Background
Major depressive disorder (MDD) is the leading cause of disability worldwide [
1]. Due to a lack of mechanistically anchored quantitative tests for identifying the correct intervention for individual patients at their first presentation, treatment choice is often a years-long trial-and-error process. One reason for the unpredictability of therapeutic response is the heterogeneity of MDD, both clinically and in terms of underlying neurobiology [
2]. While definitions vary, pharmacoresistant or treatment-resistant depression is defined as lack of response to at least one antidepressant trial of adequate dose and duration [
3]; up to 50% of MDD patients meet these criteria [
4]. Residual depressive symptoms are associated with a higher risk of recurrence, worse functioning, and increased personal and economic burden [
5]. Furthermore, pharmacoresistant depression can be life threatening: 30% of patients have one or more lifetime suicide attempts, which is at least twice the rate of those with non-resistant depression [
6].
Neural circuits (hereafter circuits) consist of vast numbers of interconnected neurons comprising the anatomical and functional networks of the brain [
7]. Circuits involved in cognitive control are promising targets for pharmacoresistant depression. Dysfunctions in the cognitive control network and reciprocal pathways linking this circuit with attention and default mode networks (DMN) are characteristic of MDD [
8] (for review [
2,
7]). Depressed patients who do not remit on commonly prescribed pharmacotherapies show dorsolateral prefrontal cortex (DLPFC) hypoactivation along with hypoconnectivity between the DLPFC and anterior cingulate cortex [
9]. Corresponding frontoparietal attention network hypoconnectivity is also observed in MDD [
10,
11] and correlated with behavioral indices of poor attention, such as false alarm errors on cognitive testing, in related anxiety disorders [
12]. Some degree of DMN dysfunction is observed in persistently unwell MDD patients [
13] and can identify MDD patients who do not remit on antidepressants [
14]. Reflecting the close interplay of cognitive control and attention networks, pharmacoresistant patients with MDD are also characterized by impaired connectivity of the DLPFC and precuneus component of the attention network [
15]. As an interposed area involved in the interactive effects of cognitive control and other circuits, subgenual anterior cingulate cortex (sgACC) impairments are also thought to exacerbate broader circuit dysfunction in MDD (for review [
10]). Pharmacoresistant MDD patients show persistent hypoactivation and connectivity involving the sgACC [
16] in analyses of brain metabolism using positron emission tomography.
Repetitive transcranial magnetic stimulation (rTMS, hereafter TMS) was cleared by the U.S. Food and Drug Administration (FDA) for pharmacoresistant MDD in 2008 and has become an important treatment option in clinical settings. While the putative therapeutic mechanism of TMS remains under study, recent neuroimaging studies provide insight into brain activity changes associated with therapeutic TMS of the DLPFC. Neuroimaging studies of TMS in both healthy subjects and in MDD have focused mostly on circuits probed during resting conditions. In healthy subjects our prior work has shown that DLPFC stimulation induces an inverse correlation between resting connectivity of the DLPFC (middle frontal gyrus) and the medial frontal region of the DMN [
17] as expected for flexible circuit organization. Neuroimaging studies implicate pre-stimulation baseline DLPFC-sgACC connectivity in the mechanisms of clinical action of TMS [
18,
19]. In 13 patients with MDD, more intact negative (i.e., “anticorrelated”) DLPFC-sgACC resting state functional connectivity prior to DLPFC stimulation was associated with superior amelioration of clinical symptoms [
18]. In a small subset of controls and two patients scanned post-TMS, individual differences in DLPFC-sgACC connectivity were highly reproducible [
19]. These findings suggest that suppression of the subgenual anterior cingulate cortex via DLPFC stimulation may be an antidepressant mechanism of TMS, and that baseline connectivity is a viable imaging biomarker to optimize TMS at the individual level. When imaged after TMS, responders (
n=5/12) showed improvement in the negative connectivity of DLPFC and sgACC, whereas non-responders (
n=7/12) did not [
20]. In a complementary study, 17 MDD patients were found to show attenuation of abnormally positive sgACC-DMN connectivity, along with reduced DLPFC to medial prefrontal connectivity, but not of the DLPFC and sgACC [
21]. However, in this latter study, TMS-related connectivity changes were not associated with clinical improvement. Most recently, Weigand et al. [
22] demonstrated that sgACC-DLPFC connectivity could predict clinical response to TMS; this study incorporated two datasets, inclusive of 25 participants who received unblinded TMS, and 16 participants who received sham stimulation and open-label stimulation from Taylor et al. [
23]. Together, findings to date suggest that TMS selectively modulates functional connectivity both within and between the cognitive control network and interconnected regions of the frontal cortex and DMN, and that modulation of these interactions by the sgACC may play an important mechanistic role in predicting the effect of TMS on alleviating depression. The results also highlight the need for systematic investigation using imaging biomarkers in samples with greater power for statistical inference.
Drawing on this evidence, a premise of our study is that TMS of the DLPFC will have antidepressant efficacy via direct effects on cognitive control processes that contribute to regulatory functions and that involve interactions with attention and default mode networks.
Despite the wide scale adoption of TMS, we still lack mechanistically-driven biomarkers designed to identify who is most likely to respond, and why; these measures are crucial for broader adoption of TMS and are possible with near-term discoveries. In our recent multisite trial of TMS in pharmacoresistant depression [
24], standard clinical measures did not predict remission [
25]. While there are relatively few side effects from TMS as compared to other neuromodulation techniques, undergoing a full course of this treatment when it will ultimately not lead to remission can be discouraging for the patient and psychiatrist, prolongs suffering and is economically inefficient. For these reasons, there is an urgent need for well-powered multisite clinical trials that advance a biomarker-driven approach to identifying which patients will benefit from TMS and through which mechanisms. Furthermore, existing rubrics can be immediately used to translate observed findings into clinical practice [
26]. Our findings will, more broadly, also lay important foundations for the systematic experimental manipulation of stimulation protocols and parameters in future mechanistic trials.
Our study objective is to systematically evaluate cognitive control network connectivity and behavior as response biomarkers for the effect of TMS in pharmacoresistant MDD, and the extent to which connectivity and behavior are predictive of clinical symptoms, function, and suicidality outcomes. We strive to meet this objective by exploring the following aims:
Aim 1
To evaluate a response biomarker of the effect of TMS on promoting cognitive control. We will assess whether activation and functional connectivity of the DLPFC-anchored cognitive control network, and interactions with sgACC, attention and default mode regions involved, change in a session (akin to dose)-dependent manner. Our broad hypothesis is that functional connectivity will change from the pre-treatment baseline to reassessment early after commencement of TMS (targeting 5 sessions) and later, post- completion of treatment (targeting 30 sessions). Related, we hypothesize that the early change will be most pronounced for patients with intact connectivity at baseline; later change post-treatment will be observed for those with more impaired baseline connectivity. First, we will address the mechanistic question of whether early changes in circuit connectivity are necessary, if not sufficient, for subsequent circuit and clinical changes observed post-TMS. Second, we will systematically test whether the extent of change in connectivity is related to the extent of dysfunction at the pre-stimulation baseline. Third, we will probe whether observed changes in connectivity increase as a session-dependent function of the total number of TMS sessions. In addressing these issues, we will incorporate a methodological technique to quantify the site of DLPFC stimulation with anatomical precision.
Aim 2
To assess whether the extent of change in a DLPFC cognitive control network connectivity response biomarker is related to corresponding change in behavioral performance. Our broad hypothesis is that the extent of connectivity change will be related to the extent of change in behavioral performance, and that this relationship will be most pronounced for patients with relatively intact connectivity at baseline; later change will be observed in patients with more impaired baseline connectivity.
Aim 3
To identify if pre-treatment functional connectivity of the DLPFC cognitive control network, involving interactions with the sgACC and regions of the attention and default mode networks, and behavioral performance, along with early change in connectivity and behavior, are predictive biomarkers of clinical outcome. Our hypothesis is that baseline connectivity and behavior, and early and later changes in these measures will predict who at post-treatment have the greatest change in symptom severity, suicidality, and quality of life.
Statistical analysis
Power calculation
The sample size was calculated based on a conservative scenario in which we estimate a main effect for TMS on targets of interest of small effect size (Cohen’s d of approximately 0.25), and use a within-subjects linear model with session as a repeated measure, at least one moderator of interest (extent of baseline connectivity dysfunction) and three covariates. It is possible that effect sizes are larger. Dependent measures are connectivity, behavior, and self-report measures (3 measurement domains), and change in these measures, assessed in separate models. With an alpha level of 0.05 (two-tailed), power of 0.875 and an anticipated correlation of 0.5 for repeated sessions we require at least 98 veterans. To target 100 veterans and allow for the potential for dropout over the 3 sessions we aim to recruit 125 veterans. If a greater effect size is obtained, this would lead to greater power.
Data analysis plan
We will pursue a stepwise analysis plan that starts with a focus on our a priori regions of interest and builds to a model based on machine-learning of these regions, and then an exploratory phase, as follows:
For Aim 1, we will quantify functional connectivity in the resting and task conditions by computing correlation coefficients between the a priori regions of interest and converting these coefficients to Fisher Z scores. In task-evoked conditions we will also use gPPI to quantify connectivity incorporating the task contrast. Multivariate linear mixed models will be used to test the hypothesis that extent of connectivity is a response biomarker determining extent of post-treatment change in both connectivity and in clinical measures after 1 Week of treatment and Post-treatment (within-subjects). We will include both binary between-subjects and continuous moderators to assess whether veterans with intact pre-treatment connectivity show connectivity change after early sessions whereas veterans with impaired pre-treatment connectivity show change after completion of sessions. Correlational analyses will be used to test whether the extent of early change is proportional to the extent of baseline connectivity impairment.
For Aim 2, we will use the connectivity values established under Aim 1. Linear mixed models, with behavioral measures included as dependent variables, will be employed to test the hypotheses that extent of connectivity relates to extent of behavioral performance and that change in connectivity predicts change in behavioral performance (within subjects). We will include binary and continuous moderators to test if these relationships differ as a function of degree of baseline connectivity dysfunction after 1 Week and after Post-treatment sessions.
For Aim 3, we will again use the connectivity values established under Aim 1. Linear mixed models, with symptom, function, and suicidality measures included as dependent variables, will be employed to test the hypotheses that extent of baseline connectivity predicts severity of symptoms, functional disability and suicidality and that change in connectivity predicts changes in symptoms, function, and suicidality (within subjects). We will include binary and continuous moderators to test if these relationships differ as a function of degree of baseline connectivity dysfunction for after the 1 Week and Post-treatment sessions.
Interactions, covariates and cross-validation
Under each of these aims, we will evaluate if the interaction of the DLPFC-anchored cognitive control network with resting attention and default mode networks, and the sgACC, further moderates these relationships. In each analysis, we will model sex, medication, medication change, comorbidity, substance use, and premorbid function as covariates. We will employ cross-validation techniques, as used in our prior pharmacotherapy and imaging trials [
14,
73,
74], to assess the rigor of our inferences.
Exploratory analyses
We will pursue the following additional exploratory options: 1) covariation due to stimulation site quantified by our gel capsule method, 2) canonical correlations to quantify dimensional relations between connectivity, behavior and symptom, function and suicidality measures, 3) predictive regression models to further interrogate our hypotheses that circuit-behavior measures are response biomarkers of TMS response and predictive markers of functional/suicidality outcomes, 4) machine-learning methods to discover how our data form naturally organized clusters of TMS response trajectories. We will use principal components analysis for data reduction, clustering algorithms (e.g., hidden Markov models) to identify cohesive subgroups defined by circuit dysfunctions, and GLMs to assess if clusters are differentiated by behavior-symptom-functional-suicidality profiles and TMS-related clinical outcomes and, 5) a whole brain voxel-wise approach to quantify circuits and regions within circuits that might be missed by using a priori circuits and regions of interest.
Data management
Behavioral and cognitive data management
Data, once acquired, will be coded and given a generic indicator (e.g. 001). Individuals who are listed on the protocol will have access to all coded study data. Coded data will be shared with participating sites for data analysis. All research staff will undergo training from the lead investigator at each site, including the means through which confidentiality is maintained, the proper procedures as dictated by study protocol, and a review of any operating procedures that are important for data collection and veteran safety and security. Standard operating procedures will be shared with participating sites and an overall training log will be kept up to date to ensure collaborating sites are collecting data and running veteran visits in a standardized way. All information regarding prescription of the treatment parameters are monitored throughout the course of treatment and captured in the VA National Clinical TMS Program Quality Improvement Project. VA HIPAA authorization approved by IRB and embedded in the consent form allows for access to this clinical data.
Shared, coded data will be transferred through a secure file transfer software. The sharing of any PHI, if necessary, over the course of the study, will follow the coordinating sites guidance for best practice. The sharing of VA PHI will happen as permitted by VA HIPAA authorization embedded in VA consent.
MRI data management
MRI acquisitions will be transferred from each facility to the central facility at Stanford through secure transport. All MRI data will be anonymized including removing sensitive subject information and defacing structural images. The data storage system can only be accessed securely by certain investigators using two factor authentication. The storage system is scalable to large datasets and snapshots are recorded over time to prevent any possibility of data loss.
Data monitoring and safety reporting
A Data Safety and Monitoring Plan will be in place, consistent with standard protocols at the participating sites. Veteran recruitment, protocol compliance, and adverse events (AEs) will be tracked for each site on a semi-annual basis to monitor veteran safety, study progress, and efficacy; and to make recommendations for study continuation. All AEs will be recorded on standard forms and will indicate the severity, date of onset, and likelihood that the AE is related to a study procedure. The PI will ensure that all measures necessary to resolve the SAE are taken and that the Institutional Review Board is notified as soon as is practical in accordance with local institutional policy.
Discussion
Despite the wide scale adoption of repetitive transcranial magnetic stimulation, we still lack mechanistically-driven biomarkers designed to identify who is most likely to respond, and why. The identification of more precise solutions for MDD patients is imperative given that pharmacoresistant depression can be life threatening. Our study addresses this need through a systematic evaluation of brain circuit biomarkers in patients taking part in the VA Clinical TMS Program. We use a prospective design to evaluate cognitive control network connectivity as a predictive biomarker of the clinical effect of repetitive transcranial magnetic stimulation, and as a response biomarker of change with TMS.
Strengths
Innovations in our study design include 1) adequate power to interrogate imaging markers, 2) standardization to minimize variability, 3) implementation of a longitudinal design to quantify TMS-related changes in imaging markers, 4) integration of task-evoked and resting state imaging markers, and 5) establishing the foundations for expanding lessons learned to additional diagnoses and parameters.
Adequate Power
Our study will be the first pragmatic, large scale mechanistic trial to use functional connectivity neuroimaging and behavioral biomarkers of cognitive control as targets for response and prediction of outcomes in pharmacoresistant patients. Reflective of the emergence of TMS research, previous neuroimaging studies of TMS have employed small samples (mean n= ~ 24) [
10]. Our target of 100 patients will ensure statistical power to test our hypotheses.
Standardization to minimize variability
Our study will rigorously standardize TMS and neuroimaging protocols and the analytic pipeline (including stringent motion correction). Drawing conclusions about the utility of neuroimaging biomarkers has been difficult from current knowledge given the juxtaposition of small sample sizes and variability in neuroimaging/connectivity analysis, methods, and TMS delivery. By implementing a standardized approach in a well-powered sample, we will be in a unique position to parse variance due to biomarkers of interest versus variance due to other factors.
Implementation of longitudinal design
We will be obtaining fMRI scans at three time points during the course of TMS (baseline, after 1 Week and Post-treatment), while other TMS neuroimaging studies have obtained images at only one (Baseline) or two time points (Baseline and Post-treatment). Our approach will allow us, for the first time, to investigate whether change in functional connectivity of particular neural circuits in response to TMS may serve as an early biomarker (i.e., after only a few TMS sessions) of the subsequent effect of TMS. This information could enable clinicians to discontinue an intensive therapy for certain patients early in the TMS course, allowing the right treatment to be identified more quickly, aborting unnecessary side effects, and lowering the risk of patients disengaging from care due to frustration.
Task-evoked and resting state imaging markers
The majority of TMS neuroimaging studies have relied heavily on resting state imaging [
10]. Our cognitive control measures will allow us to probe our biomarkers of the DLPFC-anchored cognitive control network, elicited during GoNoGo and working memory tasks, and its connectivity with resting circuits involved in regulation. Additionally, if our broad hypothesis is correct that connectivity and behavioral performance changes are correlated, then behavioral measures may be used as proxies for neuroimaging data in clinical practice. Such a finding would offer a scalable TMS response biomarker that complements our mechanistic understanding based on neuroimaging measures of circuit connectivity.
Expanding lessons to additional diagnoses and parameters
Given dysfunction in cognitive control and associated circuits are transdiagnostic [
75], our findings will be a foundation for expanding to other psychiatric disorders in future trials. Our proposed sample will be sufficiently representative of the comorbidities in pharmacoresistant MDD patients to facilitate a future such transdiagnostic approach.
Limitations
Our study design also presents certain limitations including 1) the lack of a control group inherent in the observational design, 2) the definition of biomarkers, 3) the co-administration of neuromodulation and psychotropic medications, 4) presence of cognitive control dysfunction in our study population, 5) use of the Beam-F3 method for stimulation targeting vs. neuronavigation methods, 6) choice of stimulation site and parameters, and 7) the overrepresentation of older, male veterans in our population.
Lack of control group
Our study is observational and therefore lacks a control group. Veterans receive the standard TMS protocol for MDD as part of their clinical care through the VA’s Clinical TMS Program.
Definition of biomarkers
We follow the BEST (Biomarkers, EndpointS, and other Tools) resource of the FDA-NIH Biomarker Working Group for defining biomarkers [
76]. Although we do not have a means to randomize to a treatment control in the current pragmatic design, our stratification in the analysis of veterans based on extent of cognitive control meets the broad definition that a predictive biomarker identifies individuals who are more likely than similar individuals without the biomarker to experience a favorable or unfavorable effect from TMS.
Co-Administration of Neuromodulation and Psychotropic Medication
The combination of neuromodulation and medications used for pharmacoresistant major depression adds a degree of complexity to the current trial. We considered recruitment of medication-free veterans, but requiring veterans to be medication-free would not be feasible or ethical. Combined TMS and medications have been shown to be safe and efficacious in veteran patients [
77]. Thus, following prior TMS trials, clinical interventions will be stable for at least 6 weeks prior to TMS and during the study. Based on our prior experience, we anticipate medication changes during TMS will be limited. Should medication changes occur during the TMS treatment course, they will be recorded and post-hoc explorations will be performed to evaluate the effect on outcomes.
Presence of cognitive control dysfunction in our study population
Patients with pharmacoresistant depression may be the very individuals who demonstrate impaired cognitive control function. Consistent with a dimensional approach, we will undertake analyses based on the continuous degree of cognitive control dysfunction as well as seek to binarize the sample in subsequent analyses.
Use of the Beam-F3 method for stimulation targeting
We considered several approaches to target the DLPFC and elected to use individual scalp landmarks to determine the site of stimulation (i.e., Beam-F3 Method [
72]). This approach is recommended by the National Network of Depression Centers [
78] when frameless stereotaxy is unavailable or impractical. It offers significant advantages over the standard “5-cm rule” that often misses the DLPFC [
79]. We recognize that this method has limitations in and of itself, particularly if our goal was discovery of novel approaches to precision targeting. However, recent work indicates the Beam-F3 method provides a reasonable approximation compared to neuronavigation [
80]; thus, we consider it suitable for our purposes.
Choice of stimulation site and parameters
The pragmatic design of the proposed trial, the need for standardization, and the opportunity to leverage the large-scale Clinical Program necessitate a focus on reproducible parameters, namely 10 Hz DLPFC TMS, which has been the standard clinical protocol for MDD for nearly 10 years. Nonetheless, we recognize that the field is developing quickly. Thus, we anticipate planning thoroughly for future protocol expansions that would include consideration of alternative stimulation parameters, such as lower frequency, theta burst, or accelerated TMS. Our data on systematically evaluated patients will provide an important foundation from which to explore and compare new parameters.
Overrepresentation of older, male veterans
Reflecting the veteran population, we anticipate a preponderance of older male participants; however, our recruitment strategy will be targeted to ensure maximum possible recruitment of female veterans. However, because we are leveraging the VA TMS Clinical Program, our distribution will be reflective of the demographic mix within veterans referred to the participating clinics. Recent US Census data estimates the number of women veterans in the US to be approximately 9.2% of the total veteran population [
81]. We anticipate a similar proportion of women recruited for this study. Although veteran participants in TMS trials may on average be older than non-veteran participants, older age has not been found to be a predictor of poorer response to TMS in veteran patients [
82]. We will also explore secondary hypotheses that TMS-modulated brain-behavior targets are moderated by sex differences within the anticipated male/female distribution in the veteran population.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.