Workshop: Neurofeedback Intervention Development: Opportunities and Challenges

Transcript

Sessions 1 & 2

DEBORAH KRAT: It is now 10:00 a.m. Eastern time. Welcome, everybody, to the Neurofeedback Intervention Development Opportunities and Challenges workshop. This is presented by the National Institute of Mental Health, the Division of Translational Research.

Just a few housekeeping notes and reminders for today. Participants have entered into listen-only mode, cameras off and mics are muted. Please submit your questions via the Q&A box at any time during the webinar. Questions will be answered during the discussion sessions of the workshop. If you have difficulties hearing or viewing the webinar, please note these in the Q&A box and our technicians will work to fix the problem. You can also send an e-mail directly to me, Deborah Krat, at dkrat@mn-e.com.

And with that, I will now pass it over to our workshop presenters, Dr. Christopher Sarampote and Michelle Hampson. Chris, it's all yours.

CHRISTOPHER SARAMPOTE: Thank you, Deborah. Good morning, or good afternoon wherever you are. And thank you very much for joining us today.

My name is Christopher Sarampote.  I'm a program officer in the Division of Translational Research at the National Institute of Mental Health.  The mission of the National Institute of Mental Health is to transform the understanding and treatment of mental illness through basic and clinical research, paving the way for prevention, recovery and cures.
Our webinar was developed today with the goal of sharing the state of the science for neurofeedback intervention development for mental disorders with an emphasis on real-time fMRI approaches to treatment.  

We have invited researchers and federal officials, and they have generously accepted, to highlight recent developments in early treatment research, efficacy and effectiveness trials and regulatory issues relevant to the approval and implementation of device-based interventions.  We've also asked panelists to share their thoughts on the challenges for neurofeedback research as well as the opportunities that may lead to more effective treatments for mental disorders.  

I also want to introduce our Co-chair, Dr. Michelle Hampson.  Dr. Hampson is Professor of Radiology and Biomedical Imaging at the Yale University School of Medicine.  Dr. Hampson and her group are focused on developing, validating, and applying novel functional magnetic resonance imaging approaches with a particular interest in the use of fMRI neurofeedback to treat and study neuropsychiatric disorders.  Michelle, thank you very much for co-chairing this morning.

MICHELLE HAMPSON: Thank you, Chris.

CHRISTOPHER SARAMPOTE: For today's webinar, we will have three sessions.  Session one will discuss early stage studies of target acquisition and measurement and task development. Session two will present work on ongoing efficacy and effectiveness trials of neurofeedback approaches.  And session three will cover regulatory issues and the larger question of how neurofeedback interventions are translated to clinical practice.  

Following each set of presentations, we will have a discussion period moderated by NIMH staff.  Audience members are welcome to submit questions to the Q&A box.  Please note that we will be unable to answer any specific questions you might have about patients or grants or companies, but we can answer more general questions specifically and particularly those that relate to questions you may have about the talks.

This is a public meeting.  And the recording of this webinar will be available along with a summary of the meeting on the NIMH website.  And as we get started, I want to save as much time as possible for our presenters and for the discussion section so we will get rolling, but I just want to thank every single one of our presenters, my colleagues at NIMH, and Dr. Holly Lisanby for her support and encouragement to have this webinar today.  And to thank you all for your time and your good work.  

For our first session, we have three presenters, and they will be discussing the topic of early stage development.  Dr. Stephen LaConte will lead us off.  He is Associate Professor at the Fralin Biomedical Institute at Virginia Tech.  Dr. Masaya Misaki is Associate Investigator at the Laureate Institute for Brain Research.  And then our third presenter will be Dr. Michelle Hampson from Yale University.  

So, with that, Stephen, why don't you kick us off.

STEPHEN LACONTE: Okay.  Thank you very much.  Can everyone hear me and see the screen?  Well, I'm going to assume yes.  I see a thumb floating up.  

So thank you very much, Chris and Michelle, and, you know, everyone that was involved with the organization of this.  I'm really delighted to be here today to talk about early-stage real-time fMRI neurofeedback and how to design those early-stage experiments.  And so, it could go without saying even for a general audience that the heart of real-time fMRI is fMRI.  And if we had more time, I think this point would come up more prevalently.  But I would like for everyone to keep this in the back of their minds because I think sometimes it's so obvious that it may actually be overlooked by us as investigators or by us as reviewers.  And so, we inherit all of the magic of BOLD fMRI as well as all of the limitations of fMRI as well.  

Nevertheless, if we take this block diagram as kind of a prototypic typical fMRI experiment, we have an individual in the scanner who is generally guided by some stimulus-based tasks.  And the fMRI is collecting a 3D movie of the brain in action.  Now, of course, the part of real-time fMRI is that during this data collection, we can extract signals through a variety of different methods.  Prominently region of interest, for example.  Or looking at functional connectivity across multiple regions of interest.  Or in my own work, trying to use machine learning models to decode what the sensory, behavioral or psychological state of the individual is on image-by-image basis or ongoing basis during the actual experiment.  

And regardless of what signals we are extracting, we can think about those signals as control signals.  And for today's topic, those will be controlling generally neurofeedback types of interfaces.  And I think you will see a wide variety of those interfaces as the day goes on.  
More generally, though, just trying to weave in some plugs for basic science, these signals that we are extracting can really be control signals for virtually any component of the experiment that we might want to think about.  And so really as we think about the experimental flexibility that real-time fMRI gives us, again we inherit all of the components of fMRI.  So we need to think about field of view, resolution, sampling rate.  

But then we get a tremendous and I think novel degree of flexibility from this closed-loop aspect of the experiment.  And I will try to highlight that as the slides go on.  And then not neatly fitting into this block diagram; but, of course, there are many parameters in the pipeline both before, during and after a neurofeedback experiment that go all the way from, you know, our inclusion/exclusion criteria to how we are instructing participants before they go in the scanner to exactly what we are trying to measure in terms of change and how we are trying to measure that change after a neurofeedback experiment.  

And so, this kind of leads to the parenthetical subtitle of my talk, which is real-time fMRI neurofeedback has a lot of knobs.  And I credit this quote to Sue Whitfield-Gabrieli.  A number of years ago, she just said this, and I instantly had this sort of flashback to myself as a college student trying to get an oscilloscope to give me the waveform that my circuit was supposed to be giving me.  And, of course, this is a fairly complicated interface for receiving just one or a few signals at a time.  

But actual control interfaces are massively complex when you are talking about things like aircraft carrier engine control rooms or, you know, I think what comes to a lot of people's minds are things like the flight instrumentation on the airplane.  And so this is my attempt to capture the daunting magnitude of the experimental flexibility that we have.
And, you know, this experimental flexibility I think promises great potential for future innovation, but this flexibility has also imposed a big challenge to us, a small enthusiastic community, because in one way or another we have all come to the realization that despite decades of neuroimaging data and even more decades of clinically relevant behavioral data, we are not yet ready or able to use first principles to design neurofeedback experiments.  

And so, this struck me a number of years ago when I was designing a queue reactivity study with my collaborator, Pearl Chiu.  And a number of years ago, there were not as many substances use disorder real-time fMRI studies published.  And we were really hard-pressed to find unambiguous compelling guidance from the literature.  

So, for example, should we be using images or videos?  And what types of images?  Should we design a crave-don't crave experiment or one that focused only on suppressing craving?  Could we even think about jumping in blindly and trying to change craving at an early stage?  And, you know, what regions of the brain would be most potent for changing craving and so on and so forth.  

And so, you know, of course we're not alone, and it turns out that porting basic science for evidence-based medicine is a well-known issue in implementation science.  And it extends well beyond our specific real-time fMRI setting.  So here I'm just showing the cover from this report that talks about the quality chasm between scientific knowledge and the state of scientific knowledge and the quality of clinical care.

And so, what are some of the experimental parameters that are open to us as investigators?  And, you know, because of the limits of time, I won't be able to share all of them.  I just kind of want to impress upon the fact that there are a lot of knobs.  And as you go through today's talks, kind of pay attention to the variety and breadth and depth of applications within the neurofeedback results.  And this is just kind of a small slice of what we are going to be able to talk about today.

So, one resource I ended up relying on a lot, I ended up throwing out my own bullet points and using a review article from 2020 by Samantha Fede when she was in Reza Momenan's lab.  And, of course, as a review article it is going to be locked in time and missing several important recent studies.  But they covered a lot of important topics, and I think this paper covers them very thoughtfully.  And also, I like the graphics because it kind of reinforces this idea of knobs and dials.  So, I think I was sort of drawn back to this paper as I was thinking about this oscilloscope again.  

And so, you know, one of the knobs are the types of tasks that we want to do in the scanner.  And again, we have to do them in the scanner and so we have to think about the environment, which is somewhat claustrophobic, highly magnetic, and also loud.  And so, it's possible to do auditory tasks.  It is even possible to do a limited number of motor tasks, usually manual types of tasks.  There's no dancing allowed in the scanner.  

But by and large, you know, the feedback interface is probably going to be visual.  You could do tactile as well.  Anything is possible, but it takes some engineering challenges sometimes.  
And then the current state of the literature shows that most real-time fMRI studies to date have really been focused on healthy populations.  And I'm going to tie this into kind of the concept of positive controls because a lot of these papers have been technological development or sort of intended as steppingstones to jump into other clinical populations.  

And, you know, I mentioned briefly that there are many algorithms for extracting signals of interest.  And, of course, if you are an fMRI researcher or a neuroimaging researcher or just a data analyst in general, you know that there are lots and lots of algorithms that are available.  But to date, really most of the tractable first hypotheses have come from regions of interest types of questions.  And this has actually been very productive because now you can see we are starting to grow a literature that is much more catered to our needs of trying to understand what regions are really accessible to participants to be able to use neurofeedback and even look at relative signal strengths that we might expect which would help for experimental planning and for planning statistical tests, for example.

You know, another area is for these ROIs or for a classifier that is detecting crave-don't crave, do you want to go up and down or do you want to just be modulating in a suppression direction or enhancement direction.  And those depend on the scientific question and the clinical context of what you are doing as well.
And then for the feedback display, you know, there are a lot of things that we could talk about in terms of feedback.  And I think I will save some of that for kind of the later talks.  I would love to talk more about feedback, and I think feedback timing is important.  But one area that we may have missed in this discussion is that we can also control stimulus timing.  Because we have real-time access to the brain, we might be able to control when we present stimuli so it's at a time when the brain is most vulnerable or most resilient, for example.  

And then the issue of control groups is something that I'm going to go into a little bit more in this talk with the limited amount of time that I have.  So in the face of this complexity, how do we move forward to take advantage of real-time fMRI's potential.  And I think that there are at least three interrelated ways to do this.  
The first is to embrace the NIH's experimental medicine framework.  Especially the idea of target engagement.  The second is to leverage the idea of positive scientific controls.  And the third is to utilize and remember that real-time fMRI is really useful as a basic science platform and goes hand in hand with our efforts to look at neurofeedback for clinical populations.

So, the NIH's experimental medicine framework is more than just about target engagement, but I would like to talk about the importance of target engagement through kind of an example.  And in the interest of time, I'm not going to go through this example carefully.  I'm happy to backtrack during the discussion.  But essentially what it caused us to do is go beyond the traditional efficacy testing and instead have in addition hypotheses about mechanisms of action.  

And so that then becomes the target for us to test.  And we can kind of reframe our questions with an intermediate target that is mechanistic in its nature in mind.  And so then the issue becomes how do we measure the target.  And the great thing about real-time fMRI is that we can think about using fMRI as -- now my -- I apologize, my slides for some reason have paused.  Okay.  

We can use fMRI as a target and/or as an assay.  We could actually be using the target for feedback and assaying changes in parallel even.  And so, again, just to highlight some of the experimental flexibility.  I'm not going to read all of this about experimental medicine and target engagement, but I will say that a major part of the NIH's rationale is that studies that demonstrate target engagement are meant to be mechanistically informative with this really nice feature that the information should -- value should be there whether the results are positive or negative.

And so, the other sort of motivation for this is that really the goal is to come up with generalizability and synergistic breakthroughs, recognizing that, you know, many unhealthy behaviors co-occur in different clinical populations and may have shared causal factors as well.

Okay.  So really quickly, through positive controls just as kind of a preview and I think dovetailing with some of what Michelle will be talking about, she's going to talk about early, middle and late phases of testing as kind of cast as a foraging exercise.  
And so in that early to late progression we are going to be obtaining stronger and stronger evidence for both target engagement and clinical efficacy.  And so, we need to be more and more concerned about the specificity of neurofeedback in those results.  And so that's where the idea of negative scientific controls.  For example, the classic example is a placebo group comes into play.

And so, to talk about controls, a really useful article -- again, my -- I apologize, my slides are not advancing so I'm probably going to extra advance.  But I'll keep talking.  
So, there's a terrific resource in a 2019 NeuroImage paper by Bettina Sorger, Frank Scharnowski, David Linden, Michelle Hampson and Kym Young.  And it's titled "Control Freaks."  But it's not too controlling and it's not too freaky, even though my slides are freaking me out here now.  

And so, you know, I will just kind of go through and eventually my slides will catch up to me.  But essentially the flip side of negative controls which are testing for all of the variables that could be confounding or kind of enhancing.  So just going into the scanner might have an effect on participants regardless of whether or not they are getting true feedback.  And so you want to control for those with negative controls.  

Positive controls are kind of the safe launchpads in which to explore your next sort of experimental set of parameters.  So positive controls are really useful.  I think of them as sanity checks or debugging.  But essentially you want some evidence that your system is working.  And that could be as basic as whether or not your real-time setup is correct, and all of the computers are talking together.  And/or whether or not your task is actually appropriate or not for that.  And so you may need to run control populations where you think that accessing the target might be easy or doable.  And then move on to clinical populations where you hope to target the targets.

So, I apologize for both taking a while and -- oh, now my slides are moving again.  So just as kind of the final thing to say, in the face of this complexity -- sorry, I really didn't do positive scientific controls justice.  But, you know, the idea of using real-time fMRI as a basic science platform I think is a third thing that I -- that we should be doing.  

And so, within the experimental medicine framework, we have to remember this ability to learn from successes and failures.  And so, the fMRI part of real-time fMRI I think demands post-acquisition analysis including revisiting all of our instructions and trying to improve future targeting and mapping new potential targets.  And, you know, I'm not going to go through all of the creative experiments here.  But real-time fMRI can sometimes get typecast as only a tool for neurofeedback.  But this closed-loop aspect really unlocks a whole new level of experimental flexibility.  

And so I will just kind of end there.  There's nothing profound in these conclusions.  But I will just leave them up and people can look back through the video as well.  I would just like to acknowledge my lab and my previous funding.  And there are many, many collaborators over the years which, you know, I deserve to thank.  But for today, I'm just going to thank some of my active real-time fMRI collaborators.  Thank you very much.

CHRISTOPHER SARAMPOTE: Thank you, Stephen.  And thank you for your talk.  
We have for our next presenter Dr. Masaya Misaki from the Laureate Institute of Brain Research.  Dr. Misaki, feel free to share your slides and take it away.  Masaya, you're muted.

MASAYA MISAKI: I'm sorry.  Yeah, okay.  

So, I'm going to talk about -- discuss how we gain a mechanistic understanding of neurofeedback for precise intervention using real-time neuroimaging.  
So, okay, as Dr. LaConte has nicely introduced, neurofeedback involves many adjustable parameters and there is no single best setting for them as they largely depend on the specific context or goal of the intervention.  So I believe that a deeper understanding of the mechanisms behind neurofeedback can help us identify effective parameter settings.  So my talk will focus on the mechanistic understanding of neurofeedback treatment.  

So, here's what I'll cover in today's talk.  So first, I will start with a brief introduction to real-time neuroimaging technology, focusing on fMRI.  So this will build on the points from the previous presentation.  And next, I will move on to the mechanism of neurofeedback training.  And finally, I will discuss intervention parameter settings based on mechanistic understanding.  

So real-time neuroimaging is a technology that analyzes and visualizes brain activity as it happens.  So this enables the live view of ongoing brain activation and enables immediate feedback and interaction with ongoing cognitive process.  So this technology allows for highly precise and personalized brain interventions both in terms of location and the timing.  So, with fMRI, we achieve spatial resolution down to a few millimeters, enabling us to pinpoint brain activation.  And also, immediate feedback and interaction can target the current or ongoing cognitive state.  So, this precision is expected to significantly enhance the effectiveness of the intervention.

So neurofeedback is one of the most popular and promising applications of real-time neuroimaging.  So in neurofeedback, we present brain activation signals to participants so they can learn to regulate them on their own.  So regulation is typically voluntary, but we can use rewards like monetary incentives for implicit reinforcement learning.  

And also, the application of real-time neuroimaging extends beyond neurofeedback.  So we can combine this technique with various brain stimulation methods such as TMS, tECS, FUS and even with audio/visual stimulations.  So, by integrating these methods, we are able to fine-tune stimulation parameters while monitoring their effects in real-time.  

So this integration allows for precise adjustments of parameters based on immediate feedback so we can use a closed-loop iterative optimization of brain stimulation parameters.  So, for instance, in our proof-of-concept study, we successfully optimized tECS parameters using real-time fMRI so demonstrating the technological feasibility of this approach.

So real-time neuroimaging can also facilitate brain-based communication.  SO in this application, interaction partners such as clinicians can engage with participants while actively monitoring their cognitive or brain states.
So we have implemented this system to demonstrate its potential for enhancing therapist-patient communications.  So, it's still very early proof-of-concept stage.  But their technological feasibility has been demonstrated.  

And also, regarding recent technological advancement, the current real-time fMRI system enables comprehensive image processing of whole brain voxels comparable to offline analysis in real-time.  So, for example, in our system, it is capable of applying comprehensive image processing, including a physiological noise correction in less than half a second.

So in summary, real-time fMRI technology has significantly advanced, enabling the precise real-time extraction of brain activation signals.  However, it is important to acknowledge the challenges like the temporal delay in the BOLD response and the limitations in statistical inference that still persist.  And also, several advanced applications of real-time fMRI are now technologically feasible.  However, they still require high-level system integration.  These systems often need in-house development and are not widely available.  So this highlight the pressing need for more accessible and flexible real-time fMRI solutions.  
And despite these technological advancements, our understanding of how online stimulation or adjustment of brain activation affects cognitive functions remains limited.  So in the next part of my talk, I will explore this issue further.  

Okay.  Let me recap the mechanism of neurofeedback training.  So neurofeedback involves learning to control brain activation through self-regulation of mental states and reinforcement learning.  So, it's important to highlight that this process isn't just a focal intervention, even when targeting specific brain regions.  So, rather, it involves engaging large-scale brain networks through self-regulation, mental state monitoring, and reinforcement learning.  

So, this involvement of large-scale brain networks has been demonstrated in several studies.  For instance, a meta-analysis of neurofeedback studies targeting various brain regions revealed commonly activated areas.  So, these regions are generally associated with skill learning and mental regulation tasks.  So, these activations are not limited to neurofeedback training.  

So also our studies focused on increasing left amygdala activation in veterans with PTSD have shown that the effects of the training extend beyond the targeted area.  So we observed that DMPFC and the MCC activations were associated with the training effect.  And also, the training effect were evident in the resting-state functional connectivity between the SMA and dorsal anterior cingulate cortex as well as the hippocampal volume changes.  So these regions are distinct from the targeted regions of the left amygdala.

Okay.  Given that the neurofeedback effects were not confined to the target region, it suggests that the impact on cognitive functions might be mediated through modulation in other brain areas.  So this complexity indicates that establishing a simple causality between target modulation and treatment outcomes can be challenging.    So, for a more detailed discussion on this subject, I recommend referring to this paper "Can Neurofeedback Provide Evidence of Direct Brain-Behavior Causality?"  So, it is a very nice paper for review.   So at this point, it is crucial to consider the appropriate outcome measure of training success.  What is the appropriate measure of the training success?  So should we solely focus on the regulation of the target brain signal, or should we also consider changes in cognitive functions or symptoms?  

So I believe that the primary goal of the brain interventions, including neurofeedback, should be to modulate cognitive functions.  So, to achieve this, we select the target brain region that is believed to be linked to the desired cognitive function.
And assuming a straightforward causal pathway where the target brain activation directly influences targeted cognitive functions, successful regulation of the targeted brain region could be thought of as an intermediate measure of training success.

However, we do know that neurofeedback training engages multiple brain regions beyond the target.  And it may be the activation or network connections between these regions that actually drive changes in cognitive function.  Therefore, successful regulation of the targeted brain region may not correlate directly with cognitive changes.  So, this means that regulation of the target brain signal alone cannot be considered a reliable measure of training success.  

Also, it is important to note that even if this is the case, the target brain activation can still serve as an effective feedback signal for training if it reflects the brain activations that is causal to the cognitive functions.  However, if this is the case, we need to examine the whole brain process behind the observed cognitive change to fully understand the training mechanism and validate the effects of the intervention.  

So at this point in this scenario, real-time fMRI is an ideal tool for investigating these mechanisms.  So it not only supplies a neurofeedback signal for training self-regulation, but also offers a comprehensive view of whole-brain activation during the training sessions.
So real-time fMRI provide two valuable outcomes from a single experiment, so making it a cost-effective approach for studies.  We should view real-time fMRI not only as a method for intervention but also as a valuable tool for neuroscience research.  

So, I would like to share some of our recent findings from this line of research.  So in one study, we did neurofeedback on the functional connectivity between the precuneus and the right temporal parietal junction.  Those region connectivity were positively correlated with the brooding rumination in mood and anxiety participants.  So we trained the participants with MDD to decrease this connectivity with the neurofeedback.

During the training we observed that the connectivity between these regions, indeed, reduced.  But there was no significant difference in the reduction between the active and the sham control groups.  So however, interestingly, we did find a significant difference in the changes in brooding rumination scores between these groups.  So, a statistically significant change in symptoms was observed only in the active group, not in the sham control group.

So we conducted regression analysis to correlate the symptom changes with whole-brain connectivity patterns using connectome-based predictive modeling.  And we discovered that the interaction between connectivity patterns during the regulation task and in response to the feedback signal were critical for predicting changes in symptoms.  And furthermore, these predictive connections were not confined to the target brain regions but spanned across extensive brain networks.  So, these findings indicate that the symptom reduction was not solely explained by the regulation of the target brain region.  

And in another study we conducted a whole-brain analysis for a neurofeedback study targeting the left amygdala.  The participants with MDD were trained to upregulate the left amygdala signal by recalling positive memories.  So, the details of this protocol will be presented in the research session by Dr. Young.  So, I believe this approach is one of the most successful fMRI neurofeedback treatments for depression.  

So I conducted a second analysis of the dataset.  And we observed a significant reduction in depressive symptoms in the active group, which was not seen in the control group.  And additionally, the left amygdala signal was higher in the active group compared to the control group.  However, the main effect of the interaction was not significant.  

And furthermore, when we focused on individual variability, left amygdala activity during training did not explain individual differences in depressive symptoms.  So we had no significant associations between the changes in MADRS scores and changes in the left amygdala signal during training.  So, this suggests that the symptom relief may be mediated by brain activations other than those in the left amygdala.

So here I assumed that the variability in brain activation during the training could be linked to individual differences in therapeutic outcomes.  So to identify variable brain activation patterns during the training, we classified whole-brain activation patterns into subtypes among active group participants.  So this process involved feature selection, dimension reduction, and clustering analysis, complemented by repeated cross-validation to find a reproducible solution.  

And we identified two subtypes of brain activations during the regulation block activation.  So, the first subtype was characterized by activations in the control regions, like the lateral prefrontal region and supplementary motor areas.  The second type not only showed activations in the lateral prefrontal region but accompanied by suppressive changes in the default-mode network regions such as ACC and the medial prefrontal cortex.

And interestingly, significant symptom reduction was observed in participants categorized as Type B, who demonstrated suppression of the default mode network activation.  
And we also evaluated event-related response to the feedback signal and identified three types of brain activation.  In this case, we evaluated the brain based on feedback signal to event-related response.  So, type A showed a very limited response to the feedback signal.  But Type B brain activation showed a negative correlation with the feedback signal.  And the Type C participants showed a positive correlation with the feedback signal.  

And again, significant difference in symptom reduction was observed between the subtypes Type B and Type C who showed the brain activation that correlated in positive or negative way to the feedback showed that significant symptom reduction one week after neurofeedback training.  However, Type A we saw no significant response, saw no significant improvement.

So we also noted a significant interaction between these subtypes, which further explained the variability observed in the active group.  And these maps illustrate the specific pattern combinations within the active group that led to significant symptom reduction.  So, the presence of multiple effective combinations indicates that there are several potential pathways for treatment, suggesting the way to depressive symptom relief is not limited to a single pathway.  

In summary, the subtypes of brain activation patterns during neurofeedback training explained the variance in treatment effects.  And the symptom reduction cannot be solely attributed to the successful regulation of the target region.  And instead the way in which participants regulate is critical to effectiveness of the treatment.

Okay.  So, to conclude my presentation, let's consider how we might design interventions based on these observations.  So, when determining the appropriate outcome measures in neurofeedback studies, we should prioritize cognitive changes as these are the ultimate goal of the intervention.  However, regulation success can be considered a valid outcome if it closely relates to the targeted cognitive function or if it is a part of a proof-of-concept study aimed at examining the controllability of specific brain activations and its effect on cognitive function.  So this outcome would also be valid.

Okay.  So our previous findings also raise a critical question about the optimal target for neurofeedback interventions.  So should we focus on modulating a specific brain region, or should we consider large-scale activation patterns?  
The signals from a specific regional target are easier to extract.  However, this approach generally lacks specificity for particular cognitive functions and may require explicit instructions of mental strategy to achieve the desired cognitive changes.

And in contrast, large-scale activation patterns typically show higher specificity for cognitive functions so potentially increasing the effectiveness of neurofeedback.  And I'd also like to emphasize that targeting detailed whole-brain activation patterns is a unique advantage of the real-time fMRI neurofeedback that is difficult to achieve with other focal stimulation methods.  

However, associating these patterns to specific cognitive states requires creation of a machine learning model, which requires additional sessions to collect the sufficient data for training.  And the effectiveness of the intervention depends on the quality of such a decoding model.

So, regarding pattern-based neurofeedback, several innovative approaches have emerged for training the applications.  So among these, Decoded Neurofeedback is particularly notable for its success.  

So, this method the neurofeedback based on decoding signal for cognitive or perceptual states that are derived from multi-voxel patterns.  DecNef typically uses a reinforcement learning approach often incorporating monetary incentives as rewards and it also utilizes implicit training methods which means participants are not given specific instructions on mental strategies.  

Another approach is connectome-based neurofeedback.  So it is a feedback signal derived from the strength of thousands of functional connections instead of activation amplitude.  So connectivity patterns are a more reliable marker of cognitive state so this method could prove more effective than traditional activation-based interventions.  

And also, semantic neurofeedback is another innovative method that uses multi-voxel patterns.  So unlike decoding neurofeedback, this technique extracts relative relationships between multiple cognitive states from these multi-voxel patterns.  So, feedback is provided by presenting the representational similarity state between this state and the current state.  So, subjects are navigated to move from the current state to the ideal state.

So the selection of the target is also a challenge for neuroscience.  So clarifying the brain activity associated with cognitive function through ongoing neuroscience research will enhance the efficacy of neurofeedback interventions.  So a significant advantage or another advantage of real-time fMRI neurofeedback is its ability to immediately translate neuroscience knowledge into brain interventions.  Thus, neurofeedback can evolve as neuroscience evolves.  

So there are many other feedback and the scan parameter and control conditions, but due to time constraints I cannot cover everything today.  So let me conclude my talk.  
So, there is no one-size-fits-all solution for neurofeedback parameters and utilizing a mechanistic understanding allows us to more effectively explore and optimize these parameters.  And the treatment effect cannot be attributed solely to successful regulation in the target regions.  So conducting whole-brain analyses provides invaluable insight into the mechanism of neurofeedback training.  

So in closing, I would like to express my gratitude to my colleagues.  Especially to -- my deepest thanks to the late Jerzy Bodurka, who was instrumental in establishing our real-time fMRI systems and initiating our pioneering feedback studies.  Thank you for your attention.  

So yeah, I think of Dr. Michelle Hampson from Yale School of Medicine.

CHRISTOPHER SARAMPOTE: Thank you, Masaya.  And Michelle, we are a little ahead of schedule so you have time.  And we -- allows more time for the discussion.  So take it away.

MICHELLE HAMPSON: Thank you, Chris.  So, I just want to start by thanking all of the staff at NIH who put all their work into organizing this workshop.  And also, to express appreciation in general that NIMH supports these workshops which I think are a great avenue for communication between program staff and extramural researchers.  Thank you.

So, I'm going to be talking now about what I see as the promise and challenges in early-stage development of fMRI neurofeedback interventions.  So, first of all, the positive.  The promise.  What I am most excited about with fMRI neurofeedback is the data suggesting that it really works in allowing us to do targeted alteration of mental function.  So, it's kind of a unique tool that gives us great flexibility for really intervening in humans and changing their mental function in a targeted manner.  

Many of you may be skeptical of that because biofeedback and neurofeedback have to some degree been seen as sort of a hippie, you know, not really evidence-based kind of stuff.  But there is really pretty solid basic science literature now supporting this idea that fMRI neurofeedback is a tool that can enable targeted alteration of mental function.  I'm talking here about controlled studies that demonstrated changes in targeted aspects of mental function with large effect sizes.

So just to give you an idea of what some of those are like.  There was a study where they improved perception of a specific line orientation in subjects without affecting their perception of other line orientations.  The subjects were completely blind to what was being trained.  And they counterbalanced the line orientation trained across subjects, but they got very targeted effects.  

They have also shown that you can alter higher level aspects of cognition using this tool.  A sort of creepy study was one where they trained people so as to manipulate their facial preference.  They trained one group of people to like a neutrally ranked set of faces more and another group of people to like neutrally ranked faces less.  And they got changes in the targeted direction.  Again, people were blind to what was trained.  

So, these are just some examples.  Actually, there's quite a large base of science literature now that shows that fMRI neurofeedback can be successfully used to change mental function in people in a targeted manner, certain aspects of mental function, I should say.  
Just because you can change some aspects of mental function in a targeted manner does not mean you can do this with any aspect.  And the critical question is can we do this with aspects that are relevant for mental health.  And I think you're going to hear a lot about the work that's going on in that front in this workshop today, and I think the early results have been quite promising, but it is still in the early stage.  

But I think that basic science research really gives us hope that this is, you know, really going to work potentially.  The second thing I really love about fMRI neurofeedback is it's not just a maintenance therapy where you are going to have to keep getting neurofeedback, you know, regularly for the rest of your life to stay healthy the way many pharmaceuticals work, for example.  I think it has the potential to really shift people on to a healthier trajectory.  
So why do I think this?  Well, studies that have looked at whether the effects induced by neurofeedback tend to persist after the training have, indeed, reported that they do seem to still be there months to years later.  But even beyond that, we had a study where we were running neurofeedback in patients with OCD.  And we were looking at whether the neurofeedback was improving their symptoms as measured by the Y box, which is the gold standard clinical instrument for OCD.  
A decrease in the Y box is an improvement in symptoms.  And we were happy to see that during the period of time when the intervention was applied, symptoms improved in the subjects getting real neurofeedback and didn't change much in the subjects getting shame.  
But we were really interested in whether those effects persisted over time in the following weeks.  At the two-week mark, both groups showed improvement probably because placebo effects were kicking in at this time point.  The sham group started to regress back to baseline, as expected for placebo effects.  But we were a little bit surprised that the real neurofeedback group continued to improve.  So we actually -- we got quite substantial improvements in symptoms in the real group that occurred after the intervention had ended.  

And we were surprised at this at first.  We thought maybe it's something, you know, idiosyncratic to OCD because OCD is a disorder where if you resist your urges that's a form of therapy for people and so then they can get more confidence and start on a healthier trajectory.  

But we did have another study we were running at the same time in Tourette's Syndrome.  And we thought well, we'll just look in that data and see if we see any suggestion of the same pattern.  And, indeed, we saw exactly the same pattern.  And around the time I was going out to lunch with Nick Turk-Browne, who was another neurofeedback researcher at Yale.  And I was telling him about these data.  And his first reaction was, oh, my goodness, we saw exactly the same pattern in an open label study of depression that we ran.  We didn't know what to make of it, but if you go and get the figure from that paper, you'll see it's there.  

And sure enough, that's exactly what his paper showed, symptoms were continuing to improve for weeks after the neurofeedback training had ended.  So, you know, then we started searching through the literature in general and we saw suggestions of this sprinkled throughout.  And so, we wrote a paper about this which was published in 2018 in NeuroImage.

The upshot of it is it seems that changes in symptoms induced by neurofeedback not only persist after training but in many cases continue to grow. And so, this is really exciting from a clinical perspective because it really seems like we are able to push people on to a positive trajectory to lead to lifelong benefits.

So the third thing I want to highlight about neurofeedback is that it seems to me to be a great intervention to pursue for treating youth. We're in the middle of this youth mental health crisis.  And here we have an intervention which is safe and medication free. It's based on natural learning so it can leverage the brain plasticity in young brains.  And also, as I just talked about, it has this potential to shift on to a healthier trajectory which in young people would lead to a lifetime of benefit.  So, I think it's a really promising avenue for pediatric and especially youth mental health development.  

And the last thing I want to highlight is what has already been really discussed by both Stephen and Masaya is how beautifully fMRI neurofeedback contributes to experimental therapeutics kind of framework and leads to this closed-loop development cycle which, as Masaya so nicely put it in his talk, allows neurofeedback and neuroscience to evolve together.  

So, for all of these reasons, I'm very excited about fMRI neurofeedback as an avenue for future therapeutics development for mental illness.  And my group has been very focused on it for a decade and a half.  I currently have a very small team, but absolutely wonderful young people working with me.  And we collaborate with many other people across Yale on different lines of research.  

So we have a line of research focused on OCD.  I'm showing some of the people who have contributed very dramatically to this research.  There is actually a humongous team, I can't represent everyone here.  But these clinical trials take a lot of work and a lot of people to conduct.  So, I'm sorry I can't acknowledge everyone, but this is just a subsample of people who are working on this project.  

We also have a line of research focused on PTSD.  These are some of the people involved.  We are currently running a clinical trial in adolescents with Tourette's Syndrome, hoping to improve their tick symptoms.  And we're hoping to begin soon a study in borderline personality disorder.  This would be in collaboration with Sara Feinberg and Susan Whitfield-Gabrieli.  

So, all of these lines of research have received funding from various sources.  And a lot of these grants have allowed us to get new lines of research started.  But the bread and butter and large support for all of this line of research in my lab has come from NIMH.  And I'm really grateful for all of the support that NIMH has given for what I see as a very promising line for therapeutics development.  

I'm not going to talk about the results from these specific clinical trials we're running because they are all publicly available.  We've published all of our papers, and they are available on PubMed Central.  So, if you just go there and search for Hampson, the disorder you're interested in and neurofeedback, you will find our papers in the results.
But to give you, you know, my nutshell summary of it, I think that the results have been promising.  But it is also clear that more work is needed to optimize these interventions and really get the best bang for our buck.  And that I'm going to talk about more now in the second half of my talk.  

So, I'm switching over to talk about challenges in fMRI neurofeedback early-stage research.  Actually, I'm just going to focus on one challenge because I think it's the major one of interest.  And to understand this challenge, you have to think a little bit about what you think a healthy research pipeline should look like.  

This is what I think a healthy research pipeline should look like.  In the early stage, we should be completely focused on exploration.  So, these are studies where we're going to sample parameter space broadly.  We want to find the most promising options.  We don't want to worry about false positives because they're going to come out in the wash later on, but we want to look everywhere.  

The main thing is we don't want to miss something really promising at this early stage.  And these are usually small sample-sized studies.  Because nobody wants to, you know, throw billions of dollars at something you've never tried before.  So, you expect a lot of small sample size studies in early-stage exploration.  

Then moving on to the middle stage is where you're going to separate the false positives from the real.  So let me give an analogy of this.  So, I thought there was some island in a good climate zone that might have therapeutic plants that have a special compound.  And you can take those plants and test them, investing lots of money and time in doing so to determine if they actually have the compound.  

But first I wanted to sort of find out if this island has promise for harvesting these plants.  And I said I want these people to go and explore the island, right.  What we know about the plants is they like to live in swampy areas, they have spade-shaped leaves, they like a certain type of soil.  You know, please, you know, go to this island, explore and see how much promise it has.  

Now if you were one of these explorers when you landed on the island, what would you do?  Is the first thing you would do go and pick the plants you see right getting off the boat and take them back and start doing time-consuming tests on them?  Probably not.  
Probably the sensible thing to do would be to get the lay of the land and walk around the island and find out where are the swamps and does the soil look like the right type if that is your expertise.  And, you know, did the plants have spade-shaped leaves?  So, you know, every -- all these people are doing jaunts around the island and they're sending back reports.  Oh, it looks like there's a swamp in this area.  And maybe that's a false positive, maybe that swamp doesn't have any promising plants.  

But that's okay because the second stage of research is going to winnow out the false positives from the real.  So, you're going to get a whole bunch of reports saying oh, there's a swamp on the northeast corner of the island and the northeast corner of the island has the right kind of soil.  And look, all of the plants have spade-shaped leaves.  And then you start to say this northeast corner seems really promising.  So, this is the middle stage of research where you're looking for findings that crop up repeatedly.  We're looking for convergence of reports across routine contexts.  Meta-analysis can be very helpful in this stage.  But essentially, you're trying to figure out what is the most promising thing to take into the late-stage clinical trials.

And so maybe you focus on plants from the northeast corner of the island.  You go there, you get your samples, and you take those back.  And that's when you do the time-consuming testing with rigorous statistics to avoid false positives and using very large sample sizes to ensure you don't have false negatives.  So that's classic late-stage research.  So, what I'm getting at here is the pipeline should involve a continuum where the early stage of the continuum is exploration focused.  And it is not focused on power.  It's not focused on statistical rigor.  It's focused on making sure you don't miss something exciting.  

And then, you know, as you move along the pipeline you get more interested in power and statistical rigor.  And when you get to the late-stage research, you are completely focused on power and statistical rigor.  You're testing one thing and not exploring at all.  That's late-stage clinical research.   What I see as the big problem right now -- well, I'll get into that in a second.  Let me just highlight a bit more this continuum.  Because I think it's an unappreciated point that power and exploration tradeoff against each other.  For a fixed number of resources, you can either test one thing really thoroughly with a very well-powered study or you can look at a lot of different things and you're not going to be well powered for everything that you're looking at but you're just getting the lay of the land.  
So, if you're going to be well powered you are not able to do exploration and vice versa.  And so what I see as the big problem at NIH right now is that they put the clinical trial mechanisms on as a blanket over all therapeutics development.  And the clinical trial mechanisms as they are structured pressure the work to conform to what is healthy for late-stage clinical research, but it's extremely unhealthy for early-stage exploration.

So there, you know, the clinical trial mechanisms don't really allow us to explore.  Everything is considered a clinical trial that we do with neurofeedback that is intended for therapeutics development.  And clinical trials are forced to be well powered.  

There is actually a power section required in the human subject section as well as a statistics section.  So, when you put those sections as required parts of a grant, it tells reviewers that these are important criteria that should be considered in ranking the grant.  And so, you can't even submit a grant that is not well powered with any hope of getting it funded.  

And yet to do exploration, you have to accept that it's not going to be well powered, that you're going to use your resources in these small studies to explore.  So basically, power trades off against exploration.  We're not really able to explore.  And so much of our parameter space that, you know, both Stephen and Masaya, as they talked about, we have a humongous parameter space to explore.  So much of it is unexplored and this is a huge shame.  

So you might say, well, isn't this a general problem, like why is this specific to neurofeedback?  I think this is a general problem.  I think that the way that the clinical trial mechanisms at NIH are structured is unhealthy for clinical research in general and clinical therapeutics development in general.  But the other kinds of therapeutics development often have their own avenues for exploration.  So, for example, the pharmaceutical development relies very heavily on animal models.  They can refine exactly what compound is best and, you know, address a lot of exploratory questions in animals, which obviously we don't do.  

And in the psychosocial therapeutics’ development world, they rely heavily on clinical wisdom which is obtained by clinicians who spend, you know, years working with patients and try a lot of things and get a sense of what works better.  And that's a form of exploration.  So, they can kind of distill their clinical knowledge that they've gained from that process.  We also don't have that avenue to explore.  

So it hits us particularly hard in the neurofeedback research world.  So, you know, whereas I think a healthy development pipeline would look something like this for us where we're doing a lot of exploration in the early stage.  And then in the late stage we are just doing confirmatory testing.

The reality is -- and we are -- you know, because we are doing a lot of exploration in the early stage in an ideal pipeline, we would be finding the best interventions and they would be identified in the middle stage and end up going into tests, the confirmatory tests.  
But, unfortunately, instead what is really happening is because we can't explore, we land on the island, and we have to just take one of the plants that is in our field of view.  So we look around to see if any of them have spade-shaped leaves, and we grab them and we take them back and put them in to late-stage confirmatory testing styled projects.  And, you know, this is a huge shame.  

Because we are running confirmatory trials on suboptimal interventions.  Better interventions are probably out there, but we haven't identified them.  And this is going to look really bad for neurofeedback because most likely a lot of the confirmatory trials are going to be disappointing because the best interventions were not entered into them.  
So in general I see this as a huge shame.  I'm going to throw out a few possible solutions that occur to me.  One is that NIH could distinguish between early and late stage clinical trials and their funding mechanisms.  

Early stage clinical trial research should prioritize exploration and deemphasize power considerations.  So this would mean explicitly removing the power and stats from the human subject section.  To be honest, the entire human subject section, this idea of a rigid protocol where you can't change anything as you go is implied by having to define everything in advance in the human subject section.  It's all designed for late-stage research.  It's beautiful for late-stage research and it is totally counterproductive for early-stage research.  

So you know, you could just get rid of the whole human subject section.  And that would actually be an improvement in my mind.  
Another thing is explicitly prioritize exploration over power in the wording of the mechanism and ensure SROs communicate this to the panels.  Because at this stage, we have an engrained mindset where everybody thinks that power is important for everything, and large sample sizes are important all the time.  And it is just I think really counterproductive to early-stage research.  And we need to change the engrained mindset now.

Another obvious potential solution is to increase the size of R21s which are the NIH's flagship mechanism for early-stage research that haven't increased in size for decades.  Because of inflation, they are now worth a fraction of what they used to be worth.  And they are just not large enough to run -- you know, to do much exploration.  At least for fMRI neurofeedback studies.  You can't do much exploration with the amount of money in an R21.  

And then the third suggestion is that we can consider mechanisms that provide greater flexibility.  That fund investigators to pursue lines of research rather than specific projects.  And one of the, you know, concerns about going this route is that a lot of the senior investigators might get all of the money, right, and I think that would be deleterious to younger investigators.  And that has to be, you know, actively avoided.  
A beautiful model to look at is NIGMS Neuro R35.  So they fund lines of research, but they make an explicit effort to make sure that they are funding people at all different career stage.  So they are also protecting the career pipeline.  So I thought that was really a nice grant mechanism when I read it.

So that wraps up what I wanted to say.  I just wanted to take a second to talk about some things that may interest viewers.  So, first of all, for those that do fMRI neurofeedback, the big meeting is called Real-time Functional Imaging and Neurofeedback or rtFIN.  And it's going to be held next in Heidelberg, Germany in early November.  We would love to see you there, rtFIN 2024.  

Also, if you are new to fMRI neurofeedback, you're a researcher thinking of moving into this field, we have recently published a textbook which is designed just for that purpose.  This was an international group effort with a lot of the world's experts on fMRI and neurofeedback contributing.  And it's really designed to be didactic and to, you know, help people fill in gaps because people come from all different interdisciplinary backgrounds into this field and so everybody has got gaps somewhere.  So if you think you might be interested, check it out.  Thank you.

CHRISTOPHER SARAMPOTE: Thank you, Michelle.  And thank you, Stephen and Masaya.  
We have a lot to cover.  And you guys have given us a very nice kind of overview of early-stage research and some of the complexity that's involved and also the challenges.  The types of research, securing funding for that research, but also just the methodological hurdles to overcome.  

We've gotten a lot of -- a number of questions in the chat.  And thank you to the audience members and viewers for submitting those.  I'm going to try to hit as many of these as possible.  But also, I want to open up the discussion to our panelists who may also have questions and want to riff on this a little bit and provide their input.

The other thing, I just wanted to give folks a heads up is that we will -- that we anticipate a lot of overlap between the sessions because I think a number of the researchers that you see here today are conducting research both at early, mid and later stage investigations.  And so, I think there is going to be a lot of input there.

But one of the first questions that I wanted to ask, and this is specifically for Stephen, but I suspect other people may want to comment on this as well is that we've received a number of questions in the chat about control conditions and the selection of control, the pros and cons of choosing controls and the types of controls you choose.  

And, Stephen, I wondered if you could go back and expand a little bit on that and the issues that you see, particularly from an fMRI researcher in how they relate to the choice.

STEPHEN LACONTE: Okay.  Yeah, thank you very much, Chris.  And first I think I should say what Michelle was saying captured virtually everything I wanted to hit in terms of positive controls in terms of early-stage exploration and sampling the parameter space very broadly.  
Once you start to sample that space, what I wanted to say was then you can branch out from there and kind of have a safe zone.  When you start -- in this case I will use the slightly different metaphor.  When you plant the seed and it doesn't grow, there is nothing to control for.  You don't have to control for the rain or the soil or the sunlight because you first need to sort of find these promising results.

And so, I could share my screen again and just kind of go into that "Control Freaks" paper which has this -- I just happened to pull it up because I suspected there might be a question because there seemed like a lot in the chat on this.  What is shown here is a nice flowchart.  And that dovetails -- I'm using dovetail a lot -- but that meshes perfectly with this table.  And then, of course, if you read the paper, it goes into a lot of detail.

I think at the end of the day, the message of this paper, which was not very controlling or super didactic, I found it to actually be very thought provoking and sort of open to flexibly interpreting what you need.  At the end of the day, what I would like to say about controls is once you start to see results and start to get into the later stages -- and Michelle and Kym may want to weigh in on this -- then you want to control for as many things as you can think of and afford that may be sort of negating the fact that the real-time fMRI neurofeedback itself is sort of causing these very positive effects that you are seeing.

But as reviewers and as investigators, we also have to take a step back and take responsibility for figuring out what those controls should be.  And so, these flowcharts are very helpful if you -- if you use them as kind of an exercise that you generalize on your own.  If you start to really cement them or concretize them without taking a step back and figuring out the context of the experiment and what its claims are, then you start to get into this formulaic view that I personally feel is a little bit dangerous and kind of reminiscent of using P values of .05, which is kind of, you know, an arbitrary thing.  And then people just take a break from thinking and just say it didn't surpass .05 statistical significance, so it is not significant.  

Without going the step back and thinking about what positive and negative, you know, what false positives mean and what power means in those contexts.  So that's my take on controls.  I think you as a -- as an investigator should be responsible about thinking about what else could be causing the control.  Once you learn something from an experiment, that doesn't mean that just because real-time feedback was not important, you may actually learn how to do some synergistic things outside of the scanner through your real-time control experiment.  

So, I would just say, you know, try to take on some responsibility.  That's for me as well as everyone else.  We need to think about the context.  Michelle, do you have any –

MICHELLE HAMPSON: Yeah, I mean I'll echo this idea that this obsession with P .05 and statistical rigor is really not terribly appropriate for early-stage research.  Where just your information you're getting, the lay of the land is really what we need to get in the early stages.  I agree.

MASAYA MISAKI: And I also think that unspecific -- so there are so many outcome of neurofeedback.  So we cannot address only the one control condition, but every specific effect.  

So, it is a kind of -- control condition is a kind of a null hypothesis.  So addressing the counter-position may prove the active conditions.  But we also need to find the active condition is really working or not.  So it is a kind of mechanistic understanding of the active condition is also important to validate the interventions on neurofeedback.

STEPHEN LACONTE: I will say that time helps with some of Michelle's exploration.  It's not -- it's not going to fix the sort of NIH mechanism approach or the clinical trial structure.  But as we grow a community, I think that is also an important effort and benefit of meetings like today.  And thanks again to NIMH for hosting this.

CHRISTOPHER SARAMPOTE: Thanks, Stephen.  Let me ask just a follow-up question because there was one -- and this is a question that we received in the chat that relates to this idea of exploratory research.  

And the question is this:  Exploration is lovely if we can detect results when we have them.  I like the recommendation to allow more mechanisms which provide for exploring in larger samples.  I would like to know more about how doing under-powered research even in an exploratory context allows us to detect interesting potential results if that is recommended.  
I see how if we want to restrict ourselves to detected large effects, we can power for detecting those with smaller samples.  Or how can we do repeated measures testing and allow for having different families of tests, but I'm not yet seeing how abandoning power lets me make scientific conclusions.

STEPHEN LACONTE:  One thing comes to mind immediately is we need to push to share data.  You know, as we share actually raw data, we can -- and, you know, I also think that these are challenges within the publication realm as well.  You know, how do we publish data that are underpowered and convince reviewers or even convince ourselves that we have data that are worth publishing.

But as we share data, we can go back and scrutinize those data, compile those data, amass them across different centers to do it.  So that is one thing that comes to mind.  Obviously, the question is quite a challenging one, and we may not have exact answers for all of it.  Michelle, do you?

MICHELLE HAMPSON:  Well, the question is undervaluing the importance of replication in the early stage.  So, I think that yes, you're going to get false positives.  You don't conclude anything from a single false positive in an early-stage research study.  But you look to see, you know, what emerges as showing up again and again in literature.  To me, that's how you find what is real and separate it from what's not real in a healthy early-stage pipeline.  

So, I think there is this misguided notion now that we have to know if it's real based on a single study and that you can't publish anything unless you know that that's real with great confidence.  And I think that's a problematic mindset.  We need to say no, no, you can put -- you know, there are going to be false positives and that's a healthy part of the pipeline that you accept.  But you have to also read the literature with recognition that these are small sample studies, and they have to replicate a lot before I invest in a study and go forward with that as a late-stage trial.

MASAYA MISAKI:  False positive is a part of science.  So, of course, we need to validate later with more information.  But we can't start without some finding, if it was a false positive or a real positive, so yeah, we should start with that.

STEPHEN LACONTE:  I mean, what do we think about false negatives?  That's also a challenge, and then we're --

MICHELLE HAMPSON:  And I think that's worse in the early stages.

STEPHEN LACONTE:  Yes, much, much.

CHRISTOPHER SARAMPOTE:  Another question that we received, what are your thoughts on screening potential participants on their ability to up and/or downregulate brain activity prior to enrolling them in a study?  

Considering that several studies show wide variability in subject's ability to regulate neuroresponse.  Reddy, et al. in 2020 reported that attention, motivation and mood can be important influencers of neurofeedback performance.  
Should these factors be assessed in all neurofeedback studies even if they aren't necessarily the specific phenomena of interest?

STEPHEN LACONTE:  Well, one thing that comes to mind is in the past when we have done multi-session neurofeedback studies, on the initial assessment day -- this was work Pearl Chiu and I were doing -- on the initial assessment day we actually put people in the scanner for half an hour and have them do a really simple motor neurofeedback task.  And essentially, we had stage enrollment in the study.  

If they were not staying still in the scanner and couldn't control kind of an interface using really simple motor tasks, then we didn't pass them on to the next stage of the study.  So that was used as kind of a filter.

But also kind of an exercise in training participants about kind of the time sluggishness of neurofeedback given the hemodynamic delays that are inherent in this neurovascular coupling.  So that's one approach would be to sort of screen participants based on their ability to be still in the scanner and teach them what neurofeedback -- what they can expect in a neurofeedback experiment.

CHRISTOPHER SARAMPOTE:  Thanks.  Another question.  A large issue in fMRI is related -- related to the knobs is the choice of software.  
Could you talk a little bit more about software packages that you use and whether any comparison has been made between these to address reproducibility.  This may be just as much of a problem as the choice of control condition.

STEPHEN LACONTE:  Well, I do think again in terms of reproducibility, offline analysis is useful.  So, you know, my group-built a -- my group started off directly programming the Siemens scanner and then we realized that that was impractical, and we were going to have to reinvent the wheel.  

We built a plug-in for AFNI which already had real-time capabilities.  And so our plug-in in AFNI is running during the experiments.  But once you are done with the experiments those data can be reanalyzed in, you know, SPM or FSL or any other package that you want.  
Now, you know, there are a lot of software options which I think are good, and it ties back to growing the community as well as we need more software that is available.  One cool thing that AFNI has is the ability to replay data through your real-time pipeline.

And so, if you have a NIfTI file, you can essentially feed it through a command called RTFeedMe and simulate the scanner collecting those data.  And so that is -- you know, that's one part of an answer to that question.

MASAYA MISAKI:  We implemented our own system for real-time neurofeedback, but it's also based on the AFNI system.  And I think between the availability or accessibility and the flexibility is a tradeoff.  So, some systems are accessible, very easy to use, but it is smaller room to optimize or customize for each individual experiment.  
So there is no standard way for the neurofeedback yet or no standard software.  Yeah, there are many products.  Many are using their own in-house implementations.

CHRISTOPHER SARAMPOTE:  Another question that we received, so we talked about the sharing of data as being important.  And just for simple, where do researchers in neurofeedback share their data?  Is it in OpenNeuro?  What other repositories might exist or, again, can be encouraged?

STEPHEN LACONTE:  Yeah, I think OpenNeuro is a great option.  We have -- with Cameron Craddock we've shared data through kind of the Rockland Study site as well.  I think if you share the data they will be findable.

MICHELLE HAMPSON:  I will chime in about the whole data sharing thing, though.  

STEPHEN LACONTE:  Yeah.

MICHELLE HAMPSON:  Because I feel like there is a big emphasis on data sharing and there's a lot of people now whose careers are pretty much focused on analyzing data that is publicly shared and they're not doing data collection.  

And it's much easier for them to -- you know, they can just download this data, analyze it, publish.  It makes it very hard to encourage young people to learn to do the data collection and to learn to design well-designed data collection experiments.  It is almost a sacrifice on their careers because they can't compete, of course, in terms of the publication rate with just downloading and analyzing publicly available datasets.  

So, I think there also has to be an effort to try to get credit to the people who collected really any data.  It's difficult to collect data.  They have to be included somehow as authors, it has to further their career; otherwise, we are not going to be supporting that kind of ability to do this data collection.

CHRISTOPHER SARAMPOTE:  Another question that I have -- I'm actually going to combine two questions that just simply use the same word, and that is replication.  

And the two-parter of that is, first of all, how do you decide how many replications are needed before the field has established that a signal is real, positive, reliable and significant?  

And then flipping that, do you have any suggestions or comments on how many numbers of sessions of neurofeedback training are sufficient as fMRI scans are quite expensive.  So, there is the reliability of a signal, replication of the signal, and then there's the replication and reliability of treatment.  And I know I'm cheating by putting these together, but could you comment on those aspects?

MASAYA MISAKI:  FMRI neurofeedback need fewer number of sessions compared to the EEG neurofeedback.  But it depends, of course, on some target populations.  So, for example, in our studies, targeting the participants so the very -- so quick response.  So may comment about that.  

But if we apply the same protocol to the PTSD participants, it takes more sessions to control that, it depends activity and the population and approach.

KYMBERLY YOUNG:  And I'll just throw in as a kind of a preview, I will be addressing how we establish the number of doses that we give for our real-time neurofeedback intervention at my talk in the second session.  And some of the issues that neurofeedback researchers can determine more broadly when determining dose.

CHRISTOPHER SARAMPOTE:  Thanks, Kym.  Another question -- actually, you know, you talked about sharing data.  Are you also able and willing to share AI algorithms and model carts for researchers?

STEPHEN LACONTE:  For our AFNI plug-in, that's shared.  So we have access to AFNI's GIT repository, and we do a medium job of keeping it updated.  But certainly, the crux of the capabilities is shared.  And then there are some specific communication parts that are specific to, you know, what the -- what each site has available.  And also, if it is a GE, Phillips or Siemens site.  So it tends to be a little bit difficult to get sites up and running even with the free -- with the free software to do so.  But we -- then we do it on kind of a site-by-site basis to provide support.

MICHELLE HAMPSON:  Yeah, I have to say I think the fMRI neurofeedback community is really great about sharing software.  Seems like everybody is willing to share their software and they are very helpful, you know, if you ask questions and there's a lot that's publicly available just on the web that you can take.

MASAYA MISAKI:  Yeah, I share my software on Git Hub.  Documentation is not so great.  So, most of the time --

MICHELLE HAMPSON:  That's usually the challenge.

CHRISTOPHER SARAMPOTE:  Another question.  Thank you.  

And this, by the way, other folks I think might want to comment as well.  And that is, given the evidence that multiple neuro network strategies might result in improvements in behavior and symptoms, how might we leverage a technology like tDCS in conjunction with real-time fMRI given that tDCS modulates broad brain networks in a way that other neuromodulatory methods do not?

MASAYA MISAKI:  So, yeah, we can combine the real-time monitoring of the brain activation with the online stimulation to optimize the parameters.  So yeah, I think not only the tDCS but also focal intervention like TMS also affects the whole-brain activation patterns.  So, I think we need to see the whole-brain activation patterns where other regions affect even in the focal interventions.  So yeah, there is a possibility of combining such a stimulation, external stimulation and real-time neuroimaging to optimize such parameters, I think.

MICHELLE HAMPSON:  We were really interested in the idea of using tDCS to try to amplify neurofeedback to see if you can, you know, make those areas, you know, learn quicker because of the tDCS.  So, I think there are a bunch of kinds of combined applications that would be really interesting.

SUSAN WHITFIELD-GABRIELI:  I agree.  And I think the whole idea of combining a lot of these neuro modulatory interventions to prime the neurosystem that you subsequently are going to be intervening in or just sustaining clinical benefits.  

For instance, real-time fMRI neurofeedback could be used synergistically with, say, Ketamine, for instance, when we know it doesn't have a sustained clinical benefit necessarily.  Or you can use, to your questions, tDCS to prime the neural system to real-time fMRI to increase treatment efficacy.  

So, I think that's a fabulous question, and I think a lot of people who do neuro modulatory interventions might be siloed.  But we -- I think we need to combine them because simultaneously -- which we will talk a lot about, Michelle and Kym will be talking about it, too; or synergistically in a serial fashion to prime and sustain benefits.

CHRISTOPHER SARAMPOTE:  Following up on what Michelle said about having pharma research having animal models as a starting point.  

Do you think animal studies could contribute to neurofeedback research?  For example, guiding exploratory trials or validating imaging studies?

KARINA QUEVEDO:  I dare to say yes.  I'm just going to jump in.  I know that there are already technologies that can actually somewhat intrusively monitor animal behavior in real-time.  
I mean there are microelectrodes that you can basically insert in the brain of the rat, and you can actually, you know, specifically provide the kind of stimuli that will give us in-depth mechanistic circuitry pathways and very much time-by-time exposure to a stimulus to circuitry.  And some of that research has even been done, and I know some of the people that do that at the University of Minnesota, if they are still doing that.  But I know they were a while ago.  So those were my two cents.

CHRISTOPHER SARAMPOTE:  Another --

STEPHEN LACONTE:  I mean I think on the physiologic side to understand what is possible is good.  But I do think animal studies are challenging.  You know, first of all, you need to have an animal model of -- of whatever population that you are interested in.  
And then how do you coach them in terms of cognitive strategies or, you know, how -- how homogenous or how conserved are the networks and even regions that are involved for that.

And even just the -- you know, going back to Michelle talking about the idea of it being natural learning, neurofeedback being natural learning.  You know, what are the reinforcers then for an animal model versus a human model of the disease.  And so, you know.

MICHELLE HAMPSON:  Yeah.

STEPHEN LACONTE:  The brain as behavior is a powerful thing that gets unlocked with real-time fMRI.  But the, you know, difference then going from human to animal models is -- the gap is huge, I think.

MICHELLE HAMPSON:  Yeah.  I think there are some interesting things that you could do with animals that would be relevant for fMRI neurofeedback, but I don't think the bread-and-butter early-stage exploration can be done in animals.

STEPHEN LACONTE:  Yeah, but physiologic limits I think would be useful and a good example.

MASAYA MISAKI:  And also the difference between the animals and humans.  In human neurofeedback, you have the moderation with the brain process.  But you cannot do in animal.  So, in animal model we do enforce, but it is kind of an implicit learning.  So that it may be different from the human.

CHRISTOPHER SARAMPOTE: I have a question for Michelle and Masaya about mechanisms.  
What are the mechanisms for acute versus long-acting effects of real-time fMRI neurofeedback?  Do we have longitudinal studies that can tell us this?

MASAYA MISAKI: For me, I don't have so much data.  We have only a week or a few weeks after the training.  So as Michelle presented, so interestingly the treatment effect is enhanced as the time goes on.  So yeah, it -- so Michelle can comment.

MICHELLE HAMPSON: Yeah, I would say that we don't have an answer now, but it's a super-interesting question that a lot of us are interested in and we're now trying to collect data.  At first, we weren't collecting data, you know, at follow-up, brain imaging data.  But now we're often scheduling brain imaging scans at follow-up to get information about, you know, what has changed in the brain since the neurofeedback ended.  So, we get it right after neurofeedback and then get it maybe a month later.  So, it's a very interesting question that we would all like to answer to.

CHRISTOPHER SARAMPOTE: Great.  Questions are coming fast and furious, thanks, everybody.  And this is one that we likely -- other panelists will want to -- will be talking about later.  But we often see practice effects or challenges with near versus far transfer and cognitive training.  

How do folks think about over fitting in the -- over fitting in the training context?  Are there unique challenges and opportunities related to near and far transfer effects in fMRI-based neurofeedback?

MICHELLE HAMPSON: Can I just clarify what you mean with near and far transfer effects?  Does this just mean transferring to something that is very, very similar versus transferring to a task that is somewhat different?

CHRISTOPHER SARAMPOTE: Yes.  By the way, and since this question came from one of our other members, Luke Stoeckel at the National Institute of Aging.  Luke, you are welcome to unmute and ask a follow-up if that's helpful.

LUKE STOECKEL:  Yeah.  No, you're right, Michelle, that was -- that was it.  Yeah, training the same thing.  I was also thinking about the outcomes themselves.  Often, you know, we're using similar outcomes where there might be practice effects.  

So, I wonder how those things are accounted for in the context of real-time fMRI and if there is anything unique about real-time fMRI neurofeedback related to those issues.

MICHELLE HAMPSON:  I mean I think that these are like all things that require more work for us to map out.  It -- you know, I was really surprised by some of the basic science studies which suggest very targeted effects which in some ways is really impressive that we can -- we can have these very precise narrow targeted effects and not affect other things so that we are not like drugs affecting the whole brain.  

But, on the other hand, you know, sometimes we want things to kind of have a little broader impact, so they affect you in a, you know, more general circumstances.  So, I think it's a -- there is a sweet spot, and it is really an empirical question, you know, how well we can -- how well we can tailor it to that.  But I think one of the features of neurofeedback is this wonderful development, you know, loop cycle where if something is too narrow we can try to think of ways to broaden it and look at the neuroscience for that.  So, there's this -- sort of chronic ongoing optimization that we can engage.

CHRISTOPHER SARAMPOTE:  Thanks, Michelle.  One follow-up question.  So earlier on, Michelle, you talked about the challenges.  And I think all three of you guys talked about the challenges of early-stage research and the challenges that come with power and also finding suitable mechanisms that kind of match up to that.

So at NIMH we have a UL1, which is a cooperative agreement mechanism for early-stage device development.  And I'm wondering if those kind of cooperative agreements or early-stage device notices might be applicable to the kind of work you're doing?  And if not, then what are -- I think you were highlighting the other challenges, too, that go on.  And maybe you could expand on that a little bit.

MICHELLE HAMPSON:  It is hard for me to comment on those because I'm not super familiar with the mechanisms.  And are you thinking neurofeedback would be considered a device in that case or that the mechanisms would be broadened to include?  Because currently we are classified as psychosocial.

KYMBERLY YOUNG:  That was going to be my point is that we're currently considered not -- we are not considered device-based intervention and not considered part of the device-based community.  At least we have not been up until this point, and I will discuss that a little bit in my talk as well.

HOLLY LISANBY:  I will just hop in here and welcome you to the device-based community.  As we will be hearing in session three, neurofeedback is a device, whether it uses fMRI or EEG.  And actually, that is part of the purpose of this discussion.  

And the UL1 mechanism, just for folks to know, does specifically address the early stage needs for both drug and device development.  And the emphasis is on exploration, the emphasis is on discovery.  Some of the work is first in humans so it is very, very early stage.  And we will be happy to put a link to that in the transcript from the meeting.

MICHELLE HAMPSON:  Is it a large grant?

HOLLY LISANBY:  Yes, it is not limited in size.  I will also point out the -- our -- some of our mechanisms when you need to exceed $500,000 per year, permission can be requested in advance for that.  

I will also point out you were mentioning the R21 as a high-risk/high-reward or early-stage work.  And for studies that are seeking to develop an intervention, those actually need to go to the R61, R33, at least at NIMH, which does not have a cap on its funding, whereas, the R21 does have a funding cap.

MICHELLE HAMPSON:  So one of the things I struggle with, you know, you used the term first in human.  You know, because most of the exploration has already been done when you get those mechanisms.  
So that's that problem for us with those kinds of mechanisms, we are competing against people who have done tons of exploration and have a lot of data supporting their intervention and we haven't been able to explore.  That's difficult for us.  We need mechanisms that actually just support early-stage exploration.  They need to be big enough that we can actually do some exploration, but we can't be competing for these really large grants where you are expected to have a lot of data already supporting you.  That's kind of the challenge for us, I think.

HOLLY LISANBY:  Yeah, so I will point out -- I will encourage you all to look at the UL1 mechanism.  It is meant for early stage.  And when you're dealing with a device that is not FDA cleared for the treatment of a psychiatric condition, which currently real-time fMRI neurofeedback is not, that is considered early stage because it is not currently clinically approved.  

And, you know, much of that work is to figure out how to dose the tool.  You know, I think that we learned today that there are a lot of knobs, there are a lot of parameters.  And the early stage mechanisms are meant to help investigators work out those dose response functions.  To figure out are you going after a region of interest or are you going after a network, how do you develop and optimize the machine learning or AI approaches to extract and decode the signal, how to do the timing of the feedback.  

We learned from Stephen LaConte's talk that timing the feedback is very important, how to set the controls.  So those are all considered early stage questions which we do want you to work out before.  Then using the tool to interrogate a target.  And the R61 or R33 is really about target engagement.  And we want to know before you get to that stage that you have good evidence that the tool is likely to be able to engage that target.  And that's where the early-stage work is meant to be done before you go to R61/R33 phase.

CHRISTOPHER SARAMPOTE:  Thanks, Holly.  Thanks, everyone.  We are actually ahead of time, and but I wanted to just -- and so maybe we can have a little bit of a longer break.  
But I did want to ask a more general question.  We received some questions from folks who are asking kind of this general question, which I think a number of you have already answered.  But I would like to return to that and ask the question of why fMRI approach, why fMRI-based approach is now?  

And if you could expand a little bit more on the benefits or pros of fMRI approaches particularly as they may not only inform treatment development but also they might inform this larger question of helping us to understand the mechanisms of disorder and how this tool contributes.  

And so, I guess that is another question.  There have been other approaches along the way, but why fMRI now?  Why is this exciting?

MICHELLE HAMPSON:  Well, I think fMRI allows you to translate from, you know, basic neuroscience.  It's very easy to figure out how basic neuroscience translates to fMRI.  
You can be informed by a huge rich body of information.  And then it also -- you know, as people have discussed in the session, it gives this incredibly rich mechanistic data as you're running it, so it really allows you to do continued refinement of your intervention and targeting it better and better in a closed loop kind of development cycle.  More so than other imaging modalities typically do.  So that's my take.

MASAYA MISAKI:  And neuroimaging study with fMRI and neurofeedback, fMRI neurofeedback is increasing with its development.  So I think that, as I said, so neurofeedback involved with neuroimaging evolved, I think.  So that is why fMRI is now interesting.

MICHELLE HAMPSON:  You are muted, Chris.

CHRISTOPHER SARAMPOTE:  Curses, technology.  Thanks.  

Michelle, you had mentioned that there is a value in real-time fMRI neurofeedback for youth.  Can you tell us how much work has been done in children and adolescents?  And also, I suspect others may be commenting on this as well.  

MICHELLE HAMPSON:  There has been a fair amount of work.  Karina, I'm sure, is going to talk about some of her work in the next session.  And there has been some work in the UK on kids with ADHD.  We're doing adolescents with Tourette's Syndrome.  Of course, most of the work that's going on is older kids because it is very hard to get little kids to stay still in the scanner.  But that's -- those are the things I'm most aware of.  I don't know if others can chime in and mention other things.

KARINA QUEVEDO:  Yeah, I just wanted to reiterate what Michelle has stated that the initial steps were primarily done in ADHD in King's College London.  And there seems to be a healthy balance of both positive and not so positive results.  And that is just a way of generating valuable data.  

But I believe that the field in children and adolescents is very much expanding to other conditions.  And I'm very excited and honored to be here later.

SUSAN WHITFIELD-GABRIELI:  I'll add also to what Karina said in that in collaboration with Randy Auerbach's wonderful lab and in Columbia we're trying to target repetitive negative thinking like rumination in teens with depression.  So, I think the field is growing quickly and we're quite excited about it.

CHRISTOPHER SARAMPOTE:  Thanks.  And it's a little preview of what's to come.  I did have a question that was submitted for all.  

So, Masaya argued persuasively that understanding mechanism is key to figuring out how to kind of set the knobs that Stephen told us about.  What do you see as the highest priority areas for research to address unanswered questions about how real-time fMRI neurofeedback works?

MASAYA MISAKI:  Yeah.  At least for my personal interest is mechanistic understanding, or for more neuroscientific interest for the neurofeedback effects.  But, of course, it's important as an intervention method.  So that we should focus on the outcome of the cognitive functions or symptomatic scientific effects.  

So yeah, so I -- I'm not sure that it is universal interest, but at least for me it is mechanistic understanding is my current priority.

STEPHEN LACONTE:  I agree.  And, you know, maybe I'm slightly adjacent to what Masaya is saying.  

But I'm really excited about real-time fMRI's capability to further refine our understanding of functional roles of regions and networks in the brain.  
And so what you end up with is an experiment that can fail.  And so you have a stronger falsifiability through this closed-loop experimentation than you do in a sort of open-loop fMRI setting.  And so what that means is that you can start to scrutinize aspects of your task or your pipeline and understand how that affects ability to neuromodulate.  

So if you give -- if the literature points to 10 or 12 cognitive domains for anterior cingulate, you can start to take tasks or task strategies from those different domains and do head-to-head testing to see how that maps onto neuromodulation.  And then you can study your data post talk as well and see, you know, network or broader downstream changes that happened in that case.  So that's --

CHRISTOPHER SARAMPOTE:  Yes, Talma.  Dr. Hendler.

TALMA HENDLER: I would like to address here something important to remember that more feedback is actually establishing an association between the brain state and the mental state.  

And as far as it is important to study the neuromechanism, there are two levels of learning that are occurring simultaneously actually.  The brain learns, so to say.  And the mental -- the mental state is also being modified.  And I think these are two paths that are very important to study.  

And to put effort in, in order to improve the efficacy and outcomes of neurofeedback, we should remember that we have these two levels.  And I think for the neural level I would like to contradict what was said.  Actually, animal work might be very, very valuable but also repeated sessions in humans might allow us to learn a little bit more about reinforcement learning or other learning mechanisms that take place.

But at the same time, the mental operation that is being formed while we are trying to affect the brain is also something that is being learned.  And that is very open yet to our understanding.

CHRISTOPHER SARAMPOTE:  Thanks, Dr. Hendler.  We have time for a couple more questions.  I do want to ask one, and then we will go back to the chat and try to grab some, too.  We talked about a little bit about software and algorithms.  I guess this may be for Masaya and Stephen.  What work needs to be done in the software space?  What are the key questions to be addressed about signal abstraction to train the modulation of distributed networks?

MASAYA MISAKI:  The fMRI signal itself is very slow, so has some -- fMRI has five second’s delay.  So at least it should be shorter than the one volume.  So after that, the faster is not so much advantage for the fMRI signal, I think.

STEPHEN LACONTE:  And, you know, something Masaya was saying earlier I would like to kind of expand on which is sort of the documentation and the complexity of the software.  
And so, you know, getting this to be more plug and play would help us to expand the community.  We didn't get into the nuts and bolts of setting up a real-time fMRI system.  But at the very least you need computers that are communicating with each other and, you know, having access to the data very quickly after acquisition, either reconstructed or unreconstructed form.

And so right now it is -- it's -- you know, it's best to have a dedicated system or machine that's there.  But a lot of groups, you know, because of policies with the scanner may have to resort to bringing a laptop in and then you have to test all of these things.  
It just adds a much more experimental complexity and kind of a sense of well, this experiment might fail or it's way too hard to get done six months later, you know, after we set it all up.  

So, I think on the software side, getting things a little bit more plug and play would help a lot with everybody being able to do these, making this accessible.

MASAYA MISAKI:  Yeah, so we need to work with MRI vendors for synergy to implement the real-time data exportation.  So that part I think is the bottleneck of most real-time systems.  So, data transfer from the scanner to the external processing computer is actually the most time-consuming part now.   So I hope the MRI vendors will help us to implement a standard protocol or some more time efficient real-time implementation with the data export.

STEPHEN LACONTE:  Yeah.  We have a very tight coupling with the Siemens system and the AFNI system. So, we actually make it so that the operator of the scanner can just run it and as much as possible is automated.  

But that requires a lot of scripting.  And so then those scripts become hard for other groups to read and maintain, even if we are sharing them.  You know, that's a lot of work.  And so they can take what we develop, and we can help them to try to generalize it, but it is a challenge.  So the software side of things is a challenge.

CHRISTOPHER SARAMPOTE:  We are actually at time for a break.  And so, I apologize I can't get to all of the questions.  

But I did want to ask just one more question, or at least -- and that is a number of folks have asked about the role and the long history of EEG-related neurofeedback.  And one of the decisions we made in putting this meeting together was to specifically think about and talk about fMRI-informed real-time.  Real-time fMRI neurofeedback approaches specifically because it is where we are seeing a lot of new research and a lot of new applications.  
And I'm wondering if you could comment a little bit on the history of EEG research.  And I also -- and how you think about this modality as it relates to other modalities like that.  

And I would also just for those people who were asking, I just want to give a little preview, too, for session three when Dr. Talma Hendler is going to talk about her work which actually thinks about the use of fMRI, how it informs EEG-delivered neurofeedback.  

So general question about EEG and how it relates to your work in real-time fMRI neurofeedback.

MICHELLE HAMPSON:  Well, I'll just say there's a ton of interest in the field, you know, of doing what Talma Hendler has demonstrated, which is taking fMRI neurofeedback and eventually being able to translate it to other modalities and use it effectively with cheaper more accessible modalities for the purposes of clinical dissemination.  It's a kind of a gold ring in the field.

MASAYA MISAKI:  I think that it is also way to translate fMRI neurofeedback knowledge to translate to the EEG implementation is one way, too.  So more accessible neurofeedback approach.  

So, I think the EEG neurofeedback has a long history, but in most cases usually the signal is kind of limited compared to the fMRI neurofeedback.  The fMRI neurofeedback can target specific brain region in various parts, but EEG usually use some like a frequency measure or something like that.  

So, it is very limited compared to the fMRI stimulus.  So if we can use such a fMRI neurofeedback signal, translate into the EEG neurofeedback, it could be more accessible and more effective to achieve.

CHRISTOPHER SARAMPOTE:  Thank you.  Thank you all very much.  This has been a great discussion, and I really appreciate your presentations.  We are at a break.  
Is that right, producer Deborah?

DEBORAH KRAT:  Yes, until noon.

CHRISTOPHER SARAMPOTE:  Let's aim for noon.  So, we'll take an 11-minute break, and we'll start session two at noon.  So thank you all very much.
    
(Break)

ALEX TALKOVSKY:  I can jump in while Chris gets his audio together.  

So, I'm Alex Talkovsky.  I'm also a program officer here in the Division of Translational Research.  And it's my pleasure to moderate this second session here today.  We will be hearing from three presenters before we have our discussion.  

They will be in order, Susan Whitfield-Gabrieli, Karina Quevedo, and Kymberly Young.  I will ask them each to pass the baton to the next presenter when they finish up with their slides before we go into discussion after Dr. Young's presentation.  

So, Susan, whenever you're ready, you can please go ahead and share your screen and take it away.

SUSAN WHITFIELD-GABRIELI:  Thank you, Alex.  First, I wanted to give a tremendous thanks to NIMH for hosting this phenomenal workshop.  I think it's a really terrific opportunity to share ideas about what I think is a really transformative intervention, and that is real-time fMRI feedback.  

Today I will talk about the challenges and opportunities of personalized network-based neurofeedback and neuromodulation in psychiatric disorders.  I think it is fair to say we are facing a tragic youth mental health crisis.  And not that long ago, the U.S. Surgeon General and leading pediatric health organizations have all declared a national state of emergency in adolescent mental health.  And it's clear that we need urgent innovative personalized treatments that target the poor mechanisms underlying mental illness.  

When I take a step back and I think about critical barriers that are preventing us from moving the needle in mental health treatments, I'm thinking of four such barriers.  One is that the mental healthcare system can be still primarily still oriented towards remediation instead of prevention.  Treatments are often trial -- given on trial and errors and we know that there is great heterogeneity in treatment response.  

Often, gold standard treatments are not very effective.  Only 30 to 50% of patients often respond to pharmacological and behavioral interventions and they often have negative side effects and high attrition.  And when they do work, we don't necessarily know why.  So, the pathophysiology of mental illnesses is unknown, some drugs were accidentally discovered, so on and so forth.  But I think that real-time fMRI neurofeedback can help address some of these critical barriers.

And today I will be talking a little bit about how we can potentially use it for neuroprevention, neuroprediction, neuromodulation and in the case of real-time neurotriggering.  So we and others have showed that baseline human brain neuroimaging can predict progression of symptoms such as anxiety and depression years later, even in a normative pediatric sample in children ages at seven can predict worsening of, say, anxiety and depression internalization for years later.  

And it would be great if we knew in advance who might be more high risk.  Then we could actually offer things like real-time neurofeedback and help mitigate symptom progression and possibly even avert illness.  We and others have also shown that baseline neuroimaging can predict treatment response.  But the advantage of real-time neurofeedback is that we can also look at neuroflexibility and neuromalleability and look at the change of resting state networks pre/post to intervention.  That might be an even better prediction of long-term treatment efficacy.  

And, of course, in the real-time neurofeedback framework we are hoping that this real-time feedback people will learn how to modulate their networks and have a subsequent modulation and mitigation of clinical symptoms and increase in cognitive performance.  
However, going through all of this process, it's really important, as we heard about in the first session, to try to understand more and more about causal mechanisms.  And we can do this in many different ways.  We can do this by multimodal neuroimaging, say with MRS and fMRI.  We can, say, look at how GABA and glutamate may be changing pre/post some of these real-time neurofeedback interventions.  

We can also use real-time in a different context.  And that is rather than using real-time fMRI for feeding back activation or connectivity, we can use real-time fMRI to detect brain states at which point we can trigger either fMRI test stimuli or experience sampling.  And in that way, we can actually build predictive models so that we can identify brain behavior and brain symptomatology in a causal way.  So, understanding causal mechanisms, of course, will help refine these interventions and ultimately tailor our treatments.  

So how does this all work?  Of course, we are all using an fMRI scanner, which means we can look at brain structure and brain function.  Today, I will be talking about more of the intrinsic functional architecture of the human brain as revealed by resting state functional connectivity.  And here we are looking at ultra slow frequency fluctuations that cohere on interfunctional networks and we call them resting state networks because they cohere in the absence of any task whatsoever.  

So, in the red, you'll see regions of the brain even if they are anatomically far apart that rise and fall in temporal synchrony.  And importantly, intraindividual differences in resting state networks actually give great insight into brain health and disease.  Luckily, these brain networks are plastic and can be modulated with interventions like real-time fMRI neurofeedback.  

And what I hope to excite you about today is that real-time neurofeedback can actually use these entire network and network interactions in the context of neuromodulation.  So we can build what we are calling network therapeutics.  And so we believe that being able to modulate large-scale networks and their interactions with other networks will really provide a more effective method of neuroregulation than neuromodulation involving a single lesion or anatomically unspecific pharmacological interventions.

But I'm going to take a pause here and just mention that this is really a tour de force.  And, you know, this is -- this real-time neurofeedback offers many opportunities.  But, as you know, it comes with many challenges, many of which you've heard about already.  And I can say I think with some confidence that there are many challenges that we have run into over the last 20 years.  We've been doing this since the early 2000's.  And we're still really in the exploratory phase in many circumstances.  We are still learning, we're developing new approaches, new algorithms, we're tweaking, we're optimizing acquisition parameters, analysis and feedback methods.

So but the idea here is that when you are funding the real-time fMRI neurofeedback community, you are also not only funding for that particular project, you are also funding people who are developing a lot of really novel neuroimaging tools that can be used in basic science modalities and I think address gaps and barriers in the field, open up new windows to understanding brain network organization and function and help in disease and really pave the way for us to innovate biologically-informed approaches including real-time fMRI neurofeedback but not limited to that.

ALEX TALKOVSKY: Susan.  

SUSAN WHITFIELD-GABRIELI: Yes.

ALEX TALKOVSKY: I'm sorry to interrupt, but for the sake of our closed captioning could you please just slow down a little bit?

SUSAN WHITFIELD-GABRIELI: Oh, yes, it's a biological problem.  I'm going to get real-time feedback to try to slow down, but I'm working on it.

ALEX TALKOVSKY: Thank you.

SUSAN WHITFIELD-GABRIELI: So in our case, we have something called Vsend which is a patch developed by Paul Wighton and Andre van der Kouwe at MGH Martinos.  And they have been developing this for many years and they're maintaining it for us.  But it takes a lot of work.  In fact, right now they have been working on it for quite some time to work interoperatively with the Siemens upgrade, which many of you may have had some experience with.  Nontrivial.  

And so, once you get the data, which you have to do very quickly, then you need to clean the data.  Because, as you know, artifacts like motion artifacts can have a huge deleterious effect on the activation and connectivity that we're trying to look at.  In particular, like motion artifacts can substantially affect things like the default in the network.  

Once we clean the data, then of course we have to analyze the data in a way that we can feed it back quickly like incremental GLMs.  We may look at activation.  We may look at connectivity.  But as you know from the first session, there are many different knobs to turn.  And you do this and put it all together in this feedback scenario.  Ours is an open source called MURFI that -- initially built by Oliver Hinds back at MIT.  And we're continually modifying this and upgrading it.  

And we are using a system to do neurofeedback but also neurotriggering.  And I will tell you a little bit more about that.  But then eventually, of course, we want to scale this intervention and so that requires simultaneous EEG and fMRI.  We are trying to build something called a mind-balance training platform.  But again, that's a huge amount of technical development that's going on behind the scenes.  

So because it takes so much effort, it really takes a village.  So in our case, it started in the early 2000s when Christopher DeCharms came to Stanford and asked can people learn to regulate brain activation in specific regions using real-time fMRI neurofeedback.  We didn't know and so we tried it.  And after a long time of tinkering, we were able to train people how to modulate their motor cortex.  Then he published the first paper in 2004.  

Then a more perhaps interesting question after you're convinced of that is does learned regulation of activation or connectivity of these regions influence the mental processes mediated by those regions?  And I think that is what a lot of us are working on now.  When I moved to MIT, I had the good luck in running into Eden Evins and Luke Stoeckel who did a lot of work with real-time neurofeedback and addiction.  

And they organized a beautiful conference out of which came this nice optimization real-time fMRI neurofeedback for therapeutic discovery and development paper.  Many of us are using real-time fMRI neurofeedback to help clinical conditions, and in fact, with Margaret, she -- we were trying to help people with schizophrenia to reduce their auditory hallucinations.  We got an R21 to do this.  

But as Michelle was saying, an R21 doesn't necessarily cover the costs that are needed to turn all of the knobs, so to speak.  And so, we were working on this for quite some time.  It turned out that we actually needed to combine real-time neurofeedback with mindfulness meditation.  And that was really the sweet sauce for us.  And currently we're using that modality -- the combined modalities in mindfulness meditation and real-time neurofeedback in the case of depression with our colleagues from Randy Auerbach's lab.
    
But why is it that we need this mindfulness meditation?  Well, it all comes down to the default mode network.  So this is a figure from Fox 2005.  We know the two core medial hubs are the default mode network are related to self-preferential processing and mental time travel.  So in a healthy state of mind we may be reminiscing about the past.  If we're depressed, we may be ruminating about the past.  If we're in a healthy state of mind, we may be planning for the future.  But if we're anxious we may be obsessively worrying about the future.  

And we and many others have shown that the degree to which these regions are engaged in the form of hyperconnectivity or hyperactivation is associated with more psychopathology across many different disorders.  Interestingly, the regions in blue here are showing negative or anticorrelations.  And the magnitude of these anticorrelations is related to executive function like working memory performance which was first shown by Michelle Hampson, and many others have replicated that and extended it.

So, we have two main hypotheses -- well, really it actually boils down to one.  That hyperconnectivity, the default mode network, is a targetable transdiagnostic mechanism underlying mood and psychotic disorders.  And so, what we want to do is use real-time fMRI neurofeedback to quiet down these regions that are apparently hijacked in many different clinical disorders.  And not only do we see the hyperconnectivity within these self-reference nodes, but we also see that the self-reference nodes kind of rope in, if you will, for a colloquial term, other disease in symptom-specific regions.  

Like in the case of psychosis or schizophrenia, the default mode network kind of ropes in the superior temporal gyrus which may view self-relevance to auditory hallucinations.  In the case of depression, it may rope in the subgenual anterior cingulate which may be associated with internalization.

And this hyperconnectivity we and others have seen precedes illness, indicating that this may not be a manifestation of the disease but rather a biological trait which we can target with real-time neurofeedback early in life.  The other form of default network hyperoconnectivity is expressed as an underrepresentation of default mode network anticorrelations which we believe is a transdiagnostic, again targetable biomarker of cognitive impairment.  

And so, this magnitude of these anticorrelations kind of follow an inverted U curve.  They selectively grow with the executive function.  And of course with typical development they peak in young adulthood.  And then they trail with aging.  We have seen these significantly reduced in many clinical populations.  And when you're looking at intraindividual fluctuations of intentional issues, you can see that it is predicted by these anticorrelations as well.

So, there are many different ways to address this.  And when I looked into the literature, I saw that actually from more invasive to less invasive interventions, whether it's DBS, ECT, TCS, Ketamine, antidepressants, many are mitigating the hyperconnectivity in some way.  
So, it would be nice if we could do that with lifestyle choices such as exercise and mediation that can be augmented with real-time neurofeedback in terms of meditation.  So, for instance, let's take TMF.  Many people target the region of the dorsolateral prefrontal cortex that is maximally anticorrelated to the subgenual ACC, anterior cingulate.  And that -- although, you know, it doesn't work for everybody, it is a common technique to treat depression at the moment.

And what we found is that this treatment target, the subgenual ACCDLPFC connectivity actually predicts worsening of anxiety and depression even in teens who are diagnosed with anxiety and depression.  And even in children who are at familiar risk.  So, if you look at familiar risk population, you can see that the connectivity between the subgenual and DLPFC predicts worsening of anxiety and depression and internalization and predicts conversion to illness.  

Even in young children who are not preselected to be at risk for any kind of mental disorder, this subgenual DLPFC connectivity, the anticorrelations predict worsening of internalization and anxiety and depression four years later.  

So if we can backtrack back all the way to a young normative population, why not try something a little bit softer like mindfulness meditation where you draw attentional focus to the present moment and mimic cognitive awareness and ongoing thoughts.  So, there's a vast literature indicating that mindfulness meditation downregulates the default network, increases GABA, BDNF and across many different clinical populations decreases loss of different clinical symptoms.  

And we and others have shown that it can decrease self-perceived stress in school students and increase sustained attention.  Other people have published recent papers showing that the mindfulness-based stress reduction is actually non-inferior to SSRIs to treat anxiety and is relatively similar to CBT for people experiencing depression.  

So, with this information, it seems to me that it would be really excellent to try to get younger and younger kids to do mindfulness meditation.  The problem is the very people who might be afflicted with anxiety and depression may not find it easy to do mindfulness meditation.  And there is no feedback, so you don't know how you're doing.  

So this is our current substantiation of real-time feedback.  We have our patients practicing meditation.  We put them in the scanner and then we identify their own individual networks, the default network and the frontal parietal control network.  As soon as they start doing mindfulness meditation that downregulates the default and upregulates the frontal parietal network, we convert this signal into a game which is this ballgame.  And the target is to move the ball up.  And the ball moves up when the default is being suppressed lower than the frontal parietal network.  And the ball moves with a hop size proportional to the magnitude of the difference between these networks.

When the person starts mind wandering and is not doing mindfulness meditation successfully, then the DMN goes up again and the ball goes down.  So essentially the people are moving their brain -- their mind -- using their mind to push the ball up and down when they are oscillating from mindfulness to mind wandering.

And we really like this technique because it is noninvasive and we are identifying the specific intrinsic networks, the functional networks which is a whole another hour discussion about the importance of that.  But -- and it's adaptive.  So, this ball if you can't do it at all, the ball size will adapt to your ability.  And most importantly, I find that this is a network level intervention.  So not only are you mitigating the hyperconnectivity within the default network but also for free get a mitigation to the other regions like STG in the case of schizophrenia, subgenual ACC in the case of depression.  

So, targeting these large-scale networks has a lot of different clinical benefits and is what I think Michelle was alluding to earlier.  And one of the things that I think is best about all of these real-time neurofeedback interventions is that the patients feel agency.  Many different treatments the patients are on the receiving end whether it's DBS, TMS, whatever.  What's really beautiful to watch the transformation in patients is mindset transformation while rather than being a recipient of an intervention, they are actually the actor in the show.  And they see how the brain networks are changing and they feel how the clinical symptoms are mitigated.  And I have to say there is a lot of excitement amongst the patients.  

So, I'll just give you two examples.  Schizophrenia.  We know schizophrenia is a severe psychiatric disorder that has disturbances of thought, perception and neurokinin deficits.  And we know hallucinations are a key issue there.  And we know STG is implicated.  So, we have seen this in many different studies.  We see the degree to which MPFC and PCC are correlated and predict outcome.  And we see that the degree to which MPFC and STG are kind of correlated during rest also predict outcome early before clinical onset.  

And we know that there is a high comorbidity with anxiety and depression and that anxiety and depression are risk factors for the onset of illness.  And so, we can actually see, if we look at a clinical high risk, we can see a double association such that MPFC PCC goes with depression and anxiety, whereas the MPFC STG goes with hallucinations.

So here is an example of a patient with psychosis moving the ball up when he is successfully doing mindfulness meditation; and the ball moves down when he's starting to mind wander.  And what we see is a mitigation of the MPFC PCC hyperconnectivity goes down.  And we see an increase in the MPFC DLPFC anticorrelation, so they have anticorrelations at first, but they go down afterwards.  

And when we tried this with the sham condition of motor cortex, we don't see the modulation of these networks.  And importantly, it is the MPFC STG that most correlated with the reduction of hallucinations.  So now we're trying to do neuromodulation in schizophrenia.

With targeting, rather than the DMN, let's try the STG.  So we tried the STG, the superior temporal gyrus and then found in the real condition, in the real feedback a mitigation of the MFPC PCC and a mitigation of the MPFC STG which were associated with auditory hallucinations reduction.  

So here I have shown you two different examples to target the network and then get reduction in AC -- in superior temporal gyrus.  Or you can target superior temporal gyrus and get a reduction in network.  So then you might ask why go for the network?  And I think the reason to go for this large-scale network is you get other downstream effects that may have relationships to, say, internalization of anxiety and depression.  

So, for instance, we know that the hyperconnectivity in depression, there is the DMN kind of ropes in the subgenual.  We see this in teens with anxiety and depression, and we see it even in kids who are feeling at risk.  So there is hyperconnectivity of the self-reference node, subgenual ACC and the -- each of which is correlated with internalization in anxiety and depression.  

So if we have teens with anxiety and depression, we can mitigate that hyperconnectivity both within the network and within the network in the subgenual ACC.  And then we also saw an increase in state mindfulness.  So, we're excited about this because we think that this mindfulness-based neurofeedback may be a transdiagnostic intervention for both mood and psychotic disorders that have downstream clinical benefits.  

So going forward, we have a number of clinical trials doing this.  We're really excited again for early detection and early intervention, as Michelle was saying earlier, to really, you know, capitalize on the plasticity of the brain younger and younger.  And we also are trying, as Michelle mentioned, to augment DBT that has a component of mindfulness.  
But we want to scale it so we're trying to do simultaneous EEG-fMRI to capture the DMN and DMN anticorrelations.  That's with Aaron Kucyi.  And then with Clemens, he is trying to passively collect the EEG while doing the real-time feedback in the patients with schizophrenia so that we can scale it up. 
And then we also know that the GABA anticorrelations tend to go -- GABA increases with anticorrelations.  And so now with Margaret we have -- we are going to be trying to understand mechanism.  So, we have MRS which is local -- voxels are localized with the resting state.  And we're going to look at GABA and glutamate pre/post the feedback in clinical high risk.  

We're just launching off a big clinical trial with Randy Auerbach's team where we're going to compare just mindfulness alone versus mindfulness feedback.  One big question that many people have is how you know that neurofeedback is actually augmenting skill acquisition and utilization, maybe you're just looking at the effects of mindfulness.  
Well, we're setting out to actually look at that.  Other people have looked at TMS, the cerebellum, and found a reduction of negative symptoms.  And so, we have another grant where we're looking at the comparison between neurofeedback to mitigate positive symptoms and TMS to mitigate negative symptoms.  And with Christian Webb we are asking for whom and for when.  We know that exercise and neurofeedback both decrease depressive symptoms and so we're trying to look at the within and between person moderators.  

I will close up now by talking about some of the other options -- opportunities for real-time neurofeedback.  Usually people like use baseline predictors of outcomes.  So, for instance, the baseline DMN anticorrelations in our case predict the patient's ability to modulate their networks as well as mitigate symptoms.  But the nice thing about the neurofeedback is that we can get neuromodulation predictors.  We can actually measure the neuroflexibility or neuromalleability, if you will, based on the change of resting state networks pre/post the day one intervention.  And that can be a protonation prediction.  And then when you look at the final change in network, that can be the mechanism.  

And so, you can also think of kind of an analogy between cardiology and psychiatry.  Where in cardiology you're looking at imaging the heart as the organ of interest during rest and during stress.  And we are doing the same thing with imaging the brain during rest and during stress.  And you can imagine going to the gym to gain brain fitness.  You know that when you're doing cardio -- you know, exercising whether it's cardiovascular exercising or pumping iron, you're going to get an increase in fitness.  And we can do the same thing with the real-time neurofeedback.  

So finally, I will just say that it is really important to identify these individual -- these within subject fluctuations and we can do a lot of understanding causal relationships between the brain and the mind as Talma was talking about through experience sampling.  Because then we can actually find out what the person is experiencing.  So we have done this within subject prediction fluctuations of attentional issues.  We can do the same thing with rumination.  

And so, I'll just show you -- finish by showing you one example of where Clemens identified resting state networks.  And he triggered, he did biological trigger experience sampling where you ask the subject what they're experiencing.  And we're triggering -- the blue areas mean we think they're mind wandering.  The subjective is responding.  So clearly more of a mind-wandering state here.  And then they go to the gym and do their mental fitness.  And afterwards they are more in a mindful state.  And so here when you do the biological experience sampling, which is much more efficient than experience sampling, you can get a much better mental state.  

So hopefully we can, you know, get the electrophysiological correlates of this DMN and then scale it in the real world.  Because once we're able to, you know, identify optimal/suboptimal brain states, then we can combine it, do phenotyping, and trigger just-in-time interventions.  

So I wanted to thank everybody here.  And Chris, for really doing a phenomenal job organizing this.  And Michelle has put a lot of effort into this and all of the program officers in NIH and NIMH in general.  Thank you so much.  

I'm going to now pass it off to Karina Quevedo, Associate Professor in the Division of Child and Adolescent Psychiatry at the University of Minnesota.

KARINA QUEVEDO:  Thank you.  Let me start sharing my screen.  And then hopefully this will work out.  Thank you for your patience.  Okay.  

So, actually, Sue teed off to me beautifully because she actually talked about the importance of taking advantage of the inherent flexibility of a pediatric population.  So the talk of my -- the topic of my talk today is how to hit a moving target.

ALEX TALKOVSKY:  Karina, if I may, we see the presenter view.

KARINA QUEVEDO:  Sorry about that.  I had that issue before, and I now need to fix it.  My apologies.  They helped me during the dry thingy, and I can't fix it now.  Give me a second here.

DEBORAH KRAT:  Karina, go to display settings again, the dropdown.

KARINA QUEVEDO:  Let me just disconnect from this.  And then okay.  So display settings at the top.  Sorry about that.  Okay.

DEBORAH KRAT:  Yeah, click the dropdown arrow under display settings.

KARINA QUEVEDO:  So the top?

DEBORAH KRAT: Yes.  And then click the arrow.

KARINA QUEVEDO: There is something else here so I can't --

DEBORAH KRAT: Okay.

KARINA QUEVEDO: I apologize.  Okay.  So is that the one?  All right.  My apologies.  All right.  I think I will see it now.  Okay.

DEBORAH KRAT: And swap presenter view.

KARINA QUEVEDO: Swap presenter, there you go.  Come on, buddy.  Thank you for your patience.  It is not working.  All right.  Let me do it again.  All right.  I think I need to project first and then swap.  My apologies.  I'm sorry about the technical difficulties.  
I need to -- so how about if I stop here for a second?  I wonder if Kym would like to go ahead of me so I can solve this?  Is that possible?  Kym, do you feel comfortable doing that?

DEBORAH KRAT: No, that's okay.  Stop sharing for a second, and I will go ahead and share.

KARINA QUEVEDO: Okay.  Thank you.

DEBORAH KRAT: Okay.  Karina, you may go ahead.  I'm sharing your slides with you so you may go ahead now with your presentation.  
Karina, if you are still there, if you come off mute, please.

KARINA QUEVEDO:  I'm right here.  Sorry about that.  I had technical difficulties.  So you are sharing the slides from your setup?

DEBORAH KRAT:  Yes, that's correct.

KARINA QUEVEDO:  Let's do that then.  That's fine.  Thank you.  Sorry for the delay.  
So basically, we are going to work today on to how to hit a moving target.  And essentially yes, the topic of my talk has to do, first of all, with essentially what I developed throughout my career.  And the first thing that I tried to do was to define and localize sensitive targets in pediatric populations.  

Then we went on to establish whether the target was engaged and via what processes.  And then finally we are on to the adventure of whether the target is basically linked to behavior or to psychopathology.  To that, we are going to basically do this through four different studies that are concatenated and related.  

We are going to look at a sample of depressed and suicide-attempted youth that underwent tasks that had to do with self-processing.  And there were 122 of those youth.  And then a sub-sample of those youths, that were 81 of them.

And then we are going to look into are the targets engaged and via what processes.  And that is basically a feasibility study of neurofeedback that we did with very little basically funds but that brings us to the point that Michelle was trying to make about exploratory studies.  And finally, I'm going to give you a preview as to whether the targets are linked to behavior or to psychopathology.  And that is phased innovation study that basically had 50 participants per target.  Next slide, please.  

All right.  So, I'm going to introduce a heady concept to all of you.  And to the idea that basically self-processing or self-awareness has the basic function of organism survival.  And I'm going to orient you to the fact that this very key function has developed both epigenetically and ontogenetically to reach us.
And that has endowed us with a capacity for symbolic thinking.  And these are the process that allows for abstract and symbolic representation of the organism.  Next slide, please.  
So, basically, symbolic processing relies on the use of images, concepts, abstract relations, language, numeracy, artistic or ritual expression.  Only human beings are able to do this as far as we know.  It also allows us to act contrary to self-preservation because it allows us to separate the immediate concrete circumstances of the organism from a symbolic representation of the self.  Next slide, please.

So non-coincidentally, suicide risk changes along development.  It is noted in adolescents during the age of 14-20 years, there is an increase in suicide risks.  And at the same time, symbolic self-processing consolidates.  For the first time in development, individuals are able to represent the self via symbolic abstract schemas of the self that are separated from concrete self-representation.  

And one such form of symbolic processing is the self-face images, which as you all know adolescents spend numerous hours curating in TikToK and in Facebook and in SnapChat.  So really in some ways, social media has turbocharged their ability to extend their self into symbolic ways to both shape who they are and to communicate to others.  
So I was originally interested in the seat of self-awareness.  And by studying suicide, it allowed me to return to my original interest and to the hardcore question of who we are and how do we visualize ourselves and how do we project ourselves into the future.  Next slide, please.

All right.  So with that goal, our lab designed this task.  It's a very simple task, but it's a very complex task.  And that is the emotional self-other morph task.  Adolescents are inside of the scanner, and they see around 150 or 200 faces and they have to recognize the self or the face of a stranger.  These faces are also morphed to basically add a certain degree of difficulty to the task.  

And basically, they do this for long periods of time to basically map in the localization and the intensity of activation in areas that have to do with self-representation.  And it goes really fast.  People can do this extremely fast, which is an advantage of the task.  Next slide, please.

What we discovered.  So, what we discovered was -- and I need to brag a little bit about it -- we basically came up with a first biomarker of suicide attempts using precisely this task of self-recognition.  Hit enter, please.  So that you can see that exactly.  
So what we found was that depressed-attempting youth showed higher left amygdala to rostral anterior cingulate cortex connectivity when they are recognizing the self and other faces.  And you can see that there is almost a dose response there in terms of how impacted the youth is and how emotionally engaged the amygdala is.  

We specifically interpret this connectivity marker as a marker of compensation and potentially a marker of psychopathological state.  Can you please hit return, please.  And I was particularly interested what was happening in the other connectivity circuit which was the connectivity between the right amygdala and the subgenual ACC.  Hit enter, please.

Because what we found was that high ideation youth showed higher right amygdala to rostral anterior cingulate cortex.  And interestingly, this did not differ from youth who had strong suicide ideation but not actual attempts.  So I was interested in this marker, and I decided to follow this through either longitudinal or with experimental therapeutics.  Next slide, please.  

So, in the meantime, I'm going to summarize this.  We found basically an anterior cingulate to right amygdala circuitry there.  And that this was associated with activity and connectivity during self-face recognize.  Next, please.

We are also exploring a subset of this sample, 81 adolescents.  And we found a very beautiful bilateral cluster in the medial temporal areas, and it included hippocampus, parahippocampus, fusiform, and amygdala.  And you can see it very clearly there that the depressed youth that are in green had a lower level of activity in this bilateral cluster when they were recognizing their own self-face compared to the control healthy youth who had a very robust activation on this bilateral cluster when they were seeing their own self-face.  And this is using the same task that we basically used for the suicide attempters.  And you can hit enter and you will see that the task is the same.

So, when we localized this cluster, we decided to carry forward and said OKAY, let's use bilateral amygdala and hippocampus, provide neurofeedback training to these youth, and see if we can move the cluster in order to change how the individual perceives and symbolizes the self.  And that is exactly what we did next.  Next slide, please.

So basically, our target was the bilateral amygdala and hippocampus.  The processes that we were hypothesizing had to do with memory, and specifically with the kind of memory that happens very fast but also very much in reaction to the stimuli of the self.  And we paired neurofeedback to the self-face in 53 adolescents, controls and depressed.  And again, our target was the bilateral and hippocampus cluster and pairing the neurofeedback to the self-face images.  

So, I'm going to show you how the general outline of the study was in the next slide, please.  All right.  So they arrive for a first session.  They told us their story.  We established diagnosis.  And then in the second session right there in the scanning room, they came up with happy memories.  And then inside of the scanner we provide a neurofeedback using the MURFI software that Sue introduced to the bilateral amygdala and hippocampus paired to the self-face.  Next slide, please.

They had pre and post tasks.  So they completed the same self-recognition task at the baseline.  Then they had neurofeedback.  Please hit enter.  And then they completed the self-task afterwards.  So, let's give you a preview of the results as we go into this journey of trying to locate the target.  Next slide, please.  

All right.  All right.  So these two minutes into the task, you can see that the youth were looking at their own faces and trying to upregulate that bar of color that turned green when the target was high.  And they would turn red and go down when the target was low.  And they were supposed to count backwards to the stimuli of a different face.  And the idea here was to engage basically working memory but also the circuitry that is associated with face recognition which is very robust and very fast.  And this gets to the idea of "Control Freaks."  You do what you can.  And what you can do with control conditions with neurofeedback.  But this is what we chose as a control condition.  Next slide, please.

All right.  So what we found was a very robust whole-brain network that was engaged by the neurofeedback condition which as Misaki and others have alluded to is associated very much with any strong learning task.  Can you please hit enter.  And then what we notice is that there was this very interesting engagement of the anterior cingulate cortex for that specific coordinate which was again reminiscent of what we were seeing with the suicide-attempting youth.  So, we began to say, you know, we need to begin looking at the anterior cingulate cortex, maybe not the subgenual but let's hit the dorsal here because there appears to be a cognitive control aspect here which is very much the function of the dorsal ACC.  Next slide, please.

We also found that the youth were able to regulate the amygdala and hippocampus versus the control conditions.  This is, by the way, controlling for multiple other covariants that we enter into this mixed linear model.  And that also baseline conditions of the target, specifically amygdala and hippocampus activity during the self-task was associated with what was happening during the neurofeedback task.  So that idea of repeated measures is very much an important aspect of neurofeedback research design.  And we confirmed this.  Next slide, please.  

All right.  So then we decided to deep dive into this same population.  And we looked at the suicide attempters.  So, in this population of, you know, the small sample of 53 youth, we began to look at the connectivity patterns in the suicide attempters.  Next slide, please.  And found again a left amygdala story.  Please hit enter.  As you can see, we replicated the finding during neurofeedback of a hyperconnectivity of the left amygdala with a number of new cortical areas that were associated with our task.  But again, the right amygdala did not differ between the depressed attempted and the non-attempting youth.  So, this was basically a moving target upon which we zoomed in by its absence.  Because it didn't differ between key psychological and pathological conditions, I was particularly interested in it.  Please hit enter.

So exactly, it wasn't there, it didn't differ between the groups.  So I figured we can begin to bring that right amygdala online and see if we can find an interesting curative effect here.  Please hit next slide or enter.  So finally, we zoomed in into our two moving targets.  The right amygdala circuitry which we know that is associated with fast implicit and holistic process.  And the goal was to strengthen the fast, implicit holistic self-preservation.  Because we know that the acts or the actions that lead to suicide in this population are impulsive decisions that are not well thought out.  The role was to strengthen cognitive control, flexible attention towards self-preservation.  Next slide, please.

All right.  So that brings us to what we are doing now.  And into the question of what to do, you know, with the time that we have and with the questions that we posed.  What targets do we look to?  What is important once we have designed this?  And this is the phased innovation study that I have had the privilege and the effort to conduct.
And basically, we put the dorsal ACC and the right amygdala in a race.  We made them compete against each other.  We provide the youth two days of neurofeedback with about 22-minute session.  And we decided to say may the best target win.  And we also sampled behavioral measures of emotion regulation, self-processing and psychopathology after neurofeedback.  

And this is where things get really interesting, right?  What is the right outcome to hit your milestones.  Are you interested in pathology or interested in behavior?  Please, next slide.  So again, the youth were seeing their own face and elevating when they saw their own face and decreasing where they saw a different face.  But basically, this is how the display shows when they're engaged in neurofeedback.  Next slide, please.  

And this is already very much preliminary data.  We have concluded our sample collection.  And we had an amazing success given the fact that this took place in the middle of COVID.  And I'm just going to give you a taste of what comes next.

If you want to talk more about the timings, I want you to notice that we had follow-ups at around three weeks and five weeks after the two neurofeedback interventions.  Next slide, please.  And this is what we found.  So we have initially very good news.  Both targets, the dorsal ACC which is in black on the left is associated with significant decrease of suicide ideation over the three time points, as well as the amygdala has a significant decrease of suicide ideation over time.  There is a gentle trend but not entirely insignificant for the amygdala to be associated with a faster, linear decrease of suicide ideation in the youth.  Next slide, please.  

I posed the question what do we want to target?  Do we want to target psychopathology, or do we want to target behavior?  Keep in mind that in youth and pediatric research we are often leaping from adult research that sometimes is not unconfirmed or where it is not entirely full information in the field.  I basically sprung from my colleague Kym Young's research.  There is only one publication linked to amygdala neurofeedback in depressed adults.  And this was the springboard where I basically reached my milestones.  So, I'm going to give you a taste of what we found.  Next slide, please.

So, we basically administered this task that is very interesting.  This is basically a dot probe task.  And it yields a vigilant score of sustained attention for a given stimuli.  And you do that by subtracting reaction times during congruent trials from congruent trials to different emotional faces.  And the faces are negative, neutral, or positive.  Next slide, please.

Here is the little present that I'm going to leave you with.  I had the pleasure of replicating my colleague Kym's finding whereby amygdala neurofeedback was associated with increased vigilance to positive faces.  And as you can see, when the target was the dorsal ACC, the error bars are very wide and there doesn't appear to be a strong effect when it comes to attention to positive faces.  The other thing that we replicated was the decreased attention to negative faces, but I didn't want to make the talk very long.  But we replicated that also.  Next slide, please.

So I would like to then leave you with this question.  I am very honored to be here.  Just remember that when you are looking for target engagements in children, you are literally looking for a moving target, a target that is incredibly malleable, full of possibilities.  But is it based on neurofeedback for adult psychopathology often enough?  It's often based on cross-sectional research.  

However, the excitement we can offer is precisely the nature of the target's plasticity.  Our research has the possibility of not only discovering new brain mechanisms but also confirming adult findings.  Next slide, please.  

This takes a village.  I wouldn't be able to do any of this without my wonderful students, my wonderful collaborators and research assistants.  And they will let you know that this was truly a labor of love over two years that we were pretty much in lockdown, but we got it done.  Thank you very much.  

Those are just follow-up slides in case that people have questions I can go back to the follow-up slides to show you.  I have a lot of hidden slides to answer questions.

KYMBERLY YOUNG:  All right.  And it's my turn now.  I want to thank the organizers for putting together this wonderful workshop and for giving me a chance to talk to you all about the opportunities and challenges as we move into efficacy trials and beyond.  

And so I am going to talk about opportunities and challenges that we have run into with moving fMRI neurofeedback into the clinic.  And these are -- this is the list of topics I will be talking about today.  And I'm going to start with classification and regulatory challenges.
And this is not specific to my neurofeedback protocol.  This is specific to the fMRI neurofeedback community in general.  I was specifically asked by several members of the organizing committee and several members of the panel to discuss this.  And these slides on classification and regulation were put together in collaboration with many experts from the neurofeedback community and aren't just my personal opinion.  

So I think the biggest challenge facing us right now as neurofeedback researchers is the issue of where do we belong?  And we've touched upon this in the earlier sessions today.  Are we a device-based or are we a psychosocial intervention?  And currently, all of our applications go to the psychosocial therapeutic and preventive intervention RFA.  But there's a question as to whether or not it would be more appropriate if we went to the device-based intervention RFA.  

We are not actively doing anything to the brain.  We are not adding anything or subtracting anything.  We are using processes that are already there.  Participants are shown activity, and they are taught behavioral and mental strategies to learn to control it.  The therapeutic effects are induced by a learning-based processes.  And what is really important and really differentiates us from the device-based community is that the device is not the intervention.  The device is augmenting a behavioral intervention such as savoring, mindfulness and emotion regulation.  

Furthermore, all of our studies on ClinicalTrials.gov are registered as behavioral interventions.  And so moving us to device could lead to general confusion or inconsistencies.  Now the case for device-based is we would be evaluated by experts who understand issues that are particularly unique to us related to mechanism, control conditions and power.  The challenge with that is that experts in the device-based community in general do not consider us a part of that intervention.

And so, they -- putting us together with device-based interventions feels like it's misclassifying us.  It's like asking a fish to climb a tree.  And this has historically been the case.  So, in this particular review by Dr. Lisanby, you can see that the definition of non-invasive neuromodulation does not include neurofeedback.  It is device-based interventions that apply electrical or magnetic field to modulate neurofunction.  

The search terms for the review that were included in this were TMS, tDCS, cathodal and nodal stimulation as well as electroconvulsive therapy.  And here are some more recent literature reviews and meta-analysis on neuromodulation techniques.  And again, I would like to draw your attention to the fact that they do not include neurofeedback.  They include TMS, tDCS, ECT.  And this graphic right here again is a -- based on a 2020 article by Dr. Lisanby that shows all of the devices that are considered device-based neuromodulation.  And you can see that fMRI is not included because, again, we are not doing anything actively to the brain.  The device is augmenting cognitive strategies.

And so, this -- we are worried that if we go to the device-based study section that reviewers would not see us as responsive or fitting in with the RFA and our applications would not be discussed.  And that could lead to reduced funding or even stop funding for novel neurofeedback interventions.  

And this leads -- also being part of a device-based community leads to the question of whether or not we require FDA approval, which is one of the challenges that I can tell you all fMRI neurofeedback researchers fear.  And right now, fMRI neurofeedback is currently very well regulated in the research domain.  There is FDA regulations governing scanner usage and IRB oversight of research.  

There is concern that once it moves into the clinic and it's no longer covered by IRB oversight it could be used improperly.  And that is when we would love to work with the FDA to develop the best regulatory strategies.  But we argue again that this is more not a device-based intervention, we are more akin to a FitBit.  A FitBit shows you the number of steps you have taken and then you can make behavioral changes to change that number of steps.  The watch will even alert you when you have taken too few steps to get you to engage in those strategies.  And I argue that fMRI neurofeedback is much more similar to measuring steps, measuring time than it is to deliver a current to the brain to change neurofunction.  

We haven't sought FDA approval in the past because the FDA does not regulate behavioral therapy programs.  And we have up until now been very strongly considered psychosocial therapeutic intervention.  And the FDA mission is largely to protect the public health from devices and drugs that present a risk to the health, safety and welfare of the subject.  And there has been numerous neurofeedback, fMRI neurofeedback studies published.  And there have been absolutely no adverse events reported.  

So, the concern that there would be consumers purchasing neurofeedback equipment without appropriate equipment supervision or training is really not applicable to the fMRI.  Because the MRI scanners are FDA-approved medical devices and Joe Schmoe can't set up an fMRI in his basement and open a neurofeedback clinic.  I have been engaging in conversations, my colleagues and I, with the FDA for more than the past decade.  And in the last communication we had with them, which was a little bit over a year ago, right before my confirmatory efficacy trial was funded, they cited -- one of my colleagues had a phone call with the FDA.  And they cited a specific title, chapter and section that said that once an MRI machine is approved for use, if we're using it in an approved way, the specific uses are not regulated.  

So the function of MRI is to produce images of brain activity using magnet and radiowaves and to see how certain factors affect this function.  Neurofeedback is determining how mental strategies affect brain function using stock pulse sequences.  We are not doing anything that MRI wasn't originally intended or approved for.

The FDA also states they do not regulate the practice of medicine, which is how you use the FDA-approved tool.  Even if we do use the device differently from labeling, it is considered the practice of medicine, which is not regulated.  And so, one question that I really hope our friends at the FDA answer in the next session is why now are we being considered device-based and why now are we being considered for as needing FDA approval?  

It is also unclear what needs approval.  So, is it the fMRI device that needs to be approved?  If so, is it just neurofeedback in general and just one of us as researchers need to sacrifice ourselves on the altar of bureaucracy and work with the FDA to get neurofeedback at large approved?  Or is every neurofeedback protocol going to require approval?  

Currently, if you look at ClinicalTrials.gov, there are 95 active neurofeedback protocols.  At the estimated time frame of three months and $30,000 to get approval, you're looking at 24 years of estimated labor and $3 million that is going to be put onto researchers.  Not to mention the amount of resources the FDA is going to need to contribute as well.  

And this might not be a barrier for someone more established, but especially for junior investigators who don't have the same number of resources as those of us who are more established, this could kill the future of neurofeedback trials.

The other question that we could be regulating a software.  And as you have heard from the first session, open software that is open source and flexible is critical to neurofeedback development.  We are constantly changing our software, constantly improving our software.  And that is one of the keys to the advancement of neurofeedback.  

And the FDA says that software functions are exempt if they are intended to log, record, track, evaluate or make decisions related to developing or maintaining general health or wellness.  We argue that neurofeedback software allows individuals to track their brain activity in order to make behavioral and cognitive adjustments.  

There is also the question as to whether or not the software should be considered a medical device.  But the FDA defines this as software that is to be used for one or more medical purposes that performs these purposes without being part of a hardware medical device.  And neurofeedback software can't be used without the hardware medical device of an MRI machine.  

Furthermore, it is unclear when we need approval.  If, for example, software is what -- the neurofeedback software we are using is what needs approval, the FDA said that it meets the definition of a device but poses minimal risk to patients and consumers they will not expect manufacturers to submit premarket review approval.  There have been no adverse events reported, suggesting this is a very low risk device.  

Also, there is the question of whether or not we need an IDE or exemptions.  And the FDA specifies that MRI specifically is a nonsignificant risk device because there is no potential for serious risk to the health, safety or welfare of the subject.  They state that if the IRB determines a study is a nonsignificant risk device, the IRB may approve that study without the submission of an IDE application to the FDA.  

So, to suggest that if we do require approval, we don't need it premarket.  The FDA did just approve one of the first neurofeedback interventions for PTSD that is EEG-based, and it was approved in July of 2023.

And you will hear from Dr. Hendler in the next session.  But she started out doing real-time fMRI neurofeedback and she did not get FDA approval.  And she did all of these preliminary trials on EEG neurofeedback intervention without FDA approval.  It was only once she was ready for commercialization and to start her own company that she sought FDA approval.  
And so, we argue that it is premature to seek approval when confirmatory efficacy trials are ongoing, and we don't know if the intervention will go forward as currently designed.  
So that was just my discussion of I think the most important challenge facing us.  Now I'm going to talk about some other opportunities related specifically to our real-time amygdala neurofeedback intervention.  

And one potential challenge that and opportunity that we had was dose determination.  Hence, it's critical for neurofeedback researchers as the proposed treatment parameters including dose have to be included in applications for funding, including R61s.  And traditionally the way to measure dose response has been to deliver the intervention, see the clinical change, deliver another dose, see if the clinical change, and continue administering the intervention until the desired clinical change is reached.  

Now, as Michelle alluded to earlier in her talk, one of the exciting things about neurofeedback is that the clinical improvements grow in the weeks following completion of the intervention.  So here in this 2018 paper, you can see that two months after the final neurofeedback intervention, clinical symptoms continue to improve.  We have seen this in our own study.  
And so these are the -- those current results from our R33 looking at amygdala neurofeedback in treatment-resistant depression.  And what you can see is that at 12 weeks you are seeing significant improvements still in the active group.  And, indeed, if we just looked at clinical change after the first two interventions, we would not conclude that the control intervention was -- or that the active intervention was any more effective than the control intervention.  

It's only once we get out to the extended time periods where we see that the individuals in the active group are continuing to benefit, continuing to improve, whereas those in the control condition are not getting better.  And, indeed, they are not dropping into below the moderately severe depressed range.

So, optimizing the number of sessions by embedding assessments assumes that symptom improvement between the adjacent sessions is driven by the neurofeedback rather than the delayed effect of earlier sessions.  So, this should really inform our neurofeedback designs and suggests that crossover designs may be contaminated by significant carryover effects as there is no assurance that symptoms will stabilize from the active intervention even when the arms are spaced weeks apart.  

So instead, I propose that we should be looking at neural target engagement.  Not whether or not the symptoms are changing or how much the symptoms are changing, but whether or not the brain region that's being targeted is changing and whether or not there's a limit to that change.  

In our R61, we provided five sessions of real-time fMRI neurofeedback.  And what you can see when you look at the performance on neurofeedback across visits is that there's a very nice learning curve that occurs on the first day.  Afterwards, not so much.  And, indeed, after the second session the only blocks that are significantly different from that initial baseline are the baseline and transfer blocks where no neurofeedback information is provided.  

It suggests that after the learning occurred, feedback actually interfered with learning and performance more than enhancing it.  And this is true with other studies of domains of feedback learning that once you learn what you are being taught that the feedback will actually interfere with further learning.  

We can also look at neurofeedback success, which is traditionally defined as the difference between baseline and transfer.  Instead, we looked at the difference every day.  So where did they start and where did they end at each session.  And what you can see is, again, most of the change occurred during that first neurofeedback visit.  And by the fifth neurofeedback visit, the change is negative, again suggesting that too much neurofeedback is interfering with learning.

Finally, we can look at when the amygdala activity asymptotes.  And so we look -- when looking at that information, what we see is that the amygdala seems to asymptote at visit two.  So, based on all of this data that shows that you really see learning effects on the first visit and then by the fifth visit the feedback does not seem to be helping, we have decided that two visits are what we are going forward with in our clinical trials.  Two visits are sufficient to engage in learning and result in the neural changes that we need to see for the clinical improvement to occur.

Another potential challenge but also an opportunity is predicting individual responses.  Because, as we know, not everyone is able to learn how to regulate their brain.  And so this right here is a combination of the experimental group participants from three different clinical trials of amygdala neurofeedback.  And what you can see is that about 23% had a neurofeedback success value of zero or less.  Meaning that they did not learn to increase their amygdala response with training.  

We collected demographic and clinical characteristics, the same characteristics of all three samples and then combined them to look at what was related to neurofeedback success.  And I just want to draw your attention here to the symptom change to show that we have repeatedly shown very large decreases in symptoms with almost a 50% average decrease with the active intervention.  

So, we threw everything in together and said OKAY, what's correlated with neurofeedback success.  And we only found two variables that were correlated.  The first is sex.  And our females were more successful than males.  But the second was baseline amygdala activity.  And this was how active your amygdala was during positive memory recall before the neurofeedback training.  The less you bring it online prior to training the more successful you are at bringing it online during the training.  

And then we throw all of these variables into a regression model.  The best model that explains the most amount of variance only includes that baseline amygdala activity.  So just to touch on the sex differences for a moment.  We do in fact see that females are more successful than males when it comes to regulating their amygdala.  

But when we look at the baseline, what you can see is that the females have this negative amygdala response, and the males -- that the males do not have.  Suggesting that it is not necessarily the male/female difference, it's the baseline amygdala activity difference.  And, indeed, there's a very strong correlation between the baseline amygdala activity and neurofeedback success.  And this suggests that the intervention that's targeting a specific deficit requires individuals to have that deficit.  And this also suggests that we could do pre neurofeedback scans to see if the individual has the deficit and would benefit from the training.  

Finally, in my last minute here, I'm just going to talk about implementation challenges or how we can move into part of the clinical world.  And one topic that always comes up -- and I saw it earlier in the chat -- is a note on costs.  It is always declared that fMRI neurofeedback is prohibitively expensive.  

But I want to draw your attention to comparisons with other standard and novel interventions.  This table shows a course of treatment.  So two neurofeedback sessions is a course of neurofeedback treatment.  Sixteen sessions is a course of CBT.  And what you can see is that both with conventional interventions and novel interventions like TMS or Ketamine, a course of real-time fMRI neurofeedback is not any more expensive than these interventions.  

And so stating cost is a barrier is not necessarily reality of the situation.  Now there are other barriers to the clinical implementation.  And this is a conversation we are currently working with an implementation scientist, Dr. Kelsey Dickson at San Diego State University, where we're going to be talking to people who are in positions to help fMRI neurofeedback become a clinical reality and ask them what barriers and issues, they face that we can begin to address as researchers.  

And our planned focus group are MR Center Directors, MR Technologists, as well as insurance company representatives.

Just to sum up, we face a lot of challenges and opportunities with our regulatory challenges.  We have up until today been considered behavioral and not a device-based intervention.  Because, again, we are not actively doing anything to the brain with the device.  And the FDA does not regulate behavioral interventions or the practice of medicine.  We've had no adverse events reported with the fMRI neurofeedback.  And this suggests that FDA approval is not needed, especially pre-market approval.

When looking at dose determination, rather than considering clinical improvement, what we should be looking for is target asymptote.  And when we're looking at who is going to respond to neurofeedback, in our case our deficit targeting intervention is only effective in those who present with a deficit.  

Also suggests that men with depression might have different mechanisms underlying their depression that require different interventions.  
And finally, implementation is an ongoing conversation that we are excited to engage in.  And I just want to thank all of the members of my lab.  And all of the funding that I have received from NIMH and all of you for listening.  And now I will turn it back over to Alex.

ALEX TALKOVSKY:  Thank you, Kym, Karina, and Susan, for these great presentations.  I wish we could talk about these all day.  Unfortunately, we have got until 1:30.  

I want to start with a question that recently just popped up in the chat.  Where did it go.  Dr. Young makes a compelling case for unique aspects of fMRI-based neurofeedback to make it a difficult fit for FDA regulation, at least at this time.  Earlier speakers discussed the tradeoff between exploratory studies to help develop the technology in the field and really focused late-stage clinical interventions.  

How would additional regulatory requirements for all fMRI-based neurofeedback impact the exploratory type of research that is so critical to the continued advancement of the field?

KYMBERLY YOUNG:  I think that's a great point.  And that is one of the concerns that we have as neurofeedback researchers.  It's already difficult to set up and get neurofeedback research funded.  

If we start to have to deal with issues of regulation, especially premarket regulation, you are going to see a lot of younger investigators who won't have the resources to devote to that.  And again, with exploratory studies the goal is not to determine that this intervention is what we're -- is the final say is what we're going to go into the clinic with.  It's to help us explore.  And in putting those regulations on us I think would really stifle neurofeedback exploratory research.

ALEX TALKOVSKY:  And before our session continues to answer, I also want to invite Doe Kumsa to join our conversation.  She's going to be presenting in session three.  And I think this would also be a good place for her to jump in and welcome her in with us.

DOE KUMSA:  Okay.  Thank you.  Yes, we will -- I am from the FDA.  We will be giving our presentation later on so we will add more information about what you asked, Kym.  
But I think there is some -- there is some confusion about the regulatory pathway and the things that are available.  So we do have a definition for device.  So we have a regulation for -- a rule for device.  So it could be an instrument, machine, implant, software including any component part or accessory.  And if it is intended for use in the diagnosis of a disease or other conditions or in the cure, mitigation, treatment or prevention of disease, it could be a device.  So I will put the regulation in the chat so that people can also see that.

And I also want to highlight that something could be significant risk or nonsignificant risk, but that does not mean it will not need marketing authorization to be on trial.  So significant risk means you will have to come in and do an investigational device exemption.  But if it is not significant risk, it doesn't mean you don't have to come in for an IDE but that is a separate situation to coming in for a marketing authorization.  

And a significant risk determination at the beginning lies between the sponsor, investigator and the IRB.  And they determine if it is significant risk or not.  And sometimes the IRB could say go talk to the FDA.  So at the end of the chain in terms of like determining what is significant risk, it does end up with us.  But we leave it up to the IRB to determine if something is significant risk or not.

So I hope that is clear.  And okay -- and more about the regulatory pathways that are available.  Indications, if it is target -- you know, if it's treating a specific indication, it will likely have to come in.  So, I will go more into those, but I just wanted to give, you know, some highlights.

KYMBERLY YOUNG:  Thank you.  Again, it is just providing us with just some guidelines.  Because do we need -- do we need premarket approvals, do we need IDEs?  We don't know really right now, and we've been told for the past 10 years that we don't.

DOE KUMSA:  Right, right.  And we have mechanisms to actually give you feedback as well.  So there are mechanisms where you come in and ask us questions, do we belong.  You can even submit a study as a determination, you can ask us questions about, you know, what kind of study do I need, is my device even regulated?  There are mechanisms for that which I will go into, but yeah.

HOLLY LISANBY:  And just to give a preview, I do hope that everyone will stay for the presentations during session three after the break during which we'll go much more in depth into FDA regulations concerning research as well as FDA regulations concerning the approval of devices for the treatment or diagnosis of specific conditions as Doe was referring to.  

I do just want to clarify that for research if it is nonsignificant risk, then an investigational device exemption is not needed.  And everything that has been presented today was nonsignificant risk.  And so, this doesn't slow down research, but we are using this opportunity to clarify that these are actually devices.  And we hope that they will have clinical value and benefit and have a pathway to get into the clinic.  And oftentimes, the FDA plays an important role in that translational pipeline to get such promising devices into the clinic and make them available for patients.  

But I would like to focus on saving further questions about the regulatory status to session three so that we can also delve deeper into the really important science and results that were presented by the other presenters here.

ALEX TALKOVSKY:  If there are no objections to that, we do have plenty of questions to get into in the remaining time here for session two.  

One I would like to get into is actually based on something Luke Stoeckel put in the chat for us sort of relating to both I think what Susan and Karina talked about.  
I know, Karina, you had mentioned that the translation of protocols across different age groups.  So, Luke posted a question about the sort of content of your area of work, Susan.  He posted a link to a study combining mindfulness meditation and exercise in older adults from UDSD and Washington University related to dementia and cognitive decline.  But it seemed like they didn't really get positive results from that.  

So how could -- or perhaps you could tell us a little bit more about the process that either Karina may have gone into translating protocol to a different age group or that may sort of apply to translating some of these paradigms to work with older adults because that seems like that's a population, we didn't necessarily cover that thoroughly here today.

KARINA QUEVEDO:  So I would strongly advise -- and I hope that you notice that I only presented three studies.  I would strongly advise, you know, get your preliminary preclinical data clear, right.  What works for this population with this specific developmental history and this -- at this stage of development.

So, in terms of translating upwards in development, it would be really important to have preclinical target identification in older adults.  And I know that they are at risk, for example, for suicide, especially when they have physical and mental comorbidities that put them at high risk.  And then once that takes place then it is really important to then do a small feasibility study.  And then you can open up to exploration.  

Because let's be -- you know, let's have a degree of humility here.  We are essentially studying the most complex structure in the universe, as far as we know; there is no more complex structure than this one.  And there are well-known developmental changes that happen well into senectud, and they are all not bad.  For example, emotion regulation gets better as individuals get older.  Because those preponderant impulses that have to do with sexual fitness or even fight for territory and basically jostling with your conspecifics, they become less preponderant when we are already established members of our basically group.  

So I would say target critical developmental tasks for the age range in question.  And identify your target via empirical research.  So, I recently saw some really interesting work targeting the dorsal anterior cingulate cortex in adults, in older adults.  And it was very preliminary, and it wasn't very well written, but it was very -- it was very enticing to look it up.  So that would be my general recommendation to extend into older populations.

ALEX TALKOVSKY:  I will add something else from the chat that sort of relates to a common question we get around fMRI-based neurofeedback.  
It was specifically coming from your presentation, Karina, but I think this is something that really applies to this work broadly.  

Somebody asked how you could validate your work with EEG.  And I will open that up to all three of our presenters since that seems to be a common question that all of our neurofeedback researchers get.

KYMBERLY YOUNG:  So yeah, that's a great question.  And that is one direction that the field is moving forward with.  As Sue mentioned and Dr. Hendler will be talking about is if we can do concurrent EEG-fMRI can we then use machine learning to find these signatures.  

Because EEG can't get at the amygdala.  And so can we get at, you know, these signatures of amygdala activity and have people regulate those as a proxy.  Dr. Hendler has been very successful in that.  Our lab has attempted to create our own EED signature based on amygdala upregulation and have not been as successful as her group.  But it is something that we are all considering as how we could move this into EEG.

SUSAN WHITFIELD-GABRIELI:  And just a follow-up to what Kym was saying.  It is not surprising that you are not getting it immediately.  I mean, Talma has had fantastic success which is really -- really, you know, a tour de force.  But it's really complicated, and this goes back to some of the challenges we've been talking about.  

The simultaneous EEG-fMRI, you know, it's been around for many, many years.  And the orders, you know, the artifacts that you get in the EEG signal are order of magnitude larger than the signal that you are trying to look at.  And it is very, very difficult to do this.  
But we have a few projects where we're trying to capture the electrophysiological correlates of the default mode network.  So rather than a region, we're trying to get the correlate of a network, the default network or its anticorrelations, which is complicated.  

And then another thing to do is look at just passively collect the EEG and just to see what's changing when you're doing real-time neurofeedback.  In the case of schizophrenia, that's what Clemens is doing.  But the other thing that is really helpful is to possibly use EEG and/or fMRI to trigger brain states.  And if you are using EEG in that modality, you have the higher temporal resolution that you can capitalize on.  So, there's a lot of really exciting ways both to understand and elucidate mechanisms and also to scale the intervention with EEG and fMRI.

ALEX TALKOVSKY:  Someone else in the chat asked about the validation across culture.  So I was wondering if any of you had any data about cultural differences in neurofeedback response.

KARINA QUEVEDO:  So that's a really interesting topic.  I suspect that there isn't a lot about that.  

But I can tell you because of my area of study that the self is a construct.  The way in which we visualize ourselves is very much content dependent in the -- basically the signals that our culture gives us regards to how we perceive ourselves.  So we know, for example, that cultures that are more focused toward fitting the individual into community have a larger overlap in the areas of the brain that have to do with your own face and the areas of the brain that have to do with your conspecific.  So with people of your same community.  
Whereas, in more individualistic cultures there is much more separation between the self and the others.  So there is likely to be a cultural component not just with regards to something as, you know, as global as our self-representation but also with regards to how a culture shapes the way in which our brain develops.

So, there are cultures that highly value basically very tight control of emotion regulation because, you know, they are basically -- they don't have a lot of space.  And so, they -- getting along with your neighbors, being extremely polite, very much fitting into your community is highly valued and rewarded.

So, there are other cultures like mine which are very, very extemporaneous and very -- and it is not because -- we also have no space, but it has to do with our cultural history because we come from Spain, and we are loud and we're expressive and we talk with our hands.  And all of those things do have foundational roots in your brain.  

So, I think that that is a fantastic question to begin to look at.  And it speaks to having a higher representation of diverse communities and diverse experiences in neuroscience and neurofeedback specifically.

KYMBERLY YOUNG:  I also think it points to the importance of collaboration, international collaboration.  We have so many wonderful neurofeedback researchers that are in Europe and in Asia.  And so, combining our data to then look at whether or not there are cultural differences I think would be a very powerful direction.

KARINA QUEVEDO:  I really agree with that, Kym.  Having cross-cultural research in different countries that -- that get at specific processes that very culturally that's the way to go.

ALEX TALKOVSKY:  Well, this may be the last question we have time for before we break for 1:30.  So I will introduce it.  
I've gotten a couple of questions and comments about insurance reimbursement, particularly how the FDA may be involved with CPT codes.

So maybe starting with Kym and then going to the others, how would we get insurance to start reimbursing for fMRI neurofeedback?

KYMBERLY YOUNG:  So that is one of the specific aims of my confirmatory ROM is we are specifically engaging in focus groups with insurance company representatives to ask them what they need from us, including whether or not they need from us FDA approval.  

So that's -- I don't know the answer to that, and that's something I really hope that more people will chime in on, and especially in the next section.

DOE KUMSA:  I wanted -- Kym, do you mind repeating what you said just now?

KYMBERLY YOUNG:  Just that we are talking to insurance company representatives and have focus groups planned with them so we can ask them specifically what they need from us as researchers to be able to approve this and fund this.  And whether or not FDA approval is part of that.  That is one of our questions.

DOE KUMSA:  Oh, gotcha, okay.  It will likely depend on the indication.  So, for instance, if you're specifically targeting a psychiatric indication, let's say you're treating a major depressive disorder, to market that it would likely need an FDA approval, yeah.

KYMBERLY YOUNG:  So, this also quickly brings up the point of the R doc approach and transdiagnostic nature of a lot of our symptoms.  

And as Sue has pointed out, she is targeting the same region but for many different disorders.  And it is not necessarily disorder specific.  My research has been.  But a lot of neurofeedback research has not been disordering specific and applies to a wide range of mental health issues.

DOE KUMSA:  It could be.  So as long as it is trying to go into market, if it is targeting specific disease and treating it or diagnosing it, it will have to come in for clearance for grants.

ALEX TALKOVSKY:  Well, I am sure this will not be the last we hear in our conversation about fMRI and FDA and insurance and other third-party payors.  I encourage everyone to stick around for session three.  I'm sure this will come up again.  

In the meantime, that concludes session two.  Susan, Karina, Kym, thank you so much for your presentations.  Tremendously informative.  I really enjoyed getting to be a part of the discussion.  For now, it is 1:30.  So we have a 30-minute break on our agenda for lunch.  So we will resume at 2:00.  Thank you very much for an exciting first two sessions.
        
(Lunch break)