Todd Constable “Networks in brain-behavior modeling”

Transcript

• 00:05Here in the sessions meeting.
• 00:08And I'm, I'm. I have the luxury
• 00:10because I'm going last because I
• 00:13saw some other people's slides.
• 00:15I've adjusted my talk. But.
• 00:18So I'm going to talk about
• 00:20networks and behavior modeling,
• 00:21and some of this is like, I don't know.
• 00:26The slides aren't advancing.
• 00:27Or are they no OK?
• 00:33After what?
• 00:37In the middle of what? There we go.
• 00:40So I I'm introducing some objectives.
• 00:43Sage and her talk said that
• 00:44we needed to have, you know,
• 00:46be clear on what our clinical objectives are.
• 00:49And so some of the objectives
• 00:51here in terms of bringing behavior
• 00:53modeling that I'm interested in
• 00:55are the clinical applications and
• 00:57understanding brain disorders.
• 00:58So what can this tell us about the brain?
• 01:00And when the brain doesn't work,
• 01:02the brain alone when we just look
• 01:04at the brain, doesn't really
• 01:06provide sufficient information for,
• 01:08you know, diagnostic categorization.
• 01:10Let's say mental illness and
• 01:12behavior alone also doesn't do that.
• 01:15But the hope is that these brain
• 01:17behavior models will someday do this,
• 01:19and they don't do that yet.
• 01:20But that's the goal.
• 01:23In Europe we have the IC D2 manual
• 01:27for classification of diseases.
• 01:29In the US it's the DSM 5.
• 01:32Neither of these work very well,
• 01:34and they've been around for 40
• 01:36years or 50 years.
• 01:38But they're not biologically
• 01:40based and they yield these very
• 01:42heterogeneous diagnosis with multiple
• 01:45combinations of symptoms that are
• 01:48insufficient really to provide to
• 01:50provide a good diagnosis and or to.
• 01:53Uh.
• 01:55To assign the correct treatment
• 01:57to an individual,
• 01:59so to make progress in this requires
• 02:01more transdiagnostic clinical data.
• 02:03We're collecting such data now.
• 02:06And there's not a lot of
• 02:09this data available just yet.
• 02:11But the ultimate measure of success,
• 02:12if we're going to categorize patients,
• 02:14the ultimate measure of success is
• 02:16do you assign the right patient to
• 02:17the right treatment or the right
• 02:19treatment to the right patient.
• 02:21And that's going to take some time to get to.
• 02:23So unlike maybe Simon and Sage,
• 02:27I haven't given up on the potential
• 02:29of this to someday be there.
• 02:31And I don't think we have the answer
• 02:33yet because we don't have all the data.
• 02:36So we'll discuss this in the
• 02:38context of the NIH research domain
• 02:40criteria when we get to that point.
• 02:42There's also brain objectives.
• 02:44I'd like to understand the brain
• 02:46a little bit better and I think
• 02:48you know functional connectivity
• 02:49and functional MRI in general are
• 02:51potential avenues at understanding the brain.
• 02:53So there's practical problems
• 02:55of no definitions.
• 02:56We saw some of that from Rudy's work and or
• 02:59edge definitions as we just saw from Rick,
• 03:02there's no consensus on which
• 03:04Atlas in my mind,
• 03:05there's no consensus.
• 03:06I think you know a lot of groups
• 03:08have their own consensus on
• 03:09what what the ideal Atlas is,
• 03:11but some of the results in
• 03:13terms of these modeling methods
• 03:15are actually Atlas dependent.
• 03:17And Wendy here is has a poster
• 03:20on consensus driven analysis.
• 03:23So this says that rather than choose a a
• 03:27random Atlas from your favorite group,
• 03:30we we actually choose thousands
• 03:34of random atlases.
• 03:37That are then that we choose
• 03:40thousands of C points and make
• 03:43atlases out of each of those,
• 03:45and then we find overlapping voxels
• 03:48that correspond to behavior.
• 03:51So what she's doing in this
• 03:52work is actually using
• 03:53the behavior. This is kind of.
• 03:55Randy, you touched on this a little bit,
• 03:56I'm going to come back to it at the end.
• 03:58But this turns around the problem and
• 04:00says we're going to be modeling behavior.
• 04:02What's what are the voxels that are searching
• 04:05with that behavior and not so much.
• 04:06The nodes and their associated with
• 04:08that behavior so kind of reverses.
• 04:10It uses the behavior to
• 04:12inform the choice of Atlas.
• 04:14So if you want to learn more about this,
• 04:15Wendy will be presenting
• 04:17this tonight over drinks.
• 04:20So there's a practical
• 04:21problem of no definition.
• 04:22I'm not going to talk about that.
• 04:25We can also figure out what
• 04:27data is most important.
• 04:28And by data, again the earlier talks touched
• 04:31on this data could mean brain state.
• 04:33So if we talk about temporal
• 04:36fluctuations or dynamics,
• 04:37brain state could be important.
• 04:39And this is some of our data.
• 04:42There's 16 behavioral measures across
• 04:44here collected under different activation
• 04:47conditions and the red box is highlight
• 04:51which which task runs best predicted.
• 04:55Which behaviors?
• 04:56And you can see that there's
• 04:57no clear winner there,
• 04:59although the movie condition does very well,
• 05:02as does the eyes,
• 05:04the emotional eyes task.
• 05:06I used to think when we started on this
• 05:09trajectory of green state manipulation
• 05:11in order to build better models.
• 05:14I used to think that this is
• 05:15like a cardiac stress test.
• 05:16So if you want to map attention,
• 05:19you give an attention task and that
• 05:22should help differentiate attention.
• 05:23That doesn't seem to be panning out.
• 05:26The attention task, for example,
• 05:28doesn't necessarily give you
• 05:29the best model for attention.
• 05:31The other is that this is just
• 05:33a synchronization,
• 05:34so you're removing a lot of
• 05:36temporal variance.
• 05:37It's just adds minimum wage.
• 05:40By synchronizing everybody,
• 05:41you're kind of removing a lot of
• 05:43variance and then you can get to the
• 05:46intrinsic connectivity that you want.
• 05:47But that's clearly not the case either,
• 05:49because if it was just synchronization,
• 05:51one of these tasks would win.
• 05:53And I was thinking that the movies
• 05:55are good at the synchronization,
• 05:56but they should win on all
• 05:58aspects there and they don't.
• 05:59So it's not simply synchronization either.
• 06:02And so this is a very large space
• 06:03and we're still exploring this
• 06:05and I'm not actually going to
• 06:06talk about that today either.
• 06:08Another type of data and and and
• 06:10I think Sage touched on this
• 06:13with you know we're expanding our
• 06:15datasets and actually a lot of these
• 06:17things can be compounds,
• 06:18but we're getting better and better at
• 06:20measuring things and we're learning
• 06:21more and more about how these things
• 06:24interact with the data that we have.
• 06:25And so I think as we move forward
• 06:28we'll start to understand these a
• 06:30little bit better and then maybe
• 06:32we will give up as Simon suggests.
• 06:34Yeah, but one of the,
• 06:36what I am going to talk about
• 06:38today is can we learn more from the
• 06:40connectome and I'll talk about that.
• 06:42And where I'm going to really talk about
• 06:44this is the sense of can we break up
• 06:46the connectome into networks and and
• 06:48Thomas who is left for a ski lesson.
• 06:52Let's see how it has the yellow
• 06:54networks that we all use and.
• 06:56I told him he could watch
• 06:58recording and says OK.
• 07:00Yeah, but there are other ways
• 07:02to break up the Alice and or the
• 07:05connectome into networks, and I'm
• 07:07going to talk about that a little bit.
• 07:09And then there's this feed forward.
• 07:11And I get this is where we
• 07:12go back in any expression.
• 07:13Again, there's this feed forward
• 07:15feedback sort of thing where,
• 07:17you know, we're using behavior to
• 07:18learn something about the brain,
• 07:20but we can also use the brain to
• 07:22learn something about behavior.
• 07:23And I'm going to talk about that.
• 07:24So those are my 3 objectives
• 07:27in the brief time that I have.
• 07:30So the clinical applications,
• 07:31how do we learn more about the
• 07:34connectome and what do we learn about
• 07:36behavior from this sort of modeling?
• 07:38And so basically we all kind
• 07:40of know the method or I think a
• 07:42lot of people will methods.
• 07:44So you start with a sample,
• 07:45let's say 100 subjects,
• 07:471000 subjects.
• 07:48You have some behavioral test scores,
• 07:50you have resting state or task data,
• 07:52you build the connectome and then maybe
• 07:54you specify some networks and you look
• 07:57at those components of those networks.
• 07:58But the whole idea, you know,
• 08:00we talked about predictive performance a lot.
• 08:02I mean the goal of this work, right,
• 08:03is not to predict fluid intelligence
• 08:05because you can do that with a test, right.
• 08:07The goal of this is to use.
• 08:09We have to verify that the circuits
• 08:11you're identifying that various
• 08:12function of that behavior are the
• 08:14circuits that you're interested in.
• 08:15And so that's the chief thing is to
• 08:18identify these specific circuits,
• 08:20develop normative models and the
• 08:22Donders group is you know publishing
• 08:25a bunch of stuff on this allow
• 08:28identification of particular subtypes.
• 08:30So we're particularly,
• 08:30you know when it comes to mental illness,
• 08:33we're really interested in what
• 08:34if we can do brain behavior,
• 08:35phenotyping and one of the things
• 08:38that comes up and it came up.
• 08:40Any questions is can you identify
• 08:42other strategies?
• 08:43So are people using different circuits,
• 08:45there's an assumption here that
• 08:46everybody is using the same circuits
• 08:48and there are ways to kind of get
• 08:50at that question I think and that's
• 08:52related to the normative model.
• 08:53And then count pounds as as Simon
• 08:56already expressed and we can not
• 08:58only are there countdowns that
• 09:00screw up this modeling,
• 09:02but we can identify through the brain
• 09:04there are countdowns in the behavioral
• 09:07studies that we're using as our gold.
• 09:10Standards build build these models,
• 09:12and in fact this is where the brain
• 09:14can tell us about the behavior
• 09:16or the behavioral models.
• 09:18And this can therefore provide
• 09:20insight into these systems,
• 09:22the behavioral measures you're tapping.
• 09:23So this is kind of the overarching frame.
• 09:25So we built models.
• 09:27I'm going to quickly go through how we
• 09:30do the our connectome predictive model.
• 09:32I think most people know this,
• 09:34but this is very general audience.
• 09:37So we have a connectivity
• 09:38matrix from each individual,
• 09:40we have a behavioral score.
• 09:41From each individual,
• 09:42we correlate those,
• 09:44we find the edges that vary positively,
• 09:46negatively.
• 09:47Uh, with that behavior and that yields
• 09:50the sort of network or set of nodes
• 09:53or edges that vary or that support
• 09:57performance and that behavioral measure.
• 10:00And we don't stop there
• 10:01because as Merrick showed up,
• 10:03showed these,
• 10:03this is a simple association study.
• 10:06It's not reproducible.
• 10:07But if we can build a
• 10:09predictive model out of this,
• 10:11then that tells us that it is predicted,
• 10:13that it is meaningful and predictive
• 10:15modeling tends to be much more.
• 10:17Reliable then if if we
• 10:19just stop at this point,
• 10:20so how do we build a predictive model,
• 10:22we take these edges,
• 10:24we calculate some network score,
• 10:27we build a model that fits the network
• 10:29score to the observed performance
• 10:31in a group of subjects and then we
• 10:34test that in an individual that we
• 10:37bring in and we can look at how well
• 10:39we do and observed versus predicted.
• 10:41And if we do really well then
• 10:43we're happy with the model and
• 10:44we can kind of cycle through.
• 10:46We first did this with leave one out now.
• 10:47We cabled it's best to do completely
• 10:50independent samples if you have the data.
• 10:52And you can do this with a positive
• 10:54edges or the negative edges and
• 10:56kind of doesn't matter whether
• 10:57you do both or one or the other.
• 10:59And then we can measure model performances,
• 11:01mean square error or the R-squared
• 11:05between the predicted and observed.
• 11:08So this yields, you know,
• 11:09these complex networks that
• 11:11we're all interested in that tell
• 11:14us something about the brain.
• 11:16Most standard task based MRI people
• 11:19hated these when these came out
• 11:22because they're crazily complicated.
• 11:24But the brain is kind of complicated.
• 11:25So I don't actually have a problem
• 11:28with our representation of the
• 11:29brain being complicated.
• 11:31And this is a small subset of the
• 11:33number of available edges, right?
• 11:34So this is a very small fraction
• 11:36of the connections.
• 11:37So the complexity likely reflects the
• 11:39true complexity of the brain at the scale.
• 11:42And to be honest, we're doing,
• 11:44you know, in our predictive models.
• 11:46Our meeting everybody,
• 11:48not just us.
• 11:50People that are doing this predictive
• 11:52modeling are getting some of the best
• 11:54pretty good models that are in the
• 11:55field and you know most of the other
• 11:58methods don't can't predict at this level.
• 12:00So what's what I find fascinating,
• 12:02really cool is that you know this
• 12:04is this is a 268 by 268 matrix.
• 12:06So that's what the actual
• 12:08connectome looks like.
• 12:09But it's really cool that
• 12:11there's information on tons of
• 12:12different things in here right.
• 12:14And we're just trying to what we're
• 12:16really doing is learning how to extract
• 12:18that information and figure out what
• 12:19this piece.
• 12:20Connections tell us about the brain,
• 12:22and so that's kind of fascinating.
• 12:26What we've been doing is we've
• 12:28been collecting data.
• 12:29We have a behavioral battery that we
• 12:31do outside the magnets about two hours.
• 12:33We're doing a transdiagnostic sample.
• 12:36We have a bunch of clinical scores as well,
• 12:40which I don't have on here,
• 12:43but I'm going to focus on these 16
• 12:46behavioral measures and we can extract,
• 12:48we can build models for each of
• 12:50these measures and this is the
• 12:52performance of those models.
• 12:53So these are the 16 behavioral
• 12:55measures and this is the.
• 12:56R-squared performance of these are
• 12:59our values starting on our squared the
• 13:02R values of these across the sample,
• 13:05and this is in a K fold analysis
• 13:09and talking about count fans.
• 13:12You know some of these things.
• 13:13If you regress out certain compounds,
• 13:15the model does better.
• 13:16Sometimes you regress out things
• 13:18in the model disappears,
• 13:19but it's important to check
• 13:21for various compounds.
• 13:22But the point here is that the connectivity
• 13:24matrix contains information on all of these.
• 13:26Trades.
• 13:27There's also,
• 13:27in terms of dynamics that are state
• 13:30based changes in the connectome as well,
• 13:33which you can identify and
• 13:35those are super interesting too.
• 13:37So, oh, there's the clinical measures,
• 13:39clinical tests we have as well.
• 13:42So we can build models for all
• 13:43of these things.
• 13:44And then in Group analysis,
• 13:45we can identify the circuits,
• 13:47we can develop normative models
• 13:48for what those circuits look like
• 13:49as a function of score and that
• 13:51may look something like this.
• 13:52So we have that edges that are
• 13:54involved and we have the performance
• 13:56of the model or the relationship
• 13:59between the behavioral score and the
• 14:01network score and we can start to
• 14:03look at who look for whom the model fails,
• 14:06for example and.
• 14:07Bringing individuals.
• 14:09So I think in addition to the group data,
• 14:11it's really interesting to look
• 14:13at individuals and we have this
• 14:15nature paper last summer where we
• 14:17looked at the individuals outside
• 14:18of these distributions and we show
• 14:21that in fact those individuals
• 14:22were failing the model,
• 14:24not because of something to
• 14:26do with those individuals,
• 14:27but because the behavioral test was
• 14:29biased against certain populations and
• 14:31the brain was actually telling us that.
• 14:33And I think that's actually really cool.
• 14:35So there is a feedback feedforward
• 14:37sort of loop.
• 14:38Between what the our models tell
• 14:40us about the brain and what they
• 14:42also tell us about behavior.
• 14:44And there hasn't really been a lot
• 14:47of feedback on behavioral measures.
• 14:48And I think that we have a an
• 14:51opportunity here to contribute to the
• 14:54knowledge about these behavioral tests.
• 14:56So the open questions are what does
• 14:58this tell us about behavioral measures
• 15:00besides identifying biases and the measures?
• 15:03I think Ruby showed the green and red dots,
• 15:08the green dots.
• 15:09They're from tests.
• 15:09And the red dots,
• 15:10you know the other way around,
• 15:11the red dots are from tests and
• 15:12the green dots were subjective,
• 15:14you know self report and things.
• 15:16There is a aspect of behavior
• 15:18that you know for example,
• 15:20we can't model anxiety.
• 15:21We've tried you know dozens of
• 15:22different measures of anxiety and
• 15:24we don't get any performance,
• 15:26we cannot seem to model anxiety in in
• 15:28different populations and things like that.
• 15:30And anxiety is kind of a subjective measure,
• 15:33right, where it's not a test score
• 15:35on a reaction time or performance
• 15:37accuracy and it turns out that.
• 15:39A lot of subjective measures are
• 15:41self report measures are difficult
• 15:42to model and what that means is they
• 15:45don't have a direct brain for that.
• 15:47So what you think about yourself
• 15:49doesn't necessarily isn't necessarily
• 15:51reflect reflected in your brain.
• 15:53Which is a problem,
• 15:55because psychiatry generally
• 15:56relies on self report.
• 15:58They don't rely on reaction times or,
• 16:00you know, accuracy on some little quiz.
• 16:04They rely on these sort of
• 16:07subjective measures of brain state,
• 16:09which it turns out we are poor at modeling,
• 16:12but maybe in a sort of feedforward
• 16:15feedback loop we can get better at the
• 16:18questions we ask and improve that.
• 16:20I mentioned that this doesn't
• 16:22have to be resting state.
• 16:23I already gave you the answer that,
• 16:25you know, we can actually do better.
• 16:27Ruby showed nice data,
• 16:29but kind of for predictive power
• 16:31because she's using resting state and
• 16:34I think that if that was combined
• 16:36with I know she went skiing too.
• 16:38So I can say this.
• 16:40That combined with task data might
• 16:43actually do a lot better actually.
• 16:45The other thing about Rudy's study is
• 16:48that she's using HP behavioral measures.
• 16:51They would like 58 behavioral measures,
• 16:52but a lot of those measures are at
• 16:54ceiling and then you need a good
• 16:56distribution of behavior in order to
• 16:58get good models and that's partly why
• 17:00her performance measures were kind of low.
• 17:03So the, again, the,
• 17:04the type of data that we're relying
• 17:06on to build our models is important,
• 17:08but we can do, you know,
• 17:10so I just want to throw in this thing.
• 17:11We can do all these different tasks.
• 17:14These tasks were designed to
• 17:16tap different systems.
• 17:17You know,
• 17:18the emotional system.
• 17:19Grad CPT is an intention task and
• 17:22back is working memory tasks.
• 17:24So these are designed to task,
• 17:26to tap different components of cognition
• 17:29and see how those components map on to
• 17:33the different behavioral tests that we have.
• 17:36And what was the result.
• 17:37I showed you the beginning.
• 17:38It's kind of all over the place.
• 17:39So we're still unpacking this.
• 17:42But the.
• 17:45The cool thing is that.
• 17:49We can take another step in this one.
• 17:51I'm going to talk about now this
• 17:53breaking the connectome model
• 17:54into components and measure the
• 17:57contribution of these components to
• 17:59the system and so this is trying
• 18:00to be responsive to NIH.
• 18:04Desires for this sort of getting away
• 18:06from the DSM until they get a more
• 18:10biologically based diagnostic system.
• 18:11We they introduced this rdoc
• 18:13and this is actually 14
• 18:15years ago or something they introduced this,
• 18:17but we're now just getting to the
• 18:19point where people have data on this
• 18:21and can actually start testing this.
• 18:22But the assumptions of our dog is that mental
• 18:25illness is the circuit disorder problem.
• 18:27Cognitive constructs should reflect
• 18:29some of those disordered circuits.
• 18:31And so if we can identify these
• 18:33disordered circuits, the circuit with.
• 18:34Commission maybe that is going to tell us
• 18:37something about what's wrong in the brains
• 18:40of people suffering from mental illness.
• 18:42And they the aspect here,
• 18:44the assumption here is that is this
• 18:46trans diagnostic assumption that
• 18:47we're all on the spectrum, right?
• 18:49So you know,
• 18:50we all have some level of paranoia and
• 18:53we all have any symptom that you have.
• 18:56You know,
• 18:57there's somebody along the spectrum there
• 18:59and the idea is that you can look at a
• 19:03transdiagnostic population and extract.
• 19:05Uh, components that go
• 19:07across everybody and that.
• 19:09I find that kind of an attractive theory.
• 19:11So they define the they got a bunch
• 19:13of psychologists together and they
• 19:16defined these six cognitive constructs,
• 19:18attention, perception, memory, language,
• 19:20working memory and cognitive control.
• 19:23And what we did was we went and
• 19:26built networks for each of these,
• 19:29and these networks then become
• 19:31components in our predictive modeling.
• 19:34And so we can.
• 19:36Specifically look then at what these
• 19:39components contribute to each task.
• 19:42So,
• 19:43so far we've been to,
• 19:44we've considered this data-driven
• 19:45approach where we take the whole
• 19:47connectome and we build our models.
• 19:49What we're going to do now is going
• 19:50to break up these connections.
• 19:51We're going to take components of them and
• 19:53rather than just getting a single slope here,
• 19:56we're going to build a model that
• 19:58looks at multiple components.
• 19:59And so here we can look then at the
• 20:01contribution of each of these or we
• 20:02could do this separately where we
• 20:04just take each of these components.
• 20:05Individually and build a model like that.
• 20:08And so we've done that and it's kind
• 20:11of cool that you can actually look at.
• 20:14So if you look at the memory,
• 20:15working memory construct,
• 20:16you can then look at, well,
• 20:19what's the,
• 20:20how,
• 20:20what's the contribution of that
• 20:22network to this behavior.
• 20:24And and across these 16 behaviors,
• 20:26we can rank them according to how
• 20:28much they rely on the working
• 20:30memory system for performance.
• 20:32And I don't think this has really
• 20:34been done in evaluating these
• 20:36behavioral measures previously.
• 20:38And so we can get this rank and
• 20:40there are basic assumptions around
• 20:41what these things are measuring.
• 20:43But this is now a way for us to test if
• 20:46in fact those basic assumptions are true.
• 20:48And if you really want to task that
• 20:50it's going to have working memory,
• 20:51you would look at the verbal fluency 2 test,
• 20:54I guess in this case.
• 20:55And we can repeat this for each
• 20:58of the constructs.
• 20:59And so I'm emphasizing again the
• 21:01speed forward feedforward thing is
• 21:03that we can identify the circuitry,
• 21:05the feedback, so in in a population,
• 21:07let's say feedback.
• 21:08Thing is, we could be a priority,
• 21:10define the system and then measure
• 21:13the components of that system.
• 21:16And so we've done this for the
• 21:19language that kind of control
• 21:21the working memory circuits and
• 21:23we've done it for the other three.
• 21:25The attention,
• 21:26the perception of declarative
• 21:27memory and the rankings on along
• 21:30the Y axis here change with
• 21:31each of these constructs and we
• 21:33can look at the distribution.
• 21:37And the reason I was telling
• 21:39Thomas he shouldn't go skiing is
• 21:41because I wanted to throw this in.
• 21:43But there's the seven canonical yo networks
• 21:46and you can define it this way as well.
• 21:50You can look at the contribution
• 21:52of each component of that network
• 21:54and get the same sort of results.
• 21:57So if we do this in terms
• 21:59of a normative framework,
• 22:01we have a distribution of scores,
• 22:03we get a model coefficients for each term.
• 22:06So we get distributions of those
• 22:07and we can get distributions that
• 22:09we have single components and
• 22:11distributions of the typical network
• 22:13associated with each behavioral score.
• 22:15And then we can bring in a new
• 22:17subject in contrast to their score.
• 22:19And if we have this these sorts
• 22:21of normative models,
• 22:22we can see potentially are they
• 22:24who are the same performance,
• 22:26they may be enhancing working.
• 22:28OK.
• 22:28They may be relying on working memory
• 22:30more to achieve that performance and
• 22:32that's something that we can start
• 22:34to break up with this approach.
• 22:35And so here's the summary of
• 22:37the whole system.
• 22:38So we can build the system,
• 22:39we have the circuits,
• 22:40we can divide,
• 22:41we can get innovative models
• 22:43for each behavior,
• 22:45we can see who's outside of that framework,
• 22:47and we have the predicted and observed.
• 22:51We can do this for all of
• 22:53the behavioral measures.
• 22:53So this is really a high dimensional
• 22:55space that we have 16 measures or more.
• 22:58Actually Rudy's talk there were 58 measures.
• 23:01So it's a high dimensional
• 23:02space that we can look at,
• 23:04but we can bring all this data
• 23:06together and look at these components.
• 23:08So we're currently increasing
• 23:10our sample size where we're
• 23:12finding these network definitions.
• 23:14So we defined like the working
• 23:17memory definition network we
• 23:19define using neurosystem.
• 23:21And there are other ways obviously of
• 23:24defining these networks and I think
• 23:26there's going to be some work around
• 23:28that how to segregate these networks,
• 23:31but getting.
• 23:32Getting back to Randy's question.
• 23:37This slide.
• 23:39It was ready.
• 23:41Getting back to Randy's question,
• 23:45there's a couple of ways you could do this.
• 23:47So you could start with these 16 networks.
• 23:49I mean they they're 6, they're not.
• 23:51Let's say that was sixteen.
• 23:53We could actually,
• 23:54we're starting to do this where we pull up,
• 23:57we look at the,
• 23:58you know,
• 23:58maybe there's six principal
• 23:59components in these networks that
• 24:01we can reverse engineer and we can
• 24:04define these networks in a different
• 24:06way and then see if those networks
• 24:08outperform the systems and there's
• 24:10yet another way that we could.
• 24:12Just this around which is the whole
• 24:14you know in psychology and ontology
• 24:16is the concept of working memory
• 24:19and attention and things like that.
• 24:22Those are ontologies that we use
• 24:23in psychology that define features.
• 24:25But you can do a data-driven sort
• 24:28of analysis where you define the
• 24:30way you define the components of
• 24:32cognition that span some behavioral
• 24:34space and then build models across
• 24:37that and see if those models
• 24:39do better than the
• 24:40the classical ontologies.
• 24:42Working memory and attention
• 24:43and stuff like that.
• 24:44So that's kind of where
• 24:45we're going with this.
• 24:46It may all be, you know, compounds,
• 24:48but I think this is potentially
• 24:51interesting and it it kind of uses.
• 24:54The data and keep forward feedback
• 24:56networks where the behavior
• 24:57drives some of the imaging data
• 24:59and the imaging data can drive
• 25:01the behavior and I think that's
• 25:03a super useful approach to to do.
• 25:06So. I'm going to wrap up,
• 25:09we're interested in this clinical objective.
• 25:11We want to learn more about the
• 25:13connectome and how it relates to
• 25:15behavior and but the connectome
• 25:16can also tell us about behavior
• 25:18and I think that's really powerful
• 25:20and really kind of exciting.
• 25:22So thank you.
• 25:30We'll start with Bandy first bandy.
• 25:34I'd like to read it all up. Publication.
• 25:47My question.
• 25:51Yes. Yeah. Subject.
• 25:58Yeah, on both sides of the intersection.
• 26:03Yeah. No, we're very interested in this.
• 26:05And you know that was a creep representation,
• 26:08OK, but we are doing that that sort of thing.
• 26:12Todd was next and then I'll go to Julia and.
• 26:17It was bizarre to me the way you
• 26:20came up with the networks for
• 26:22those cognitive funds drugs.
• 26:24Do you think neuroscientists bizarre?
• 26:27Just because it's, you know,
• 26:30looking at you could walk,
• 26:31we do analyze the last base
• 26:34and all of those tasks that
• 26:37are mentioned there already,
• 26:38you can actually analyze those tasks
• 26:40and they're summarized in neuroscience.
• 26:45No. Yeah. Yeah. Yes. So that's true.
• 26:47Networks that you know,
• 26:48turns out that I know that this is not in
• 26:52any way except for the popular opinion.
• 26:54But it turns out that there is a
• 26:56set of networks that I said many
• 26:58times here before that spans pretty
• 26:59much all past that are available.
• 27:01So if you take,
• 27:02you know all the working memory,
• 27:04massive publicly available
• 27:05or shared between people,
• 27:07cognitive control tasks, attention pass,
• 27:10whatever other types of constructs,
• 27:12you're fine.
• 27:15Across those paths and then a,
• 27:18you have the networks and B,
• 27:20that cognitive function that may not
• 27:22match the ones that have been decreased
• 27:25by cognitive psychology or clinical
• 27:27nurse oncology to us might start to
• 27:29emerge by observing the way those networks.
• 27:31So yeah,
• 27:32I agree fully and and in my defense,
• 27:35let me say that we're super excited
• 27:37about the thinking about bringing
• 27:38this way and I think there's a lot
• 27:40of details still to be worked out on
• 27:42how we define these networks and.
• 27:44You know,
• 27:44these six cognitive constructs
• 27:46are the ones the NIH came up with.
• 27:48You know, that may not also be the best,
• 27:50but the whole data-driven approach
• 27:52to it I think may actually be better.
• 27:54But Julie was next.
• 27:56Now, but I was just wondering,
• 27:59you said that you know these big
• 28:02contracts with disabilities and
• 28:03I could argue that pension is.
• 28:07Everywhere.
• 28:10Yeah.
• 28:13Yeah. Well, that's why you do the
• 28:17data-driven approach and maybe we
• 28:19can discover better ways of dividing
• 28:21that data rather than relying on
• 28:24these kind of old formalisms, right,
• 28:26that that if we can do better in
• 28:28terms of predictive models and then
• 28:30feed forward feedback then then I
• 28:32would say that's a good argument for
• 28:34using a different circle anthology.
• 28:38Sorry, you were next.
• 28:42OK, so.
• 28:45So I I guess kind of related to when
• 28:49this poster if it would take too long.
• 28:53Like you can define the passports
• 28:56with behavior. It's like,
• 28:58well I guess it's kind of running to do that,
• 29:00but like if you define the networks with
• 29:03the behavior and let's say that all,
• 29:05I predict my behavior better even
• 29:07though it is like cross medication,
• 29:09wouldn't that be kind of silly? I see.
• 29:12So the question was if you're at the
• 29:14beginning I I mentioned helping the
• 29:17behavior to define a parcellation so
• 29:19we don't have to rely on a particular
• 29:21Atlas and so the question is,
• 29:23is that circular?
• 29:24I don't think it's that circular if
• 29:27they're an independent populations,
• 29:29but we are and I think actually what
• 29:32we're getting at there is that,
• 29:34so I I'm a I know this is maybe heresy
• 29:37and some of the some of the items here,
• 29:40But I'm not as strong.
• 29:42Believer in uh,
• 29:43an aerial area that you know doesn't
• 29:45function and has, you know,
• 29:47clearly defined by the site of architecture.
• 29:50I, you know,
• 29:51there's tons of evidence that these
• 29:54functional regions engineering
• 29:55spoke about this a little bit,
• 29:58that these functional regions,
• 29:59you know,
• 30:00changing more and stuff depending
• 30:01on brain state and their tasks.
• 30:03And one of the cool things about
• 30:06what Wendy's doing is that when
• 30:08you when you define these nodes in
• 30:10this kind of Atlas free framework.
• 30:12Um,
• 30:12you'll get overlapping nodes and and Thomas.
• 30:15You'll have this in an ontology
• 30:17paperback and cerebral cortex where
• 30:20he showed overlapping brain regions
• 30:23responsible for different tasks or
• 30:25essentially different behaviors.
• 30:27As to the circularity,
• 30:28I don't know whether people think
• 30:30if you define this in a in a
• 30:32separate group do is that circular
• 30:34is a completely circular?
• 30:35I think the the brain is the
• 30:38behavior is helping us zero in on
• 30:42the voxels that are consistent.
• 30:45And and then we choose those
• 30:47voxels and I think that's,
• 30:49I think that's legit.
• 30:52Yes, David. I would have a question.
• 30:58Basically we use.
• 31:0330.
• 31:06Stop so.
• 31:09What is the difference?
• 31:12These things. Well, again, it's,
• 31:15you know, these things are.
• 31:16So yeah which you know how
• 31:18does what's the physiological
• 31:20manifestation of depression?
• 31:22I mean I think the idea is the
• 31:24assumption is that that there's
• 31:25disordered brain circuits and they
• 31:27may be reflected in somebody's
• 31:28behavioral measures that we've seen.
• 31:30So it's not necessarily saying
• 31:33here's the depression circuit,
• 31:34but here's a disrupted circuit and
• 31:36maybe that's a target for therapy or
• 31:38something like that or maybe that's
• 31:40a a metric for measuring response
• 31:41to therapy and biological basis for.
• 31:45That makes sense.
• 31:48Ohh, there's a I OK, I gotta read.
• 31:51There's an online question,
• 31:52then I did somebody else have a
• 31:54question while I look that up.
• 31:56Ring of going back to the six domains
• 31:58if you saw that it wasn't a paper,
• 32:00but it was like a working group thing
• 32:04from 2014, about 6 different domains.
• 32:09Yes. How much reliability has been how,
• 32:12what is the state of reliability and
• 32:14validity of these? Oh, yeah. Yeah.
• 32:15So summarize all of them. Yeah.
• 32:18Well, we're actually looking into that.
• 32:19I have some stuff to bring on that right now.
• 32:22You know the we're using the
• 32:23behavioral measures is like a gold
• 32:25standard and we're building our
• 32:26models on these behavioral measures.
• 32:28But others have already published
• 32:30papers showing that really reliability
• 32:32of many of these measures is poor.
• 32:34Again, that's a fee for feedback thing.
• 32:36I mean if you put garbage in,
• 32:37you're not going to get model and
• 32:39I think some of the variations.
• 32:41You see and the ability to model things
• 32:43might be directly related to that.
• 32:44So yeah, we can't, you know,
• 32:47we don't have any other,
• 32:48nothing else to do, right.
• 32:50I mean we have to use what we can,
• 32:52but I think we need to be aware of
• 32:54that reliability of those things that.
• 32:56Ohh, there's online question here.
• 32:58Can you guys read that?
• 32:59But doesn't show up for you, OK?
• 33:03I am Brandon from Beck Todd's lab at UBC.
• 33:07What might be the benefits,
• 33:08disadvantages of evaluating the
• 33:09contributions of networks and normative
• 33:11distributions during different cognitive
• 33:13tasks compared to looking at normative
• 33:15versus normative representations,
• 33:16emotionally relevant stimuli and specific
• 33:18regions of interest during these tests?
• 33:20Well, looking at specific
• 33:21regions of interest during,
• 33:23you know,
• 33:24tasks doesn't seem to be
• 33:26very predictive of stuff.
• 33:27I mean that's the bottom line.
• 33:29You kind of have to these,
• 33:32as I said at the beginning.
• 33:34These kind of connection based models
• 33:36tend to be much more predictive
• 33:38of performance or or traits than
• 33:41F MRI based activation as I think
• 33:43that's kind of what the question is.
• 33:49Any other comments? One more so.
• 33:53People confuse on the brain.
• 33:57Is that like there's obviously
• 33:59like variation from college trial,
• 34:00like investigators reaction time like?
• 34:03You know, sometimes it's easier, sure.
• 34:07And then there's obviously some
• 34:09sort of average dynamics, you know,
• 34:11and like how could these people like
• 34:13because the input change the connectivity,
• 34:15so like yes, yes and no.
• 34:18So when you're speaking
• 34:19your face is changing,
• 34:21but I can still recognize you.
• 34:23And I think, you know when we're doing stuff,
• 34:26our connectome is there.
• 34:27It changes like a little bit.
• 34:30I think in our original fingerprinting
• 34:31paper we showed that we could still
• 34:33identify people by their connectome
• 34:35under different task conditions.
• 34:37See you tweak the connect them a little bit,
• 34:39but you don't change it completely
• 34:41and you still look the same.
• 34:42And so that's the whole state
• 34:44versus trade thing there.
• 34:46You can do state manipulation and change
• 34:48various edges and in the connectome,
• 34:50but you still mostly look the same
• 34:53and hopefully you're getting trades
• 34:54out and that's we're advocates for
• 34:56like 20-40 minutes of data so that you
• 34:59actually average out I think a lot of
• 35:02those moments and moment fluctuations,
• 35:03but I think Joanna is going
• 35:06to talk about dynamics.
• 35:07And there's tons of stuff going
• 35:09on at the moment to moment level
• 35:11and you can look for that and you
• 35:13can get things out of that.
• 35:15That's not to say that there's
• 35:17not a grand average that we're
• 35:18dealing with kind of thing so but
• 35:20does that answer your question?
• 35:22I'm just confused how you do.
• 35:25Well, they're both there.
• 35:26We're ignoring the the moment to
• 35:27moment fluctuations and we're looking
• 35:29at the ensemble averages of activity.
• 35:31So if you wanted to look at moment
• 35:33to moment fluctuations,
• 35:35you could some of this still works.
• 35:38It gets noisier,
• 35:39you know,
• 35:39the less data you have,
• 35:40the more noisy it gets.
• 35:43But you know,
• 35:44these things still hold for the most part.
• 35:48OK.
• 35:48So I'd like to thank all three
• 35:50speakers for this session.
• 35:52Thank you.
• 35:57And we have.