2+ antidepressant trials).They delivered stimulation 6 cm anterior to the motor cortex hand area and obtained MRI with a marker at the target point, which was then projected onto the cortical surface to determine the targeted brain co-ordinates.The mean target for non-responders was in DMN and did not show significant sgACC FC (they used a sham-controlled design, finding that the target location was significantly different between responders and non-responders to active stimulation, but did not differ between "responders" and "non-responders" to sham).Cash et al. (2019a) (1.10) found that greater negative FC between sgACC and the left DLPFC target was associated with improvement on the MADRS (N = 47, moderate-severe depression, 2+ antidepressant trials).Their average target location was within the ECN.Cash et al. (2021) (1.20) also showed that closer proximity of the stimulation target to an "ideal" target, defined as the point that showed greatest baseline anticorrelation with the sgACC, was predictive of response, a finding also shown by Kong et al. (2022) (1.27) using a similar approach (N = 18, moderate), and by Stöhrmann et al. (2023) (1.32), using a bilateral stimulation protocol with iTBS to left DLPFC and cTBS to right DLPFC (N = 15 active treatment, moderate depression, 2+ antidepressant trials).Stöhrmann et al. used a sham-controlled design but did not compare baseline predictors between active and sham (N = 5) groups.

Connectivity between the sgACC and ECN
In contrast, in the second-largest included study, Hopman et al. (2021) (1.22) did not find significant baseline FC differences between short-term (immediately after twenty sessions) responders and non-responders (on the MADRS) to once-daily 10 Hz rTMS but did find differences when examining response two months post-treatment (N = 63, moderate depression, 1-4 previous antidepressant trials).Longer-term responders showed greater baseline FC between sgACC and the left DLPFC target as well as between sgACC and frontal/parietal ECN.
Hopman et al. speculate on differences in participant ethnicity as the cause of their discrepant findings (Hong-Kong Chinese sample in their study versus primarily Caucasian samples).A recent study by Elbau et al. (2023) (1.31), and the largest included in this review (N = 295, moderate-severe, 80 % taking antidepressants, 1+ antidepressant trial), used data from the THREE-D clinical trial (conducted in Canada) to address sources of variability in relationships between baseline target-sgACC FC and clinical improvement.This confirmed an association between clinical improvement (measured with Quick Inventory of Depressive Symptomatology, although consistent results when HDRS-17 used) and more negative baseline FC between the sgACC and left DLPFC target (mean location in ECN).However, the effect was weaker than in the prior, smaller sample studies.They showed that substantial between-study variability in the size (and direction) of relationships between baseline FC and clinical improvement would be expected if small sample sizes were used.In their full sample, they found a significant relationship only when the sgACC seed was individualised for each patient based on anatomically informed modelling of the distribution of current from the TMS coil, and only when the overall, "global", brain signal was regressed out of the data (a commonly used approach for studying negative FC relationships, but an approach that creates challenges for interpretation (Murphy and Fox, 2017)).They further showed that the relationship was strongest in those patients with greatest fluctuation in this global brain signal, and specifically in those with signal fluctuations consistent with a "burst" breathing pattern (Lynch et al., 2020).They speculate that this finding reflects either: there is a sub-group of patients with a tendency to burst breathing patterns and for whom baseline sgACC FC strongly determines outcomes with rTMS; or that burst breathing occurs at the time of certain time-lagged, high-amplitude, fMRI signal events that may make negative FC relationships more apparent.

Baseline predictors of symptom improvement.
In the only study using arterial spin labelling (ASL) data, Wu et al. (2022) (1.29) derived a measure of baseline connectivity (covariance of cerebral blood flow) between sgACC and left DLPFC targets, which were primarily in the DMN.They found this connectivity to be predictive of HDRS-17 improvement (N = 41, moderate, no current psychotropics apart from regular benzodiazepines, 1+ antidepressant trial).They used a sham-controlled crossover design, and distinguished responders to active versus sham stimulation in the analysis.Their methodology did not allow them to distinguish positive from negative connectivity.
Returning to fMRI data, Liston et al. (2014) (1.2), with an open-label design, found more positive baseline FC between sgACC and DMN regions was associated with greater improvement to 25 sessions of once-daily 10 Hz rTMS (N = 17, three depression in context of bipolar disorder, severe, 2/3 taking antidepressants/mood stabilisers, 2+ antidepressant trials).Baeken et al. (2014) (1.1) reported conflicting results using 20 Hz rTMS in an accelerated protocol (four days of five sessions/day).They found more negative baseline FC between sgACC and an anterior DMN (aDMN) or an SN region of superior frontal gyrus in responders than non-responders, and responders showed greater increase in sgACC-aDMN FC post-stimulation (N = 20, severe, taking only regular benzodiazepines, 3+ antidepressant trials).Baeken et al. used a sham-controlled crossover design, but their FC analyses collapsed across both active and sham time points.The opposite direction of their effect may be due to differences in rTMS frequency, medication status (most taking antidepressants in the former studies versus none in Baeken et al.), or number of sessions per day (once daily versus an accelerated protocol).

Associations between change in connectivity and symptom
improvement.Philip et al. (2018) (1.8), in an open-label design, used an average of 33 once-daily sessions of 5 Hz rTMS and examined people with MDD and co-morbid post-traumatic stress disorder (PTSD).They found that reductions in sgACC FC with the precuneus/posterior cingulate (posterior DMN), dACC (SN) and sensorimotor network were associated with improvement on the Inventory of Depressive Symptomatology (N = 33, severe, 2/3 taking antidepressants, 1+ prior antidepressant trial).DMN and ECN 3.3.3.1.Baseline predictors of symptom improvement.Four studies used once-daily 10 Hz rTMS, in people with MDD, the majority of whom were taking antidepressants.Two open-label studies examining patients with moderate-severity MDD on average, reported consistent results.Ge et al. (2020) (1.17, N = 50, moderate severity, 1-4 antidepressant trials), reported above, found that greater baseline FC between rostral ACC (DMN) and inferior parietal lobule (ECN) was associated with greater improvement on the HDRS-17 following 20-30 sessions.Hopman et al. (2021) (1.22,N = 63, moderate, 1-4 antidepressant trials), reported above, found that MADRS responders at two months following twenty sessions showed more positive baseline FC between the left DLPFC target (mean co-ordinates in ECN) and a posterior DMN region.

Connectivity between the
One study examining patients with, on average, severe MDD, reported opposing results.Rosen et al. (2021) (1.23, N = 23, severe, 2+ antidepressant trials), reported above, found that TMS responders showed more negative baseline FC between their (on average, ECN) left DLPFC target and clusters of the anterior and posterior DMN (but more positive FC between the target and SN, ECN, and a sensorimotor network).They used a sham-controlled study design, although this analysis did not contrast effects with those of sham stimulation.

Associations between change in connectivity and symptom
improvement.Moreno-Ortega et al. (2020a) (1.18), in an open-label study, found that HDRS-17 improvement following 36 sessions targeted to an ECN region of left DLPFC was associated with increase in FC between the target and a DMN parcel, in non-responders to previous rTMS (N = 10, severe, 1+ antidepressant trial).P.M. Briley et al. 3.3.4. Connectivity between the SN and ECN 3.3.4.1.Baseline predictors of symptom improvement.Consistent results were obtained in open-label studies that reported FC between the SN and ECN and used either 10 Hz rTMS or iTBS.Fu et al. (2021) (1.21) found that greater improvement on the HDRS-17 was associated with more positive FC at baseline between anterior insula (SN) and an ECN left DLPFC region (r s = 0.66, N = 27, moderate, no current antidepressants and minimal treatment history).Of note, Fu et al. also found that greater baseline white-matter structural connectivity (quantified using "fractional anisotropy") between these regions was also associated with greater improvement (r s = 0.46).Rosen et al. (2021) (1.23), reported above, also found that more positive baseline FC between the left DLPFC target and clusters of the SN was associated with greater HDRS-17 improvement (N = 23, severe, 2+ antidepressant trials; used a sham-controlled study design but this analysis did not contrast effects with those of sham stimulation.).Iwabuchi et al. ( 2019) (1.13) found that greater "net outflow" from rAI to left DLPFC at baseline was associated with greater improvement on the HDRS-17 following sixteen once-daily sessions of iTBS or 10 Hz rTMS (N = 27, moderate, 75 % taking antidepressants, 1+ prior antidepressant trials, participants were randomised to either stimulation type, but there was no sham control).The left DLPFC target was identified from within the ECN.Net outflow was defined as directed, "effective", connectivity (EC), from right AI to left DLPFC minus EC from left DLPFC to right AI. Unlike FC, which is calculated from the correlation of the activity time courses between brain regions, EC refers to the directed influence of one region on another, calculated using Granger Causality Analysis.

Connectivity within the SN
3.3.5.1.Baseline predictors of symptom improvement.Two studies that used once-daily 10 Hz rTMS, in people with MDD, most of whom were taking antidepressants, obtained consistent results regarding FC within the SN.Ge et al. (2017) (1.7), in an open-label design, found that responders on the HDRS-17 after twenty sessions had shown higher baseline FC within the SN (anterior insula, dorsal anterior cingulate) (N = 18, moderate, 1-3 antidepressant trials).Fan et al. (2019) (1.12) found that improvement on the MADRS following twenty sessions was associated with greater functional segregation of the SN at baselinethat is, greater within-network FC and lower FC between the SN and all other networks (N = 32, mild, 1+ antidepressant trial).Fan et al. used a sham-controlled design but did not identify significant differential predictors of response to sham or active treatment.
In contrast, Iwabuchi et al. (2019) (1.13), reported above, found that improvement following sixteen sessions of 10 Hz rTMS or iTBS (targeted at the voxel within ECN left DLPFC showing greatest baseline effective connectivity from right anterior insula), was associated with lower baseline within-network connectivity of the salience network (N = 27, moderate, 1+ antidepressant trial).This study differs in TMS type (10 Hz rTMS versus iTBS) and targeting method (pre-selected MNI co-ordinates versus connectivity-based targeting).

Associations between change in connectivity and symptom
improvement.Godfrey et al. (2022) (1.26) found that improvement on the MADRS after twenty sessions of open-label 10 Hz rTMS was associated with reductions in FC within the SN (N = 26, moderate, 2+ antidepressant trials).In contrast, in people with puerperal-onset MDD, Zhang et al. (2022) (1.30) found that improvement on the Edinburgh Postnatal Depression Scale (EPDS) was associated with increase in FC between left and right insula (SN), following open-label iTBS delivered in an accelerated protocol (ten sessions/day, five consecutive days, targeting based on sgACC connectivity) (N = 31, fifty sessions, severe, not taking antidepressants).
3.3.6.Connectivity within the DMN 3.3.6.1.Baseline predictors of symptom improvement.Mixed results have been obtained regarding baseline within-network DMN connectivity, in people with MDD, most of whom were taking antidepressants.Using the same data as Cash et al. (2019aCash et al. ( ), (2019b) ) (1.11) found that lower baseline FC within a DMN network (and within an "affective" network) was associated with greater improvement on MADRS following 20 once-daily sessions of open-label 10 Hz rTMS (N = 47, moderate-severe, 2+ antidepressant trials).Similarly, Taylor et al. (2018) (1.9) found that baseline FC between posterior (posterior cingulate cortex) and anterior (inferior frontal gyrus) DMN was lower for HDRS-17 responders than non-responders to twenty once-daily sessions of 10 Hz rTMS (N = 32, mild, 1+ antidepressant trial).Responders also showed lower baseline FC between posterior DMN and SN (right AI).Although Taylor et al. used a sham-controlled cross-over design, their baseline analysis collapsed across active and sham phases.Ge et al. (2017) (1.7), discussed above, produced opposing results, finding that HDRS-17 responders showed higher within-DMN FC at baseline than non-responders (N = 18, moderate, 1-3 antidepressant trials, open-label)..3.6.2.Associations between change in connectivity and symptom improvement.Tang et al. (2021) (1.24) found that response to open-label iTBS delivered in an accelerated protocol (ten sessions per day, five consecutive), was associated with pre-to post-stimulation increase in FC within the DMN (as well as decrease in FC between the ECN and SMN) (N = 15, severe, all with suicidal ideation).In contrast, Moreno-Ortega et al. (2020a) (1.18), discussed above, found that HDRS-17 improvement in a group of TMS-naïve participants, was associated with FC decrease between the (on average, DMN) left DLPFC target and a DMN atlas parcel following treatment (N = 22, severe, 1+ antidepressant trial, open-label).pre-frontal cortex, whose baseline connectivity distinguished responders and non-responders in a graph theoretic analysis (N = 47, including 9 bipolar, moderate, 1+ antidepressant trial).There appeared to be an effect of laterality in their findings.Responders on the HDRS-17 showed more positive baseline FC between VMPFC and ECN (DMN-ECN, consistent with 1.17 and 1.22 in the studies that stimulated left DLPFC), but also VMPFC and SN, and primarily left-sided DMN areas.Whereas responders showed more negative baseline FC between VMPFC and other SN regions, as well as primarily right-sided DMN areas.

Connectivity studies targeting DMPFC
Salomons et al. (N = 25, including 4 bipolar, 2+ antidepressant trials) examined FC with a DMPFC seed region, finding that more positive baseline FC between the DMPFC and a cluster incorporating the sgACC and VMPFC was associated with improvement on the HDRS-17.Reduction in FC between DMPFC and sgACC from pre-to posttreatment was also associated with improvement.Interestingly, more positive baseline FC between sgACC and ECN/SN regions of the DLPFC was associated with improvement (opposite to most studies that targeted left DLPFC).
In two sham-controlled studies with overlapping samples, Persson et al. (2020) (3.3) andStruckmann et al. (2022) (3.4) studied functional connectivity changes following ten days of twice-daily iTBS to DMPFC bilaterally.Persson et al. found that improvement on the MADRS was associated with more negative baseline FC between sgACC and precuneus (posterior DMN), as well as between precuneus and the DMPFC target (N = 23, including 2 bipolar, moderate, approx.90 % taking antidepressants).Increase in FC between precuneus and the target from pre-to post-stimulation was associated with improvement, an effect present only following active TMS.Finally, Struckmann et al. (N = 34,including 3 bipolar) found that greater improvement on the MADRS was associated with greater reduction in FC between an SN left DLPFC area and a sensorimotor left insula area post-stimulation.Again, this effect was present only following active TMS.

Discussion
Most included studies stimulated left DLPFC with a 10 Hz rTMS or iTBS protocol, and measured improvement using the HDRS-17.Nevertheless, there was considerable heterogeneity in the findings.No consistent relationships were identified in the attempted meta-analysis.This may have been due to: in some cases, a focus on specific brain regions; low statistical power (most studies included between 17 and 63 participants); methodological and sample differences.The absence of pre-published analysis protocols also raises concerns around false positive findings.Connectivity relationships involving the sgACC featured prominently and are considered first below.

Connectivity of the sgACC
There is evidence from positron emission tomography (PET) studies that depression is associated with elevated metabolic activity (increased cerebral blood flow and glucose metabolism) within the sgACC and that effective antidepressant treatment is associated with reductions in this metabolic activity (Drevets et al., 2008).More recently, Argyelan et al. (2016) studied a measure of brain activity derived from rsfMRI -"fractional amplitude of low frequency fluctuations" (fALFF) -which has been shown to be correlated with PET-measured glucose metabolism (yet has higher spatial resolution) (Aiello et al., 2015).They found greater fALFF within the sgACC in people with depression than healthy controls at baseline, which normalised across a course of electroconvulsive therapy treatment.
Due to the PET (and later fALFF) findings, the demonstration by Fox et al. (2012) that TMS efficacy is related to more negative baseline FC between the sgACC and the stimulated region of left DLPFC within the ECN, suggested a plausible mechanism of the effects of TMS in MDD.Under this rationale (and assuming that negative baseline FC is underlain by negative ECdirected influencefrom left DLPFC to sgACC), excitation of the superficial (and, thus, TMS accessible) left DLPFC target could exert suppressive effects on the deeper sgACC (which is inaccessible to TMS), normalising its activity.The relationship was initially demonstrated using connectivity calculated from separate, large-scale, connectome data, rather than individual-level connectivity.Such studies do not meet inclusion criteria for this review.Later work showed the relationship using individual-level connectivity data, as included in the review (Cash et al., 2021(Cash et al., , 2019a;;Elbau et al., 2023;Ge et al., 2020;Rosen et al., 2021), but the relationship is not reported in all studies that use a whole brain approach, and there is opposing data (Hopman et al., 2021).
There has been complexity in interpreting the findings around metabolic activity of the sgACC.There are reports of reduced metabolic activity in MDD measured with PET, which have been attributed to inadvertently measuring from surrounding tissue areas due to a reduction in sgACC size in MDD (a so-called "partial volume" artefact) (Drevets et al., 2008).This could remain an issue in fMRI analyses using fixed-volume regions-of-interest across participants.In addition, the sgACC lies close to an air-filled sinus, and air-tissue interfaces create artefacts in MRI.Further, there is ongoing debate about the meaning of anti-correlations in FC analyses (Murphy and Fox, 2017).This is because they primarily (although not exclusively (Chai et al., 2012;Chang and Glover, 2009)) arise when removal of the overall, "global", brain signal is employed as a pre-processing step (Murphy et al., 2009).This step serves to reduce the influence of widespread non-neuronal signals (such as changes in arterial CO 2 concentration).Mathematically, the step enforces zero-centring of connectivity values across the brain.Modelling studies indicate this can introduce spurious anti-correlations (Saad et al., 2012), yet anti-correlated BOLD signals were, for the most part, associated with anti-correlated neuronal signals when measured from the same people in a study using recordings from subdural electrodes that had been implanted for seizure monitoring (Keller et al., 2013).
The most recent (and largest) included study, from Elbau et al. (2023), confirmed the relationship between clinical improvement and more negative baseline FC between sgACC and the left DLPFC target, but found this was weaker than reported in studies with smaller sample sizes.Strong relationships were observed in patients showing substantial variation in overall, global, brain signal, which was associated with a specific, burst, breathing pattern.They speculate that: either the global signal variations and burst breathing patterns are associated with fMRI signal (blood oxygenation level dependant, "BOLD") events that facilitate the measurement of negative (anti-correlated) functional connectivity relationships, or that there is a subset of patients, who have this specific breathing pattern, for whom sgACC FC is most critical to rTMS efficacy.Further work is clearly needed to understand the implications of their findings.
Despite the underlying rationale of modulating the sgACC via stimulating left DLPFC (a motivation for the targeting approach in the increasingly popular SAINT protocol (Cole et al., 2020)), there has been limited demonstration that the sgACC is effectively modulated by the approach.Recent TMS/fMRI studies have examined immediate changes in the BOLD response from sgACC following single pulses of left pre-frontal stimulation.In two studies, suppression of sgACC activity was observed, which, in people with MDD, was greater when targeting cortical regions with more positive baseline FC with sgACC (Duprat et al., 2022;Oathes et al., 2021).In a third study, targeting the DLPFC location with greatest negative baseline FC with the sgACC led to increases in sgACC activity (Tik et al., 2023).Given that suppression of sgACC over-activity is desired, these findings from single-pulse studies stand in contrast to findings that rTMS to the DLPFC location with greatest negative baseline FC with the sgACC is associated with greatest clinical improvement.One explanation could be that repeated induced excitation of sgACC via DLPFC may, over the course of an rTMS session or over multiple rTMS sessions, lead to longer-term sgACC suppression.Alternatively, multiple sessions of rTMS may first act by increasing connectivity between DLPFC and sgACC, before then suppressing the activity of the sgACC itself.There was some evidence for an association between clinical improvement and modulations of sgACC connectivity in the studies identified in this review (Ge et al., 2020;Philip et al., 2018).
Interestingly, one included study that targeted a DMPFC region within the DMN found that more positive baseline FC between the target (which was within the DMN) and the sgACC was associated with greater response (Salomons et al., 2014).In studies using left DLPFC targets, better response was associated with more positive baseline FC between sgACC and DMN regions in a non-accelerated paradigm (Liston et al., 2014), but more negative baseline FC in an accelerated paradigm (Baeken et al., 2014).Understanding the dynamics of sgACC modulation within and across treatment sessions, and how this relates to treatment parameters and baseline FC between the sgACC and the targeted region, or between the sgACC and other ECN/DMN regions, would appear a valuable next step in understanding the variability in findings and in understanding how to optimise treatment approaches.

Other connectivity relationships
The recently completed BRIGhTMIND trial of connectivity-guided iTBS versus standard rTMS for treatment-resistant MDD (Morriss et al., 2024) used connectivity involving a region of the anterior insula, based on preceding pilot work (Iwabuchi et al., 2019(Iwabuchi et al., , 2017(Iwabuchi et al., , 2014) ) and meta-analytic evidence that the activity of the anterior insula, which resides within the SN, could distinguish responders to different treatment options for depression (McGrath et al., 2013).There was some evidence in this review that greater baseline FC between the SN and ECN was associated with greater response to (ECN) left DLPFC targeted treatment (Fu et al., 2021;Rosen et al., 2021) (also found for structural connectivity (Fu et al., 2021)), and that greater baseline FC between the SN and DMN was associated with greater response to (DMN) DMPFC targeted treatment (Downar et al., 2014).The findings of our BRIGhT-MIND trial (analysed and published after the inclusion cutoff date for this review) suggest that the balance of influence (difference in effective connectivity), at baseline, between the SN (anterior insula) and left DLPFC target may predict treatment outcomes (Morriss et al., 2024).More positive effective connectivity from the left DLPFC target to the anterior insula, in particular, was associated with greater response.Evidence identified in this review of associations between clinical improvement and baseline FC within the SN, or changes in FC within SN, highlight the potential importance of modulating the activity of the insula and its associated network (albeit again with inconsistency of effect directions) (Fan et al., 2019;Ge et al., 2017;Godfrey et al., 2022;Iwabuchi et al., 2019;Zhang et al., 2022).Thus, anterior insula and the associated salience network may represent alternative indirect treatment targets for cortical TMS.
One of the most consistent functional connectivity differences in MDD identified in a large meta-analysis was abnormally elevated connectivity between the ECN and DMN, a finding that could reflect intrusion of DMN-mediated internal-world processing and rumination on ECN-mediated external world processing and task performance (Kaiser et al., 2015).Included left DLPFC studies showed evidence for associations between baseline ECN-DMN FC, and change in ECN-DMN FC, and clinical improvement (Ge et al., 2020;Hopman et al., 2021;Moreno-Ortega et al., 2020b;Rosen et al., 2021), as well for associations between improvement and baseline FC within the DMN or change in within-DMN FC (Cash et al., 2019b;Ge et al., 2017;Moreno-Ortega et al., 2020b;Tang et al., 2021;Taylor et al., 2018).The BRIGhTMIND trial found that greater reduction in ECN-DMN FC (specifically, left DLPFC and DMPFC) from baseline to follow-up was associated with greater improvement on self-reported measures of depression (Morriss et al., 2024).It remains to be determined whether such changes are driven by other aspects of clinical improvement, such as improvement in concentration and attention.
It should be noted that several studies identified relationships between clinical improvement and baseline FC, or change in FC, involving subcortical regions such as the striatum or amygdala (Avissar et al., 2017;Chen et al., 2020;Du et al., 2017;Kang et al., 2016;Philip et al., 2018;Salomons et al., 2014).Whilst reported in the Tables, these have not been the focus of our Results or Discussion due to specific connectivity relationships being reported in isolated studies.Given the role of the striatum in reward processing (Delgado, 2007), and of the amygdala in the experience of fear and anxiety (Davis, 1992), approaches that successfully modulate these regions might be able to treat specific facets of MDD.Further work is needed on the relevance of connectivity relationships involving these areas to the effectiveness of TMS.Our recent systematic review of connectivity features associated with anxiety symptoms in people with MDD suggested that re-establishing connectivity between the amygdala and other key brain networks may be an important treatment goal (Briley et al., 2022).Perhaps in line with the few relationships identified involving amygdala connectivity in the current review, there is evidence that left DLPFC rTMS or iTBS has less effect on anxiety symptoms than other symptoms associated with MDD (Kaster et al., 2023).
Finally, studies that move beyond pair-wise connectivity metrics and use graph theoretic approaches to capture the complexity of network changes following stimulation are needed.Only two included studies (1.14/2.1),examining left DLPFC stimulation, used such approaches (Caeyenberghs et al., 2018;Klooster et al., 2019) (one additional study, 3.1, examining left DMPFC stimulation, used a graph theoretic measure to determine the region-of-interest for subsequent analyses (Downar et al., 2014)).

Limitations
Whilst searching for studies examining connectivity measured with any MRI imaging approach, and either task or rest paradigms, the included studies are dominated by those examining resting-state FC with fMRI.Despite the large amount of information on brain network, and network abnormalities, that has been provided by resting-state studies, "rest" is not a homogeneous state.Connectivity during tasks may provide greater interpretability (Finn, 2021).More work is needed examining structural connectivity relationships as, despite some overlap in information provided by functional connectivity measures (Greicius et al., 2009), structural change represents an enduring influence of neuromodulation approaches (Damoiseaux and Greicius, 2009).Most studies examined single follow-up time points, shortly after the final stimulation session.There is little information, therefore, on predictors of longer-term or sustained clinical improvement, or on changes in relationships between predictors and outcomes.
Most included studies were small, and under-powered to detect all effects of interest.Some of the heterogeneity in findings likely reflected a study's focus on specific brain regions (partly as a strategy to mitigate having to correct for multiple comparisons with a small sample size).Some of the differences in effect directions may be due to differences in participant characteristics.MDD is known to be a heterogeneous disorder (Goldberg, 2011).It is conceivable that some subtypes of MDD may respond differently to stimulation due to differences in the underlying pathophysiology.In addition, the primary measure of MDD used in the included studies (the HDRS-17) is itself multi-factorial (Nixon et al., 2020).The included studies report change in total score on the HDRS-17, which is not, therefore, trivial to interpret.Improvement in one factor may conceivably be accompanied by decrements in another.Studies identifying associations with improvement on specific groups of HDRS-17 items, may be valuable.Some work using connectivity features to identify sub-types ("biotypes") of MDD has been conducted, although the meaning of these is under debate (Dinga et al., 2019;Drysdale et al., 2017).
A small number of patients with a primary diagnosis of bipolar disorder were present in a few of the included studies (three studies that delivered TMS to DLPFC - Liston et al. (2014) 1.2, Avissar et al. (2017) 1.4, Godfrey et al. (2022) 1.26 had between two and three people with bipolar disorder, whilst all studies that delivered TMS to DMPFC had people with bipolar disorder).As per our pre-published protocol, we included studies in which at most 20 % of the sample had bipolar disorder, as we were aware that early TMS studies had accepted a small number of patients with bipolar disorder if they had predominant depressive episodes.Study authors reported that the inclusion of these patients did not affect their results.Nevertheless, this caveat should be borne in mind when viewing the link between DMN and sgACC FC in Fig. 2, and the results for studies that used DMPFC stimulation.
Finally, whilst several studies used data from sham-controlled trials, only a few were analysed in such a way as to be able to distinguish the contribution of non-stimulation-specific effects to observed relationships between clinical improvement and either baseline connectivity (Fan et al., 2019;Klooster et al., 2020;Rosen et al., 2021) or change in connectivity (Baeken et al., 2017;Chen et al., 2020;Kang et al., 2016;Klooster et al., 2019;Persson et al., 2020;Struckmann et al., 2022).Recently, Wu et al. (2020) have specifically examined baseline FC predictors of clinical improvement following sham TMS.Greater improvement was associated with greater baseline within-DMN FC (between rostral anterior cingulate cortex, rACC, and precuneus / posterior cingulate cortex) and, potentially, DMN-ECN FC (between rACC and MFG, though MFG is a more heterogeneous region).

Conclusions
We bring together studies on brain connectivity predictors of improvement, and brain connectivity changes associated with improvement, in MDD following primarily left DLPFC stimulation.Some relationships show promise, such as those involving the targeted area or its associated network, and the sgACC or anterior insula (or the relationship between the ECN and the DMN).This is consistent with the hypothesis that TMS to superficial cortical areas acts by modulating the activity of connected deeper (and thus TMS inaccessible) areas, a hypothesis that is only now being studied directly (Duprat et al., 2022;Oathes et al., 2021).Progress will depend on understanding the transmission of the effects of TMS from the targeted area to other brain areas, how these effects change within and across sessions, and how they manifest in clinical improvement.Most of the identified studies were small and there was considerable heterogeneity in reported effects.This may partly reflect differences in patient characteristics, stimulation protocol and analysis method.Notably, there were no studies that used a targeting approach designed to optimise any specific baseline connectivity relationship at the individual level.Replication of the findings in Fig. 2 are needed in larger sample sizes, ideally facilitated by collaboration and synthesis of data at the patient level across studies.A move towards pre-published analytical protocols is essential for reducing researcher degrees of freedom and the risk of false positive findings.Further work is also needed on non-stimulation-specific contributors to improvement (placebo, and sensory, effects).The recent demonstration that some of the heterogeneity in sgACC findings may be due to between-patient differences in contributors to the global brain signal needs replication and further exploration (Elbau et al., 2023).Further work using task-based approaches, or techniques beyond fMRI, is needed, given limitations in the interpretability of resting-state fMRI findings (Finn, 2021).Greater understanding of brain activity changes following TMS (measured by changes in cerebral blood flow with PET or ASL) and of connectivity changes measured with alternative techniques such as magnetoencephalography, is also needed to provide a stronger framework for interpreting FC changes measured with fMRI.We conclude by summarising a list of recommendations to enhance confidence in the findings of future studies and to understand the identified heterogeneity (Table 4).Whilst TMS is an effective treatment for MDD at the group level, connectivity-informed approaches for predicting, or optimising, treatment response, to reduce response heterogeneity and ensure that more patients can gain maximal, and rapid, benefit from these techniques, are still needed.

Declaration of competing interest
None.
3.3.1.1.Baseline predictors of symptom improvement.Consistent results were reported in four studies (one of which included comparison with sham stimulation) that used 20-30 sessions of once-daily 10 Hz rTMS in people with MDD, the majority of whom were taking antidepressants.In an open-label design, Ge et al. (2020) (1.17) found that lower baseline FC between sgACC and right DLPFC (assigned to ECN) was associated with greater improvement on the HDRS-17 (N = 50, moderate severity depression on average, as per thresholds of Zimmerman et al. (

Fig. 2 .
Fig. 2. Summary of relationships between baseline functional connectivity and clinical improvement following a course of TMS to left dorsolateral prefrontal cortex, for connectivity relationships reported in at least two included studies.sgACC: subgenual anterior cingulate cortex, DMN: default mode network, ECN: executive control network, SN: salience network.Networks assigned as per nearest neighbour in the Power et al. (2011) atlas.Numbering corresponds to that used in the Tables and indicates the studies contributing to each relationship (with the direction of the relationship for each study represented by a minus or plus sign).Relationships for which most studies indicate a positive direction (i.e., greater, or more positive, baseline connectivity associated with greater clinical improvement) are indicated by green solid arrows.Red dashed arrows are used when most studies indicate a negative direction (lesser, or more negative, baseline connectivity associated with greater improvement).Grey dotted arrows indicate equipoise between studies in effect direction.Highlighted studies used a sham-controlled design and report a relationship that differed between the sham and active stimulation arms.

Table 1
Associations between functional/effective connectivity and clinical improvement in studies delivering TMS to left DLPFC.

Table 2
Associations between structural connectivity and clinical improvement in studies delivering TMS to left DLPFC.

Table 3
Associations between connectivity and clinical improvement in studies delivering TMS to left DMPFC.