Methods

Some text in this chapter is reproduced with permission from Morriss et al.18 and Pszczolkowski et al.83 These are Open Access articles distributed in accordance with the terms of the Creative Commons Attribution (CC BY 4.0) licence, which permits others to distribute, remix, adapt and build upon this work, for commercial use, provided the original work is properly cited. See: https://creativecommons.org/licenses/by/4.0/. The text below includes minor additions and formatting changes to the original text.

Study design

This study was a multicentre parallel group, double-blind, randomised controlled trial, to test the efficacy of cgiTBS versus neuronavigated rTMS without connectivity guidance, in patients with a primary diagnosis of moderate to severe MDD, which was treatment resistant in their current episode (TRD). This study was carried out at five sites across the UK National Health Service: Nottinghamshire Healthcare Foundation NHS Trust (Nottingham site), Northamptonshire Healthcare NHS Foundation Trust (Northampton site), Cumbria, Northumberland, Tyne, and Wear NHS Foundation Trust (Newcastle site), Camden and Islington NHS Foundation Trust (London site) and Pennine Care NHS Foundation Trust (Oldham site). MRI scans were performed at Nottingham (for Nottingham and Northampton), Newcastle, London (baseline only) and Manchester (for Oldham). Participants were recruited through secondary care services across the five study sites, self-referrals, and patient identification centres compromising primary care services and secondary care services at neighbouring NHS trusts near the main study sites.

Ethical approval was granted by East Midlands Leicester Central Research Ethics Committee (ref: 18/EM/0232) on 30 August 2018. The study was registered on 2 October 2018 (ISRCTN19674644) under the public title ‘BRIGhTMIND: brain imaging guided transcranial magnetic stimulation in depression’. The trial protocol and Magnetic Resonance Imaging Protocol were published.18,83 Of note, the latter protocol provided details of changes to main study design because of COVID-19 (see Appendix 1 for more details). The trial was sponsored by the Research and Evidence Team, Nottinghamshire Healthcare NHS Foundation Trust.

Trial management

Study meetings with the research team occurred monthly for the duration of the BRIGhTMIND trial and was attended by the chief investigator, trial statistician, trial manager, patient and public involvement (PPI) lead, principal investigator (PI) from each centre and the site-specific research assistants (RA).

Data Monitoring and Ethics Committee

The Data Monitoring and Ethics Committee (DMEC) reviewed unblinded accumulating data on trial conduct and participant safety and reported their recommendations with regard to the trial continuing to the Trial Steering Committee (TSC). The DMEC had entirely independent membership. The trial statistician presented the study data unblinded in closed sessions, and the blinded progress and demographic data was reviewed in open sessions attended by the chief investigator, trial manager and imaging lead. In total there were eight DMEC meetings between (3 October 2018 and 25 January 2023).

Trial Steering Committee

The TSC would meet following the production of the DMEC reports. Although the majority of the TSC had independent membership, the meetings were attended by the chief investigator, trial statistician, PPI lead, trial manager, imaging lead and PI from each centre. In addition, the TSC and DMEC also reserved the right to meet independently to draw conclusions about the progress of the trial and deliberate their recommendations. In total there were eight TSC meetings between (2 May 2018 to 22 March 2022).

Internal pilot

The reporting of the internal pilot consisted of providing DMEC and TSC with the recruitment rates. Stopping guidelines were based on recruitment rate, with sites expected to recruit 3 participants per month within the last 3 months of the internal pilot (1 April 2019 to 30 June 2019) for a total of 36 participants. The study succeeded in meeting the recruitment target of 36 participants. The DMEC reviewed progress with recruitment and safety in July 2019 and recommended continuation of the trial, with the committee looking for an increase in recruitment with each site aiming for four participants per month.

Data management and quality assurance

All source data collected on participants case report forms (CRFs) was entered onto a MACRO system trials database (Elsevier UK, London, UK), set up and managed by the Leicester Clinical Trials Unit (LCTU). The MACRO data could be accessed by the LCTU statisticians, trial manager, site research staff, and TMS staff. Of note, unblinded information could be accessed only by the TMS staff and LCTU statisticians. For MRI scans, the subject digital imaging and communications in medicine (DICOM) session files were uploaded onto an XNAT (Washington University School of Medicine; https://www.xnat.org/) database infrastructure using anonymised subject numbers. The MRI scanning data were then downloaded to a University of Nottingham research drive that could be accessed only by the MRI imaging team. Qualitative interview transcripts were stored on a Nottinghamshire Healthcare NHS Foundation Trust research drive, with the research drive accessible only to the researchers at the Nottingham site.

The trial manager carried out monthly quality checks, alongside queries produced by the LCTU statistician regarding the data entered onto MACRO. Furthermore, quality assurance checks were completed for each of the research sites by the trial manager to ensure compliance with good clinical practice and scientific integrity. This included 4 checks for the London site, 7 checks for the Nottingham site, 6 checks for the Newcastle site, 4 checks for the Northampton site and 2 checks for the Oldham site. This included both physical and remote monitoring checks.

Protocol amendments

The substantial amendments made to the BRIGhTMIND protocol are provided in Appendix 1, Table 27. Amendments made that are important to note here firstly include substantial amendment 1, which occurred prior to project start. Due to a technological breakthrough by the TMS Company, neuronavigation was implemented rather than the manual BEAM methods that were proposed in the funded grant application. This neuronavigation system allowed for more reproducible accuracy of TMS stimulation and without this change the trial might have been criticised for not being as precise in terms of target stimulation as it was possible to be.

After recruitment to the trial commenced, there was a second substantial amendment for a change to the exclusion criteria. Initially, any benzodiazepine use was an exclusion and only the occasional use of other hypnotic drugs zopiclone, zolpidem, zaleplon and promethazine was allowed. This changed to the following exclusion criteria: ‘Daily prescription of benzodiazepine above 5 mg diazepam equivalents, zopiclone above 7.5 mg, zolpidem above 10 mg or zaleplon above 10 mg. These drugs should not be used intermittently in the 2 weeks before baseline assessment’. There was growing clinical opinion that intermittent hypnotic drug use rather than continuous hypnotic drug use might be associated with seizures or syncope following TMS.

In the third substantial amendment, the definition of treatment resistance was amended to reflect the scoring of the MGH-S. This changed from ‘who have not responded to treatment with at least two antidepressants in their current episode’ to ‘patients with TRD who have a MGH score of 2 or more, who are not responding to 2 different antidepressant regimes, antidepressant augmentation strategies, high prolonged dosages of antidepressants or ECT in their current episode’.

In the fourth substantial amendment, baseline imaging and associations with baseline clinical characteristics and changes in cognition, and a qualitative substudy to explore the reasons and meanings why participants request a copy of their MRI images, were added but are not reported on here.

In the fifth substantial amendment, the original sample size was reduced, primary clinical outcome was changed, and the study end date extended by 16 months. The sample size was reduced to 266 participants (133 per arm). This was due to site set-up and recruitment being delayed by 8 months to implement the neuronavigation technology. In addition, because recruitment to the study was paused for 5 months (March–August 2020), due to the COVID-19 pandemic and the funds and time available, it would not have been possible to recruit the original sample size of 368. Therefore, the primary outcome was also changed to ensure sufficient power (89.3%) with the new sample size stated. The primary outcome changed from ‘cgiTBS is more efficacious at 16 weeks than standard rTMS in patients with TRD as assessed by the proportion of patients who show a response [50% reduction in depression symptoms from baseline on the Hamilton Rating Scale for Depression-17 (HDRS-17)]’ to ‘cgiTBS is more efficacious in reducing the mean HDRS-17 score over 26 weeks compared to standard rTMS in patients with TRD’.

The Northampton site was closed as it was not possible to open the site after the COVID-19 pandemic and in the sixth non-substantial amendment, a fifth site ‒ Oldham, Greater Manchester ‒ was added instead, and commenced recruitment in August 2021. Study timelines are reported in Appendix 1, Table 28. The report from this point on refers to the final approved protocol.

Participant inclusion/exclusion criteria

Inclusion criteria

• Adults ≥ 18 years.
• Diagnosis of current MDD (defined according to the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition), that is treatment resistant defined as scoring 2 or more on the MGH-S.
• Have a HDRS-17 score of 16 or more (moderate to severe depression).
• Capacity to provide informed consent before any trial-related activities.

Exclusion criteria

• History of bipolar disorder (due to risk of mania) or depression secondary to other mental disorder.
• Neurological conditions for example, brain neoplasm, cerebrovascular events, epilepsy, neurodegenerative disorders, and prior brain surgery.
• Standard contraindications to MRI, that is irremovable metal objects in and around the body (e.g. cardiac pacemaker or implanted medication pump) and pregnancy (any doubt was resolved by a pregnancy test, with women of childbearing age taking precautions against pregnancy). This includes other potential complicated factors such as red tattoos, which consist of iron on the head, neck and back, and claustrophobia.
• Major unstable medical illness requiring further investigation or treatment.
• Change in prescribed medication 2 weeks before baseline assessment.
• Prescription of lamotrigine, gabapentin or pregabalin in the 2 weeks prior to baseline assessment.
• Daily prescription of benzodiazepine above 5 mg diazepam equivalents, zopiclone above 7.5 mg, zolpidem above 10 mg or zaleplon above 10 mg. These drugs should not be used intermittently in the 2 weeks before baseline assessment.
• Current substance abuse or dependence (defined by DSM-5 criteria).
• Prior TMS treatment.
• At risk of suicidality.
• Potential complicated factors relating to the TMS treatment, such as hairstyles that would impair magnetic transmission and piercings. Participants would be excluded only if they chose to not make the changes required to ensure effective treatment.
• Involved with any other clinical trial at the time of consent or 6 months prior.
• Unable to read or understand English.

A pre-screening questionnaire was used to telephone screen interested participants first, with potentially eligible participants invited to attend a baseline assessment. At the baseline assessment, all participants gave written informed consent and study eligibility was determined. Diagnosis was assessed by the Structured Clinical Interview84 for the Diagnostic and Statistical Manual of Mental Disorders (SCID), Fifth Edition.85 Treatment resistance was measured by the MGH-S86 (adapted for new treatment options ‒ see Appendix 1, Table 29). Depression severity and the primary outcome measure were assessed using the Grid version of the Hamilton Rating Scale for Depression-17 (GRID-HDRS-1787). The Childhood Trauma Questionnaire (CTQ88) was used in place of the post-traumatic stress disorder on the SCID. Information on sociodemographic was also collected at baseline assessment. Furthermore, to assist with determining study eligibility, medical and psychiatric history including a detailed assessment of treatment resistance were be obtained from secondary care mental health service case files where available and primary care notes.

Randomisation and blinding

Randomisation was conducted via a web-based randomisation system (Sealed Envelope, www.sealedevelope.com) and conveyed by the trial manager to the unblinded TMS nurse practitioners and healthcare assistant delivering TMS at each study site, immediately prior to the start of the participant’s first intervention session. Baseline eligibility was assessed by blinded outcome assessors who conveyed the information to the trial manager for randomisation and determination of treatment allocation. Participants were randomly assigned in a 1 : 1 ratio into the rTMS and cgiTBS arms. Randomisation was stratified by study site and minimised on severity of depression (HDRS-17 score 16-23 moderate or ≥ 24 severe) and degree of treatment resistance [low 2–3.5, medium 4–6, high (≥ 6.5) as assessed at baseline].

Participants, referring clinical teams and the outcomes assessors were kept blinded with respect to the treatment protocol assigned and administered, until after the participant’s final follow-up assessment. Any unintended unblinding of outcome assessors were recorded, with other assessors completing all further assessments for that participant. At each follow-up assessment, the outcomes assessor was asked to guess the participants’ treatment allocation.

Participants were to be randomised within 4 weeks of the baseline assessment and within 2 weeks of the baseline MRI scan. If the time window between baseline assessment and randomisation exceeded 4 weeks, then the HDRS-17 interview and MGH-S were reassessed prior to randomisation taking place. If the reassessed HDRS measure showed a total score < 16, treatment was still to be administered with intention-to-treat (ITT) analysis.

Interventions

A 70 mm figure-of-eight coil (E-z Cool coil) and a Magstim Horizon Performance Stimulator with StimGuide Navigated TMS Package (Magstim Company, Whitland, UK) was used for all treatments. A total of 3000 pulses were delivered in each rTMS or cgiTBS session, with 20 TMS sessions delivered over a 4- to 6-week period. Each session for both cgiTBS and rTMS lasted approximately 37.5 minutes in total for the purposes of blinding participants and assessors of outcome. MultimodalMRI was acquired for each participant at baseline (before TMS treatment) with T1-weighted (structural MRI) and task-free functional MRI (rsfMRI). The protocol paper outlines the brain imaging acquisition and analysis used to create personalised TMS targets for the BRIGhTMIND study.83 Interventions were delivered by a nurse and HCA at each session. They were trained to deliver the intervention and use the neuronavigation by study PIs and staff from Magstim plc. Treatment was delivered in specially adapted intervention rooms with a separate waiting area at each site. During the pandemic, no waiting was allowed, and additional hygiene and COVID-19 protection measures were employed.

Connectivity-guided intermittent theta burst stimulation treatment arm

Participants assigned to cgiTBS received bursts of three pulses (80% motor threshold) at 50 Hz applied at a frequency of 5 Hz (i.e. every 200 ms). Each 10-second cycle consisted of 10 bursts (consisting of 2 seconds of stimulation and 8 seconds’ rest) with a total of 20 cycles performed per run. The pulses were repeated for a total of 5 runs with 5-minute rest intervals between runs, with a total of 3000 pulses per session. The cgiTBS brain target was defined based on the Granger causality analysis of maximal strength of EC between the rAI and the lDLPFC from the participant’s rsfMRI and T1-weighted structural MRI scans. The StimGuide Navigated TMS Package computed the nearest location for TBS stimulus on the scalp from an individualised head model based on the structural MRI and three fiducial points ‒ the nasion, left preauricular and right preauricular sites.

Neuronavigated repetitive transcranial magnetic stimulation treatment arm

Participants assigned to rTMS followed the standard protocol approved by the US Food and Drug Administration. Stimulation was at 120% motor threshold with 75 × 4-second trains of 10 Hz interspersed by 26-second intertrain intervals, with a total of 3000 pulses per session. The rTMS site brain target was determined by the participant’s structural and rsfMRI scans by computing the voxel in the brain parenchyma that is closest to it [i.e. Montreal Neurological Institute (MNI) co-ordinates: x = −41.0, y = 43.0 mm, z = 32.0 mm] to target the F3 site over the IDLPFC. As with the cgiTBS treatment arm, the StimGuide Navigated TMS Package was used to compute the F3 site over the lDLPFC for TMS from the same individualised head model and three fiducial points mentioned above.

Percentage rMT was determined at the first treatment session and tested again on the sixth treatment session for both treatment arms.

Standardised steps (below) were developed for participants who were unable to tolerate the cgiTBS or rTMS protocols as per the MRI data: (1) StimGuide reset to 1 cm upwards or backwards from the site (within 2 cm diameter the original site); (2) StimGuide reset to 1 cm forwards of the site (within 2 cm diameter of the original site); (3) rMT reduced with StimGuide reset at the original site (as per MRI data.) including details of the new percentage rMT; (4) rMT reduced with StimGuide reset at 1 cm upwards or backwards from the site (within 2 cm diameter of the original site) including details of the new percentage rMT; (5) rMT reduced with StimGuide reset to 1 cm forwards of the site (within 2 cm diameter of the original site) including details of the new percentage rMT; and (6) revert StimGuide to standard F3 site (ensure allocated TBS or TMS is delivered). If treatment protocols were not administered as per the MRI data, then protocol deviations were recorded.

Further deviations were recorded if participants completed the 20 treatment sessions after 6 weeks and if there were gaps of more than 4 days between treatments in the course of 20. If MDD pharmacotherapy or psychotherapy changed prior to the 16-week follow-up being completed, these were also recorded as protocol deviations.

Follow-up intervals and assessments at each visit

The assessment of outcome measures was completed at 0 weeks (baseline) and 8-, 16- and 26-week follow-ups (± 1 week from randomisation). Researchers at each site collected the local data listed in the schedule of study assessments below. The primary outcome was observer-rated with all secondary outcomes self-rated. The baseline MRI scan was completed within 2 weeks of the baseline assessment, and the follow-up MRI scan was to be completed within 14 days of 16-week follow-up assessment, except at the London site where the 16-week scan was not completed because of the cost of the scan at that site (see Figure 1). Travel expenses were covered for participation in the study along with a £10 shopping voucher at 16- and 26-week follow-up assessments as a mark of respect and gratitude for the time and input of the participants to the follow-up aspects of the trial. Protocol deviations were recorded for any assessments completed out of the correct time frame or if components of assessments were not completed.

FIGURE 1 Study flow chart. (Figure not reproduced here; see the original article.)

Clinical efficacy outcomes

Hamilton Rating Scale for Depression-17

The primary outcome measure was assessed using the GRID-HDRS-17 over 26 weeks at 8-, 16- and 26-week assessments. The HDRS-17 is a widely used interview measure of depression symptoms and the GRID form was utilised given evidence of improved inter-rater reliability.87

The total score for the HDRS-17 was also utilised for secondary outcomes, assessed separately at each assessment at 8, 16 and 26 weeks. In addition, secondary outcomes also included the exploration of HDRS-17 response rate and remission rates. Remission rates at 8, 16 and 26 weeks were defined as a score of 8 or less on the HDRS-17. Responder rates at 8, 16 and 26 weeks were defined as a 50% or greater reduction in HDRS-17 score at each of the specified time points from baseline. Sustained response at 16 and 26 weeks was defined as a continuing response as defined above following a response at the previous time point.

Other clinical secondary outcome measures

To answer all further secondary outcome clinical efficacy objectives, we collected the following self-report measures at baseline and all three follow-up time points (8-, 16- and 26-week follow-ups), unless specifically stated otherwise below (see Table 2).

Depression measures included the Beck Depression Inventory, version 2 (BDI-II89), Patient Health Questionnaire-9 (PHQ-990), and Quick Inventory of Depressive Symptomology (QIDS-SR1691). Of note, the QIDS-SR16 was collected at all time points except for the 26-week follow-up and only from March 2020, as this measure was included specifically for a mechanistic substudy (substantial amendment 3). Results for QIDS-SR will therefore not be included in this report, which does not cover the substudy. The Generalised Anxiety Disorder Assessment (GAD-792) was used to measure anxiety symptoms, the Work and Social Adjustment Scale (WSAS93), was used to assess impairment and functioning, and the EuroQol-5 Dimensions, five-level version (EQ-5D-5L) visual analogue scale (EQ VAS94) was used as a measure of health outcome. Additionally, we adapted the patient global impression of change measure after review by the Lived Experience Advisory Panel (LEAP), which recommended from a 7-point scale with 2 ratings for being worse, 1 unchanged and 4 for improvement, to a shortened 5-point Likert scale (1 = much worse to 5 = much better with as many rating points for worse and improved mental state) to assess participants belief about the efficacy of treatment. This was assessed after each TMS session and at each follow-up time point.95

The THINC-Integrated Tool (THINC-iT) assessed subjective and objective cognitive functioning (THINC-it Task Force, http://thinc.progress.im/en). The THINC-it tasks were presented to participants on an iPad, with responses to be made by touching the screen. All tasks were defined as continuous outcomes. ‘Spotter’ is based on a choice reaction task (CRT96) viewed as a measure of sustained attention, with the primary outcome being mean response time (RT). ‘Trials’ is based on part B Trails Making Task (TMT97) viewed as a measure of executive function with a primary outcome of the total time to complete the task. CRT was recorded in milliseconds and TMT recorded in seconds, with better performances indicated by lower scores. The ‘codebreaker’ is based on the Digit Symbol Substitution Test (DSST98) whereby performance is considered dependent on the functional integrity of various cognitive skills, including working memory, attention, and executive function. ‘Symbol check’ is based on the one-back paradigm (N-back99) viewed as a measure of working memory. The primary outcome of interest for these two tasks was the total number of correct responses in 2 minutes, with better performances indicated by higher scores. Finally, the subjective Perceived Deficits Questionnaire, 5 domains (PDQ-5-D100) is a measure of subjective cognitive function, and the outcome of interest was the sum of scores from the 5 questions, with each question scored 1–5, with 1 being no concerns of any problem, to 5 being very significant problems. As a result, better performance is indicated by lower scores. From this point on, the more conventional cognitive task name or abbreviation will be used.

Each of these tasks were ran once at each of the participant’s assessments, except for the CRT, which was completed twice. The purposes of this were to facilitate ex-Gaussian analysis, in order to assess intra-individual variability (IIV) of sustained attention. Intra-individual variability was assessed only for the CRT task, with this considered a secondary outcome. Three parameters of an ex-Gaussian distribution (a combination of a Gaussian and an exponential function) are fitted to the distribution of RTs for a given individual.101 The three ex-gaussian parameters of IIV allow for a better separation of mean RT (mu), variability in RT (sigma) and abnormally slow responses (tau).102 In terms of interpretation, generalised slowing of RTs is reflected by higher values for all three parameters. We also analysed individual participant standard deviations (iSD) and coefficient of variation (CoV) – iSD normalised by mean RT, for comparability with other studies.

Adherence

The following variables were used to measure adherence with treatments:

1. Number of TMS sessions delivered as per protocol or as per changes allowed in the TMS standard operating procedure (SOP) within 6 weeks after randomisation.
2. The number of breaks between TMS sessions of more than 4 days as well as the average break length between TMS sessions of more than 4 days (measured in days).
3. Number of participants whose existing psychotropic medications or psychological interventions were not kept stable for 16 weeks for the duration of the trial, except for those at risk to themselves or others.
4. Number of participants whose daily prescription of benzodiazepine was above 5 mg diazepam, zopiclone was above 7.5 mg, zolpidem or zaleplon was above 10 mg from baseline assessment to the end of TMS treatment.

Power calculation for primary hypothesis for the clinical efficacy analysis

As agreed with the funders, TSC, and DMEC, we planned to enrol 266 participants, with 133 participants per treatment arm. A sample size of 266 participants would provide 89.3% power to detect a mean difference of 3 points in the HDRS-17 over 26 weeks between the groups at the 5% two-sided significance level assuming a SD of 8 informed from both the pilot study and a multicentre RCT in chronic persistent depressive disorder led by the chief investigator,58 a correlation between follow-up measures of 0.7 (1 baseline measure with correlation of 0.27 to the follow-up measures) and 20% data loss/drop-out. NICE defined 3 points as a clinically important difference in outcome on the HDRS-17 in its NICE Clinical Guideline for Depression in 2004.103 Given the uncertainties of recruitment to the study due to the COVID 19 pandemic, we note that under the same assumptions, a sample size of 253 would have yield 87.8% power and 232 participants 85.1% power.

Clinical efficacy analysis

The statistical analysis plan for the clinical efficacy is published (https://doi.org/10.6084/m9.figshare.21271140.v1) and we summarise the plan here rather than providing the detailed plan in its entirety. Statisticians at the LCTU conducted the analysis for all clinical outcomes detailed above, except for cognitive functioning as assessed by the THINC-iT. These data were analysed by the lead PI and researchers at the Newcastle site.

Primary analysis of the primary outcome

The primary analysis of the primary outcome was conducted on the ITT population with the multiple imputation technique being implemented to deal with missing data in instances where participants were missing any of the HDRS-17 scores. To ensure the population remained truly ITT, variables such as gender, ethnicity, age, centre, baseline HDRS-17 score and degree of TRD were used as predictors of the primary outcome to substitute the missing data with the predicted values from a multivariate normal regression equation. A total of 20 imputations were estimated.

Secondary and sensitivity analysis of the primary outcome

A completer’s population analysis and a per-protocol analysis of the primary outcome were completed using an available data approach. Sensitivity analysis included a Missing Not at Random (MNAR) assumption analysis was implemented using a control-based imputation approach, a pre-/post-COVID-19 period analysis, and a centre as random effect analysis, both of which were conducted on the ITT population using the imputed datasets from the primary analysis of the primary outcome. Furthermore, moderator analyses were conducted on the ITT population using an available data approach, with age, baseline CTQ, baseline HDRS-17, baseline GAD-7 and degree of treatment resistance as continuous moderator variables, and the number of TMS sessions completed (20 sessions or < 20 sessions) as a categorical moderator variable. Further details on these analyses can be found at https://doi.org/10.6084/m9.figshare.21271140.v1.

Mixed linear regression models were used for all the aforementioned analyses. All models were adjusted for centre (stratification variable), baseline HDRS-17 score, and baseline MGH-S for TRD (minimisation variables), visit number and a categorical variable for treatment arm (rTMS arm as reference).The pre-/post-COVID analysis also included a categorical variable for the pre-/post-COVID-19 period and an interaction term between the pre-/post-COVID-19 period and treatment arm and the moderator analyses also included two interaction terms: one between moderator of the primary outcome and treatment arm; and another between each of the follow-up time points and treatment arm. Participant identification number (ID) was included as random effect for all analyses, except for the centre as random effect analyses, whereby the centre was included as the random effect and not as a predictor variable. The variance-covariance matrix for the random effects was assumed to be unstructured and the models were estimated using restricted maximum likelihood. Treatment comparison estimates are presented for all models as adjusted mean differences between the treatment arms, with two-sided 95% CIs and p-values, with statistical significance considered at ≤ 5%.

Analysis of secondary outcomes

1. Analyses were conducted on HDRS-17 measured separately at 8, 16 and 26 weeks, conducted on the ITT population using an available data approach.
2. Analyses were conducted on the proportion of responders at 8, 16 and 26 weeks, proportion of remitters at 8, 16 and 26 weeks, and proportion of sustained responders at 16 and 26 weeks, on the ITT population using an available data approach.
3. Analyses were conducted on the total scores of the BDI-II, PHQ-9, GAD-7, WSAS and EuroQol-5D-5L (VAS score) across 26 weeks for the ITT population using an available data approach.

Binary logistic models were used for the responders, remitters and sustained responders’ analyses, and linear regression models used to assess HDRS-17 at the separate 8-, 16- and 26-week time points. All models were adjusted for centre (stratification variable), baseline HDRS-17 score and baseline MGH-S (minimisation variables) and a categorical variable for treatment arm (rTMS arm as reference). Mixed linear regression models were used for the other secondary continuous outcome variables, with each model adjusted for their respective baseline measure. Treatment arm, treatment centre (stratification variable), baseline HDRS-17 score, and degree of treatment resistance (minimisation variables) were fitted as fixed effects in the models, while the participant ID was fitted as a random effect. Treatment comparison estimates are the same as reported for the primary analyses of the primary outcome, except for the binary logistic models reporting adjusted odds ratios.

As reported in Table 2, prior to the suspension of recruitment to the study due to the COVID-19 pandemic, the THINC-iT was measured at the baseline assessment and all three follow-up time points. Between this suspension date and the reopening of sites to recruitment (30 April 2020 to 1 August 2020) no face-to-face assessments were completed, with the THINC-iT not collected at any follow-up time points during this time period. From 1 August 2020 onwards, the THINC-iT was collected at the baseline MRI scan and 16-week follow-up MRI scan, due to the COVID-19 mitigation restrictions in place. Therefore, the analyses conducted and reported in Chapter 3 only includes the THINC-iT outcomes at baseline and 16 weeks.

Linear mixed models (LMMs) were used for the cognition analyses. The cognition outcomes were the dependent variables, and the independent variables of interest were the THINC-it time point (baseline and 16 weeks), treatment group (rTMS and cgiTBS), baseline GAD-7, baseline HDRS-17, and change in HDRS-17 score between baseline and week 16. These models also included three interaction terms: treatment group*time point, time point*change in HDRS-17 and change in HDRS-17*treatment group*time point. Confounder variables included age, gender, site, and MGH group. Any confounder variable not found to be significant on initial testing were removed from the models and the analysis re-run, with the results from these re-run analyses presented in Chapter 3.

The analyses were conducted on the ITT population (all participants randomised to treatments) using an available data approach. However, important to note from a quality control (QC) perspective, outliers were removed for any scores on the CRT, DSST or N-back (primary outcomes) that were four times the SD for the whole cohort at that time point on that task.104 Similarly total time to complete the TMT of over 5 minutes was removed. Additionally, if any individual did not have at least 3 of the 4 objective tasks scores meeting quality control (QC) (or not completed), then they were completely removed from that time point in the analysis.

Finally, if a participant only completed the CRT task once, the data from the completed task was used with the following provisos:

1. The RT results are reported as described above based on the one running of the task. The participant was not included in analysis comparing the first and second running of the task.
2. The number of correct responses obtained was doubled prior to inclusion in the group analysis.
3. The participant’s data at this time point were not included in the IIV analysis due to the limited number of responses available.

If the ex-Gaussian curve fitting routine was unable to fit a participant’s data at a specific time point, then values for mu, sigma and tau were set as missing data for that time point, with a record that these missing data were as a result of an inability to fit the data rather than missing or excluded CRT data.

Finally, regarding participants’ global impression of change, these were defined as categorical outcomes, reported descriptively, and listed by treatment arm and overall.

Safety, acceptability, facilitators and barriers

Safety outcomes

1. Internationally agreed definitions were adopted for AEs (i.e. any untoward medical occurrence in a clinical trial subject administered TMS whether or not it has a causal relationship with TMS) and SAEs (any AE or adverse reaction that results in death, is life-threatening, requires hospitalisation or prolongation of existing hospitalisation, results in persistent or significant disability or incapacity, or is a congenital anomaly or birth defect).
2. Common TMS side effects (i.e. headaches, neck pain, scalp discomfort, tinnitus, dizziness, jaw ache, nausea, and watering eyes) as well as reports of any uncommon side effects were recorded on an adverse checklist after every TMS session and at the 8-week follow-up assessment.
3. Syncope was recorded as an AE, unless participants were admitted to hospital, in which case it was defined as an SAE.
4. In the event of a seizure, treatment was stopped, and no further treatment was administered. Seizures were reported as SAE.
5. Any participant found to be at risk to themselves (suicide, neglect) or others, or developing a SAE were referred to relevant clinical services. A review by a clinical expert in TRD was offered to any participant whose depression has become more severe at 16- and 26-week follow-ups for safety reasons.
6. The DMEC reviewed all unblinded accumulating data on trial conduct and participant safety and reported their recommendations with regard to the trial continuing to the TSC.

Facilitators, barriers, and acceptability outcomes

Patient acceptability was assessed with a purposively designed 5-point Likert measure rated from 1 = unacceptable (negative effects outweigh benefits) to 5 = acceptable (beneficial effects outweigh the negative effects). This was assessed at every TMS session and for all follow-up assessments. Participants were also given the option after each TMS session and each of the follow-ups to provide any additional open-ended comments about their experience of the TMS or the study so far, which were utilised for the qualitative analysis.

Semistructured topic guides were developed to qualitatively assess facilitators, barriers, and acceptability of TMS. For participants who were randomised to treatment and either completed all follow-ups or withdrew during or after treatments, topic guides included five broad questions: (1) how they found out about the trial; (2) their initial reaction on hearing about the trial and treatment involved; (3) the factors that motivated their decision about participation; (4) whether they had any concern about taking part in the trial; and (5) the acceptability of the treatment. Those who were not randomised to treatment and declined to participate in the trial had a topic guide which included the first three questions above, and whether they thought there would have been any potential positive outcomes from taking part in the study (see Appendix 1, Table 30).

Sample size for qualitative facilitator, barriers and acceptability analysis

We aimed to achieve a purposive sample of 25–30 participants or interviewing until saturation of data was achieved (i.e. no more themes were emerging in subsequent interviews). We aimed to include a variety of participants, reflecting a mix of demographic characteristics, consent, or non-consent to participate, adherence and non-adherence to treatment and follow-up.

Safety and acceptability analysis

Statisticians at LCTU conducted analysis of the safety outcomes and quantitative acceptability. Qualitative analysis was conducted by the qualitative PI and researchers at the Nottingham site.

Common and uncommon side effects were summarised by treatment arm and overall using data collected at 8, 16 and 26 weeks as well as at each TMS treatment session. The overall proportion of participants with any side effects in addition to the frequency and duration (in days) of all side effects were also reported by treatment arm and overall. All safety data were presented according to the safety population. All AEs were listed by treatment arm and overall based on the following characteristics: seriousness; duration (measured in days); relatedness; severity; action taken and outcome.

Quantitative measures of patient acceptability were defined as categorical outcomes reported descriptively and listed by treatment arm and overall.

All qualitative interviews were completed by BRIGhTMIND RAs or members of the clinical research network. Interviews were anonymously recorded on an encrypted digital voice recorder then transcribed using the professional transcription service DICT8 (https://www.dict8.com/). Qualitative interviews were combined across the two treatment groups and an inductive thematic analysis using a grounded approach was adopted.105 The first 8 transcripts were read and coded by two researchers to ensure they were immersed in the data.106 One researcher read and coded all further transcripts. The two researchers agreed consensus that data had reached saturation as no new or meaningful codes emerged after 12 transcripts.107 However, some further interviews were conducted to see if any new data was identified due to the COVID-19 pandemic. Initial codes were then summarised into broader categories. Consultations with the qualitative PI took place, whereby the overarching framework of themes and naming of themes were produced and refined. The framework of themes and supporting quotations were also presented to the BRIGhTMIND PPI team across the sites, with PPI contributors aiding the interpretation of said themes. This allowed us to check the credibility of the analysis and results, further improving the trustworthiness of the results.108 Coding and themes were input into a framework matrix using NVivo 12 (QSR International, Warrington, UK).

Inductive thematic analysis was also used to analyse the optional comments that were recorded on the quantitative acceptability measures mentioned above, in order to determine if any further subthemes/themes could be identified outside of the qualitative interviews. These comments were read repeatedly and coded by one researcher, with a second researcher reviewing these codes for reliability. Following this, the framework and naming of themes was further refined and approved by the qualitative analysis team.

Economics evaluation

The economic evaluation was designed and analysed by health economists at the University of Nottingham. The health economics analysis plan can be found at https://figshare.com/articles/online_resource/BRIGhTMIND_Health_Economics_Analysis_Plan/22093148 and in Appendix 1, Table 31. The CHEERS checklist is also found in Appendix 1, Table 32.

Resource use and costs

The viewpoint for the study is broad as such data were collected on health care and social care at baseline and during the treatment and intervention regime, but in addition sought to collect societal costs such as productivity (days off work where appropriate) and the direct patient cost of illness. Resource use data were collected at baseline, 16- and 26-week follow-up time points using a purposely designed patient-informed resource use questionnaire with PPI input. Baseline forms recorded resource use from 3 months prior to trial entry. The resource use questionnaire is collected directly in clinic and has the advantage that research nurses can assist patients should any confusion arise. Follow-up forms collected resource use between assessment periods (e.g. collection at week 16 represents resource use utilised between baseline and week 16). Costs reflected Great British pounds at 2020–1 prices and were divided between NHS and personal social services (PSS) (health service) perspective costs (in accordance with NICE guidance109) and broader societal costs. Health service costs include intervention-related costs; inpatient hospital service costs; outpatient service costs; primary care costs; community service costs, and mental-health-related medication costs. Broader societal costs include participant out-of-pocket expenses from travel, over-the-counter medication, and private services, together with societal productivity costs.

Intervention costs were calculated on an ITT using a micro-costing framework comprising three elements: imaging costs; staff costs, and equipment costs. Imaging costs included all structural and resting-state functional MRI scans provided to participants before randomisation.1 MRI scans were pooled between arms given their a priori nature (i.e. by definition unrelated to allocation). MRS scans were deemed unrelated to treatment (done for research purposes only), hence not considered in the costing. Since rTMS does not necessitate MRI scans, a scenario analysis considers removing all imaging costs from rTMS. Staffing costs equated to the number, band, and time of healthcare professionals dedicated to treatment appointments, as recorded from site reports. Staff costs for each participant were calculated by multiplying the number of treatment sessions attended (protocol: 20 sessions) by the treatment-specific average staff appointment cost. COVID-19-related expenses (i.e. advanced personal protective equipment, cleaning, etc.) were omitted due to their exceptional nature and general equality between arms. The staff and equipment logs of delivering treatment were checked using a study log completed by a purposive sample of nurses and healthcare assistants. Since trial protocols insisted on consistent contact times, a second intervention-related scenario considered a 30-minute time saving for cgiTBS relative to rTMS informed via published expert opinion.62 A third scenario jointly considered removing all imaging costs from rTMS and the 30-minute time saving between cgiTBS and rTMS appointments. Treatment equipment costs were informed via financial records from the trial and were distributed across all participants on an ITT basis using the average site throughput (n = 51). Participants randomised to cgiTBS and rTMS received treatment from the same machines, incurring the same equipment cost. Base-case intervention costs only differed between arms via differences in treatment attendance and staff requirements observed. All intervention-related unit costs are provided in Appendix 1, Table 33.

Inpatient, outpatient, primary care, and community services were costed according to the number of recorded attendances, multiplied by relevant unit costs from NHS Reference costs and Personal Social Service Research Unit (PSSRU).110,111 Unit costs were inflated to 2020–21 prices using the PSSRU pay and prices index where necessary. All ‘other’ recorded cases in each healthcare setting were costed on a case-by-case basis. All procedures were assumed to be provided by the NHS unless stated otherwise. Medication costs considered all ‘current’ mental-health-related medications participants recorded at baseline, and at 8-, 16- and 26-week follow-ups. Since it was unknown when changes in participants’ medication schedules occurred between assessment periods, medication costs were interpolated between follow-up periods using drug costs from eMIT and BNF databases.112,113 The analysis assumed no drug wastage and that medications taken ‘as required’ were applied fortnightly.

Broader societal out-of-pocket travel costs, over-the-counter medication costs, and private service expenses were taken from the costs directly reported by participants during the trial. Participants driving to and from healthcare settings were assumed to incur a £0.45 cost per mile.114 Productivity costs were calculated using the human capital approach, specifically combining self-reported depression-related absenteeism with salary costs from the Office for National Statistics.115 In the absence of occupation or earnings data for informal carers; informal care was not costed but instead considered as a secondary outcome (see Chapter 2, Economic outcomes).

Table 3 provides an overview of all the resource use and costs considered in the economic evaluation. All specific health service and broader societal unit costs can be found in Appendix 1, Table 34.

Economic outcomes

The primary outcome used in the cost-effectiveness analysis was QALYs, a generic health measure in which the benefits, in terms of length of life, are adjusted to reflect quality of life.116 One QALY is equal to 1 year of life in perfect health. The health-related quality of life (HRQoL) of trial participants was assessed using the EQ-5D-5L, a questionnaire asking individuals to rate their health according to five broad health dimensions (mobility, self-care, usual activity, pain/discomfort and anxiety/depression) at five levels of severity (no problems, some problems, moderate problems, severe problems or extreme problems).117 The EQ-5D-5L has 3125 possible health states, each defined according to a unique combination of severity ratings.118 The HRQoL associated with each health state is calculated through a preference-based weighting derived from a representative sample of the UK population.117 Study outcomes were assessed at baseline and at week-8, week-16 and week-26 follow-ups.

Consistent with UK guidance for economic evaluations, HRQoL weights were estimated from a published mapping of EQ-5D-5L responses onto HRQoL values based on a UK population survey for the EQ-5D-3L instrument.109 Mapping was conducted using the Stata^® (StataCorp LP, College Station, TX, USA) package eq5dmap.119 To evaluate differentials and dynamics in HRQoL during the trial, EQ-5D responses were descriptively assessed at every assessment period by domain and severity strata. Average observed and imputed HRQoL preference values were also compared between arms. QALYs for each participant were calculated using an area under the curve approach using linear interpolation in HRQoL between time points. Base-case cost-effectiveness analyses considered QALYs using mapped EQ-5D-3L preference values.109 Cost-effectiveness findings using EQ-5D-5L preference values will be considered in a scenario analysis.

Informal care requirements were considered as a secondary outcome and were assessed via descriptive comparisons of informal care needs during the trial. Informal care was defined as care support received via friends or family only and was assumed constant between follow-ups.

Economic evaluation

The economic evaluation compared health service and societal costs both within resource categories and those aggregated into total costs, alongside EQ-5D responses, QALYs, and informal care. In line with UK guidelines on assessing cost-effectiveness, health outcomes were expressed as QALYs and costs measured in Great British pounds (2020–1) from health sector and broader societal perspectives (i.e. considering factors beyond the health service). Cost, QALY and cost-effectiveness comparisons are presented using unadjusted (using observed data only) and adjusted analyses (considering imputed data with econometric modelling controlling for patient co-variables). Cost-effectiveness was assessed using incremental cost-effectiveness ratios (ICERs) and incremental net monetary benefits (INMB).120 Uncertainty was evaluated using probabilistic sensitivity analysis and explored further via scenario and threshold analyses.

Incremental cost-effectiveness ratios are a measure of the additional cost per additional unit of health gain (QALY) generated by one intervention compared with another. INMB represent the monetary value from the health gains of a strategy less that which would have otherwise been generated elsewhere, had the same resources been allocated for alternative purposes. This opportunity cost is defined by the cost-effectiveness threshold (k) and represents the cost per unit of health gain (QALY) at which a technology represents an efficient use of limited NHS resources, otherwise described as the marginal productivity of the healthcare system. Interventions with positive INMBs and those with ICERs below the threshold are deemed cost-effective when compared to the next best relevant comparator.120

To increase efficiency and reduce the risk of bias, missing cost and outcome data were imputed.121 To best estimate predicted values, and to keep imputed values within their natural bounds, missing values were populated using multiple imputation by chained equations with predicted mean matching.122 The imputation model included variables that are associated with both the missing data and CEA outcomes (cost-level and EQ-5D-level variables), and includes all covariates used in the adjusted cost-effectiveness analysis (see below).121 The number of datasets (n = 20) was informed via the proportion of missing data. All subsequent analyses of multiple data sets followed Rubin’s rules.123 Both observed and imputed outcomes and costs were evaluated. Observable total costs and QALYs were presented on a complete case basis (i.e. data required across all follow-up periods). The time horizon of the analysis was 26 weeks, the length of the BRIGhTMIND study. Costs and QALYs were not discounted as they accrue within a 12-month period.

Cost-effectiveness was assessed using two perspectives, the first from a health service perspective in line with the NICE guidelines (NHS and PSS), and the second from a broader societal perspective incorporating productivity costs and all travel, private service, and over-the-counter medication costs borne by participants. Cost-effectiveness analyses were also conducted using two separate analytical approaches: an unadjusted analysis using observed trial data; and an adjusted analysis using multiply imputed data and regression methods to estimate costs and QALYs while controlling for a set of relevant participant co-variables. From these approaches, the adjusted health service perspective analysis was considered the primary analysis, from which scenario and sensitivity analyses was extended. Table 4 summarises the four analytical approaches taken for assessing the cost-effectiveness of cgiTBS and rTMS.

Unadjusted cost-effectiveness analysis

Unadjusted cost-effectiveness findings compared mean total costs and QALYs observed in BRIGhTMIND over the trial horizon. To represent the uncertainty about observed mean values, 5000 bootstrap replications were produced reflecting the variability in mean estimates.124 The sampling distribution for total costs and QALYs consisted only of observed complete cases.

Adjusted cost-effectiveness analysis

For the adjusted cost-effectiveness analysis, costs and outcomes were estimated using multiply imputed data and regression methods. Regression analyses controlled for treatment allocation, participant age, gender, ethnicity, site, and baseline depression status (HDRS-17 severe vs. moderate). To control for broader imbalances in baseline HRQoL between arms (i.e. beyond that measured by the HDRS-17), QALY regression analyses also controlled for baseline EQ-5D preference values.125 Ordinary least square regressions were applied for the QALY analysis while generalised linear models were used for the cost analysis, accounting for the non-normality and typically skewed nature of the cost data (gamma family distribution selected on the basis of superior model fit). Total broader societal costs were modelled separately on the log scale. Uncertainty in the estimated regression coefficients was estimated using Monte Carlo simulation with 5000 simulations, assuming multivariate normality of the regression coefficients and applying a Cholesky transformation to the variance-covariance matrix to account for correlation.126 Considering the statistical analysis identified no moderator variables (factors identified as interacting with treatment effectiveness), no subgroup analyses were considered.

Cost-effectiveness

For each analytical approach cost-effectiveness was assessed according to ICERs and INMB. The probability that an intervention is cost-effective was calculated via the proportion of simulated costs and QALYs that would be judged as cost-effective. The probability of being cost-effective and INMBs are calculated at cost-effectiveness thresholds of £20,000 and £30,000 per QALY, the range used by NICE.109 The probability that cgiTBS would be considered cost-effective was plotted in a cost-effectiveness acceptability curve with threshold values up to £50,000 per QALY considered.127 The expected value of perfect information was also evaluated (i.e. the expected net monetary value from attaining perfect information to inform decision-making).120

In the absence of a placebo/sham arm in BRIGhTMIND, a hypothetical scenario analysis considered pairwise comparisons of cgiTBS and rTMS with an artificial placebo. The artificial placebo was assumed to maintain average baseline HRQoL (mapped baseline EQ-5D-3L) and incur the same health service costs as its comparator, thereby contrasting intervention costs with the improvements in HRQoL observed in the trial (compared to what could be considered broadly representative HRQoL in the absence of intervention). Conclusions from this scenario must be interpreted with caution given the largely hypothetical nature of the artificial placebo.

All costs, outcomes and exploratory analyses considered in this analysis were reviewed and informed by the LEAP, comprising members with personal experiences of depression and treatment.

Mechanism-of-action outcomes

Magnetic resonance imaging acquisition

Six MRI scanners operating at 3T were used during the BRIGhTMIND study: Newcastle – Achieva dStream (Philips); London – Prisma (Siemens); Oldham – Achieva (Philips); Nottingham and Northampton used the following three scanners at Nottingham (sequentially) over the course of the trial owing to a scanner upgrade carried out midtrial: (1) Discovery MR750 (GE Healthcare); (2) Ingenia (Philips); and (3) Premier (GE Healthcare). Participants’ baseline and 16-week MRI scans (London site underwent scanning at baseline time point only) were carried out at the same site and using the same scanner platform, using a core protocol across the treatment sites. Details of the multimodal scanning sequences can be found in the BRIGhTMIND Magnetic Resonance Imaging Protocol.83

All participants underwent high-resolution structural T1-weighted scans using sagittal fast-spoiled gradient echo BRAVO (or equivalent) sequences with 1 mm³ isotropic voxels covering the whole head from the vertex to the neck. All participants also underwent eyes open rsfMRI scans using a gradient echo EPI sequence aligned with the anterior commissure-posterior commissure line, with acquisition covering from the vertex downward (repetition time [TR]/echo time [TE] = 2000/32 ms; flip angle = 77°; 35 slices; voxel size = 3 mm³; slice gap = 0.5 mm; field of view = 192 × 192 mm; interleaved bottom/up; 240 volumes; phase encoding direction = posterior > anterior). Directly before the rsfMRI, forward- and reverse-phase-encoded blood oxygenation level dependent (BOLD) sequences were also acquired with the same image dimensions as the rsfMRI, to facilitate distortion correction. Additionally, participants at Nottingham, Newcastle and Northampton also underwent a Mesher-Garwood Point Resolved Spectroscopy (MEGA-PRESS) MRS scan acquired using the same schema described by Mikelsen et al. (voxel dimensions are 45 × 30 × 20 mm for anterior/posterior [A/P], left/right [L/R], inferior/superior [I/S] directions, respectively; TR/TE = 2000/68 ms; 320 averages).128

Magnetic resonance imaging analysis

The mechanism-of-action analysis was completed by researchers at the Nottingham site. Statistical analysis plans for the functional and effective outcomes are available online at http://doi.org/10.17639/nott.7251 and the MRS outcomes available at http://doi.org/10.17639/nott.7275.

Functional and effective connectivity processing

For QC of the structural and functional MRI scans, we used the MRIQC v0.11.0 QC XNAT container. Exclusion criteria included: average and maximum framewise displacement of BOLD images > 1 mm and 3 mm, respectively; BOLD images showing long-lasting intensity changes or image artefacts; and T1w images that did not contain whole head coverage and contained incidental findings or image artefacts.

Pre-processing of the structural and functional MRI scans was carried out using the Sir Peter Mansfield Imaging Biomedical Research Centre (SPMIC-BRC) pipeline, version 1.5.5, 23 August 2021 (github.com/SPMIC-UoN/BRC_Pipeline). Pre-processing steps were the same as those used during the trial for computing cgiTBS target co-ordinates, as detailed in the trial MRI protocol.83 This is except for the use of advanced Eddy current and movement correction tool (EDDY) slice-to-volume and volume-to-volume motion correction,129 which was used in place of the Motion Correction using FMRIB’s Linear Image Registration Tool (MCFLIRT) six degrees-of-freedom volume-to-volume correction used in the trial. The SPMIC-BRC pipeline is based on tools from Statistical Parametric Mapping toolkit version 12 (SPM12), the functional and structural MRI software library (FSL), and Freesurfer. EDDY motion correction uses the NVIDIA produced accelerated computing toolkit (NVIDIA CUDA). Further details on the structural and functional pre-processing steps can be found in the resting-state fMRI analysis plan (http://doi.org/10.17639/nott.7251).

Functional connectivity analyses were conducted using the FSL and Granger causality analysis was computed using REST (resting state fMRI data processing toolkit) software.130 Correlations of interest were Fisher’s r-to-z transformed with the converted Z-scores serving as measures of FC. Effective connectivity values were also converted to z-scores using Fisher’s r-to-z conversion, with the mean z-score output across all voxels in y (when considering the influence of x on y) or x (when considering the influence of y on x) serving as the measure of EC. The rAI-to-lDLPFC EC (rAI ‘outflow’) minus the lDLPFC-to-rAI effective connectivity (rAI ‘inflow’) was used to calculate net out flow from right AI to IDLPFC, as used in Iwabuchi et al.43

For all connectivity analyses regions of interests (ROIs) were 6 mm spheres with MNI co-ordinates reported. For the EC analysis, the seed regions for the rAI was (x = 30, y = 24, z = −14131) and the individual DLPFC cgiTBS co-ordinates for each participant (regardless of treatment group). For the DLPFC-DMPFC FC analysis, the seed regions were the anterior DLPFC (x = −44, y = 40, z = 2967), posterior DLPFC (x = −44, y = 22, z = 36132) and DMPFC (x = −7, y = 49, z = 18133). For the DLPFC-sgACC FC analysis, the seed regions were the sgACC (x = 6, y = 16, z = −10134), and the target co-ordinates used based on the treatment participants received. For participants who received rTMS, their rTMS co-ordinates were projected 6-mm deeper into the brain,135 giving MNI-co-ordinates x = −38, y = 40, z = 29.

Magnetic resonance spectroscopy processing

We originally planned to perform the entire MRS analysis for GABA and Glx estimation using linear combination model (LCM) as per the published protocol.83 However, during pre-processing it was noted that data from only one scanning vendor was compatible for the in-house analysis pipeline. Therefore, due to resource and time constraints, we instead used GANNET MATLAB (multi-paradigm computer programming language) package (V3.3.0) for GABA estimates with the default setting.136 For Glx estimates for best-practice estimation from off spectra, instead of using GANNET MATLAB we ultimately implemented LCM for all vendor spectra as the preferred approach for reliable Glx metrics derived from MEGA-PRESS.137

In brief for estimation of GABA+ (without macromolecule suppression), as per published protocol all spectra were visually inspected by an experienced researcher blinded to any clinical or treatment arm information, followed by GANNET processing that uses time domain signals that were frequency and phase corrected by spectral registration. A 3-Hz exponential line broadening and zero-filling with a factor of 16 was applied for filtration. Each pair of spectra (edited – non edited) were aligned, and edited spectra was subtracted from non-edited spectra to detect a GABA peak. The processed spectra were averaged and fit to Gaussian function to the GABA and Glx peak at 3.0 and 3.8 ppm, respectively. GABA peak was modelled using a five-parameter Gaussian model (2.19–3.55 ppm) and water peak was modelled using a Gaussian-Lorentzian function. Creatine (Cr) and N-acetylaspartate were modelled as two Lorentzians in the non-edited spectra. The integral Cr area in non-edited spectra were used to calculate GABA+ ratio relative to total (t)Cr. For GABA QC measurements with > 10% water referenced GABA+ fit error138 were excluded, and negative measurements were also excluded.

For Glx estimation, LCM was used limiting the analysis to off spectra. Different from edited MRS approach for GABA, Glx fit results at 3T are highly dependent on the MRS quality, especially shim and baseline results, and expected to vary between scanner platforms. We hence applied a more stringent QC by visual inspection for Glx spectra (assessing issues with line width, peak separation, water suppression, baseline model and bias in fit residuals based on LCM outputs) undertaken by an MRS experienced researcher fully blinded to the clinical details and treatment assignment of subjects but aware of the pre- versus post-status of the spectra. This in-depth QC process was limited to available spectra pairs pre- and post-completed intervention for Glx only and also included a qualitative and quantitative review of the anatomical position versus target, and achieved anatomical overlap of the spectrum position before and after treatment after co-registration of MRS with anatomical MRI in native and template space.

Further details on the MRS processing and analysis can be found in the MRS analysis plan (http://doi.org/10.17639/nott.7275).

Power calculations for the primary hypotheses for the mechanistic imaging analysis

Hypothesis: Reduction in IDLPFC – left DMPFC FC from baseline to 16 weeks will be associated with greater clinical improvement (averaged across all post-treatment time points), across both treatment groups.

A sample size of 120 achieves 96.3% power to detect a difference of −0.3 between the null hypothesis of a correlation between clinical improvement and change in FC of 0.2 and the alternative hypothesis correlation of 0.5 using a two-sided hypothesis test with α = 0.05. The null hypothesis of 0.2 represents a very weak or weak correlation, whereas the alternative of 0.5 represents a moderate correlation.139 The correlation between DLPFC-DMPFC FC change at 12 weeks and HDRS-17 change at 12 weeks in the pilot data was 0.58.

Hypothesis: Greater clinical improvement on 17-item HDRS, averaged across all post-treatment time points (8, 16, and 26 weeks), will be associated with more positive baseline EC from right AI to IDLPFC, and this relationship will be stronger in the cgiTBS group.

Sensitivity analyses are presented to evidence that the revised sample size for BRIGhTMIND will still allow for testing of these hypotheses. Sensitivity analysis based on a one-tailed test with α = 0.05, 90% power, and a sample size of 232, suggests a required effect size of |ρ| = 0.19. This correlation of EC from the insula-DLPFC with changes in HDRS-17 at 12 weeks in the pilot data was −0.26.

Hypothesis: Changes in depression symptoms on 17-item HDRS (averaged over all three-post treatment-time points at 8, 16, and 26 weeks) will be associated with change in GABA between baseline and 16 weeks (post minus pre-treatment) in both treatment groups.

Sensitivity analysis following the same formula but with a lower expected sample size of 111 (as outlined for DMPFC-DLPFC FC hypothesis), suggests a required effect size of |ρ| = 0.27. The correlation between GABA and HDRS-17 changes at 12 weeks in the pilot data was 0.68.

Functional and effective connectivity analysis

Linear mixed models were used for the functional and effective connectivity analyses. Analyses was conducted on the ITT population (all participants randomised to treatments), with participants excluded if TMS was delivered to unintended targets (primarily cases where TMS co-ordinates were uploaded incorrectly to neuronavigation software), or participants received fewer than 15 TMS sessions, or if the 20 TMS sessions were completed in > 6 weeks. All analyses included HDRS-17 time point as an independent variable and HDRS-17 treatment response as the dependent variable, with participant as a random effect. All analyses also included independent variables of baseline effective connectivity or FC, specific to the hypothesis being explored (e.g. baseline effectivity from right AI to IDLPFC). Change in connectivity analyses also included an independent variable of change in FC or EC again specific to the hypothesis being explored. For baseline connectivity analyses, models included three interaction terms: baseline EC or FC * HDRS-17 time point, baseline EC or FC * treatment group, and baseline EC or FC* HDRS-17 time point * treatment group, with the change in connectivity analyses including interaction terms based on change in EC or FC, rather than baseline connectivity. Confounder variables included age, gender, site and MGH group. Any confounder variable not found to be significant on initial testing were removed from the models and the analysis re-run, with the results from these re-run analyses presented in Chapter 3. All hypotheses were also tested with other measures of depression to test the robustness of the findings, with the HDRS-17 time point and HDRS-17 treatment response replaced by the HDRS-6 (6 of the 17-item HDRS which measures the core symptoms of depression140), BDI-II, or PHQ-9 measures.

Magnetic resonance spectroscopy analysis

Linear regression models were largely used for the MRS analyses. Analyses was conducted on the ITT population, with participants excluded if they had a major protocol deviation for receiving the incorrect treatment. The dependent variable was change in depression score (baseline minus follow-up at 16 or 26 weeks, or averaged over 8, 16 and 26 weeks), with the predictor variable of interest being baseline GABA+/Cr, change in GABA+/Cr (follow-up minus baseline), Glutamate and glutamine ratio to creatine and phosphocreatine (Glx/tCr) change (follow-up minus baseline) or GABA/Glx ratio (follow-up minus baseline). This is with the exception of the hypotheses comparing change in GABA+/Cr and change in Glx/Cr between treatment groups, in which instance the spectra variable was the dependent variable and the treatment groups (rTMS and cgiTBS) the predictor variable of interest. The primary baseline GABA hypothesis baseline GABA as a predictor of clinical improvement (16 and 26 weeks vs. baseline), was analysed using an LMM. Participant was included as a random effect, and post-treatment time points (16 and 26 weeks) were included as an independent variable. For analyses, covariates of no interest included baseline depression score, MGH treatment resistance category, gender, age, treatment group, study centre, MRS platform, and spectra fit errors (GABA+ fitting errors at baseline and follow-up where appropriate and baseline and follow-up CRLB values where appropriate). For the two primary hypotheses related to GABA change, p < 0.025 was the level considered for statistical significance (p < 0.05/2, i.e. 5% significance level with Bonferonni correction), for all other hypotheses p < 0.05 was the level considered for statistical significance. The HDRS-17 was the primary depression measure of interest, and the HDRS-6, BDI-II, and PHQ-9 exploratory depression measures. All analyses were re-ran removing non-significant covariates with these final models reported.

For missing data (e.g. due to poor spectra quality, participants’ absence for the scanning, or incompleteness of whole MRS sessions) we took an available data approach, meaning individuals with enough data for a model were included in that model’s analysis.