>> Eric Horvitz: Let's go ahead and get started. It's my pleasure to introduce Brian Caffo. He's a professor out at the Johns Hopkins University in the biostatistics department in the School of Public Health. In the group he runs, SMART Group, is really fascinating, so this is statistical methods and applications for research and technology. I think I got that right. >> Brian Caffo: Yep. >> Eric Horvitz: They do everything from wearables and biosignals through health data and the thing he's going to talk to us about today, quantitative neuroscience. Rather than get in the way, I'm going to hand it over to Brian. >> Brian Caffo: Great. Thank you. And thank you for inviting me out here. Really lacking forward to the visit. So let's see. What's the easiest way to do this? Okay. So I just wanted to mention a couple of things about the places that I'm associated with. So yeah, the group I started is called the SMART Group. It's a backronym. We actually wanted the word SMART and worked backwards until we got the acronym. But it's a really fun group. We have the domain, www.smart stats.org if you want to check us out. We do a lot of different things. It's kind of taken on a life of its own at this point. So I'm at Johns Hopkins in a very interesting place in the School of Public Health, the Bloomberg School of Public Health in the Department of Biostatistics, which is a very fun place to work. It's actually the we claim to be the first Department of Biostatistics ever, and it's it has some amount of legitimacy to the claim of being the first statistics department, though some of our colleagues from the UK dispute it when they come talk to us. And then so I also work a lot with the laboratory for neuro cognitive and imaging research, which is at the Kennedy Krieger Institute, which is a Children's Hospital affiliated with Hopkins. So about how much time do I have, about 45 minutes or something like that? >> Eric Horvitz: >> Brian Caffo: yeah. >> Eric Horvitz: You can go an hour, even an hour and a half. Hour and a half, oh, great, wow. All right, No reason to rush. >> Brian Caffo: No reason to rush, okay. Oh, great. So what I like to think about is brain connectivity. So when I think about brain connectivity, I like to think about the telephone system, you know, old school telephone system network, right. So and there's been kind of a useful sort of taxonomy of brain connectivity that, you know, some of the luminaries in the field have come up with, and it's useful for me. I think maybe the neuroscientists maybe don't like how crude it is, but I think for statisticians and mathematicians and computer scientists and et cetera, working in the field, it's sort of a very useful organizational thought process. And the taxonomy goes something like this. Basically, there's there would be three sort of forms of connectivity that people are often interested. The first is sort of the easiest to think about is sort of anatomy, right, and that's analogous to the wires, right? So in our brain that's white matter, right, that connects, you know, well, at the sort of macro scale would be white matter that connects different areas of the brain. You know, down to the smallest scale, it would be down to sort of synaptic connections between neurons, right. So anatomical connectivity is maybe the easiest to think of, and that's the sort of wiring. But again, it's that even though it's the conceptually simplest of the three ways to think about things, it still is complicated because the measure measuring it is quite complicated and, to be honest, the brain has a lot of connections. So the second way that people often talk about connectivity is so called functional connectivity. So the wiring enables the communication, but then we might want to talk about the actual communication that occurs, who tends to call whom, right. Which groups of people tend to talk to one another, and there doesn't have to be a specific, direct, physical connection between two entities for them to talk to one another. Of course, it gets routed, and that kind of connectivity would be sort of functional connectivity, how two entities how frequently two entities sort of tend to call one another. Then the final one, which is a little bit more sort of an aggressive style of thinking would be effective connectivity, which would be who tends to initiate the phone calls, right you know. Does one group tend to call another group more often than that group calls the former? So the effective connectivity gives us this idea of causality. Does one brain system tend to initiate communication with another brain system. So there's an idea of temporality and there's an idea of causality in effective connectivity. So let me just put these up here. So this is due to the person who wrote one of the biggest kind of software platforms for neuroimaging data analysis, and I think I kind of railed against it for a while, but now I quite like it. So at any rate, the three forms of connectivity, anatomical relating to neuro circuitry and direct neural connections. Functional. And functional is what we're going to talk about today. And then effective, relating to causality. And I'll give you some examples of the ways in which we try to study each of these things. They're all hard to study, right, because there's this fundamental almost sort of Heisenberg principle in studying the brain is that you're always going to have some sort of trade off in measurement, right. So if you can get you know, you can get actually really nice, very small scale study of anatomical connections, for example, in simple systems, like you can cut open mouses and mice and get down to synaptic connections and that sort of thing, but you can't really do that in humans. Or you can't do that in living humans. And then you can go to things like neuroimaging, but then you sort of zoom out quite a bit, or you can do so at any rate, in each of these different ideas, there's measurement processes, and each of those measurement processes has their own set of benefits and trade offs that you have to make to use them. And so we'll go through some examples of each, maybe discussing a little bit of their specific trade offs and benefits, yeah. >>: The thinking is that neuro connections change over time as individuals age or >> Brian Caffo: So certainly, you know, in development, of course, they develop you develop neuronal connections. And as you age, of course, you lose some. Then there's diseases which I'm particularly interested, you know, I work at, you know, Hopkins is sort of primarily not primarily, but is very well known for its hospital and our study of diseases. So what we're really what I often what motivates me is working with a lot of medical research and studying diseases. And some diseases, say, for example, multiple sclerosis exactly attacks connections. >>: And when you say attack connections, should I think about that as destroying these connections or destroying the functional aspects of them? The connections are still there, but they are left unused? >> Brian Caffo: I think you could actually so we'll talk about a little bit about both. But, you know, it depends on the specific resolution you're studying and way in which you're studying it in how you think about those think about those things. So I think we'll be sort of zooming out to a scale where it's difficult to think about that with that degree of specificity about the problem. So functional so functional connectivity, one thing that I find interesting about functional connectivity is it's sort of a biological definition where it's really defined in terms of the estimator. It's not defined in terms of an estimand, right. You know, the definition of functional connectivity involves correlations, empirical correlations. So it is a bit of a weird definition. That always bugged me. So one thing that always bugs me about this area is what if we could come up with a better definition of what constitutes a functional brain network, you know. Saying things are a brain network if there's correlations, that's sort of analogous, like, to saying there's a network if people tend to talk to one another. But there's this idea of maybe there's something a little bit more intrinsic than that. People tend to talk to one another because they're family or they're friends or something like that. So anyway, that's something that's always bugged me. But we're not going to address it because it's too hard. So let me talk a little bit about a way in which we study anatomical connectivity. So here's a way we looked at it with tract propagation. So there's a measurement technology called diffusion tensor imaging. And diffusion tensor imaging and, actually, I would say it's a new technology because it's actually kind of old now. Diffusion tensor imaging is a way in which you can image the directionality of water diffusion using the MR scanner. Most of the technology I'm going to be talking about today is using MRI. And from that, you can basically reconstruct sort of little arrows that represent the direction that the principal direction that water is flowing at that particular location in the brain. It's maybe a little bit more complicated than that, but let's assume you reduce it down to that problem. Then some very clever people said, well, what we can do is we can we have all these little arrows. We can try a little connect the dot exercise. And the idea is that the water diffuses more anisotropically in a long, kind of well defined white matter tracts. Sort of white matter tracts sort of serve as like a straw and the water tends to diffuse along the straw, like. So if you were to drop a droplet of ink in to a bowl of water, it would sort of diffuse outward isotropically. But if you were to drop it in a straw, it would diffuse anisotropically along the gradient of the straw. So the idea is that white matter, at some level, is acting like the straw, forcing diffusion in a particular direction. We can measure that diffusion at some level with gradients. We can then process it with algorithms via these tract propagation algorithms. And from that, we can get things like white matter tracts in the brain or estimates of white matter tracts. And not only estimates of white matter tracts, but estimates of how kind of how directional, how anisotropic the water diffusion is along those tracts. And that process is called tractography. So what we were interested in was sort of how to sort of quantify and analyze tractography. So take, for example, the cortical spine tract. What we were interested in is how do you analyze that as a data point? And so, you know, there's lots of instances where you're interested in things like the cortical spinal tract, but one clear one is in the study of multiple sclerosis. So in multiple sclerosis, you have white matter demyelination, DTI is sort of really good measure of white matter integrity, and so we would like to summarize this white matter integrity with some sort of tract estimate from the tracts, and we would like to build up ways in which to quantify it in analytic tools. So our approach has been kind of the following. We got really into how you you know, if you have some sort of non linear structure, collection of data points, if you have some sort of nonlinear structure, first, how do you fit a skeleton curve through it that is sort of a minimum orthogonal distance to the centroid of the point? And we've elected to do that with principal curves, which is an attempt to solve the ill defined problem of finding the curve that minimizes all of the orthogonal differences through a set of points, noisy points in three dimensional space. So, for example if you wanted to fit this spiral, you know, your principal curve would hopefully find the curve that went right down the middle. And at any rate, that's somewhat of a hard problem. So we think we have a reasonable algorithm for that now. And then after that, what we do is we chop along orthogonal planes to the curve and project the data points on to the orthogonal planes and then sort of fit sort of ellipses, Gaussian we fit Gaussians at each of those projection points and then we create this tube. This tube, then, we use to kind of summarize the fractionally isotropy, the measure of how directional the white matter is at that point along the curve, and then we use that in our statistical algorithms to say, for example, try to predict white matter I'm sorry, multiple sclerosis symptoms from, say, current white matter integrity measured by MRI. So who might try to or who might successfully have an alleviation of their symptoms based on their current imaging. So at any rate, that's the goal. That's one example of using MR technology to try and study or at least that's our example of trying to use MR technology to study anatomical connectivity. And this exercise is sort of fraught with a lot of complexity. We also, there's there's one instance where one clear, very important instance where it fails is there's probably the biggest single white matter structure in the brain not probably. Definitely the biggest white matter structure in the brain is the so called corpus collosum. This is a this is the white matter that connects the left and right hemispheres of your brain. And it looks a little like looks a little like like that. Okay. So it doesn't make a lot of sense to fit a curve through it, so now we've gotten into the business of how do you fit sort of surfaces in a way that we feel like we can come up with adequate quantification of the corpus collosum. Yes? >>: Brain's full of these white matter tracts. your analysis, isolate be one tract. How are you, in >> Brian Caffo: Okay. So that's interesting. So the tractography algorithms get at little fiber bundles that are estimated via this connect the dot algorithm. Those fiber bundles are quite noisy, and if you were to cut into someone's brain, it's not like you would definitely see a fiber. You might, but you might not. So we've kind of we do a much more kind of low brow version of tractography. So when I talk about the corpus collosum or the cortical spinal tract or things like that, these are tracts that every single person, if you were to cut open their brain, it would be there. And you would see it. And so instead of looking at fiber, individual fibers reconstructed by this tractography algorithm, we look at, I guess what I would call fiber bundles or we just call them tract. So we try to get we try to use the tractography algorithm and estimate the corpus collosum as a whole entity, which we know that everyone has a corpus collosum. It's the, probably the quantity that's best imaged by DTI. And we do that because maybe as a function of our own uncertainty in how high quality the individual fiber measurements are. So we tend to focus only on these large structures. I would say there's another way that people can get at none of these are let me get a corpus collosum. Yeah, here's a corpus collosum. There's another way you can get at these structures that doesn't involve any of this connect the dot processes. So what someone has done, a colleague of ours, is they have taken a white matter image and closely drawn out the major anatomical structures. Of course, the corpus collosum is there, the cortical spinal tract, you know, optic rays, all the big, major white matter tracts they've drawn out. Then they'll take the diffusion image and they'll morph it into this template space. So they'll shrink and stretch and squeeze and do it in a local manner so that they get it into this template space. And once you have it in the template space, you have the labels for the actual corpus collosum. Then you just pull it out. So you've lost the you've lost the anatomy, because you've morphed it into a common space. You've lost the anatomy. But the intensity, the photometric properties, you can retain. You can retain how directional. You can retain how directional things are at each location, but you've lost the anatomy. So the tractography tries to kind of a little bit retain the anatomy, and the shape and this template morphing procedure loses the anatomy at some level. You can still analyze the anatomy using it via a slightly different process. But nonetheless, we wanted to create a procedure that would honor any of processing streams for getting at these major tracts. But then, we've never done anything with individual fibers. But I would say there's a person at Wisconsin named Mu Chung who has done very analogous analysis procedures with fibers. So at any rate, it turns out so the technique we used to fit these midline structures are called principal curves and then there's an easy extension of them called principal surfaces. It's actually the problem of how do you fit kind of curves in three dimensional space is it's one of sort of like you have something and, let's say, if you have a 2D image and you want to fit a curve, you kind of drag it up into a third dimension, right? You parameterize it with a sort of a time you can imagine if this is something, the trajectory of something circling the drain, circling a drain, right. If you actually knew where it was at each time point, then fitting this would be pretty easy, right? So if you conceptually parameterize it as a function of time, right, and all you have to do is discover the time points, then you've figured out a way to fit the curve. Now, and that's, you know, the analogy breaks down when you move to more than one latent dimension. But that's the useful kind of analogy we use in to how we construct these algorithms to come up with these nonlinear structures. So at any rate, we think we have something pretty good for the corpus collosum, and our goal is to fit these kind of central structures in the corpus collosum, project the image intensity data down on to it and then maybe try to give these things now, as a summary of the corpus collosum, and in the hopes that it maintains a lot of the important tract integrity measures without discarding without discarding the key information. Now, of course, we've lost whatever information is in the orthogonal directions to this nonlinear surface we fit. But we think that hopefully we're coming up, because the just the look of this is so manifold like, we're hoping that we're retaining the important information. So now, at this point, our next step in working with this way of thinking about anatomical connectivity is we'd like to use these little pieces of paper we come up with that projects this 3D entity and use these pieces of paper as, you know, either as diagnostic tools or to put into secondary analysis engines and we're hoping that we can get rid of a lot of processing by this procedure. Yeah? >>: These are different time points? >> Brian Caffo: These are different subjects, yeah. different subjects. >>: These are Essentially, that is also a template, right? >> Brian Caffo: It is a template, yeah. It is a template. So what we're hoping is that matching the corner, say, for example, this guy, will avoid the need for normalization, for registration. That's our hope. But we have a while to prove that we can do that. Let me talk a little bit about effective connectivity. So when we first started studying effective connectivity, we had kind of a dumb algorithm, but I like it so I'll talk about it just for a minute. And I also think effective connectivity is a neat area. But I'm going to spent most of my time talking about functional connectivity. At any rate, so effective connectivity, remember, is this idea of directionality. So what I mostly study is fMRI. FMRI actually doesn't measure neurons firing, right. It measures blood oxygenation, which is the proxy for neurons firing. And so in that the function that maps neurons firing and the need for replenishment of energy via oxygen locally, that function varies spatially in a person's brain and varies between people to people. So it's actually very hard to get direction, causality and temporality in fMRI, the modality that I like to work in the most. What would be ideal is if you were to directly measure, you know, electrical fields, little electrical fields really in high, you know, high temporal resolution. And there is a way to do that, right? So electroencephalograms do that. They that's all an EEG, if you ever go put on EEG hat, right, the electrodes are just measuring little electrical fields. There's a lot of problems with EEG in terms of attenuation through the skull and these other things. And at any rate, so another way that you can study another way you can study it, but maybe in a very narrow population, a very narrow sense, is this so called electrocorticography. And electrocorticography, a person who is undergoing brain surgery for, you know, for something like severe epilepsy, they will put the electro plate on a person directly on their cortex. And so that, you know, that offers a lot of problems in terms of, you know, how generalizable the population you're studying is, how you know, what kind of spatial coverage you can get for the electrodes. You can only you know, you're not willing to, like, shove them in there, right? You've just got to put them where you can. And so on. But the nice thing is you get these incredibly, very focal, high temporally dense measurements. So we're talking about maybe a thousand or more measurements per second, with maybe a hundred or so electrodes. So you can study, you know, when one area when one area is firing, does it tend to precede another area that sort of thing. So we worked with a technique called Granger causality, which was developed by the economist Granger, who won a Nobel Prize for, among other things, his work on studying things like temporality and time series. And there's a lot more now I think since we started this, there's been a lot of development in the computer science and statistics and applied mathematics literature on how do you fit giant directed graphs. A lot of work. But when we did this, we weren't you know, we weren't so clued into that community so we said what we'll do is we'll fit these kind of parallelized Granger causal models, which basically looks at, you know, one node relative you know, looks at whether an arrow exists between two specific nodes, and you get kind of a rat's nest here. And then we said, well, let's just now kind of parse this rat's nest and try and find, for every triplet of nodes, where a connection, say, between A and C could reasonably be explained by mediating effect through B, or a connection between B and C could be caused by a confounding effect of a latent factor and then did all three variable Granger causal models and just pruned it by, you know, putting B in this relationship with A and C and seeing if the arrow goes away, as an example. And there's a lot of problems in how you search when you do that, and that sort of thing. And it would be better, certainly it would be preferable to do this as one giant thing. But it's you know, I will say to our credit it is hard to do one giant thing. You know, remember, each of these points is, you know, recording a thousand measurements per second. And you have lots of second that you've recorded. So at any rate, when you do that just that simple step in this case, you go from that graph to this graph where this is, and this was during an auditory task. This is sort of the part of the temporal lobe where you sort of hear things and this is heading in the direction of auditory processing areas. So that's an example of studying effective connectivity. I would say since then, we've kind of dramatically improved our ability to fit the kinds of models that don't require the sort of messy trade offs we made in doing this now. >>: [inaudible]. First few slides you talked about [indiscernible] tensor measurements to the tracts. And now you're talking about actually using oxygen? >> Brian Caffo: This is actual electricity. tiny little electrical fields, yep. It's measuring the >>: Okay. How are you getting precise data with electrical fields as to [indiscernible]? >> Brian Caffo: Well, okay. So there's a question of how do you know the locations. So in terms of the actual experiment, they take, you know, like a rotary saw and they cut someone's head open, right. They cut, you know, they cut they lift the top of their skull off and then they, you know, get through the dura, and then they put the electrode plate directly >> [indiscernible] of electrodes down? >> Brian Caffo: Yeah, and you know the cables kind of come out of the back of the head and they put the. >>: [indiscernible]. >> Brian Caffo: Yeah. >>: And you basically try to build this graph from that surface region? >> Brian Caffo: So yes. And I would say we did do one other thing that I'm not going to present, because it failed. Which is we were you know, I work in a School of Public Health. We want to compare across people. That's, you know, this is one subject. We want to create kind of populations. And to be honest, you know, you can't give that many people with this kind of data, right? We can get six subjects. But we'd still like to somehow average our six subjects or something like that. So what you can do is you can get an MR. You can register the template, the electrode locations via an MRI, and you can get kind of spatial, common spatial locales for each of the points across subjects. And then we spent a lot of time trying to, you know and then the registration wasn't perfect so we spent a lot of time trying to get this registration down just right. And then try to sort of smooth the points so that we could do something across subjects. But what it turns out is that these the measurements of these guys is so precise that even if you have another electrode in another subject that right nearby, it's just not comparable. It's just you can't combine them. And with an EEG, which is on top of the scalp, you know, the spatial extent of what an individual electrode is measuring is large enough that you can do a lot of smoothing. In these, you can't >> [inaudible]. >> Brian Caffo: So in this case, yeah, it was a stimulus. And it was pretty simple experiments in this case, you know. They present a, you know, a pure sign wave tone. >>: [inaudible]. >> Brian Caffo: Yeah. >>: I always find it interesting in macro stuff [indiscernible] brain this implicit but not talked about assumption that many neurons in the same region are doing the same thing. There's a [indiscernible] thing they're supposed to the alternative world, where cells are scattered all over the place. And they're all very different from each other. >> Brian Caffo: Yeah. >>: [indiscernible] get these big giant columns and they're all doing the same thing. >> Brian Caffo: Yeah. But that's a principle, right? That's the principle of what's it called? So that principle goes back to Bracha and Wernke and all these folks that put forward I mean, I guess it ultimately goes back to phrenology, right, where people posited that, you know, areas of the brain are specialized. >>: Those guys talk in terms of big areas. >> Brian Caffo: Yeah, big areas. >>: Say that the really small you have macroscopic size [indiscernible] doesn't make a difference where you put your pin if it's in the right area. That's called 12, right, which is and it doesn't have to be that way, right? Very, very structurally rich areas. Move one cell to the left and it's a whole different story. >> Brian Caffo: >>: Yeah. As they talked about. I think it's interesting. >> Brian Caffo: So we, you know, I think in the degree of focality in electro corticography gets the point where you actually can start to concern yourself a little bit with that problem. With fMRI, you know, with fMRI or EEG, you know, you're averaging at an extent where, you know, the kind of the so called specialization, the regional specialization at a macro scale is really all you've got. >>: [indiscernible] what is it X by Y, how many [indiscernible] are there? >> Brian Caffo: In these cases, so I think there were 60 some odd in this particular subject. >>: Talking about huge regions. >> Brian Caffo: >>: Yeah. Of nodes? >> Brian Caffo: Yeah, yeah. >>: That might be the right way to look at things, but people [indiscernible]. >> Brian Caffo: Yeah, you know, I think that one of the main problems is, you know, that you can kind of pick your poison. In this case, you get great focality, kind of a large spatial extent, but not whole brain. Great temporal resolution, you know, weird subject population that you're studying. So you've made several trade offs. If you want to study living humans, the trade offs are pretty extreme. But I would say even if you're doing animal studies and getting down to the cellular level, you still wind up making quite a few trade offs in >> What was the process [indiscernible] to get to the graph? >> Brian Caffo: So for developing this graph, we just used sort of ordinary Granger causality, which takes two time series and says, you know does if I look at node A, does the variation explained by it and its history, does inclusion of node B and its history explain enough of the variation that I can then draw an arrow in that direction? And Granger causality doesn't distinguish you can have an arrow going in both directions. So you can do that, but then you could also, of course, there's a multi variant version of that, and you can add a third variable. So this just does all collection of pairs and then we said, well, instead of trying to do the massive multi variant version, which is possible now, but at the time we hadn't really thought about how to do it, we said instead of doing that, why don't we just look in a smart way to see which edges we could prune. Of course, this gets rid of problems where there's a complicated fourth order interaction that we've just elected not to study. Yeah. >>: And that led to this? >> Brian Caffo: And that led to this, so at least it worked pretty well for this one subject, for this one task. Yeah. >>: Two questions. One of them is the subjects that you're pulling their skull off and doing this. >> Brian Caffo: but yeah. >>: I mean, I'm not personally doing that, Like there's something else wrong >> Brian Caffo: >>: Yeah. It is actual surgeons. Is something wrong in the first place? >> Brian Caffo: Yeah, so as far as I know, the only people who get these kinds of surgeries have fairly severe epilepsy. So maybe there might be some cancer applications. But in all the data that I have, it's all epilepsy surgery, where all the stimulation, drugs and everything have been tried and they're moving to surgery. >>: So the kind of stimulation that you would get this kind of data out, is that something like you asked that surgeon, could you piggyback this study on your surgery, or were they going to do this kind of thing anyway? >> Brian Caffo: Oh, okay. So there is a yeah, so that's a good question. There is a clinical application to the electrode placement in that the surgeons care a lot about not cutting into, say, for example, speech areas. And, you know, they would err on the side of not removing speech of, you know, they're very conservative about certain functions that they want to try. And it's, you know, you have to map those. And they do that with the electrodes which have a clinical use and then they also do that with probes. They actually probe you know, do some sort of electrode probing to try and get at, for example, language areas that they would not you know, that they would even if, you know, they were trying to balance the clinical utility versus the parent's function afterwards. So yeah, there is a clinical application. But then, you know, of course, all the patients supply informed consent and they're all, I think, from what they tell me, often quite happy to be involved in the research, you know. I think, you know, in between the pre surgical the surgery prep and actually going to surgery, I think there's some time, and they actually value being a part of the science. >>: The question I wanted to ask is about the actual timing of the connections that you can tell the Granger causality. Like if it's 300 milliseconds >> Brian Caffo: Even lower than that in our example. And I think that's a good point is how far back yeah, how this is all this is all, I guess, what I would call extremely temporally local directionality, not yeah, so >> [indiscernible] the actual neural length or the speed of the [indiscernible] or is that more like >> Brian Caffo: No it's longer than that. It is longer than that, but it's not, say, fMRI long where you're talking about seconds. It is, you know, milliseconds. But I think when we put our measurements are let's see. So it's a thousand measurements per second and we put, you know, a couple lag terms in there. So, you know, at best we go back a couple milliseconds. >>: Okay. So then those are those aren't going to be like actual neurons? Those are going to be like some >> Brian Caffo: No, no. And again, these are all measuring electrical fields. So it's still, you know, going back to Eric's point, it's fields of neurons, right. It's not yeah, yeah. Okay. So let's so anyway, let's get to my current favorite topic, which is talking about resting state functional connectivity. So one of the big kind of one of the big things that lots of people are working on is using an fMRI scanner in a particular way. So fMRI using a paradigm to get people in the scanner and have them tap, tap their finger. Yeah. Have them tap their finger, that and showing that the motor region lights up, that's pretty well established and that was a real loon for sort of so called brain mapping. Then people came up with the idea, well, you know, why don't we just study, you know, we put a person in an fMRI scanner, which measures this sort of proxy for neuronal activation and just calculate correlations between disparate regions. And I'll talk for a minute about or talk in a minute about why people might want to do that. But it explodes the problem a little bit, right. So you have maybe 50,000 voxels in an fMRI experiment collected over, say, 200 time points. And all possible correlations then is 50,000 choose two, that's a lot. That's a big number. And, you know, you only really, you know, the correlation, the matrix itself only has ranked 200, right, because it has the rank of the smaller dimension. And so you've got to do something to make that problem tractable. And so there aren't that many directions that people have figured out how to do it. One is to go down decompositions and certainly the most popular decomposition technique in this area is called Independent Components Analysis. But the idea is not all that different from most other kinds of decompositions. We like to do versions of ICA where we specify a full statistical model and we've done some computational tricks that we think are useful. The other way you can the other way you can study it is to pick specific locations or regions and average over those regions and reduce data in that way, in a kind of informed sense. So let me talk for a minute about ICA and what people have a tendency to do. So each one of these is a subject where they were in a resting resting state fMRI experiment. So over here is time, maybe every row is two seconds or so, okay. And along this axis is space. So every one of these is a voxel and we've just vectorized everything and dumped into it a matrix. So group ICA, it's a little bit more complicated than this, but conceptually all you do is you stack everyone up in this direction. And notice you have to if you had to pick between stacking them up in this direction or stacking them up in the other direction, you have to do it in this direction. And the reason is because they're just sitting in the scanner. The time for this person has nothing to do with the time for this person. See, it doesn't match up. You couldn't stick them hike that, because it doesn't match up. You might say, okay, but the space for this person doesn't match the space for this person. But what we do is we try some very what is a rather complicated process of morphing everyone's brain into the same template space and then we feel justified in matching them up. Now, that's an anatomical warping, and we're studying function, but whatever. That's the best we can do. So we stack everyone up this way, and then everyone gets their own the decomposition kind of breaks this down into smaller parts and every person gets their own little, I guess what I would call weight matrix and then across everyone they have these common spatial entities. So we like to call these common spatial entities networks. And then we think of these guys as how those networks are engaged for that specific person. So you have a person specific network measurement of network engagement and you have an across people measurement of the networks themselves. And that's the that's the so called group ICA. >>: [indiscernible] how large is the pool for this experiment? >> Brian Caffo: So let me show you. So here in this example, we did 150 subjects. But we've done it to, I think, 500 subjects. So what we've spent a lot of time doing is working on how you can do the full likelihood calculations without kind of memory creep and running it in parallel. So we spent a lot of time doing that. And we feel like we have it figured out now. The only problem is so then now the largest repository of the resting state data that we would apply this to, it has several thousand subjects. You might say, well, why didn't you just do it for several thousand subjects if you're claiming that you can? And what we found is that the variability between sites is huge, and that so all the big repositories have different sites that contribute to it. The variability between sites is so huge. And we found, as we were adding subjects, we were actually getting worse network estimation and so we've got to do something about that. We haven't exactly figured out how to handle that. >>: [indiscernible]. >> Brian Caffo: It's going to have to be something like that, yeah, absolutely. It's going to have to be something like that. But right now, so this is results for one big site where we dumped them all together. And, you know, so what we can see is we get really good kind of network estimation. This is >> What's the actual stimulus here? >> Brian Caffo: So this is no stimulus. This is just folks sitting in the scanner. So what's interesting about resting state is, you know, you can get, you know, the folks are resting. They're just sitting in the scanner. And by doing these decompositions, one of the components, one of the networks that will come out is the motor network. They're just moving. If their eyes are closed, you'll still get the visual network, right? If they're sleeping, you'll still get the motor network. If they're, you know, we can do you know, we can do this experiment on people in a coma, for example. You'll still get networks even though they're not functional with respect to that to not currently functional with respect to that what is commonly used for that specific brain network. So yeah, so we get vision networks when people's eyes are closed. We get motor networks clearly when they're not moving and so on. So that's why this technique has at least the promise of a lot of power. In that it's potentially a way for quantifying a lot of networks. And there is this idea that these kind of spontaneous, synchronous fluctuations, measured by resting state fMRI, you know, measure the network measurements might be very useful for quantifying disease and recovery and that sort of thing, especially in things like coma, where you can't have a person do a task in a coma, right? And if you want to know if you want to kind of guess who's going to come out of the coma earlier, right, it might it would be very useful, and the idea is that maybe some of these things will be useful biomarkers. And we'll see later on, there's some potential troubles. This network, in particular, this so called default mode network, is the most famous one and easily the most controversial one. So the original impetus for the default mode network is people looked at areas where in a very attention demanding task, they looked at the negative associations with the task. What went down when the people were doing this task that required them to attend very closely? And this network kept popping up. And then lo and behold when you study resting state, it's easily the most robust network that pops up. And so this is called the so called default mode network and people have postulated this is sort of a it's sort of a default state in the brain, and there's been, you know, hypothesis about its role in introspection and all these other things. At any rate, it's easily the most studied and simultaneous controversial entity coming out of resting state. But we get, we think, a pretty good map of the default mode network by stacking up, by our particular variant of ICA. And I should say we also I don't want to go into it too much, but we do our own version of ICA. So one thing that we are very interested in, so let me give you an example of the ways in which people might try to use resting state data to study disease. So one thing we're very interested in, I work a lot with a development a lab that studies developmental disorders. And one thing that's pretty common across a lot of developmental disorders is motor dysfunction. And so if you asked, for example, a kid with ADHD to hold out their hands and move this collection of fingers, right, you know, I think, you know, I can do it. I'm sure everyone in this room can do it where these fingers don't move to, right? But if you ask most kids to do that, you know, if you've got a little kid, ask them to do it. Watch. They'll move these fingers a little bit too. So and what happens is that you know, the differentiation, this lateralization of motor function tends to happen a little bit later for developmentally disabled kids than typically developing kids. And so we've focused a lot on that. So that inspires us to be interested in left left/right brain connections in the motor network. And so one way in which we've looked at that, and I'll talk about this now a lot, one thing that we've specifically looked at from a methodological point of view is exploiting that symmetry and quantifying that symmetry when using ICA. So just the idea that one interesting fact is that exploiting this symmetry, this is work of Juemin Yang, exploiting this symmetry seems to help just it just seems to help regardless if the symmetry exists. So if you were to take, for example, flags, some of which have some midline symmetry, the Canadian flag, the European Union flag, the Russian flag all have some midline symmetry, and the Chinese and U.S. flag do not, and simulate data and do our version of exploiting the symmetry versus the standard algorithm, you know, here what we find is there are they're among the symmetric flags, they do quite a bit better and much more so as you add noise. So what you see is in the top row in each case being the one that exploits the symmetry and the bottom row being the one that doesn't, you get a lot more kind of leakage between the components if you don't exploit the symmetry. And as a benefit of it, so let me show you it's really kind of simple what we're suggesting. So we take everyone's brain, we split it down the hemisphere, the mid sagittal plane, which isn't a plane, but whatever. And then we flip the, one of the hemispheres over. So now we do the same decomposition style technique. But now everyone contributes a left and a right hemisphere. So everyone contributes two bricks instead of one brick. But, you know, each part being a hemisphere. And you have to flip one of them so you've got to be careful about flipping. There is another problem that when you register brains, the brain that you register people to isn't exactly symmetric. So we were in the process of creating a symmetric brain when we found out that the Montreal Neurological Institute has released a mathematically symmetric brain. So we used that one. Who is no one's brain. So we register so the comparison space that we're doing it in is no one's brain. that. But we're at peace with And so then everyone gets their own left and right hemispheric wait matrices, and then they get hemispheric specific networks that represent a commonality across hemispheres. And so the result of this, the result of this is you can compare these guys and compare using decomposition, you can compare left and right hemisphere. So if you are interested in lateralization, which is what we're very interested in as motor network lateralization, we will compare the left and right weight matrices from ICA. >>: So ICA, as I understand, looks at temporal coherence, right? >> Brian Caffo: Yes. >>: So when you're separating the hemispheres, you're essentially assuming that both hemispheres are temporally [indiscernible]. >> Brian Caffo: So not necessarily. So it's very confusing, but what we've done is so let me rephrase your question. Your question is could this algorithm estimate a lateralized network. That's a rephrasement of your question, right? And it could. It's just that this weight matrix, if it was lateralized in a left hemisphere, this weight matrix would be zero. So take language network, for example, which is a lateralized network. But the point is you would never do this not that you would never do this. The point is to what extent do lateralized networks gum up the works for doing better estimation in non in symmetric networks? And what we found is that the symmetric networks do just as well, if not better, in the presence of lateralized networks. If you want to study the lateralized networks, then there's no point in doing what we're suggesting. We're just making it more complicated. That's why when this flag example included flags that were lateralized and flags that were symmetric and we find that you know so here I'm only plotting the symmetric ones because you only do this for the symmetric things that you're interested in estimating. But we find that, you know, the symmetric ones get estimated better even in the presence of lateralized networks. >>: So did this function in, let's say, the motor cortex, shows up as separate spatially coherent regions? Like different shape if you're doing this method >> Brian Caffo: >>: It's the same shape, yeah. So then you're losing your functional boundaries, right? >> Brian Caffo: So that is yes. So we have a different technique for dealing with that, but it's not this, yeah. So I see what you're if so it's not clear what happens if things are symmetric but, say, subsumed. One is entirely contained in the left hemisphere, yeah. So fair enough. But what we've found, one thing I would say, is we get these great I mean, this is only 20 subjects. These phenomenal network estimates out of this process if the network is symmetric. And it looks pretty, because I think our eyes visually like symmetric things. So want to just briefly touch on a couple other things. So decompositions were the kind of the big thing going for a while, and decomposition is this idea you get these nice blobs, right. But now if you go to human brain mapping graphs of the new blobs and it graphs of the new blobs, then sort of how you pick the nodes that you use to create these graphs, that's the house of cards that the whole enterprise is built upon. And just want to point out one way in which we create to answer your question, one way in which we create nodes. So we'll take, say, the motor network, and we'll do a clustering algorithm, and then we'll take these the clusters, estimated from the clustering algorithm, and use those to create the graphs. So in that case, there is no spatial constraint between the left and right hemisphere, for example. >>: This starts with atlas, right, which is very different from >> Brian Caffo: >> Oh, you would create a functional? >> Brian Caffo: >>: No, no, so we would create a functional Yeah. So it would be functional clustering. On fMRI data? >> Brian Caffo: On fMRI data. groups, not on subjects. >>: But functional clustering on I see. >> Brian Caffo: Yeah. So, in fact, we have a study that Mary Beth looked at the motor network and showed that the functional clustering for autistic kids and typically developing kids, the functional clusterings look about the same. But then as the typically developing kids get just a little bit older, they look identical to adults. Their clustering looks identical to the clustering of adults, whereas as the autistic kids get a little bit older, they still have functional clusterings that look similar to the typically developing kids. So this was confused you. This was ADHD. So I want to >> [indiscernible] actually make graphs from that data? >> Brian Caffo: >>: Yes. Where are they? Do you have examples of some? >> Brian Caffo: So let me talk about the different ways so unfortunately, there's like a million different ways to create graphs and there's a million different ways to create nodes and then there's a million different ways to create graphs. So one way you can create graphs is you can take these independent component analysis, style weights, and, you know, taking the outer product, and that, you know, each row or column of these corresponds to a spatial network and that creates a graph, right. That creates a covariance matrix then you can use to create a graph. Our preferred way to create graphs now is to take the time series that exists in some node definition and use Gaussian graphical models, which means to estimate a conditional independence graph based on the inverse covariance matrix. So but either way and so one point I would like to make is regardless of how you construct the graph, because of the incredible amount of uncertainty in this process and all the different choices you can make, a very important aspect of it is to become is to get reproducibility. So if you put a subject in the scanner and you make all these choices to create a graph and then you take them and scan them again an hour later, you would hope to get the same graph, right? So one thing we have done is come up with a measure of graph reproducibility. So for a single number reproducibility, a very famous measure of reproducibility is the ICC, the inter class correlation. And so we came up with with a basically a graph ICC. So that measures the reproducibility of graphs. Can't do very big graphs yet, but >> [indiscernible] calibrate across different sites. other problem you have one subject In the >> Brian Caffo: Yeah, so people do that. Sort of traveling test subject type studies, yeah. I am involved in a study where they're doing exactly that to try and get calibration. Calibration's a very hard problem, though. It's a very hard problem. Getting the same so this is calibration within the same scanner, within the same processing team. Everything is held fixed. It's just scanning the subject once versus the other. And even that, it's surprising, you know, it's a little hard to get reproducibility, you know, good reproducibility. You have to control those processes pretty tightly to get good reproducibility. Just because I'm, you know so I would say these are the kinds of things that we work toward, toward getting, and we do have some semi interesting results at this point where we're where we're finding that kind of the variability in the graph is a lot higher at the younger ages. So we get these less, less dense networks among kids, right. So but in visionary, they're dense and nice, so there appears to be a lot of commonality among kids in visual areas where, you know, that that develops very early. And then as we age, we get these much denser networks that are more common across subjects, across subjects. So at any rate, just in the interest of time, I wanted to talk about the last thing, which was just because it was fun. So the group of people that put out the data or a huge chunk of the resting state public resting state data that you can just download ran a competition, and it was called the ADHD 200 competition. So the ADHD 200 competition is a competition where they released let me see. So they a bunch of sites contributed. Here's all the sites that contributed, and they you have three subclasses of ADHD, control, combined and hyperactive impulsive. And they released this information as well as the scans on a large collection of subjects and then they withheld it on a bunch of subjects and then your goal was to predict withheld ones. And there's some just looking at the site demographics is fairly informative. One is they withheld one entire site, right. That was introducing Brown 100 percent was withheld. Another one was, you know, things like Pittsburgh that was 91 controls and no ADHD subjects and then nine withheld. You know, I'm going to pick control for those nine. >>: [indiscernible]. >> Brian Caffo: So these three are all different ADHD sub types. So ADHD, most of the developmental disorders are so called spectrum disorders, where they have some amount of commonality of symptoms, but there's a certain amount of distinction between there's a high variety I mean, take autism, for example. The distinction between, you know, high functioning autism and low functioning autism is enormous. And so ADHD is similar. The one, this one, the hyperactive impulsive one is there were so few subjects they just said forget about that. Don't predict anyone for that. So we have the imaging data and we had the some amount, a small amount of covariants. So we had gender, we had quality control, you know, so there's a lot of different ways in which you could the imaging could fail. You could have a lot of motion, that sort of thing. We had a nice, fa fairly nice age range. It was all kids, though. I think the largest age was 21 or something like that. Yeah. So IQ was one thing we were very concerned about was how different the sites were. So the exception of the neuro image site, IQ was about the same across sites. That was one of the ways in which we were kind of testing the demographic change, because the sites you know, the studies were all collected for other reasons so we were concerned, you know, some study was yeah, so we were concerned about the distinction between the populations. So the IQ was not that different by site. So at any rate, so then we started you know, we spent months, and this was the kind of competition where, you know, there were maybe 50 teams that entered. I think 20 people actually got around to submitting, 20 groups actually got around to submitting entries. And you really couldn't you wouldn't bother to get into this business unless you really kind of worked in the area. Because the processing was so annoying, right. I mean, there was a huge amount of processing. You had to know how to process a thousand fMRI or structural images. And so we started digging in, and one thing one thing we tried very hard to do was to use motion to predict disease type, because we figured the ADHD kids would move more in the scanner. So that has nothing to do, by the way, with anything about the brain, right? So but they were clever enough to think to try and thwart that. But we found this kind of decomposition technique called the CUR decomposition, and we just sort of blindly threw it at the data and found these weird results. And found that the CUR decomposition was kind of predictive of disease status. And what we think is the CUR detection was just kind of a little bit more of a a little bit better of a motion detector. And so we think we were getting some nice motion information. What was interesting is after this competition, there was these papers that came out that showed residual motion even among groups that were very attentive to the study of motion in fMRI. I mean, these are the top, top people. They really understand how motion can impact fMRI. And they're saying even in our studies, even in our biggest studies, motion is creeping in somehow, we think, right. So motion is a really hard problem. that And it's not unrealistic >> Would you be happy with discriminating hyperactivity if you [indiscernible] motion completely with it was just motor activity [indiscernible] hyperactive kids. >> Brian Caffo: That would so we would be fine with that. So in this case, we were just trying to win the competition. So in this case, we were just trying to win the competition. And so it gets weirder. It gets weirder. There's always these competitions are like this, right? So at any rate, here's why we took Mary Beth's atlas of the motor network atlas, and that was the one of kind of scientific thing, that was the one kind of let's see. That was the one kind of scientific thing that kind of came out was that this dorso medial, dorso lateral connectivity from Mary Beth's atlas seemed to have some kind of really weak predictability for ahead status. And we've since looked at it in autism. And it's not terribly predictive, right? Here's your control subjects. Here's your ADHD combined and here's your ADHD inattentive. I mean, you know, few look at the mean and the conference interval around the mean, it's not like this is an extremely predictive entity. But that was the one kind of sciencey thing we found. Everything else was gaming the system like this, like trying to find motion. We directly put in motion as well. >>: [indiscernible]. >> Brian Caffo: So this, to the best of our knowledge, is legitimately significant. To the best of our knowledge is legitimately significant. Though, of course, we did you know, we paid a lot of attention to overfitting. >>: And what do your behavioral people say about this result? >> Brian Caffo: They were pretty happy that this is what came out, to be honest. This does go back to so if you look at the kind of homuncular organization of the motor network, this is consistent with kind of symptomatology of ADHD. It's, you know, again, it is a just so story, right? We found this and then we searched for an explanation. But there is a reasonable just so story, scientifically motivated just so story that went goes along with it. that So ultimately, our final algorithm included the connectivity between these regions from this clustering algorithm that we used. It included this weird kind of motion detector we developed, and it included age, IQ, I think we put the quality control in there as a predictor and sight in there as a predictor. Sight being mostly a predictor for demographics, right. And at any rate, we won, but there's a story. But there's a story. So how did we win? So what we how we won and it was weird. The scoring system was very weird. So the scoring system gave you a full point if you got a control correct. And then it gave you a half a point if you declared someone ADHD and they were actually ADHD. And then a full extra half a point or they gave you an extra half a point, so to get a full point you have to get their sub type right, okay? Then it was just total number of points was what you went for. Okay. But there were clearly a lot more controls than there were ADHD in the site or in the study. So at that point, you got to be really stingy with declaring someone ADHD to do well in the competition. You have to be really stingy. And, in fact, I'm pretty confident we haven't done this full analysis. They've since released the withheld data. But back when we were spending a lot, a lot of time on this, I kind of calculated that I think you would have gotten at least third or fourth just by declaring everyone a control. And maybe higher. Maybe higher. Depends on how the withheld shake out. But I think I could calculate that you would have gotten, you know, top five or better by declaring everyone a control for sure. So but we did almost that. So we got, you know, super high accuracy on typically developing. And then very low accuracy when we declared someone ADHD. We missed a lot of ADHD kids, right. But then when we declared someone ADHD, we got their sub type right more often than everyone else. So what we did is we so that was our strategy. And that was the putatively winning strategy. However, what you'll notice if you look at point, there's one team with a higher point value. From Alberta. And what happened is they got disqualified because they didn't use any imaging data at all. >>: Scoring system? >> Brian Caffo: They used the scoring system, IQ, gender, age, sight and quality control. Just the database and no imaging. And what's interesting about this Alberta >> [indiscernible]. >> Brian Caffo: So I'm of two minds about this. One is that it's clever and I'm actually good friends with the Alberta guy. So the Alberta guy is a really interesting character, first of all. He's a computer science masters student at Alberta, working entirely on his own. That's one thing. So he's an interesting guy. And since then, we talk a lot, and we submitted a paper where he wrote an article about baseball predictions. So he's super interesting, smart guy. And he legitimately looked at the imaging data and said it doesn't it isn't very predictive. Now, all of us looked at the imaging data and thought it wasn't predictive, and all of us almost did exactly the same thing he did. Everyone who did well. And, in fact, we spent a lot of time seriously entertaining the possibility of putting in all controls as our answer. So from the competition potent of view, you know, they wanted people to sift through the imaging data, right? I mean, a person that sifts through the imaging data and uses the imaging for prediction has lots and lots of meaningless numbers that they're adding to their prediction equation that make it a lot harder. And, of course, there's a good chance you're going to wind up with worse prediction if you do that. If you add, you know, all the different ways in which you can combine the structural and functional imaging data, you know, the chance that you're going to you know, that that's going to make your prediction worse is actually pretty high. So from the competition organizer's point of view, it's fair, but, then, you know, what constitutes a legitimate use of the imaging data? I mean, we've practically called everyone a control. I mean, were we really using the imaging data? I mean, we did. But we practically called everyone a control. So at any rate, it was interesting and it was a lot of fun. All of us that worked on it kind of now talk a lot more. Everyone it got everyone to sift through a huge amount of data, you know. I'm good friends with the guy from Alberta now. So >> Is there any intuition that this is a like if you had three orders of magnitude more data, do you think the imaging would clearly dominate? Like in other words, do you think this is >> Brian Caffo: Three orders of magnitude more data with a tiny signal embedded in there? Or three orders of magnitude more data you're saying oh, I'm sorry, more subjects. >>: More subjects, more [indiscernible]. more. I don't know what, but >> Brian Caffo: I think even if you have more subjects, a lot more subjects, it would still, the signal in the fMRI data is pretty weak, and you wind up with a lot of numbers and you don't exactly know where to look either, right. You are a little bit flying blind. And asking algorithms to sift between all the possible ways in which you can take this 50,000 by 200 matrix per subject and and that's after a fair amount of processing too, I might add, and distill down into 200 predictions is a lot to ask. Yeah. So at any rate, that's the last thing I wanted to say. >>: So where is this going? >> Brian Caffo: >>: What's next in your In fMRI? Yeah. >> Brian Caffo: So I, you know, I think there's a community of us that are kind of dedicated to sort of keeping on sort of hammering away at resting state and to improve it as a methodology. So that and it has to get attacked on several fronts to get it to work. One is we have to improve the processing. The processing's very hard. And the community is working very hard on getting the processing right. There's some physics that might help too. There's biomedical physicists that are working very hard on trying to change the measurement process to account for some of the ways in which the processing's hard. Then I think there's still a lot of work to be done in terms of analysis and figuring out how, you know, what what kinds of problems this is going to be relevant for and what kinds of uses it's going to be relevant for. We're kind of highly committed. We're a little bit opinionated about the kind of specific kind of graph that we think people should be creating, and basically, what I would say is 100 percent of the computer scientists, statisticians, and applied mathematicians that work in the area all agree on a particular way in which these graphs should be created, but the field hasn't the field, the broader field hasn't agreed on it. We everyone thinks that it should be kind of inverse covariance matrices and Gaussian graphical models and that sort of thing. And the field is still thresholding raw marginal correlations. >>: I can see how the graphs are could be [indiscernible] how things work and are connected. But they also could be a source of rich features to make these discriminations that wouldn't be in the [indiscernible] data, potentially, as a high level representation. Has that been studied directly, showing that graph gives you more discrimination, the power to diagnose cohorts of patients than going to this broad lower level [indiscernible]? >> Brian Caffo: >> So people have The actual graphs they're generating and the properties. >> Brian Caffo: So yes, there is, actually, that is a big movement right now, actually, in that there are a group of people that are looking at kind of traditional graph matrix of, say, you know, they'll take a node and they'll see the the things about edge distributions coming out of the node, for example, and measures of small worldness as an example. And then show and then try and look at are these small worldness measures better discriminators of the between and so, yeah, actually, there's a big movement along that lines. I think there's a huge problem right now, and so I don't have any bearing on what's the correct solution to this because I think the graph problem is quite hard. But there seems to be what a lot of people are trying to do is to use graphs to makes statements about neurological organization, right. If you say that if you say that, you know, the graph represents, say, small world organization of the brain, that's a statement about neurological organization. And if you say it exists at multiple levels of hierarchy, that's a statement about neurological organization. But that is a tremendously hard thing to test, and the only way to get at it is that these indirect measures, like graph distributional properties of the graph, right. But to me, what you really need is something akin to a likely ratio where you're saying among the class of graphs that represent small world small worldness, here's the one that's best supported by the data. And among the class of graphs that would represent the alternative, here's the one that's best supported by the data. And here's, you know, factoring in the impact, the randomness in the process, here's how well differentiated these two hypothesis of network organization are. And to me, that problem is very hard but is one of the more interesting ones. And I think what would be nice about cracking that nut is then I think a lot of the how you would then use it to study disease would follow. But I think, you know, people have been working on, you know, how do you test hypothesis about graphs living in spaces, that's been going on for a long time, and it's apparently a pretty hard problem. So it's not my area. So it's apparently a pretty hard problem. So we tend to try to borrow a lot of the methods that have been built up in there. But it's apparently very hard. And so I think until then, I think actually doing things like just coming up with graph distributional properties and testing them, versus diseased or non diseased is the best thing going. >>: [indiscernible] neuropathology like severe FTD or Alzheimer's versus normal, you'd think that there would be big signal there's. >> Brian Caffo: So that has been done. That has been done. And Alzheimer' disease, in Alzheimer's disease [indiscernible] Alzheimer's disease, the big question in Alzheimer's disease is you want once the disease becomes very severe, the brains of Alzheimer's patients are very different from that of control patients. They've had a lot of atrophy. It's very apparent. >>: [indiscernible]. >> Brian Caffo: I think at all levels. At all levels. I mean, you have memory impairments and so all the action in Alzheimer's disease is at what level can we find precursors before the disease becomes severe. Maybe at the stage of mild cognitive impairment or even preferably before mild cognitive impairment. And so then you're looking at a much subtler difference, but it's where kind of 100 percent of the action is because if you can solve that problem, then there's a chance you can pretreat the disease. And there are potential medications to pretreat the disease. And I think Alzheimer disease is clearly the most visible of these kinds of problems. But a lot of the ways in which people want to use this technology have similar characteristics. So in the when I work with people who study traumatic brain injury and coma, what they really want to determine is who has who's going to have good function or bad functional outcomes based on their current imaging. >>: Detecting whether the concussion goes to [indiscernible]. >> Brian Caffo: Yeah, exactly, exactly. Now, in the developmental disorders, I think it's a little bit different. There, they're interested in what is the in refining the phenotype, right. So what is it but that is itself a hard problem for a different reason. And there, I think the fog is a lot, you know there's no gold standard to build things off of. >>: So my question is about the contest, the last part. Should I think about these sort of results as, you know, when you have kids with ADHD, they can [indiscernible] having them do an MRI is not very useful? Or is that conclusion too simplistic, because this seems to sort of suggest that. >>: Remember, this is a specific kind of MRI. This is a functional MRI. But to be honest, you know, people don't the doctors don't prescribe MRIs idly, so more often than not, my guess that is for diagnosing a developmental disability, they would not prescribe an MRI. The definitive diagnoses for all of these disabilities is still a physician's assessment. And I think if you were to ask most psychiatrists, for example, they would say we're comfortable that we can diagnose autism, but it's really, you know, the biological underpinnings of the disease, the real legitimate kind of characterization of the sub phenotypes is where imaging maybe would play a better role. It's not clear to me that the end goal and these the people who organized this competition are like, you know, top people in the area of developmental disorders. So they're all very aware of this. I think more than anything, what they wanted people to do was sift through a mountain of fMRI data and get at some nice ideas as to the biological underpinnings of what's a kind of really complicated disorder. But I don't think they ever had any illusions that people would use fMRI as a diagnosis machine for this particular disease. >> Eric Horvitz: Thank you very much.