Randomization and blinding.

Participants who complete the MRI localizer session (described below) are randomized within-site to receive either active or control neurofeedback (see details in Control condition section below) using a predetermined randomization schedule prepared by and known only to the site neuroimaging staff. Clinical staff and assessment raters remain blind to condition assignment. Each site has 1:1 randomization, with the constraint that the number of active NF cases is always greater than or equal to the number of control NF cases (necessary to ensure each control participant has a prior active NF participant to be matched to), and also the constraint that the number of active NF cases does not exceed the number of control NF cases by more than three (to keep the two groups reasonably balanced over time). If a participant is randomized but does not complete the NF MRI session, they are replaced (the next participant uses their randomization number) and they do not continue with study activities. Participants are debriefed at the end of the study, and at that time their group assignment is disclosed.

Mindfulness-based neurofeedback (mbNF).

Participants who meet eligibility criteria following intake and assessments complete a neuroimaging session at either Yale or the Martinos Center for Biomedical Imaging at MGH (Fig 2). The session begins with baseline structural, diffusion, and resting-state functional MRI acquisitions, which serve as the localizer for generating individualized network maps used in neurofeedback. After scanning, participants receive guided instruction in mindful describing (see details below), which they are encouraged to use in the MRI during the subsequent neurofeedback session.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Mindfulness-based Neurofeedback procedure.

Note: dMRI – diffusion MRI.

https://doi.org/10.1371/journal.pone.0338002.g002

MRI data acquisition.

All MRI data will be collected on 3T Siemens scanners with a 64-channel head coil: a Vida system at Yale and a Skyra system at MGH. Sequence parameters have been harmonized across both sites to the extent permitted by hardware and software constraints; parameters for Yale are presented in Table 2. Imaging includes a high-resolution T1-weighted structural MRI (sMRI) and diffusion-weighted MRI (dMRI) sequences. Functional imaging (fMRI) comprises both resting-state and NF sequences and includes two images with opposite phase encoding for offline fieldmap correction.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. MRI acquisition parameters.

https://doi.org/10.1371/journal.pone.0338002.t002

Localizer session and generation of personalized network maps.

At the start of the scan visit, participants complete a localizer session consisting of sMRI, dMRI, and 10 minutes of resting-state fMRI. These scans serve dual purposes: characterizing baseline anatomy and functional connectivity and generating individualized network maps to guide NF. This ensures that the feedback signal reflects each participant’s unique functional architecture.

While participants complete the mindful describing training outside the scanner, the resting-state fMRI data are preprocessed using FMRIB’s Software Library [36]. Preprocessing steps include motion correction, brain extraction, co-registration to standard space, spatial smoothing (5 mm full width at half maximum [FWHM]), and high-pass temporal filtering (0.01 Hz). Independent component analysis is applied to extract approximately 30 components. These components are spatially correlated with canonical network templates from the Yeo et al. 2011 parcellation to identify participant-specific maps of the DMN and CEN [37]. For each network, the component with the highest correlation is selected, thresholded to retain the top 10% of voxels by loading strength, binarized, and registered to the participant’s native space.

Mindful describing training.

After the localizer session, participants come out of the scanner for mental noting training (here termed “mindful describing” to cohere with DBT terminology). This semi-structured mindfulness training protocol, established in our prior work [23], is administered by trained research staff in a 45-minute in-person lesson. This lesson describes how to do the practice and introduces a behavioral and neural rationale for the approach, including introducing the goal of activating the CEN versus the DMN. The mindful describing practice focuses on present-moment awareness; participants learn to mentally label present-moment sensations (e.g., seeing, hearing, feeling). They are discouraged from focusing on thinking instead of perception, and from focusing on breath, which could confound fMRI data during the mbNF acquisition. Each participant selects an “anchor,” a part of their own body they can return to focusing on if they lose track of the practice. Each participant is trained until they and their trainer feel they are competent to use mindful describing in the scanner. This is assessed after several practice rounds: practicing mindful describing aloud to study staff, silently in their heads, and silently while hearing simulated MRI background noise. Participants are encouraged to use the skill during the neurofeedback training session and to consider during and after training how they could use mindful describing to manage distress in daily life.

Neurofeedback: Paradigm.

Following mindful describing training, participants return to the scanner for the neurofeedback session (Fig 1). A brief functional scan is first acquired to co-register the participant-specific network maps to their current position in the scanner. This is followed by 10 minutes of resting-state to establish post-training, pre-neurofeedback baseline connectivity.

Participants then complete one 2.5-minute “transfer” run prior to neurofeedback, in which the neurofeedback display is visible but static and participants view the interface without receiving real-time feedback. This allows for the assessment of whether the learned regulation strategies persist in the absence of feedback. The core neurofeedback protocol consists of five 2.5-minute runs of mbNF. During these runs, participants practice mindful describing while receiving visual feedback based on the difference in activation between the CEN and DMN, guiding participants towards decreased DMN activity and increased CEN activity. (see Neurofeedback: Implementation for details).

The session concludes with a second transfer task and 10-minute resting-state acquisition to assess post-mbNF connectivity.

Neurofeedback: Control condition.

Determining a control condition for non-medication interventions can be complex, as there is often no readily apparent placebo; neurofeedback is no different [38]. This study employs a yoked-sham control condition. Participants in both the control and active NF groups complete the mindfulness training and undergo similar MRI scan procedures. The primary difference is that the control NF participants do not receive feedback based on their own brain activity, but instead view feedback derived from a previously acquired active mbNF session from another participant. Thus, the visual display in the control group is independent of their own neural activity, ensuring that participants across groups are exposed to equivalent stimuli.

The yoked sham control approach enables participant blinding, provides a similarly engaging feedback experience, and balances the amount of positive feedback received across groups. Specifically, each control participant receives feedback indicating they are achieving the same amount of success as the active NF participant to whom they are yoked. Perceived success can influence motivation, as individuals are more likely to remain engaged when they feel they are doing well. It may also amplify placebo effects, as seeing apparent success in regulating mindfulness-related brain patterns can enhance expectations and perceived efficacy. Therefore, controlling for perceived success is critical.

An alternate control condition used in some studies involves training participants to regulate a brain region unrelated to their symptoms (e.g., from the somatomotor network). However, this approach has the disadvantage that the ease of modulating this region may differ between the control and active NF groups, as some brain areas are more easily regulated than others. This can lead to unequal perceptions of success on the task, which in turn may introduce differences in motivation and placebo effects.

One concern common to all sham-controlled designs is that participants in the control group may notice that the feedback does not correspond with their emotional state, potentially compromising the blinding. For this reason, participants are asked at the end of the NF session to report their beliefs about group assignment (i.e., which intervention they think they received, and their confidence in this belief).

Overall, by providing maximally similar procedures across active and control groups, we aim to account for potential mindfulness practice effects (e.g., learning, utilization) and non-specific scanner effects.

Neurofeedback: Implementation.

Real-time NF is implemented using the multivariate and univariate real-time functional imaging (MURFI) [39] software in conjunction with PsychoPy. A high-performance computer connected via Ethernet to both the MRI scanner and stimulus presentation system enables real-time data transfer and display. As functional volumes are acquired, MURFI computes the NF signal, which is transmitted to PsychoPy for dynamic display updates.

The NF signal is derived from a positive diametric activity (PDA) metric, defined as the difference in activation between the CEN and the DMN. For each acquired volume, an incremental general linear model estimates expected signal based on nuisance regressors, including six head motion parameters and a linear trend. The residual signal, representing neural activation, is computed by subtracting the nuisance signal from the observed fMRI intensity and z-scored relative to a 30-second baseline. Network activation is calculated as the weighted mean of residuals across voxels, with weights inversely proportional to baseline noise. The PDA metric is then computed as the difference between CEN and DMN activation and used to drive the feedback display.

During neurofeedback runs, participants view a visual interface consisting of a centrally positioned white dot and two flanking circles (Fig 3). The dot represents the current neural activity, and its position updates in real time based on the PDA metric. When CEN activation exceeds DMN activation, the dot moves upward towards the top circle; when DMN activation is dominant, the dot moves downward toward the bottom circle. This intuitive display guides participants toward decreased DMN activity while practicing mindful describing.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Participant view and corresponding activity during neurofeedback.

Note: CEN: central executive network; DMN: default mode network.

https://doi.org/10.1371/journal.pone.0338002.g003

Dialectical Behavior Therapy skills group (DBTsg).

Within ten days of completing NF, participants begin DBTsg. The groups are led by trained Masters- and PhD-level staff trainees from the UMass Psychological Services Center who have completed at least the equivalent of foundational training in DBT. Each group is led by two leaders, consistent with the DBT manual [7], who are supervised by a licensed clinical psychologist (KDG) and attend a weekly DBT consultation team. The groups follow the 20-week curriculum developed by McMain et al., which improves emotion regulation, distress tolerance, anger, and suicidal/self-harming behaviors [40]. The group curriculum draws on the standard DBT manual, and teaches skills across distress tolerance, interpersonal effectiveness, dialectics, and emotion regulation. Each of these modules is preceded by one session focused on mindfulness. DBTsg runs continuously, with rolling admission: New participants join at the beginning of each module, and graduate once they have completed all sessions. In this way, we enroll participants continuously through the life of the project because DBTsg start dates occur approximately every five weeks.

fMRI outcome measures.

The primary outcome for the R61 study is the change in two features of DMN connectivity between pre- and post-NF resting-state scans: (1) within-network connectivity between the medial prefrontal cortex (mPFC) and posterior cingulate cortex (PCC), and (2) DMN-CEN anticorrelation, defined as connectivity between the mPFC and dorsolateral prefrontal cortex (dlPFC).

The resting-state data will be processed using fMRIPrep [41] and the CONN toolbox [42]. Preprocessing steps include slice timing correction, motion correction, susceptibility distortion correction using fieldmap-based unwarping, co-registration to the T1-weighted anatomical image, normalization to standard space, and spatial smoothing (4 mm FWHM). Denoising will include regressors for the six motion parameters and their temporal derivatives, the top five components capturing signal from the white matter and cerebrospinal fluid [43], outlier volumes defined by a framewise displacement >0.5 or standardized DVARS>1.5, and a linear trend term. Bandpass filtering will be applied to retain frequencies between 0.008 and 0.09 Hz.

Regions of interest (ROIs) for the mPFC, PCC, and dlPFC will be defined based on 10 mm spheres centered on established coordinates [44]. Mean timeseries will be extracted from each ROI, and Pearson correlation coefficients will be computed between the mPFC and PCC (within-DMN connectivity), and between the mPFC and dLPFC (DMN-CEN anticorrelation). We will also explore generating participant-specific ROIs by identifying the peak voxel within each ROI, and constructing a 10 mm radius sphere centered on that peak.

Patient-reported measures.

The secondary outcome measure for the R61 study is change in state mindfulness from before to after mbNF on scan day. The State Mindfulness Scale (SMS) is a 21-item self-report measure that quantifies current-moment mindfulness of body and mind [45]. Change in mindfulness in the study will also be measured by querying self-reported use of mindfulness skills in daily life at baseline and subsequent assessment timepoints.

Several scales are used to track clinical and psychological change in the study, though these are not the focus of the R61 phase of the project. The Borderline Symptom List (BSL-23) is a 23-item scale that measures BPD symptom severity over the past one week [46]. The Beck Depression Inventory (BDI) is a 21-item scale that measures depression severity over the past two weeks [47–49]. We also track DBTsg group attendance and use of mental healthcare resources outside of the study to support exploratory analyses.

Further psychological measures are collected to support exploration of change in shame [50], social functioning [51], emotion regulation [52–54], sense of agency [55,56], and episodic future thinking [57] over the course of the study.

Additionally, several measures are collected at baseline to contribute to description of the sample and for exploratory analyses of heterogeneity of treatment effects, including demographics, measures of handedness, quality of life, and psychiatric symptoms (see details in Fig 1).

Lived experience-led program evaluation.

This study incorporates assessment of participant experience at multiple time points to support the development of an intervention that is well-tolerated by the target population.

To evaluate tolerability, study staff assess both physical and emotional adverse events before and after neurofeedback sessions. Prior to neurofeedback, participants are interviewed about their expectations about the intervention. Following neurofeedback, they complete the Intrinsic Motivation Inventory (IMI) — which measures motivation, likelihood of future participation, and enjoyment [58] — and participate in a semi-structured qualitative interview exploring their experience in the scanner and strategies used to maintain focus.

A key strength of this study design is the integration of a lived-experience consultant and structured opportunities for participant reflection. The study leadership team includes a consultant with lived experience of BPD who piloted all study procedures during setup and provided feedback on the overall research design. Participants are randomly identified to join focus groups and one-on-one qualitative interviews following completion of neurofeedback, gathering their reflections on the study experience and feedback on the intervention approach.

Embedding this element of assessment within the study design allows participants to share their perspectives in an open, low-pressure environment with an empathetic listener, without concern for meeting perceived researcher expectations. Importantly, the first round of interviews will occur before the R33 phase of the grant, allowing time to integrate participant feedback into the final study design.

Aims.

In the R61 study phase, we aim to determine whether mbNF targeting the DMN modifies connectivity and improves mindfulness in participants with BPD, relative to control NF. The primary outcome is change in resting-state DMN connectivity from pre- to post-mbNF, either by decreasing within-DMN hyperconnectivity or by increasing DMN-CEN anticorrelation. The GO criterion for continuing to the R33 phase is greater pre- to post-NF change in at least one of these targets, with a medium effect size (d ≥ 0.5) between groups. The secondary outcome is change in self-reported mindfulness scores on the state mindfulness scale (SMS) from pre- to post-mbNF. We plan to randomize 52 participants (26 per group) in the R61 phase of the study. Because the outcome measures for this cohort are collected within two hours after randomization, little or no attrition is expected.

The R33 phase analyses focus on outcomes at later timepoints (after mbNF), over the course of the 20-week DBT skills group and two month follow-up period. This phase will assess the relationship between target engagement, clinical improvement, and longitudinal outcomes, including exploratory analyses to determine whether neural and symptom changes persist after mbNF and DBTsg. More participants (72, 36/group) are enrolled in this phase of the study because as much as 33% attrition is expected over the course of the full study.

Data preparation and statistics.

Data will be examined for outliers, non-normal distributions, and non-linear associations. Preliminary analyses will examine potential covariates including site, age, sex, clinical variables, and head motion. Psychiatric service use (e.g., therapy, medication, unplanned mental health service use) will be assessed to test whether this differentially impacts (or is affected by) change in our neural targets. We will also explore baseline symptoms as a potential covariate, as differences in severity may lead to differences in target engagement. We will explore the presence of group effects (and time effects, when appropriate) across individual covariates or combined covariates using a liberal p < .05 uncorrected threshold and include any suprathreshold variable as a control covariate in subsequent analyses.

We will use a repeated measures ANOVA (or linear mixed-effects model in the presence of significant covariates) to test our hypothesis that mbNF will lead to larger decrease in DMN connectivity than control NF, with group (active vs. control) as a between-participant factor and time (pre- vs. post-NF) as the within-participant factor. This analysis will include all the R61 participants who complete the neurofeedback day (n = 26 per group). With a total sample size of 52 participants, assuming sphericity and a correlation of 0.5 among repeated measures, we will have 94% power (α = 0.05) to detect a medium effect (Cohen’s d = 0.50, Cohen’s f = 0.25). While this power analysis assumed a medium effect size, consistent with our prior work, we acknowledge that resting-state fMRI connectivity effects are often smaller and more variable in practice; thus, our observed effect sizes may ultimately differ from this estimate, and replication in larger samples will be important to confirm the robustness of our findings.

Missing data will be accommodated using multiple imputation or weighting procedures. Missing data will be handled using full-information maximum likelihood (FIML) within the LMM framework, valid under missing-at-random (MAR). Because MAR may not hold, prespecified missing-not-at-random (MNAR) sensitivity analyses will be conducted, including multiple imputation and pattern-mixture models (“copy reference” and “control-based” scenarios), to assess robustness of treatment effects. In the case of violations of statistical assumptions, non-parametric alternative tests will be employed. Conclusions will primarily be based on effect sizes, with 95% confidence intervals, rather than statistical significance to maximize reproducibility. Based on our prior studies, we estimate ~30% attrition/data loss and have planned recruitment to support data collection on primary and secondary outcomes from 52 participants.

Data management, sharing, and monitoring.

Data are stored in HIPAA-compliant servers (MR data), REDCap databases (clinical and self-report data) and TEAMS folders (interview recordings). In accordance with funder guidelines, we will share data through the National Data Archive (NDA) by assigning each participant a global user identifier and uploading data using NDA standards.