Dynamic Causal Modelling

The video is a part of the project British Scientists produced in collaboration between Serious Science and the British Council.
In the past 30 years, possibly 20 years, there’s been a revolution in cognitive neuroscience, in systems neuroscience. A revolution that’s been accelerated by the capability to look at the brain in action, to image the brain using either metabolic or haemodynamic tools like functional magnetic resonance imaging. Or indeed using electro-magnetic responses, as measured noninvasively by EEG and MEG.

The picture that is emerging of how the brain works has two aspects: On the one side, it is clear that different parts of the brain are specialized for processing particular aspects of our sensorium. for doing particular cognitive operations like, say, memory, or attention, or processing of emotions. Nearly every job that your brain does, that you can conceive of probably has a dedicated brain system or set of areas or regions that talk to each other.

So that’s the principle of functional specialization, and when that specialization becomes segregated in the little cortical area at the size of my thumb nail, for example, visual motion processing roughly in this part of the brain here – that segregation is known as functional segregation. So segregation of a functional specialization for doing a particular thing of the many things that the brain does.

We have for the first 10 years spent a lot of time with careful experimental design, rigorous data analysis trying to assign functional specialization to different brain areas to build a map of what different parts of the brain are responsible for: this is known as cartography. It’s also being criticized as neophrenology. So, phrenology was a procedure many centuries ago, whereby it was thought that literally by palpating the skull and feeling for little lumps, and knobs, and bumps, one can diagnose and infer the sort of person and the competencies and the functional processes that the brain was engaged in simply by palpating. And many people think that functional neuroimaging, functional imaging, suffers from the same philosophical shortcoming – just by looking at bumps in metabolic activity or regional hotspots, often referred to as blobs, you’re just recapitulating the same sort of conceptual error that the phrenologists of the 19th century were committing.

That criticism, I think, is greatly mitigated by the second principle; and the second principle is functional integration. So on the one hand we have segregated brain regions, sometimes referred to as nodes in a graph or distributed network, and then we have to think about how those nodes or those regions are coupled or connected, how they talk to each other, how they are integrated, how that processing is distributed over nodes or regions in a coordinated and organized and functional way. And that integration of the distributed responses, the distributed processing is a functional integration. So now you’re in the game of having established all your favorite functionally specialized areas, now you want to know how they are integrated, how they talk to each other, how they are coupled.

And over the past decade or so, it’s become clear there are two ways in which one can characterize this coupling functionally: You can either look at correlations in activity of two brain regions, so, say, we imagine we have two parts of the brain, this one dealing with visual information from one side of the visual field and this one dealing with the information on the other side of the visual field. And if they talk to each other and share information, we might expect that during our brain imaging experiments or during our EEG experiments; as the activity in one these areas goes up, so will the activity in the other area. So there’s now a correlation, or a statistical dependency, between the measured responses in each of these functionally segregated or specialized brain areas, that’s known as functional connectivity.

It’s easy to measure, it’s operationally defined and what it tells you is that somehow the processing over time of these two different brain areas are coupled in the sense that they’re likely to be doing similar things, so they are both engaged in the same distributed pattern of activity. What it doesn’t tell you is how the activity here influences the activity here and vise versa, so just knowing two things are correlated or functionally connected, doesn’t tell you about the directed influences that one brain region exerts over another, and that’s called Effective Connectivity. So Functional Connectivity: correlations, dependencies and operational definition;
Effective Connectivity: directed causal connection mediated by long axonal neuronal processes, so that you’re driving here in a way that depends upon the activity here.

So that’s where Dynamic Causal Modeling comes in. So Dynamic Causal Modeling speaks to the fact that in order to make sense of brain imaging data, for example, or EEG data, or MEG data, you have to have a model of how this part of the system influences this part of the system and vise versa. You have to do that in order to interpret the data, and put very simply, once you’ve established a model, a model of coupling, then you can ask what coupling parameters, what model parameters of that causal model, a model of the causal influences of this part of the system, or node, on this part of the system best account for the observed data, so this is, in a sense a model fitting exercise, where you’ve got this distributed pattern of activity throughout the brain and you want to fit this particular model to explain the data, and this particular model is a;; about the dynamics of fluctuations of brain regions that are causing activity in other brain regions – hence Dynamic Causal Modeling.

Technically, it’s just a state-space model, it’s a sort of model which people who engage in time-series analysis would normally call upon to understand how, say, the weather unfolds, so technically speaking these are exactly the same sorts of models you’d be using for weather forecasting. Or in economics, the fluctuations of the markets – how one event over here causes changes in an event over here and how that unfolds over time as many distributed events all cause each other in a reciprocal and recurrent way.

So that is, in essence, Dynamic Causal Modeling. It’s the technology that has been brought to bear on deep questions about functional integration, about functional architectures. So we’ve moved beyond the functional anatomy of functional specialization and segregation, and now we’re talking about networks, distributive processing and architectures that are equipped not just with where stuff is happening, but how stuff here is distributed and influences and causes stuff over here. And then there are all sorts of interesting questions about the brain network, about what has recently been called the connectome. How does that architecture inform our understanding of how the brain works.

So one simple example here would be the notion of a brain hierarchy – the idea that there are certain nodes or regions in the brain that are very close to sensory information. Say, the back of the brain in receipt of visual information, primary auditory cortex along the side of the brain, directly in receipt of hearing, or auditory, information. And these parts of the brain would be at a hierarchically lower level, and yet if we move into the hierarchy, deeper into the brain, say, towards the front of the brain for example, then we have this notion that there are parts of the brain engaged in higher level, more abstract representations, modeling of the causes of the sensory inputs. Because if you have a model of the brain as a hierarchy of interconnected regions, with some levels of the hierarchy being subordinate or lower to higher levels of the hierarchy, then that presupposes that there’s a difference between bottom-up connections and top-down connections.

So that distinction is absolutely fundamental to understand functional brain architectures. And implicit in that distinction between bottom-up, from the sensorium, from the sensory cortical areas, through to higher cortical areas, say, in the prefrontal cortex at the front of the brain, then you are talking about the difference between directed connections, which of course requires you to measure this directed effective connectivity. So many of the applications of Dynamic Causal Modeling are to understand data from imaging experiments either with FMRI or the electromagnetic sort, in terms of the distinction of bottom-up processing and top-down processing. And one important aspect of the top-down processing is to contextualize and to select the channels that can provide the bottom-up input.

So as I’m talking, then I am selecting specifically certain cues in terms of when I should say certain words and where I am in terms of the narrative that I’m pursuing. And in that selection, I am giving weight to and modulating and contextualizing the sorts of information that I need to sample to work out what I’m going to do next. So that practically, simply means, switching on some connections and switching off other connections. So what am I saying here? Well, to understand the context-sensitive nature of functional architectures in the brain, we need to understand how the connection strengths the affected connectivity between the brain areas is it self contextualized and controlled on a moment to moment basis. So that’s probably the most interesting aspect of Dynamic Causal Modeling. It is not the architecture in and of itself, although that is very important, it’s how those connection strengths, that coupling, changes as a function of what I am doing, what I am attending to, what I am intending to do. So all that higher cognitive function becomes then characterized in terms of tuning the coupling and selecting which connections are in play at any one time.

So Dynamic Causal Modeling, in conclusion, is a modeling procedure that allows one to pose questions about functional brain architectures or indeed the architectures of any coupled dynamical system to data to ask questions, not only about which connections are present and how their deployed, is this a sort of centripetal or is it a hierarchical structure, is it fully connected, is it very sparse, does it have small world characteristics. All of these characteristics are a way of understanding networks, depend upon knowing which connections are present and which are not present and which are in play and which are not in play.

Futhermore, beyond that I can equip these models with a context sensitivity, by saying, in this condition these sets of connections will be active, and in this situation, they won’t be. And I can have a connection context sensitivity built into my model, and I can estimate that and start to tell you which connections you’re using at the moment whilst listening to me.