The Glutamatergic Effects of Clinical Repetitive Transcranial Magnetic Stimulation in Depressed Populations: A Preliminary Meta-Analysis of Proton Magnetic Resonance Spectroscopy Studies

Abstract

Background:

Repetitive transcranial magnetic stimulation (rTMS) alleviates symptoms of major depressive disorder, but its neurobiological mechanisms remain to be fully understood. Growing evidence from proton magnetic resonance spectroscopy (¹HMRS) studies suggests that rTMS alters excitatory and inhibitory neurometabolites. This preliminary meta-analysis aims to quantify current trends in the literature and identify future directions for the field.

Methods:

Ten eligible studies that quantified Glutamate (Glu), Glutamate+Glutamine (Glx), or GABA before and after an rTMS intervention in depressed samples were sourced from PubMed, MEDLINE, PsychInfo, Google Scholar, and primary literature following PRISMA guidelines. Data were pooled using a random effects model, Cohen’s d effect sizes were calculated, and moderators, such as neurometabolite and ¹HMRS sequence, were assessed. It was hypothesized that rTMS would increase cortical neurometabolites.

Results:

Within-subjects data from 224 cases encompassing 31 neurometabolite effects (k) were analyzed. Active rTMS in clinical responders (n=128; k=22) nominally increased glutamatergic neurometabolites (d=0.15 [95% CI:−0.01, 0.30], p=0.06). No change was found in clinical non-responders (p=0.8) or sham rTMS participants (p=0.4). A significant increase was identified in Glx (p=0.01), but not Glu (p=0.6). Importantly, effect size across conditions was associated with the number of rTMS pulses patients received (p=0.05), suggesting dose dependence.

Conclusions and Relevance:

Clinical rTMS is associated with a nominal, dose-dependent increase in glutamatergic neurometabolites, suggesting rTMS may induce glutamate-dependent neuroplasticity and upregulate neurometabolism. More, larger-scale studies adhering to established acquisition and reporting standards are needed to further elucidate the neurometabolic mechanisms of rTMS.

Introduction

Major depressive disorder (MDD) is associated with high and rising morbidity and mortality [1], yet current pharmacological treatments have limited efficacy in some patients [2] and substantial side effects that discourage adherence in others [3]. Repetitive transcranial magnetic stimulation (rTMS) targeted at the dorsolateral prefrontal cortex (DLPFC) is a safe and well-tolerated [4] FDA-approved intervention for treatment-resistant depression [3, 5, 6] and is being studied for use in other psychiatric disorders [7]. Yet, despite its promise, rTMS clinical results are variable [8] and its underlying neurochemical mechanisms are still being uncovered [9].

rTMS is a non-invasive neuromodulatory intervention that delivers magnetic pulses via an electromagnetic coil to a task- or disease-relevant target region [9]. The magnetic pulses cause depolarization in axons [10], which may induce intracellular glutamate (Glu) accumulation by upregulating enzymes involved in Glu catabolism [11–14]. Through these and other mechanisms, rTMS can induce synaptic plasticity by selectively strengthening or weakening synapses. High-frequency protocols (>5Hz) used in most clinics favor long-term potentiation (LTP) [9, 15, 16], a process that can selectively enhance neural circuits, strengthening synaptic connections through repeated and synchronized electrical stimulation. In support of this proposed plasticity mechanism, rTMS effects are abrogated by inhibition receptors involved in LTP and LTD using glutamate (Glu) NMDA receptor antagonists such as dextromethorphan [17] and ketamine [18, 19]. Alternatively, rTMS may induce local disinhibition, a mechanism of synaptic plasticity wherein alterations in Glu- and GABA-ergic input on local inhibitory interneurons lead to increased excitability of Glu principal neurons [18]. The abovementioned plasticity mechanisms, all of which center on modulating the balance of excitatory and inhibitory neurotransmitters, are theorized to alter the reactivity of behaviorally relevant neural circuits or functional networks to achieve a desired clinical result. Indeed, rTMS may induce large-scale plasticity by means of altering functional connectivity within and between networks associated with depressive symptoms, such as the default mode network (DMN), executive control network (ECN), and the salience network (SN). However, reports thus far are heterogeneous [20–25].

Growing evidence implicates systematic, brain-wide perturbations in Glu and GABA in the pathophysiology of depression [26]. Aberrant Glu signaling and atrophy of glutamatergic synapses have been reported in MDD patients using ¹HMRS, positron emission tomography, post-mortem studies, and other methodologies [27–30]. Furthermore, glutamatergic and GABAergic alterations in the dorsolateral prefrontal cortex (DLPFC), anterior cingulate cortex (ACC), and other key nodes of the DMN, ECN, and SN may impair reciprocal activation of these networks [31–34]. This may lead to aberrant coupling associated with clinical symptoms of MDD such as rumination [35–37], reward responsiveness [38], and impaired attention [39]. Relatedly, increased cortical Glu in regulatory functional regions such as the ACC have been reported after successful treatment of MDD with antidepressants or ECT [40–42], although some have reported discordant findings. It is possible that stress-triggered neuroinflammation or excitotoxicity may play a role in these alterations in Glu and GABAergic metabolism [43–46]. Overall, growing evidence suggests that alterations in the balance of excitatory and inhibitory neurotransmitter in MDD may preclude adaptive neural signaling [47, 48], [49, 50]. This is important because rTMS may exert its antidepressant effect in part by altering the levels of these neurotransmitters.

The emerging excitation/inhibition narrative was highlighted in a recent review of rTMS studies that used proton magnetic resonance spectroscopy (¹HMRS) studies to quantify neurometabolite changes in vivo in human patients with MDD [11]. This review did not provide a quantitative synthesis but highlighted some important trends. For example, the review noted that several studies reported increased Glu [51, 52] and GABA [53–55] in patients who respond to treatment with high-frequency rTMS targeting the DLPFC. Alteration of cortical excitation/inhibition and synaptic plasticity seems to be a central mechanism of rTMS; changes in these neurotransmitters have been associated with clinical efficacy in some studies [51, 52, 55, 56], although reports are variable [57–60]. This aligns with findings from primary ¹HMRS studies in patients with MDD, a population in which cortical Glu is lower than in non-depressed populations [28], that have demonstrated normalization of Glu levels in specific brain regions after treatments including electroconvulsive therapy [42] and others [28, 60]. There is also accumulating evidence for a GABAergic signaling deficit in MDD [34], but Gonsalves et al. noted mixed results in rTMS-induced GABA changes. A more refined analysis of the influence rTMS has on excitatory and inhibitory neurometabolites may offer support for a potential mechanism of action of rTMS treatment.

Given the enthusiasm around rTMS as a potential treatment, growing evidence that rTMS alters levels of neurotransmitters, and the centrality of excitation/inhibition in the pathophysiology of depression, we aimed to provide a quantitative synthesis of the extant ¹HMRS literature to further elucidate its effects on in vivo excitatory and inhibitory neurometabolite levels in psychiatric patients. To our knowledge, the present study is the first quantitative meta-analysis of the rTMS-¹HMRS literature. We aim to synthesize the results of the extant literature to provide a more powerful estimate of true neurometabolite changes after rTMS treatment and highlight trends in the literature to guide investigators in the field. We furthermore seek to identify the influence of various potential moderators that may explain heterogeneity of the present results, including demographic (e.g., sex), methodological (e.g., ¹HMRS protocol), and neurological (e.g., brain region) moderators. We hypothesized that brain Glu and GABA would increase after rTMS treatment independent of psychopathology and anticipated that heterogeneity across studies would be explained by specific methodological choices. Overall, we propose that increased brain Glu and GABA availability after rTMS underlies the clinical efficacy of this treatment. A deeper understanding of these molecular mechanisms will lead to optimization of rTMS as a treatment for MDD and other psychiatric disorders.

Methods

The present study followed the Preferred Reporting Items for Systematic Reviews and Meta-analyses (PRISMA) reporting guidelines and was guided by the recommendations set forth by the Cochrane Collaboration. The study and protocol were not registered.

Literature Search and Data Extraction

Data collection began with a broad search through PubMed, MEDLINE, PsychINFO, and Google Scholar (See supplement for search terms). Articles published from database inception to July 31, 2023 were identified. Additional articles were collected through references in the primary literature. Data extraction was performed by two reviewers (M.P. & A.M.) and all data were reviewed by both individuals. Discrepancies were addressed in the presence of a third party (D.R.) and resolved via consensus agreement. The final dataset used for analysis is included in the supplemental materials.

Selection Criteria

To be eligible for inclusion in the present meta-analysis, primary reports (i.e., peer-reviewed manuscripts) were required to include: 1) measures of glutamate (Glu or Glx) or GABA using a ¹HMRS approach from a within-subjects rTMS design comparing at least one ¹HMRS measurement before and after a clinical rTMS intervention; 2) samples of unrelated individuals with MDD; and 3) data or statistical information that enabled the calculation of effect sizes independent of any control groups included in the study.

Moderator Variables and Meta-regression

Despite the limited heterogeneity observed in our initial meta-analytic result, moderator analysis and meta-regression were used to investigate the contribution of moderating variables to identify subgroup differences in effect size. Additional variables extracted for moderator or regression analysis included: a) number of rTMS pulses administered across treatment, b) neurometabolite, c) ¹HMRS acquisition sequence, d) ¹HMRS region of interest, e) voxel type (single- or multi-voxel), f) neurometabolite quantification method (ratio or water-referenced), g) age, h) gender (percent female), i) voxel size, j) number of rTMS sessions, and k) timing of post-rTMS scan session. Studies conducted on healthy individuals or non-MDD disorders, those using alternative neuromodulation techniques, such as transcranial direct current stimulation (TDCS), and effects for neurometabolites other than Glu, Glx, or GABA were not included in this synthesis. Clinical non-responders and those who received sham rTMS were used as control groups and were not included in moderator analyses (see Supplemental Methods for additional details).

Analysis of publication bias

Publication bias is a substantial risk in all meta-analyses because studies with larger effects are more likely to be published than those with smaller or null effects. To assess publication bias, multiple methods were employed: 1) funnel plots, which display an asymmetric relationship between effect size and study precision in potentially biased datasets; 2) adjusted rank-correlation tests following Begg and Mazumdar (1994), Egger et al. (1997), and Duval and Tweedie (2000) [61–63], in which high correlation between effect size and sampling variances suggests bias; and 3) post-hoc trim-and-fill method incorporating calculation of a fail-safe N. The trim-and-fill method, based on funnel plots, identifies, and removes outlier studies while estimating a publication bias-adjusted summary effect size by accounting for potentially missing unpublished studies. This comprehensive approach helps control for the influence of outliers and addresses potential publication bias by considering both published and unpublished data.

Analysis of data quality

The MRS-Q tool [65] evaluates the adequacy of methods reporting and appropriateness of acquisition parameters to generate an assessment of data quality for systematic reviews of MRS studies. It aligns with the recent MRSinMRS guidelines [66]. Studies that reported adequate spectroscopy parameters and analysis methods were classified as “high quality”, while studies reporting insufficient parameters and methods were classified as “low quality”. Studies with insufficient information to make a definitive classification were classified as “unsure.” Most studies in our synthesis were conducted before these guidelines were published. In our study, note that “low quality” does not necessarily indicate that the study itself is of low quality but merely reflects that the pertinent information was not available, or the study used the best available parameters at the time. The influence of acquisition parameters was assessed with moderator analyses (see above). Two investigators (MP and AM) independently evaluated the quality of each study using the aforementioned tools. Disagreements were resolved through consensus discussions involving the senior investigator (DR).

Results

The Results section should describe the most important findings of the study. The most important results should be indicated, and relevant trends and patterns should be described.

Characteristics of included studies

Of 594 articles screened and 34 articles reviewed, 10 articles were included in the present meta-analysis (n=224 total participants; k=31 individual neurometabolite effects) (Table 1; see Supplementary Tables S1–3 for additional study characteristics). Publication dates ranged from 2007 to 2021. Sample size ranged from 6 to 27, reflecting that a large-scale ¹HMRS-rTMS study has yet to be conducted in an MDD population. All included studies featured a clinical, 10–20Hz rTMS intervention targeted at the DLPFC in MDD patients, with treatment lasting an average of 4 weeks and ranging from 1 day to 7 weeks. ¹HMRS data were collected at a field strength of 3 tesla (3T) from frontal regions including the DLPFC (k=18 effects), ACC (k=12 effects), primary motor cortex (k=2 effects), and contralateral DLPFC (k=2 effects). All studies were conducted in exclusively or predominantly MDD cohorts (Dubin et al. (2016) [54]: n=21 MDD, n=2 bipolar disorder; Godfrey et al. (2021) [25]: n=26 MDD, n=1 bipolar disorder). One study [57] implemented a 4-week medication washout prior to the study; participants in the other nine studies were taking a stable regimen of psychotropic medications including selective serotonin reuptake inhibitors, selective norepinephrine reuptake inhibitors, benzodiazepines, antipsychotics, and others. The experimental group (n=10 studies, k=22 effects), consisted of 130 patients (63 female) who showed a clinical response to active rTMS. Average age was 37.9 years. Two studies reported sham data (k=2 effects, N=28 participants (7 female), average age=26.8 years) and five studies reported data from clinical non-responders (k=7 effects, n=68 participants (32 female), average age=35.9 years). Every study reported pre- and post-rTMS data for all experimental groups; thus, all neurometabolite effects were calculated in comparison to baseline values rather than in comparison to a control group. A chi square test of gender distribution across groups showed a significant difference in gender distribution across groups (X²=8.7, p<0.05). An ANOVA comparing average age showed no difference between groups (F=0.95, p=0.4).

Table 1:

Summary of Included Studies

Study ID	Control Groups	Neurometabolite	Patient Diagnosis	Medications	Number of Patients*	Age Mean (SD)	Gender (F%)	TMS Frequency (Hz)	TMS Intensity (%MT)	Number Daily Sessions	TMS Pulses/Session	HMRS Region of Interest	Voxel Size (ml)	HMRS Pulse Sequence	TE (ms)	HMRS Field Strength
Baeken (2017) [53]	Sham	Glu, GABA	MDD	Benzodiazepines	Exp: 18 Nresp: NA Sham: 18	47.2 (12.5)	66.7	20	110	20	1560	L DLPFC R DLPFC ACC	3.375	PRESS	40	3T
Bhattacharyya (2021) [58]	None	Glx, GABA	MDD	SSRIs, SNRIs, neuroleptics, mood stabilizers, and/or stimulants	Exp: 6 Nresp: 6 Sham: NA	53 (15)	66.7	10	120	30	3000	L DLPFC	8	MEGA-PRESS	68	3T
Croarkin (2016) [56]	None	Glu	MDD	Anti-depressants (class not specified)	Exp: 10 Nresp: NA Sham: NA	15.4 (1.2)	40	10	120	23.2	3000	L DLPFC ACC	8	PRESS	80	3T
Dubin (2016) [54]	Non-responders	Glx, GABA	MDD & BP (n=2)	Anti-depressants, mood stabilizers, anti-psychotics	Exp: 8 Nresp: 15 Sham: NA	37 (16)	75	10	100	25	3000	L MPFC ACC	18.75	PRESS	68	3T
Erbay (2019) [57]	None	Glu	MDD	None (medication washout)	Exp: 18 Nresp: NA Sham: NA	43.38 (11.1)	55.6	10	Not reported	20	3000	L DLPFC	4.5	PRESS	30	3T
Godfrey (2021) [25]	None	Glx, GABA	MDD & BP (n=1)	Psychotropic medications (class not specified)	Exp: 27 Nresp: NA Sham: NA	41.6	40.7	10	120	20	4000	L DLPFC R M1	26.25	MEGA-PRESS	68	3T
Levitt (2019) [55]	Non-responders	GABA	MDD	Anti-depressants, benzodiazepi nes, mood stabilizers	Exp: 12 Nresp: 14 Sham: NA	40 (13.8)	50	10	110	30	3000	L DLPFC	6	MEGA-PRESS	68	3T
Luborzewski (2007) [51]	Non-responders	Glu	MDD	Anti-depressants	Exp: 6 Nresp: 11 Sham: NA	47.5 (9.1)	16.7	20	100	10	2000	L DLPFC B ACC	8	PRESS	80	3T
Zheng (2010) [60]	Non-responders, Sham	Glx	MDD	Anti-depressants	Exp: 12 Nresp: 1 Sham: 14	26.9 (6.2)	36.8	15	110	20	3000	L DLPFC	1.5	PRESS	30	3T
Zheng (2015) [59]	Non-responders, Sham	Glx	MDD	Anti-depressants	Exp: 11 Nresp: 1 Sham: 14	26.9 (6.4)	33.3	15	110	20	3000	L ACC	1.5	PRESS	30	3T

^*Group labels: Exp: Experimental group (clinical responders); Nresp: Clinical non-responders; Sham: Group received sham rTMS

Primary Outcomes

Overall, responders to active rTMS (n=128 participants; k=22) had a nominal increase in glutamatergic and GABAergic neurometabolite concentration (d=0.15 [95% CI:−0.01, 0.30], p=0.06, Fig. 1. a) that was homogeneous (Q=21.26, p=0.4). No overall effect was seen in clinical non-responders (n=68 participants; k=7 effects; d=−0.05 [95% CI:−0.55,0.44], p=0.8; Fig. 2. a,b) or in patients who received sham rTMS (n=28 participants; k=2 effects; d=−0.21 [95% CI:−0.70,0.28], p=0.4; Fig. 2. a,b). To highlight clinically relevant neurometabolite trends, effects for non-responders and sham participants were excluded from subsequent analyses.

Fig. 1.

Clinical rTMS alters the levels of neurometabolites in patients with MDD who respond to treatment. The bar graphs depict average effect size (Cohen’s d) for the increase in neurometabolites after rTMS intervention.

a. Meta-analysis revealed a nominal increase in neurometabolites in clinical responders who received active rTMS and responded to treatment (k=22, d=0.15 [95% CI:−0.01, 0.30], #: p=0.06). Data were combined across ¹HMRS regions and neurometabolites.

b. Moderator analysis was used to compare effect sizes across neurometabolites. No significant effect was identified in GABA (k=8, d=0.171 [95% CI: −0.09,0.43], p=0.2) or Glu (k=8, d=−0.07 [95% CI: −0.34,0.21], p=0.6), but a significant increase in Glx was found (k=6, d=0.38 [95% CI: 0.09,0.67], p<0.01). (**: p<0.1, d>0) A t-test revealed that the effect size for Glx was significantly different from that of Glu (p=0.03)

c. Forest plots depict the individual effects for each neurometabolite in clinical responders. The region of interest from which data were obtained is noted for each effect. (DLPFC: dorsolateral prefrontal cortex; ACC: anterior cingulate cortex; M1: primary motor cortex). The yellow line marks the overall effect size across all metabolites.

Fig. 2.

Clinical rTMS induces no significant change in metabolites in clinical non-responders or patients who receive sham rTMS.

a. The bar plot depicts the results of a meta-analysis of rTMS-induced neurometabolite changes for clinical non-responders and participants who received sham rTMS. No significant neurometabolic effect was identified in clinical non-responders (k=7 effects; d=−0.05 [95% CI:−0.55,0.44], p=0.8) or the sham rTMS groups (k=2 effects; d=−0.21 [95% CI:−0.70,0.28], p=0.4).

b. Forest plots depict the individual effects for each neurometabolite in clinical non-responders and sham participants. The metabolite for the given effect is noted in parentheses. The yellow line marks the overall effect size.

Moderator Analyses

Moderator analyses were performed to assess the influence of methodological and participant variability contributed to effect sizes.

Neurometabolite:

Undergoing rTMS resulted in a significant increase in Glx (k=6, d=0.38 [95% CI: 0.09,0.67], p<0.01, Fig. 1. b), but no significant change in Glu (k=8, d=−0.07 [95% CI: −0.34,0.21], p=0.6, Fig. 1. b) or GABA (k=8, d=0.171 [95% CI: −0.09,0.43], p=0.2, Fig. 1. b). Pairwise comparison revealed the increase in Glx was significantly greater than the effect for Glu (p=0.03).

¹HMRS Sequence:

The influence of ¹HMRS acquisition protocol was also assessed (Supplemental Fig. S1). Across all metabolites, no significant change was detected in studies using PRESS acquisition (k=15 [Glu=8, Glx=3, GABA=4], d=0.02 [95% CI: −0.18,0.22], p=0.9). However, there was a significant increase in metabolite concentration in studies employing MEGA-PRESS (k=8 [Glx=3, GABA=5], d=0.36 [95% CI: 0.08,0.64], p=0.01). Because all Glu effects were acquired with PRESS, we repeated this moderator analysis with just Glx and GABA (Supplemental Fig. S2); the same pattern of results was seen, with an unchanged effect for MEGA-PRESS (p=0.01) and no significant effect in metabolites measured with PRESS (k=7 [GABA: k=4, Glx: k=3], d=0.12 [95% CI: −0.17,0.4], p=0.4). In both moderator analyses, t-tests comparing MEGA-PRESS and PRESS effects were not significant (p>0.05).

Neurometabolite Quantification:

The influence of neurometabolite quantification approach was also assessed (Supplemental Fig. S3). 8 of 10 studies reported metabolite levels as water-referenced concentrations. Among these studies, a nominal increase in metabolites was identified (k=19, d=0.16 [95% CI: −0.01,0.32], p=0.07). No significant effect was found in studies that reported neurometabolite ratios normalized to creatinine (k=3, d=0.10 [95% CI: −0.46,0.67], p=0.7). A t-test comparing the two quantification methods was not significant (p>0.05).

Additional moderator analysis showed no significant effect for ¹HMRS region of interest (stimulation site (DLPFC) or downstream (at another frontal region)). All studies acquired ¹HMRS data using a single-voxel technique at 3T, so no assessment of these parameters was possible. Due to reporting inconsistencies, we were not able to assess the impact of the timing of the post-rTMS ¹HMRS scan.

Moderator Subgroup Analysis:

Neurometabolite changes were compared at the rTMS stimulation site (DLPFC) versus downstream regions, including anterior cingulate cortex (ACC), primary motor cortex (M1), and contralateral DLPFC (Supplemental Fig. S9–10). Despite the low number of effects per group, there was a nominal change in Glx in the DLPFC (k=3, d=0.393 [95% CI: −0.03,0.81], p=0.07) and in downstream regions (k=3, d=0.366 [95% CI: −0.05,0.78], p=0.08). A nominal change in GABA at the DLPFC was also identified (k=4, d=0.38 [95% CI: −0.1,0.86], p=0.1). No change in downstream GABA or Glu in any region was detected (p>0.05).

Meta-regressions:

Meta-regressions were performed to assess the association between continuous study variables and neurometabolite effect size in clinical responders. The total number of rTMS pulses across the treatment course revealed a significant positive association with neurometabolite effect size (r²=0.13, p=0.05) (Fig. 3. a), suggesting dose dependence. rTMS frequency, ranging from 10–20Hz across studies was not associated with effect size (r²=0.06, p=0.14) (Fig. 3. b). Effect size was not associated with ¹HMRS voxel size (p=0.3), average age (p=0.7), rTMS intensity (p=0.9), or sex composition measured by percent female participants (p=0.6) (Supplemental Fig. S4–8).

Fig. 3.

Meta-regression assesses the impact of rTMS treatment parameters.

a. Meta-regression shows a positive association between glutamatergic effect size versus the number of rTMS pulses administered throughout the course of treatment (r²=0.13, p=0.05), suggesting that rTMS-induced neurometabolite change may be dose-dependent.

b. Meta-regression reveals no significant association between effect size and rTMS frequency. (r²=0.06, p=0.14).

To further characterize the impact of rTMS parameters on neurometabolite effect size, meta-regressions for total number of pulses, rTMS intensity, and rTMS frequency were repeated on non-responders to active rTMS (k=7 effects). Neurometabolite effect size was positively associated with total number of rTMS pulses (r²=0.4, p=0.03) and negatively associated with rTMS frequency (r²=0.33, p=0.05). No association with rTMS intensity was present (r²~=0, p=0.24). There was not a sufficient number of sham effects in our dataset (k=2) to conduct regression analyses for this control group.

Publication Bias

Potential bias in the present subset of publications was assessed using a Begg’s funnel plot (Supplemental Fig. S11) and accompanying tests for asymmetry. Begg and Mazumdar’s rank correlation test (Kendall’s tau=0.11, p=0.36) and Egger’s test of the intercept (p=0.76) suggested no funnel plot asymmetry. The Duval and Tweedie trim and fill imputed six missing values and generated an overall effect size that was lower than the original (d=0.05 [95% CI: −0.1,0.2], p=0.5), suggesting some studies with null results may remain unreported. Leave-one-out analysis showed that removing any single effect had no significant impact on the overall effect size (Fig. S12).

Data Quality

Quality assessment using the MRS-Q reveals that the majority (75%) of studies reported an appropriate level of information about the ¹HMRS protocol (Fig. 4). See Supplementary Tables 2A–C for additional details about demographics, ¹HMRS parameters, and rTMS parameters. The “low quality” label does not necessarily reflect insufficient parameters, but merely an insufficiency in reporting. All studies were included given they were published before field reporting guidelines had been established.

Fig. 4.

MRS-Q assessment shows high data quality in the majority of studies

Per the recommendations of Peek 2020 [64], classification of quality was made on the basis of reporting appropriate sequence and adequate parameters (data points 2 and 3). 7 of 10 studies were classified as “high-quality.” All studies classified as “low quality” were published before the MRSQ guidelines and did not report ¹HMRS acquisition parameters in sufficient detail. The gray square for data point 2 from Baeken 2016 [53] signifies that an appropriate sequence (unedited PRESS) used for Glu, but this was insufficient for GABA.

Discussion

This meta-analysis of proton magnetic resonance spectroscopy (¹HMRS) in repetitive transcranial magnetic stimulation (rTMS) for major depressive disorder (MDD), to our knowledge the first of its kind, aimed to characterize rTMS neurometabolic effects in depressed populations and provide insights to guide future work in the field. In preliminary support of our central hypothesis, we found that treatment with rTMS causes a nominal, dose-dependent increase in neurometabolites in patients who clinically respond to treatment. Across all conditions, a significant increase in cortical Glx, but not Glu, was identified. We found a nominal increase in onsite Glx and GABA and downstream Glx. Neurometabolite effect size in clinical responders was not associated with rTMS intensity.

While further research is required to establish whether the findings of Glu or Glx from our current study more accurately captures the true glutamatergic effects of rTMS, Glx may offer a more robust readout of neuroexcitation at 3T, as it entails fewer preprocessing steps [67]. Glx is an aggregate signal of Glu and Glutamine (Gln) driven mainly by Glu given the relative abundance of Glu compared to Gln in brain tissue [67, 68]. Although Gln is not directly involved in neurotransmission, Gln and Glu are tightly yoked within a common metabolic pathway, with Gln stored and synthesized from Glu in astrocytes and Glu stored and synthesized from Gln in neurons [69–71]. Thus, it is possible that Glx was elevated after rTMS in this meta-analysis because the aggregate Glu-Gln measure increased signal to noise ratio. If we treat Glx signal as a proxy for neuroexcitation, then the significant increase in Glx is in keeping with rodent studies where rTMS upregulates the expression and activity of enzymes and membrane channel proteins involved in Glu catabolism, including VGLUT1 [11–14, 72]. High-frequency rTMS, like the protocols included in the present meta-analysis, increases cortical excitability in humans [74] and may induce a process similar to long-term potentiation (LTP) that is Glu-dependent [19, 74, 75] and leads to synaptic strengthening [10]. This LTP-like process may be induced by rTMS-triggered action potentials within axons and leads to an NMDA receptor-mediated increase in AMPA receptors in postsynaptic membranes of excitatory glutamatergic synapses [76]. Structural evidence of synaptic plasticity in the form of dendritic spine growth after rTMS has also been reported in animals [76, 77]. Lastly, the increase in Glx, may reflect changes in glial activity that could be related to clinical improvement [56, 78]. For example, astrocyte Glu reuptake helps regulate Glu levels at the synapse, preventing excitotoxicity associated with a maladaptive stress response [30, 79, 80]. Microglia and astrocytes are also involved in the interconversion of GABA, Glu, and Gln, helping maintain a homeostatic excitation/inhibition balance [70, 81, 82]. By altering the expression of rate-limiting enzymes or otherwise directly affecting glia, rTMS may help restore homeostatic processes that directly help restore synaptic functioning.

Moderator subgroup provided preliminary evidence of distributed rTMS glutamatergic effects. We identified a nominal increase in Glx at the DLPFC and at downstream sites such as the anterior cingulate cortex (ACC). The nominal increase in GABA at the DLPFC we identified is consistent with extant reports indicating reduced cortical GABA levels in individuals with MDD compared to non-depressed individuals [83–85] and increased GABA catabolism after rTMS [9, 86–88]. While a concurrent GABA and Glx increase at the DLPFC may seem counterintuitive, it is important to note that GABA and glutamate are generated from common substrates. Thus, metabolic pathways upregulated by rTMS would increase the substrates used for both metabolites [11, 13, 14]. The distributed nature of the neurometabolic effects is also suggested by the lack of correlation between effect size and voxel size. If metabolic response to rTMS spreads through a network of interconnected regions rather than a confluent, localized area, larger voxels would not necessarily increase the signal-to-noise ratio. Distributed neurometabolic changes may reflect the induction of plasticity along anatomical tracts or functional network hubs. Investigation of the spatial pattern of metabolite changes is inherently limited by the ¹HMRS methodology, in which only one region can be assessed at a time. This limitation aside, the increase in downstream Glx, particularly within the ACC, has important implications for the treatment of MDD patients. Atrophy of glutamatergic synapses in the ACC and other key functional brain regions – potentially caused by astrocyte dysfunction [89, 90], neuroinflammation [91–93], and excitotoxicity from long-term stress [94, 95] – has been widely reported in patients with unmedicated MDD [30]. Successful treatment of MDD has been linked to increased Glu and GABA in the ACC and other cortical regions [27–29, 96], although results are heterogeneous [98], [99]. Our findings provide preliminary support for the hypothesis that rTMS may alleviate depression symptoms by altering levels of excitatory and inhibitory cortical neurometabolites. Larger-scale studies adhering to new imaging reporting standards are needed to determine how metabolite changes are specifically induced and how alterations in regulatory cortical regions like the ACC may rectify the activity of dysfunctional neural circuits and restore adaptive network decoupling.

Our findings also shed light on how rTMS parameters impact its neurometabolic effects. Meta-regression analysis of the total number of pulses administered during rTMS revealed evidence of dose dependence: administering a higher number of pulses may lead to larger changes in metabolite concentrations. If rTMS-induced metabolite changes are indeed rectifying a baseline deficit or relative excitation/inhibition imbalance, then larger neurometabolite effects may translate to greater clinical improvement. Interestingly, our study did not find a relationship between rTMS intensity or frequency and metabolite change. However, all included studies utilized high-frequency rTMS protocols (10Hz or higher) thought to enhance cortical excitation [19], [100], [101], [102]; more research is needed to elucidate how a wide range of rTMS frequencies modulates its glutamatergic effects in patients with depression. It is possible that lower intensity and frequency protocols could be equally effective and potentially more tolerable for patients. Interestingly, meta-regression also revealed similarities between clinical responders and non-responders to active TMS. Although there was no main effect of rTMS on neurometabolite concentration in clinical non-responders, effect size in both clinical responders and non-responders was positively associated with number of rTMS pulses. If neurometabolite change indeed mediates rTMS response, non-responders may require additional titration of frequency or “dose optimization” to yield a clinical effect. There was not sufficient sham data to compare this group with participants who received active rTMS. Overall, the findings suggest that the total number of pulses administered during rTMS plays a determining role on its neurometabolic effects and, potentially, on the clinical outcomes. Ongoing and future studies should report subject-level metabolite and clinical data so that future meta-analyses can assess this relationship directly. We did not have sufficient data to directly assess the association between rTMS-induced neurometabolite change and clinical efficacy, the dose dependence trend and evidence from other lines of research suggests this is a strong possibility and should be the area for urgent future research. If a mediating effect is confirmed, physicians using rTMS in clinical settings should consider extending rTMS protocols when feasible to potentially enhance therapeutic neurometabolic treatment effects.

While efforts were made to minimize limitations, some were unavoidable. Most clinical rTMS studies target magnetic pulses at the DLPFC using cranial landmarks, a method which remains the most feasible and favorable for most clinics, but which introduces some variability in the region being stimulated. Participants in nine of ten studies were taking stable regimens of psychotropic medications. While this could theoretically introduce confounding factors in the assessment of true rTMS effects, most real-world rTMS patients are concurrently taking medication; thus, the present group of studies provides a naturalistic representation of rTMS patients. In addition, while it is possible the dynamics of cortical excitation in response to TMS may change with age [103], robust clinical responses to rTMS have been reported in patients across the lifespan [104], [105]. While age introduces another potential source of variability across studies, we chose not to exclude any study based on participant age to maximize the statistical power and the generalizability of our results. Nine of the ten unique studies in this meta-analysis reported multiple effects, introducing the potential for confounding study effects. While multilevel meta-analysis is often employed to correct for study effects, our dataset was not of sufficient size or heterogeneity for this method to be feasible. However, sensitivity analyses showed minimal risk for publication bias. Because of systematic differences in experimental design in studies conducted in healthy versus clinical populations, mostly pertaining to the length of the treatment course, our synthesis only included studies conducted in clinical populations, specifically in patients with MDD. To allow for valid comparison, future studies in clinical populations should consider gathering additional ¹HMRS data points after one to five rTMS sessions, which is the typical length of rTMS studies in healthy subjects. The extant literature also reports limited data in non-responders to active rTMS and in patients who received sham rTMS. More data is needed to strengthen inferences from these group comparisons. Although our meta-regression analysis on rTMS frequency was hindered by the limited number of frequencies represented, we incorporated the continuous regression to emphasize the potential for a nuanced gradient relationship between frequency and neurometabolite effect size. Future studies should incorporate a wider spectrum of rTMS frequencies to better elucidate the impact of rTMS parameters on neurometabolite effects. Despite these limitations, it should be noted that the data synthesized in the present meta-analysis provides composite data from a sample that is roughly 5 times larger than the largest primary study in the literature. Lastly, few studies to date have collected ¹HMRS data before and after theta burst stimulation protocols, which may induce plasticity differently than conventional rTMS. Thus, these studies were excluded from the present synthesis. While necessary curation steps limited the heterogeneity of our final dataset, they allowed us to explore more targeted clinical questions of interest.

A major goal of our study was to highlight trends and future directions for the field in terms of acquisition techniques. Using the standardized MRS-Q quality assessment tool, we categorized three studies as “low quality”. However, it is essential to acknowledge that this classification primarily indicates that these studies did not meet recently established standards in their reporting of acquisition parameters. Studies using insufficient parameters must be seen in historical context considering they were published before a standardized, rigorous reporting guideline existed. We included the MRSQ assessment to raise awareness of the field’s evolving acquisition and reporting standards. These guidelines should not be used to retroactively punish studies published prior to the guidelines. Instead, we emphasize and encourage the field’s much-needed shift toward greater standardization in ¹HMRS acquisition, which will be vital for conducting large-scale multi-center investigations.

Specific findings in our study highlight potentially high-yield paths forward. The unexpected disparity between Glx and Glu effects may stem in part from difficulties parsing the Glu and Gln signals at 3T and heterogeneity in ¹HMRS acquisition sequences and echo times [106], [107], [108]. In the present meta-analysis, all studies quantifying Glu used PRESS sequences and echo times ranging from 30 to 80ms, while studies quantifying Glx used both PRESS and MEGA-PRESS sequences with echo times ranging from 30 to 68ms. Although our data quality analysis suggests that studies in the present meta-analysis utilized robust parameters to detect neurometabolite changes, the literature has yet to establish a definitive consensus on the ideal combination of acquisition sequence and echo time to optimize Glu detection at 3T [109], [110], [111]. Some have suggested that a shorter TE (<35ms) maximizes Glu/Gln separation in both PRESS and MEGA-PRESS [110], [112]. Thus, the null result for Glu in the present study may stem from suboptimal ¹HMRS acquisition parameters or insufficient field strength to detect Glu change independent of Gln. Our results suggest that quantifying Glx instead of Glu and using a MEGA-PRESS versus PRESS sequence, when indicated by MRSQ guidelines, may optimize researchers’ detection of subtle rTMS-induced neurometabolite changes at 3T. However, more research is needed to identify the ideal acquisition parameters and clarify the differential effects of rTMS on Glu, Gln, and Glx in the treatment of MDD. Most studies in our meta-analysis also utilized water-referenced values rather than neurometabolite ratios. While we did not have sufficient data to robustly assess the difference in effect size, we recommend researchers use water-referenced values given that total creatine levels may also change with rTMS intervention [51], [57]. The present study focused solely on Glu, Glx, and GABA. As more literature becomes available, future quantitative syntheses should explore relative changes in these metabolites elucidate rTMS excitation-inhibition effects in greater detail. It would also be beneficial to explore changes in other neurometabolites implicated in rTMS effects, such as NAA, Cho, and mIns [51, 57, 59, 113]. The use of ultra-high field MRI techniques like Glutamate Chemical Exchange Saturation Transfer [114–116] will likely improve the sensitivity and spatial resolution of brain glutamate measurements before and after rTMS.

In conclusion, the present meta-analysis suggests that rTMS may increase cortical glutamatergic metabolites in a dose-dependent, distributed fashion, with at least nominal metabolite increases seen both at the site of stimulation and in other frontal regions. These findings provide preliminary support for the hypothesis that rTMS addresses underlying glutamatergic and GABAergic deficiency in MDD. The present quantitative synthesis also has important implications for clinicians and researchers. The present study highlights the need for further research on the neurometabolic effects of rTMS in healthy subjects and clinical populations to better understand and eventually optimize this and other neuromodulatory interventions and add to the still limited understanding of mood disorders neurobiology.

Data Availability Statement

The data used for the meta-analysis conducted in this study has been made available in the supplementary materials. Further enquiries can be directed to the corresponding author.