WEBVTT

00:00:03.577 --> 00:00:10.497
This is the Convergent Science Network podcast. Leading researchers in the domain

00:00:10.497 --> 00:00:16.777
of neuroscience, brain theory and technology are interviewed by Paul Verschure and Tony Prescott.

00:00:19.297 --> 00:00:22.977
Paul Verschure here with the Convergent Science Network podcast.

00:00:24.417 --> 00:00:27.197
I'm here with Victor Jesra. Victor, welcome.

00:00:27.677 --> 00:00:33.637
You were speaking today at BCPT 2018 about translational neuroscience with a

00:00:33.637 --> 00:00:37.917
very strong emphasis on a more that's a whole brain perspective on epilepsy.

00:00:38.837 --> 00:00:43.617
There are two issues here. Why epilepsy?

00:00:43.897 --> 00:00:52.197
Why do you think epilepsy is a helpful lever to understand how brains work or not work?

00:00:53.997 --> 00:00:57.357
Enough. So, first of all, hello. Nice to be here.

00:01:00.217 --> 00:01:05.537
I have chosen epilepsy for a variety of reasons.

00:01:06.117 --> 00:01:13.457
But epilepsy is a dynamic disorder. It expresses itself through very characteristic

00:01:13.457 --> 00:01:17.617
features, spatially and temporally.

00:01:17.657 --> 00:01:23.317
Temporally, you have high-frequency oscillations that are, if you have the electrodes,

00:01:23.317 --> 00:01:29.837
If you happen to have the electrodes in the brain that are visible with the eye, essentially.

00:01:30.077 --> 00:01:34.597
So you have very clear data features that you can recognize that are then also

00:01:34.597 --> 00:01:42.077
linked to the semiology, to the science, to the clinical science when the seizure

00:01:42.077 --> 00:01:44.457
onsets. That makes it unique.

00:01:44.977 --> 00:01:49.297
It's also spatial because it involves a network.

00:01:49.637 --> 00:01:53.557
We talk about epilepsy spread. Spread. The seizure spreads through the network.

00:01:54.556 --> 00:02:01.416
From a perspective of someone who wants to model the brain, someone who wants

00:02:01.416 --> 00:02:05.316
to understand emergent brain function or dysfunction,

00:02:05.916 --> 00:02:10.736
here you don't have to look for the features. They're evident.

00:02:10.936 --> 00:02:15.756
They have been documented for a long time. They are linked to characteristic

00:02:15.756 --> 00:02:23.876
behavioral patterns. When you look at other diseases or disorders linked to

00:02:23.876 --> 00:02:25.216
the brain, it's much more difficult.

00:02:25.336 --> 00:02:29.396
What you do with schizophrenia, what you do with bipolarity,

00:02:29.436 --> 00:02:33.396
what you do even with multiple sclerosis where you have a structural equivalent,

00:02:33.696 --> 00:02:38.496
when you want to measure the function or the impaired function,

00:02:38.956 --> 00:02:45.436
what you look at in the data, in the measurements, it's extremely difficult.

00:02:45.436 --> 00:02:50.336
So even already at the first starting point, you run into difficulties.

00:02:50.596 --> 00:03:02.176
And epilepsy promised for me a wonderful entry point, enabling me to apply some

00:03:02.176 --> 00:03:05.296
of the tools that were under my control.

00:03:05.756 --> 00:03:08.376
So how long ago did you start with epilepsy?

00:03:09.616 --> 00:03:15.776
Epilepsy, we started seven years ago.

00:03:15.936 --> 00:03:21.096
I followed epilepsy before a little bit, but more as an interested observer,

00:03:21.416 --> 00:03:23.676
fascinated by the dynamic features.

00:03:25.056 --> 00:03:31.476
But epilepsy per se, the first time I touched epileptic data was seven years

00:03:31.476 --> 00:03:38.136
ago. So, epilepsy was among the Greeks known as the holy disease or the sacred

00:03:38.136 --> 00:03:39.716
disease, the sacred disease, right?

00:03:39.776 --> 00:03:44.616
Because it was believed that this had supernatural features.

00:03:45.576 --> 00:03:50.716
So, what are the key features in epilepsy that you think you should try to understand

00:03:50.716 --> 00:03:54.896
and control? control there are,

00:03:57.036 --> 00:04:05.256
is a number of features that come to the mind that.

00:04:08.643 --> 00:04:14.343
Epilepsy is fairly widespread in its expressions.

00:04:14.443 --> 00:04:17.883
It depends on the type of epilepsy that you look at.

00:04:18.403 --> 00:04:22.663
Today in today's talk, I showed you a patient with a frontal,

00:04:22.823 --> 00:04:26.263
prefrontal organization of the network involving the temporal lobe.

00:04:26.263 --> 00:04:30.463
And there the

00:04:30.463 --> 00:04:34.523
behavioral features are fairly expressive

00:04:34.523 --> 00:04:41.403
in the sense of behavior these were normal behavioral features you would not

00:04:41.403 --> 00:04:49.083
know yeah if you isolate some of these features you would not think that this

00:04:49.083 --> 00:04:54.283
is an epileptic seizure such as rocking off the body or crossing the legs So,

00:04:54.403 --> 00:04:58.843
in other features, you just have these muscle spasms in other seizures.

00:04:59.103 --> 00:05:10.983
One of the key features we need to get under control is when the seizure propagates

00:05:10.983 --> 00:05:12.643
and spread out through the network.

00:05:12.963 --> 00:05:18.083
It starts taking away control for the patient.

00:05:18.083 --> 00:05:22.783
This is one of the most horrible experiences for the patient,

00:05:22.843 --> 00:05:28.523
that the patient starts losing control of his or her behavior,

00:05:28.723 --> 00:05:33.503
and that is often not linked to the onset of the seizure, but actually to the

00:05:33.503 --> 00:05:34.563
propagation of the seizure.

00:05:34.563 --> 00:05:38.823
So, when you talk about some of the features we would have to get under control

00:05:38.823 --> 00:05:42.083
is to improve the quality of life for the patient.

00:05:42.163 --> 00:05:49.063
If we can limit this type of impairing features on the behavioral level and

00:05:49.063 --> 00:05:53.103
just constrain the epilepsy.

00:05:53.483 --> 00:05:59.903
The discharge from propagating, from spreading through the network and recruiting other areas,

00:06:01.123 --> 00:06:04.543
this would be a wonderful feature to get under control.

00:06:04.563 --> 00:06:07.283
Another feature that often is being.

00:06:08.672 --> 00:06:17.332
Evoked is a loss of consciousness, which is also often linked to the propagation of the seizure.

00:06:18.552 --> 00:06:24.292
This is horrible for a patient when she knows that she can lose consciousness at any moment now.

00:06:24.432 --> 00:06:31.872
When the aura appears, then the patient already knows I may be able to lose

00:06:31.872 --> 00:06:33.352
consciousness and driving a car.

00:06:33.352 --> 00:06:41.132
This is a very debilitating and constraining feature for the patient.

00:06:41.812 --> 00:06:46.292
So now you've pointed the way to an important step in your approach, right?

00:06:46.332 --> 00:06:51.352
Because then the surprising thing, I guess for the classics.

00:06:52.432 --> 00:06:54.312
The symptoms could be so variable.

00:06:54.692 --> 00:06:57.892
It might indeed be loss of consciousness, it might be vocalizations,

00:06:58.312 --> 00:07:02.512
it might be movements, right? In the end, going back to the same underlying

00:07:02.512 --> 00:07:06.552
deficit, if you only look at the surface of expressions of symptoms,

00:07:06.732 --> 00:07:08.412
indeed, it looks very mysterious.

00:07:08.932 --> 00:07:13.972
So on those grounds, you already said, look, it's a network. It's a network deficit.

00:07:14.292 --> 00:07:22.052
So would you really describe it in those terms? You would see epilepsy really as a network pathology?

00:07:22.932 --> 00:07:25.812
I would describe it as a network disorder. order

00:07:25.812 --> 00:07:29.452
um and technically

00:07:29.452 --> 00:07:32.732
speaking uh so in this sense it's not a disease all

00:07:32.732 --> 00:07:36.452
of us can have the capacity

00:07:36.452 --> 00:07:44.752
to show an epileptic seizure some of my colleagues even even say even claim

00:07:44.752 --> 00:07:50.732
that epileptic discharges or the capacity to show epileptic seizures as part

00:07:50.732 --> 00:07:55.992
Part of our dynamic repertoire is part of the dynamic repertoire of the brain.

00:07:57.402 --> 00:08:00.022
Let's take the short way for ripples and hippocampus, right?

00:08:00.522 --> 00:08:05.742
Yeah, but this is more spatially localized.

00:08:06.342 --> 00:08:15.122
This is temporal features. The approach that we have taken is a network approach

00:08:15.122 --> 00:08:19.602
and reducing it to a statement such as it's the same deficit.

00:08:20.902 --> 00:08:22.362
It's simplifying it.

00:08:23.802 --> 00:08:33.402
The epilepsies are so different. They are linked to network activations that organize themselves.

00:08:34.322 --> 00:08:40.382
The underlying mechanisms may be completely different.

00:08:40.522 --> 00:08:49.102
However, the way how the network then expresses itself, this is then linked

00:08:49.102 --> 00:08:54.662
to the feature and the subsequent semiology for the patient.

00:08:54.662 --> 00:09:00.722
So this is then our entry point. We do not...

00:09:01.462 --> 00:09:07.102
My lab is a highly theoretical quantitative lab composed of mathematicians,

00:09:07.782 --> 00:09:14.022
physicists, engineers in the institute working very closely together with signal

00:09:14.022 --> 00:09:17.022
processing engineers, clinicians, neurologists, neurosurgeons,

00:09:17.682 --> 00:09:20.482
all of us together housed in the same institute.

00:09:21.742 --> 00:09:23.242
We are...

00:09:24.535 --> 00:09:30.175
Not necessarily looking at the genesis of epilepsy and trying to identify the

00:09:30.175 --> 00:09:35.275
mechanisms underlying the genesis, at least the quantitative part of the group.

00:09:35.415 --> 00:09:38.935
But we are trying to understand once the network is epileptogenic,

00:09:39.115 --> 00:09:46.535
the organization of the network that is linked to those features that you were referring to.

00:09:46.935 --> 00:09:53.735
And which means you could argue that each epileptic network is different.

00:09:54.195 --> 00:10:03.315
However, certain concepts and principles should be obeyed at least as long as

00:10:03.315 --> 00:10:05.655
we are looking at it from the network perspective.

00:10:05.915 --> 00:10:11.155
I'm always coming back to the network because if you look at an individual area

00:10:11.155 --> 00:10:16.375
with the type of approach we have taken, I cannot make a statement about that

00:10:16.375 --> 00:10:20.415
beyond some dynamic features, but I cannot make mechanistic statements about that.

00:10:20.475 --> 00:10:26.355
I can make mechanistic statements in terms of network language about the epileptic network.

00:10:26.695 --> 00:10:30.435
Coming back to your first question is, this is where the power of our approach

00:10:30.435 --> 00:10:35.015
lies, where we can, despite the fact that we take a mathematician's approach

00:10:35.015 --> 00:10:41.495
to epilepsy, it expresses itself through network features.

00:10:42.035 --> 00:10:43.615
This is where we can contribute.

00:10:44.835 --> 00:10:52.455
For me, important things now, Because it's a multi-scale phenomenon, right?

00:10:52.515 --> 00:10:58.535
Because there's a local circuit that might generate some dynamics that has some

00:10:58.535 --> 00:11:02.875
knock-on effect on the surrounding networks. And now I go from micro to macro.

00:11:03.095 --> 00:11:05.335
So there are two levels of organization we shouldn't worry about.

00:11:07.575 --> 00:11:13.775
But this microscopic, if you want, distortion and perturbation can come in many

00:11:13.775 --> 00:11:15.215
forms, that's what you're saying, right?

00:11:15.215 --> 00:11:21.675
But that's maybe not the most important part to understand or to control the

00:11:21.675 --> 00:11:23.135
perspective of symptomatology.

00:11:23.375 --> 00:11:27.335
What you're saying is what you really want to understand is that these knock-on

00:11:27.335 --> 00:11:32.535
effects across the network that will be invariant, independent of what this

00:11:32.535 --> 00:11:35.455
microscopic deficit exactly is. So this is the consequence.

00:11:35.655 --> 00:11:39.675
So you also see a sort of encapsulation of these two levels of operation.

00:11:40.535 --> 00:11:46.055
Because it also would mean that if as soon as a network kicks in and starts

00:11:46.055 --> 00:11:51.375
to switch itself into a pathological state or dynamic, it doesn't matter anymore

00:11:51.375 --> 00:11:53.835
what you do to the local pathological circuit.

00:11:53.975 --> 00:11:57.815
The network now is pushed into this part of the state's place where it will

00:11:57.815 --> 00:12:00.255
give rise to symptoms that you don't want to have.

00:12:01.246 --> 00:12:04.626
This is more or less, actually, this is exactly what I was saying.

00:12:04.706 --> 00:12:08.866
So you introduced the language of micro-macro, so the microcircuitry.

00:12:09.306 --> 00:12:14.406
What I did not say, though, is that the microcircuitry or the microscopic understanding

00:12:14.406 --> 00:12:17.446
is not important by no means.

00:12:17.606 --> 00:12:25.486
I'm saying it's not our entry point towards the understanding of epileptic networks.

00:12:25.486 --> 00:12:27.046
It's extremely important.

00:12:27.266 --> 00:12:37.426
And in fact, if you want to intervene with the epilepsy of a human's brain on

00:12:37.426 --> 00:12:41.746
the microscopic level, then pharmaceutics.

00:12:41.746 --> 00:12:44.646
This is where the molecular entry points are.

00:12:44.746 --> 00:12:49.326
This is where you have signaling pathways where you want to intervene,

00:12:49.546 --> 00:12:56.126
that you want to find the right molecule that has the right effect and reorganizes

00:12:56.126 --> 00:13:04.066
the microcircuitry and drives the area away from its capacity of discharging. So this is important.

00:13:04.246 --> 00:13:12.406
But very often, this magic molecule that does this job, it's simply not to be found.

00:13:12.686 --> 00:13:17.086
And if you look at the development of drug history,

00:13:17.306 --> 00:13:23.466
so it goes back roughly 80 years, we have essentially four or five families

00:13:23.466 --> 00:13:31.066
of anti-epileptic drugs within multiple branching into sub-families, etc., but not more.

00:13:31.066 --> 00:13:33.446
Yeah and there.

00:13:36.226 --> 00:13:42.186
Since 30% of all epileptic patients are drug resistant there you have to find

00:13:42.186 --> 00:13:47.746
well either new drugs or other ways of intervening and this is where we are

00:13:47.746 --> 00:13:52.526
then coming in and then it's not the microscopic line yeah well on top of there

00:13:52.526 --> 00:13:53.726
are two things here and on top of that,

00:13:54.549 --> 00:13:57.249
What are the side effects of these drugs that are being used?

00:13:57.569 --> 00:14:01.189
Are they harmless in that sense, or do people pay a price for using them?

00:14:01.569 --> 00:14:05.169
I cannot tell you. This is not my expertise.

00:14:06.109 --> 00:14:09.109
I do not dare to express myself on the side effects.

00:14:09.429 --> 00:14:13.849
But the other thing is that you're saying, I understand you have to say the

00:14:13.849 --> 00:14:20.289
microscopic generator that kickstarts this whole process is important.

00:14:20.289 --> 00:14:24.569
Because this is, of course, also reflecting how the field itself is organized.

00:14:24.749 --> 00:14:31.289
A lot of effort has been put into trying to understand the local genesis of the seizures.

00:14:31.669 --> 00:14:35.169
On the other hand, as you also said yourself, over the last 15 years,

00:14:35.229 --> 00:14:38.169
we've made actually no progress whatsoever in treating epilepsy.

00:14:39.369 --> 00:14:45.049
I did not say this. Over the last 50 years, I was talking about surgery.

00:14:45.289 --> 00:14:49.269
Yeah. In treating pharmacoresistant epilepsy.

00:14:49.269 --> 00:14:57.649
And if you average across all epilepsies, then you have a fairly flat curve

00:14:57.649 --> 00:15:03.229
in the surgery success rate,

00:15:03.409 --> 00:15:09.029
which averaged across all epilepsy types is around 50%.

00:15:09.549 --> 00:15:14.229
Temporal lobe epilepsy is better. That's 70%. Frontal lobe is lower,

00:15:14.509 --> 00:15:20.149
25-30%. So let's say 50% and it has not improved. But that was surgery success.

00:15:21.754 --> 00:15:26.134
Okay, I over-generalized. However... It's important. It's important, yeah.

00:15:26.194 --> 00:15:31.274
Sure, that's fine. But still, the consequence is still that if you would have

00:15:31.274 --> 00:15:37.874
to choose today where you can have the most impact in trying to make progress in treating epilepsy.

00:15:38.774 --> 00:15:44.694
The network might be maybe a more opportune target than the local circuit.

00:15:44.894 --> 00:15:48.194
If you really want to make progress on intervention planning,

00:15:48.554 --> 00:15:51.354
treatment, treatment symptom control and so

00:15:51.354 --> 00:15:54.774
on i i if i understand you correctly you're

00:15:54.774 --> 00:15:57.734
you would have good reasons to go for the network as

00:15:57.734 --> 00:16:01.254
opposed to the microscopic generator of the sea yeah

00:16:01.254 --> 00:16:04.574
i would be very comfortable with this

00:16:04.574 --> 00:16:08.394
state i would be comfortable with this statement statement uh

00:16:08.394 --> 00:16:12.594
from the perspective

00:16:12.594 --> 00:16:16.594
of making the biggest progress

00:16:16.594 --> 00:16:23.394
um on the microscopic mechanistic level it's important to continue there's no

00:16:23.394 --> 00:16:28.454
question about this but there are also many co-dependencies yeah so this magic

00:16:28.454 --> 00:16:35.934
molecule it in the testing that is being performed There are co-dependencies on other factors, etc.

00:16:36.274 --> 00:16:43.934
So it's kind of one of these blue sky projects that we talked about earlier

00:16:43.934 --> 00:16:53.594
that we need a clear plan and agenda organizing our thoughts in order to move

00:16:53.594 --> 00:16:54.614
forward more structured.

00:16:55.674 --> 00:17:04.894
On the network level, though, there we have at the moment, we may approach a tipping point.

00:17:05.874 --> 00:17:11.074
At the moment, we get more and more technology that becomes available to us.

00:17:12.050 --> 00:17:15.350
Interact to modulate the network stimulation

00:17:15.350 --> 00:17:18.250
is one for instance different types of

00:17:18.250 --> 00:17:22.890
stimulation become available non-invasive surgery

00:17:22.890 --> 00:17:27.010
non-invasive surgery does not exist less invasive

00:17:27.010 --> 00:17:30.270
or minimally invasive surgery such as

00:17:30.270 --> 00:17:33.630
thermocoagulation or laser surgery where

00:17:33.630 --> 00:17:36.670
you enter into the brain through very little drill

00:17:36.670 --> 00:17:42.330
holes and are able to make ablations

00:17:42.330 --> 00:17:45.550
at little volumes in

00:17:45.550 --> 00:17:48.790
the brain that can help again network concept

00:17:48.790 --> 00:17:54.650
to re-equilibrate the network and there so there we have a battery of tools

00:17:54.650 --> 00:18:02.250
that allows us to give access what we need is an understanding of how to use

00:18:02.250 --> 00:18:07.270
these tools in a well-informed manner.

00:18:07.530 --> 00:18:13.410
And there we are at the beginning, really at the beginning.

00:18:13.470 --> 00:18:19.490
We have realized that the epilepto focus is actually an epileptogenic network.

00:18:19.730 --> 00:18:26.410
It's an epileptogenic zone spread, sometimes very focused, sometimes disparate

00:18:26.410 --> 00:18:30.830
with topologically non-connected elements, that the propagation network can

00:18:30.830 --> 00:18:34.350
be widespread. So it's clearly a network phenomenon.

00:18:35.030 --> 00:18:38.970
How can we make use of that without...

00:18:40.137 --> 00:18:43.017
Generating huge cognitive deficits because that

00:18:43.017 --> 00:18:46.237
is always when you interfere with the network you generate

00:18:46.237 --> 00:18:49.457
cognitive deficits can we ask questions or

00:18:49.457 --> 00:18:53.317
build decision-making software guide the

00:18:53.317 --> 00:19:00.737
surgeon of making an informed intervention minimally invasive reducing maybe

00:19:00.737 --> 00:19:07.877
just the epileptic seizure propagating and minimizing the cognitive cognitive

00:19:07.877 --> 00:19:10.477
deficit so So there are loads of possibilities.

00:19:10.657 --> 00:19:16.817
And this is where I hope that models can contribute a lot in the future.

00:19:16.997 --> 00:19:22.237
And then there will be other ways of interfering. It doesn't have to be surgically,

00:19:22.317 --> 00:19:29.417
but there may be some drugs that will be delivered locally, for instance,

00:19:29.657 --> 00:19:32.077
that can turn on off populations.

00:19:32.757 --> 00:19:36.817
So there will be other possibilities. Drosothicols, right? It might all be non-invasive.

00:19:36.997 --> 00:19:40.977
Yeah. You got nox. Yeah. But you're saying two things that really stand out, right?

00:19:41.417 --> 00:19:47.317
One, on the one hand, what you are announcing is sort of a revolution of moving

00:19:47.317 --> 00:19:49.297
away from the idea of the broken brain.

00:19:49.417 --> 00:19:51.417
Like, oh, some molecule is missing.

00:19:51.777 --> 00:19:57.417
And if I just now reinsert that molecule, then everything will restore itself to normal.

00:19:57.977 --> 00:20:04.617
And from there, you move more to a network medicine perspective on brain pathology.

00:20:04.617 --> 00:20:10.417
And I think this, you are, I think, really are one of the big examples of that

00:20:10.417 --> 00:20:12.157
movement, I think, right now in neuroscience.

00:20:12.517 --> 00:20:15.937
And it's an important one. We have to really, I think, appreciate that also

00:20:15.937 --> 00:20:21.137
see it as a very important development in how we think about neuropathology.

00:20:21.257 --> 00:20:27.117
Would you agree with that or do you think I'm really exaggerating that? know i would

00:20:27.117 --> 00:20:31.097
agree with it that

00:20:31.097 --> 00:20:34.857
the network science is definitely entering

00:20:34.857 --> 00:20:38.077
into medicine into

00:20:38.077 --> 00:20:41.877
computational medicine and the network thinking um is

00:20:41.877 --> 00:20:44.837
it a revolution it's changing the way how people

00:20:44.837 --> 00:20:48.657
think about brain disorders definitely

00:20:48.657 --> 00:20:52.997
in epilepsy but also because

00:20:52.997 --> 00:21:00.137
of the successes about the progress we made in epilepsy they start thinking

00:21:00.137 --> 00:21:07.057
about how can we think differently about other disorders or diseases also so

00:21:07.057 --> 00:21:09.137
i agree with the fact that.

00:21:10.492 --> 00:21:13.872
I wouldn't be hesitant to say it is a revolution.

00:21:14.352 --> 00:21:19.492
But it's definitely one of the hot topics at the moment.

00:21:19.912 --> 00:21:23.812
This computational medicine with a network idea, number one,

00:21:23.832 --> 00:21:34.232
and then link it to personalized medicine of being able to render a network patient specific.

00:21:34.232 --> 00:21:41.312
Specific, suddenly we talk about virtual brains of my virtual brain and it bears lots of promise.

00:21:42.352 --> 00:21:46.232
But there are two consequences now, right? So on the one hand,

00:21:46.292 --> 00:21:48.752
it implies that we move to a non-locality principle.

00:21:49.812 --> 00:21:56.332
In the past, with the broken brain view, you would also localize the problem somewhere at X, Y, Z.

00:21:56.512 --> 00:22:02.252
This is where we have to now fix the thing. while you are also showing in the

00:22:02.252 --> 00:22:07.872
results we might touch upon later that these effects can be not co-localized

00:22:07.872 --> 00:22:14.532
with the position of your original let's say pathological circuit or lesion.

00:22:14.652 --> 00:22:21.092
The real problem might be sitting somewhere else as a dynamic reorganization of a network.

00:22:21.372 --> 00:22:26.172
So we have to think about deficits in a non-local global fashion.

00:22:26.812 --> 00:22:30.952
But the second thing that you say here, and this is what you show concretely

00:22:30.952 --> 00:22:35.932
with your virtual brain project, to make progress in network medicine or network

00:22:35.932 --> 00:22:39.692
neuroscience, we must rely on computational methods.

00:22:39.932 --> 00:22:45.572
We have to start to build models and we cannot just follow simple lookup tables

00:22:45.572 --> 00:22:47.792
and heuristics to try to solve the problem.

00:22:47.792 --> 00:22:52.392
So do you really see this as the key strategic step, the strategic trajectory

00:22:52.392 --> 00:22:58.332
that we have to explore now of bringing computational models into the clinic to make progress?

00:23:00.072 --> 00:23:01.872
I subscribe to that.

00:23:05.212 --> 00:23:16.032
You brought up two points. The non-locality has been known actually for a while.

00:23:16.032 --> 00:23:26.312
This is not a new feature or lesions or injuries in the brain at a particular

00:23:26.312 --> 00:23:28.992
location causes deficits in

00:23:28.992 --> 00:23:33.752
functions that are officially localized in completely other brain regions.

00:23:33.752 --> 00:23:37.752
So this has been well known.

00:23:37.852 --> 00:23:46.052
We can invert that and actually make use of this and propose,

00:23:46.252 --> 00:23:47.892
and this is far from standard,

00:23:48.112 --> 00:23:55.072
propose interventions at a different brain region that is actually not involved

00:23:55.072 --> 00:23:58.572
in the network disorder, but then

00:23:58.572 --> 00:24:02.832
has a positive effect upon the network organization and network function.

00:24:04.372 --> 00:24:07.572
This is a logical consequence, isn't it? And the,

00:24:08.980 --> 00:24:17.040
This to find, and now I'm coming to your second point, how can we empirically find ways of doing this?

00:24:17.160 --> 00:24:22.500
For the negative parts, namely injuries in areas causing deficits and other

00:24:22.500 --> 00:24:28.240
subnetworks, we run into this through accidents empirically.

00:24:28.500 --> 00:24:36.140
But the positive, the interventional therapeutic aspect, we cannot run into by coincidence.

00:24:36.140 --> 00:24:39.200
Coincidence so what we cannot

00:24:39.200 --> 00:24:43.300
simply operate blindly on human

00:24:43.300 --> 00:24:46.180
beings we need a strategy and how

00:24:46.180 --> 00:24:49.520
do we do this in silico modeling yeah we understand

00:24:49.520 --> 00:24:52.600
the network better at

00:24:52.600 --> 00:24:55.480
least on the network level of a human being the brain

00:24:55.480 --> 00:24:58.240
network and we try to test out in

00:24:58.240 --> 00:25:01.740
silico new interventional strategies that

00:25:01.740 --> 00:25:04.980
cannot for ethical reasons performed in

00:25:04.980 --> 00:25:08.360
the human being yeah and uh for practical

00:25:08.360 --> 00:25:11.240
reasons not in the animal model because

00:25:11.240 --> 00:25:14.840
it cannot stimulate all different areas so all

00:25:14.840 --> 00:25:18.160
that remains at the moment are in silico approaches

00:25:18.160 --> 00:25:21.160
but having understood these concepts

00:25:21.160 --> 00:25:24.480
it brings up these in silico approaches

00:25:24.480 --> 00:25:27.200
as a vision for the future we have the

00:25:27.200 --> 00:25:30.620
computational powers nowadays yeah uh we

00:25:30.620 --> 00:25:33.300
have high performance computing and structures we have the

00:25:33.300 --> 00:25:36.480
data science that can support it so i

00:25:36.480 --> 00:25:41.740
think it's a no-brainer in quotation marks that the trends are going into the

00:25:41.740 --> 00:25:50.460
direction of uh brain network in silico modeling and the logic that i just proposed

00:25:50.460 --> 00:25:53.900
is completely independent of personal preferences it's uh.

00:25:55.162 --> 00:25:59.722
But surprisingly, even though you call it a no-brainer, it's not widely adopted yet.

00:25:59.822 --> 00:26:04.922
So there are apparently some obstacles we have to overcome, right? So we're not there yet.

00:26:05.182 --> 00:26:10.182
And now, of course, the crux of the matter is, okay, what makes a good model?

00:26:10.602 --> 00:26:18.322
And in your case, you made a very strong point for being field models of the

00:26:18.322 --> 00:26:21.662
whole brain with emphasis on the neocortex, right?

00:26:21.662 --> 00:26:28.422
As a way to start to get a handle on the dynamics of the brain in health and disease.

00:26:28.642 --> 00:26:32.502
So why do you believe mean field models are the way to go?

00:26:35.462 --> 00:26:43.402
On the level of organization that we look at in my laboratory,

00:26:43.402 --> 00:26:47.782
The activity of

00:26:47.782 --> 00:26:53.182
individual brain regions when it's being communicated to other brain regions

00:26:53.182 --> 00:27:05.142
that expresses itself for the communication is sufficiently described on the mean field level.

00:27:05.142 --> 00:27:12.902
We are validating this with high-dimensional microscopic simulations where we

00:27:12.902 --> 00:27:16.302
use single-neuron models,

00:27:16.542 --> 00:27:20.882
not very detailed single-neuron models, but spiking-neuron models.

00:27:21.242 --> 00:27:27.102
And when we perform these very high-dimensional simulations and mimic these

00:27:27.102 --> 00:27:32.422
activations in the large brain network, it takes a long time to simulate. But.

00:27:33.773 --> 00:27:43.713
We find so far the same consequences for the network and our understanding of

00:27:43.713 --> 00:27:46.053
the network's organization.

00:27:46.373 --> 00:27:51.373
For this reason, if you ask network questions.

00:27:52.713 --> 00:28:01.993
It is sufficient to perform the mean field modeling at least in our hands, in our laboratory.

00:28:01.993 --> 00:28:11.273
If you want to pose these network questions and direct them into other directions,

00:28:11.653 --> 00:28:17.573
probably linking to microscopic underpinnings, then you have to go beyond that.

00:28:17.813 --> 00:28:25.893
We're not doing this at the moment. We seek the proximity to the patient and the clinic.

00:28:25.893 --> 00:28:29.513
Yeah the microscopic underpinnings are

00:28:29.513 --> 00:28:32.733
in many cases performed in

00:28:32.733 --> 00:28:36.173
the experimental laboratory not necessarily

00:28:36.173 --> 00:28:42.273
in the clinic unless you extract human tissue etc etc so the b-field model allows

00:28:42.273 --> 00:28:46.533
you to collapse the microscopic dynamics across thousands or not millions of

00:28:46.533 --> 00:28:50.333
neurons into single state variables right but you say look this whole population

00:28:50.333 --> 00:28:55.953
here i can basically capture sure group one state variable that might evolve over time.

00:28:56.133 --> 00:28:58.913
Multiple state variables. Yeah, not one but multiple.

00:28:59.213 --> 00:29:02.193
Yeah, but not many. A few. A handful. Yeah.

00:29:02.893 --> 00:29:07.213
But for the question then comes, what's your benchmark? Right? So...

00:29:08.549 --> 00:29:14.089
What are for you the dominant benchmarks to validate that very abstract,

00:29:14.329 --> 00:29:16.209
compressed view of brain dynamics?

00:29:16.389 --> 00:29:22.349
What is really the benchmark, the gold standard you feel today to validate such a meaningful model?

00:29:23.609 --> 00:29:34.769
What is being done in the literature and in the community is you use paradigms such as stimulation.

00:29:37.929 --> 00:29:43.269
And stimulate a microcircuit composed of these millions of neurons and then

00:29:43.269 --> 00:29:46.689
you get these firings of action potentials of spikes.

00:29:47.109 --> 00:29:52.129
There you have certain features with the raster plots with the particular statistics.

00:29:53.969 --> 00:29:58.129
That's the way in which you visualize it or assess it, but would it be like

00:29:58.129 --> 00:29:58.989
the resting state network?

00:29:59.329 --> 00:30:03.129
Is that for you? This is where I'm going to.

00:30:03.129 --> 00:30:05.909
So this is what is being done in

00:30:05.909 --> 00:30:08.769
the community mean field mean field does not make any

00:30:08.769 --> 00:30:12.389
statement about networks yet okay yeah mean

00:30:12.389 --> 00:30:15.229
field doesn't make any reference to networks yet it's

00:30:15.229 --> 00:30:18.149
a localized population so most

00:30:18.149 --> 00:30:21.449
mean field uh modelers mathematicians

00:30:21.449 --> 00:30:24.689
that work with that uh look at these features what

00:30:24.689 --> 00:30:28.189
we do in our hands the mean

00:30:28.189 --> 00:30:32.749
field has to represent uh

00:30:32.749 --> 00:30:36.429
for uh the same propagation for

00:30:36.429 --> 00:30:39.649
instance through the network when you stimulate and then the same sequence

00:30:39.649 --> 00:30:43.309
of brain regions is being activated um we

00:30:43.309 --> 00:30:47.489
have not looked at resting state activity yeah

00:30:47.489 --> 00:30:55.349
but what we have looked at is we take a very simplified network model with a

00:30:55.349 --> 00:31:02.369
reduced architecture and implement it with detailed microscopic neuronal features

00:31:02.369 --> 00:31:06.389
and then with mean fields and then we obtain,

00:31:06.929 --> 00:31:14.429
under a parameter modulation for instance connectivity or time delays we obtain the same.

00:31:14.989 --> 00:31:18.529
Uh behaviors such as a increase

00:31:18.529 --> 00:31:24.349
of synchronicity the spatial reorganization at a particular value of the control

00:31:24.349 --> 00:31:29.329
parameter that we manipulate etc so we want to manipulate network features and

00:31:29.329 --> 00:31:36.229
then we get the same behavior this is what we have done we have not validated this in a.

00:31:37.284 --> 00:31:43.644
Full connectome-based brain network models where we have a full implementation

00:31:43.644 --> 00:31:51.984
of spiking neuronal networks with a connectome versus a mean field-based model.

00:31:52.124 --> 00:31:55.704
The mean field-based brain network model we have done, we have published with

00:31:55.704 --> 00:32:01.204
this for over 10 years and it's well established nowadays.

00:32:01.944 --> 00:32:07.564
Many labs are working with these type of concepts. There are efforts within

00:32:07.564 --> 00:32:14.464
the Human Brain Project trying to generate the neuroinformatics frame in which

00:32:14.464 --> 00:32:18.804
we can do these high-dimensional microscopic networks,

00:32:19.284 --> 00:32:22.284
but it's not done yet.

00:32:22.284 --> 00:32:27.164
There are efforts from the Diesmann Lab in Jülich. There are efforts from my group.

00:32:27.944 --> 00:32:36.404
There are mixed efforts in the sense that we built a network partly composed

00:32:36.404 --> 00:32:42.164
of mean fields and partly composed of these high-dimensional microscopic networks

00:32:42.164 --> 00:32:45.044
to demonstrate this called co-design,

00:32:45.124 --> 00:32:49.064
to demonstrate that we get the same dynamics.

00:32:49.064 --> 00:32:52.024
But this is all this is still unpublished

00:32:52.024 --> 00:32:56.764
so these are efforts for validation and

00:32:56.764 --> 00:33:00.304
it's crucial and critical but efforts for validation for

00:33:00.304 --> 00:33:03.164
which the first results will come out in

00:33:03.164 --> 00:33:09.264
one year two years three years yeah but it's all uh going in place yeah but

00:33:09.264 --> 00:33:13.664
this this sounds like validation against more microscopic electrophysiology

00:33:13.664 --> 00:33:18.624
right where you can capture that correctly but given given Given the huge amount

00:33:18.624 --> 00:33:21.224
of work already done in mean field models,

00:33:21.504 --> 00:33:26.744
I assume that within that community, there must be a sense of having gold standards

00:33:26.744 --> 00:33:30.344
for the validity of these models.

00:33:30.764 --> 00:33:34.844
Because there also have been plenty of publications of mean field models describing

00:33:34.844 --> 00:33:38.684
the whole neocortex doing something.

00:33:38.684 --> 00:33:46.204
So today, if you today would have to list such a gold standard benchmark to

00:33:46.204 --> 00:33:49.844
calibrate a mean field model of the whole brain, what would it be?

00:33:54.324 --> 00:34:02.444
Here we are running into a situation where… You have to terminate the interview. No, I'm leaving.

00:34:04.784 --> 00:34:08.804
No, it's less a model that is the problem,

00:34:08.864 --> 00:34:13.684
but it's more the metrics that you use to describe the resting state activity

00:34:13.684 --> 00:34:20.864
and the data feature that you use to calibrate and validate the model. And...

00:34:22.647 --> 00:34:28.987
And one of the standards, and since you asked for gold standards,

00:34:29.227 --> 00:34:36.187
established today is in connectomics, the notion of functional connectivity,

00:34:36.547 --> 00:34:41.887
which even there I said the notion of because there are multiple metrics available

00:34:41.887 --> 00:34:43.207
for functional connectivity.

00:34:43.207 --> 00:34:46.167
And every single one of them is questionable to some degree,

00:34:46.327 --> 00:34:56.387
because these metrics collapse much of this information into a very compressed object.

00:34:56.607 --> 00:35:01.227
And we know that mostly this very compressed object of functional connectivity

00:35:01.227 --> 00:35:02.647
requires stationarity.

00:35:02.727 --> 00:35:08.407
We know it's non-stationary. We cannot measure for longer than typically 20

00:35:08.407 --> 00:35:12.447
minutes resting state activity in the human.

00:35:13.167 --> 00:35:24.547
So we run into issues of describing correctly or quantifying with a proper metric what we measure.

00:35:25.387 --> 00:35:31.867
And then you wanted to calibrate the gold standard in order to compare the mean field models.

00:35:31.867 --> 00:35:35.127
And there are you could take

00:35:35.127 --> 00:35:37.987
some data fitting approaches that's not good enough

00:35:37.987 --> 00:35:40.707
you uh in the naive sense of data

00:35:40.707 --> 00:35:46.227
fitting you will always find a minimum yeah and uh or maximum uh depending on

00:35:46.227 --> 00:35:52.407
what you're looking for and uh if you become more sophisticated non-linear with

00:35:52.407 --> 00:35:59.667
uh uh non-linear Bayesian inference techniques that allow sampling on nonlinear manifolds,

00:36:00.427 --> 00:36:07.367
then we are not there yet that these methods can converge and provide us with good outcomes.

00:36:07.467 --> 00:36:12.467
The diagnostics is in place, but the methods may never converge in our lifetime.

00:36:12.747 --> 00:36:19.747
So at the moment, for this type of validation, it's not the mean field network

00:36:19.747 --> 00:36:27.587
model that poses a problem. It's a go-between, the metrics that shall be used for validation.

00:36:27.927 --> 00:36:35.647
How do you quantify a spatial temporal trajectory evolving in a high-dimensional

00:36:35.647 --> 00:36:43.367
space that is undergoing a random process with some deterministic features?

00:36:43.647 --> 00:36:49.507
The situation is worse than initially expected, right? Because even if you had

00:36:49.507 --> 00:36:53.327
the gold standard, you wouldn't even know how to sort of quantitatively match to it.

00:36:54.506 --> 00:37:00.986
Exactly. Or I could even pervert it even more. I build a virtual brain.

00:37:01.286 --> 00:37:04.326
I tell you it's conscious. How would you test it?

00:37:04.466 --> 00:37:10.026
You may not even be capable of recognizing. I don't want to go in direction consciousness.

00:37:10.226 --> 00:37:16.966
I just said consciousness. But how would you, what type of metrics would you

00:37:16.966 --> 00:37:19.946
have to apply in order to go there?

00:37:20.066 --> 00:37:27.326
And I think this is something that we are suffering from in system neurosciences.

00:37:27.606 --> 00:37:33.306
We cannot simplify in order to be close to the patient or to real world applications.

00:37:33.346 --> 00:37:36.006
We cannot simplify everything to a state.

00:37:36.246 --> 00:37:40.586
And then we just describe the state or the statistics of the state. It's a dynamic process.

00:37:40.846 --> 00:37:44.526
So we need a process-based signal analysis.

00:37:44.886 --> 00:37:48.886
Right. I completely get that, and I agree with you.

00:37:49.246 --> 00:37:56.366
So I'll confess that, for me, the mean field approach has never been that convincing,

00:37:56.606 --> 00:38:00.566
even though it's always presented to me with a lot of confidence, right?

00:38:01.046 --> 00:38:05.366
Because in some sense, it has always been anchored to very slow signals.

00:38:05.366 --> 00:38:10.286
The majority of mean field models have been calibrated against fMRI data,

00:38:10.606 --> 00:38:16.406
which is very coarse, a very, very, very low-pass-filtered representation,

00:38:16.766 --> 00:38:19.986
both spatially and temporally, of what goes on in the brain.

00:38:20.106 --> 00:38:24.886
So if we move to this low-dimensional state space, in some sense,

00:38:24.886 --> 00:38:27.406
many abstract models can do a good job.

00:38:27.406 --> 00:38:33.526
Well, so this is arguably a little bit overconfident at this point in time about

00:38:33.526 --> 00:38:35.506
the power of these mean-field models.

00:38:35.706 --> 00:38:40.026
I mean, you could argue that for the low-dimensional state space,

00:38:40.366 --> 00:38:43.786
they're sort of super powerful. They'll always work. They'll always capture these dynamics.

00:38:43.946 --> 00:38:47.246
For that, you have enough free parameters. Not a big deal, right? You can always do it.

00:38:48.268 --> 00:38:52.588
But, friend, if you go to a more realistic state space, for instance,

00:38:52.728 --> 00:38:53.988
we work with stroke patients.

00:38:54.508 --> 00:38:59.908
We have shown recently that they have dysrhythmia, as in Parkinson's.

00:39:00.028 --> 00:39:05.868
So you get dynamic fluctuations in the cortex that appear transiently.

00:39:05.868 --> 00:39:09.708
And if a result was playing out in the thalamus, then itself is going to result

00:39:09.708 --> 00:39:10.788
in what goes on in the cortex.

00:39:11.068 --> 00:39:17.508
So you have rapid shifts in the dynamics that's playing out in a broad frequency range.

00:39:18.268 --> 00:39:23.088
So my mainfield model would not trivially capture that just like that.

00:39:23.488 --> 00:39:27.428
I'm playing a different game now. So the concern I'm having,

00:39:27.508 --> 00:39:32.568
and so maybe you can resolve this for me, but yes, we have built all these tools and they're fantastic.

00:39:32.888 --> 00:39:35.648
If any people got tenure with these kinds of models, it's great,

00:39:36.268 --> 00:39:39.788
but it's all anchored to sort of low-dimensional dynamics.

00:39:40.208 --> 00:39:46.228
What we really want to be, as you described, is high-dimensional transient dynamical

00:39:46.228 --> 00:39:48.988
states to which they might never really generalize.

00:39:49.468 --> 00:39:54.148
So would it not be wise at this point in time to all say, well,

00:39:54.188 --> 00:39:55.888
we learned a lot using mean field models.

00:39:55.928 --> 00:40:02.248
We've learned to appreciate the real problem we're facing, but maybe mean field

00:40:02.248 --> 00:40:05.908
models are now essentially not that helpful anymore as a tool.

00:40:06.028 --> 00:40:10.788
We should move on and think more about, let's say, multi-scale dynamically configured

00:40:10.788 --> 00:40:14.748
networks that operate at micro and macro level simultaneously.

00:40:14.748 --> 00:40:18.908
Simultaneously, would you go in that direction or would you still feel that

00:40:18.908 --> 00:40:22.668
there's a lot of leverage to be gotten from the mean field approach?

00:40:23.428 --> 00:40:31.488
I feel there is still a lot of leverage to be gotten from the mean field for the following reasons.

00:40:33.968 --> 00:40:41.368
The way you argue is that the mean field modeling has found mostly application

00:40:41.368 --> 00:40:44.468
in the resting state literature of the fMRI.

00:40:44.748 --> 00:40:52.168
I would say, no, mean field models have been used in the resting state network literature.

00:40:52.948 --> 00:40:56.708
And even very phenomenological models, as you correctly pointed out,

00:40:56.728 --> 00:40:58.308
a simple model is fully sufficient.

00:40:58.548 --> 00:41:03.268
We don't need four or five dimensional models. It's just a simple phenomenological

00:41:03.268 --> 00:41:08.448
oscillator may already capture many of the features that are being observed,

00:41:08.608 --> 00:41:09.568
why it's network effects.

00:41:09.788 --> 00:41:13.128
And with that, I agree, actually. They are spatially filtered,

00:41:13.388 --> 00:41:20.908
they are temporally filtered, and if you just look at temporal or spatial configurations, it's...

00:41:22.189 --> 00:41:25.649
Probably not very insightful what you do in the resting state network literature.

00:41:26.009 --> 00:41:29.729
You have to look at spatial temporal features. And there it's becoming interesting.

00:41:30.229 --> 00:41:36.689
And there, having talked about the non-stationarities, you need nonlinear models

00:41:36.689 --> 00:41:40.929
that have certain characteristics that are then being informed by the connectome.

00:41:41.049 --> 00:41:45.949
But in fMRI, due to the nature of the signal, looking at non-stationaries and

00:41:45.949 --> 00:41:51.389
the fMRI signals, there we are already at the front line of the research.

00:41:53.809 --> 00:42:02.969
But having said this, mean field modeling has by far not just been applied to fMRI signals.

00:42:04.289 --> 00:42:09.409
This is fMRI modeling neurovascular coupling, yes, uses mean fields.

00:42:09.409 --> 00:42:14.309
But you can have multi-level, multi-scale,

00:42:14.589 --> 00:42:21.549
mean-field models using the super,

00:42:21.689 --> 00:42:27.909
infra and granular layers and assigning a mean-field or neural population,

00:42:27.989 --> 00:42:29.849
neural mass to each of these layers.

00:42:29.849 --> 00:42:33.809
Interacting with multiple scales interacting with

00:42:33.809 --> 00:42:36.669
different types of connections slower and faster

00:42:36.669 --> 00:42:39.889
connections in order to introduce a temporal multi-scale

00:42:39.889 --> 00:42:44.069
architecture in there which has been very important example is for instance

00:42:44.069 --> 00:42:48.789
the work of Fabrice Wendling who's performing following stimulation paradigms

00:42:48.789 --> 00:42:53.809
and capturing some of the response features there we're talking of a neuroelectric

00:42:53.809 --> 00:42:59.049
signals on time scales that are relevant for the time.

00:43:00.035 --> 00:43:02.715
Processing time scales we encounter in the

00:43:02.715 --> 00:43:06.015
brain for the for the spectral features

00:43:06.015 --> 00:43:10.315
that we are familiar with for the different bands rhythmic bands

00:43:10.315 --> 00:43:14.235
that we are familiar with and there these mean

00:43:14.235 --> 00:43:18.875
field models are not just simple scalar elements

00:43:18.875 --> 00:43:26.255
that are just shifted around on a linear equilibrium point and reorganizing

00:43:26.255 --> 00:43:30.195
the networks as in the resting state literature no here here Here we have to

00:43:30.195 --> 00:43:37.035
work with detailed organizations that can be stimulated,

00:43:37.235 --> 00:43:39.075
that propagate through the network,

00:43:39.355 --> 00:43:43.715
that have continuous propagation through the gray matter, that send signals

00:43:43.715 --> 00:43:45.235
through the white matter.

00:43:46.015 --> 00:43:48.875
This is being done. I agree.

00:43:50.155 --> 00:43:55.935
It is the beginning. I fully agree with that.

00:43:55.935 --> 00:44:02.895
But your question was can we still gain some leverage out of this and yes yeah

00:44:02.895 --> 00:44:07.875
what I see there's a sort of conceptual problem here that,

00:44:08.575 --> 00:44:12.135
the definition of mean field modelism is shifting right so he has this famous

00:44:12.135 --> 00:44:17.475
quote from Norbert Wiener that the best model of a cat is a cat and preferably the same cat,

00:44:18.035 --> 00:44:23.875
what he means with that is to model means to abstract right and in some sense

00:44:23.875 --> 00:44:28.195
what I hear you say is that well Well, as long as we abstract, I can call it mean field.

00:44:30.066 --> 00:44:34.266
Why do you hear this? Well, because you're saying I have mean field models that

00:44:34.266 --> 00:44:40.786
would take into account organization at different layers in the cortex, specific cells.

00:44:40.986 --> 00:44:45.506
And so it just means you change the granularity of the averaging that you perform.

00:44:46.046 --> 00:44:50.686
Is what we perform, yes. That they have certain properties such as adaptation.

00:44:51.146 --> 00:44:57.766
And then in the different layers, you have different populations and this is being reflected.

00:44:57.766 --> 00:45:03.906
But that means if you go to the mean field models of some time ago,

00:45:04.046 --> 00:45:08.206
that was more a closer link to, let's say, statistical physics. Yes.

00:45:08.726 --> 00:45:11.686
They made, I think,

00:45:11.706 --> 00:45:16.866
a stronger claim of the boundary because those mean field models also had the

00:45:16.866 --> 00:45:22.926
objective or at least a mission to collapse them into some sort of mastery equation

00:45:22.926 --> 00:45:26.906
with which you can really describe the macroscopic dynamics of that system.

00:45:26.906 --> 00:45:30.586
So we need the search for this abstraction to also, in the end,

00:45:30.606 --> 00:45:32.506
have an analytical handle on that system.

00:45:32.866 --> 00:45:37.986
Well, what you're saying now, the way I take it, which is not necessarily criticism,

00:45:38.326 --> 00:45:43.846
it just means we have to think a little bit about what we mean exactly with the mean field model.

00:45:43.846 --> 00:45:48.126
Model is, you know, as long as I'm averaging in some sense, as long as I'm a

00:45:48.126 --> 00:45:53.346
collapsing detail into state variables, I'm using a mean field model and therefore

00:45:53.346 --> 00:45:58.106
the mean field model can scale and be diversified in all possible directions.

00:45:58.626 --> 00:46:02.526
But if something starts to represent everything, it represents nothing,

00:46:02.666 --> 00:46:05.346
right? So what then do we really mean with mean field?

00:46:05.506 --> 00:46:08.206
And are we maybe we should then

00:46:08.206 --> 00:46:11.246
also would it be useful to start

00:46:11.246 --> 00:46:14.746
and then also give more specific labels to this granularity

00:46:14.746 --> 00:46:18.806
of modeling right like the number of free per map parameters we are going to

00:46:18.806 --> 00:46:25.806
allow the commitment to analytic solutions or not right so so don't you feel

00:46:25.806 --> 00:46:26.766
that you're sacrificing a

00:46:26.766 --> 00:46:32.106
little bit the specificity of the meaning of a mean field approach only to.

00:46:35.495 --> 00:46:39.575
No, we are still in my comfort range.

00:46:41.195 --> 00:46:48.135
I would not be willing... My comfort range is probably limited up to this point

00:46:48.135 --> 00:46:50.815
where we go across the individual layers.

00:46:51.215 --> 00:47:01.195
I would not go any further in terms of granularity because at some point it doesn't make any sense.

00:47:01.195 --> 00:47:09.855
But in terms of abstraction, it again reduces to what I want to explain,

00:47:10.215 --> 00:47:15.635
what type of signals I want to explain, what type of phenomena I want to explain.

00:47:15.635 --> 00:47:24.855
And the explanatory power of, I mean, field benefits from the layered organization

00:47:24.855 --> 00:47:30.515
and propagation through the gray matter, but also then through the white matter fibers.

00:47:30.515 --> 00:47:38.895
But I would not be comfortable of deconstructing it any further for two reasons.

00:47:42.655 --> 00:47:49.615
Any further deconstruction or specification may not help in the explanatory

00:47:49.615 --> 00:47:56.015
power of the signal that we want to explain, number one.

00:47:56.015 --> 00:48:06.115
Number two, at some point, this mechanism of averaging,

00:48:06.435 --> 00:48:13.635
because of all the detailed architecture,

00:48:14.315 --> 00:48:19.595
physiological architecture that is present, we have glia cells, astrocytes, etc.

00:48:19.835 --> 00:48:27.155
It doesn't make sense anymore. Other effects would have to be integrated also.

00:48:27.615 --> 00:48:34.415
It has its limits. You have to have a sufficiently coarse,

00:48:36.323 --> 00:48:45.903
object in order to average across its inner organization. Otherwise, it doesn't make sense.

00:48:46.263 --> 00:48:50.263
But this is interesting, right? Because it doesn't make sense until it does.

00:48:51.063 --> 00:48:55.983
So, for instance, if for an epilepsy case, I would find out that some sort of,

00:48:58.023 --> 00:49:00.243
infrasupergranular layer distinction would be really critical,

00:49:01.063 --> 00:49:02.863
then you would be happily embracing it.

00:49:03.523 --> 00:49:06.743
Then you would be happy to exceed that boundary. they should not draw.

00:49:09.023 --> 00:49:14.923
Pragmatically thinking I would probably I would then probably consider it I

00:49:14.923 --> 00:49:17.123
would have to be convinced that it does matter etc.

00:49:17.343 --> 00:49:22.703
Yeah of course yeah yeah yeah exactly yeah exactly and I'm not saying that this

00:49:22.703 --> 00:49:28.443
distinction does not matter I have certain objectives with the research I want

00:49:28.443 --> 00:49:30.503
to that I want to do I want to reach,

00:49:31.523 --> 00:49:40.443
real world questions in the clinic so I'm building the tools that help me to

00:49:40.443 --> 00:49:48.603
address this type of questions so I get it but this is important for the field

00:49:48.603 --> 00:49:50.103
of mean field approaches,

00:49:50.723 --> 00:49:55.083
I think it would be useful to have that discussion to also see okay maybe we

00:49:55.083 --> 00:49:59.483
should look at different types of mean field models and also start on their

00:49:59.483 --> 00:50:02.843
interrelationship again it is being actually done Paul,

00:50:03.763 --> 00:50:08.423
and And there is a community working on this. Imagine heterogeneity.

00:50:08.943 --> 00:50:13.983
Neurons in a population are not identical. If you just take the most...

00:50:14.871 --> 00:50:20.251
A simple feature of them the threshold distribution yeah how do you take it

00:50:20.251 --> 00:50:26.651
into account is actually generates if you have this diversity or no dispersion

00:50:26.651 --> 00:50:30.891
within the population which in physical space is zero dimensional but if a dispersion

00:50:30.891 --> 00:50:32.491
of thresholds it already.

00:50:33.731 --> 00:50:39.651
Generates a huge complexity in the dynamics i'm talking about the model now

00:50:39.651 --> 00:50:44.051
just in the model we We know that you can have chimera states, the same population.

00:50:45.491 --> 00:50:51.191
Actually, in this case, it's not chimera states, but you can have clustering in the behavior.

00:50:51.251 --> 00:50:59.511
And within the population, you get clusters of similarly behaving neurons due

00:50:59.511 --> 00:51:03.471
to their similarity in their physiological constants.

00:51:03.471 --> 00:51:09.711
But you cannot average over it because some are spiking regularly and others

00:51:09.711 --> 00:51:15.031
have the irregular spiking frequency that we know like a Poissonian train.

00:51:15.511 --> 00:51:21.131
So, what would I do there?

00:51:21.291 --> 00:51:26.731
And again, if I am aware of these things, I split it in multiple mean fields,

00:51:26.931 --> 00:51:29.671
which means I complexify.

00:51:29.931 --> 00:51:34.891
And this has been done. This is one mean field model that exists and it makes

00:51:34.891 --> 00:51:37.191
sense up to a certain degree.

00:51:38.131 --> 00:51:45.031
But once it becomes non-handleable anymore, then its pragmatic use,

00:51:45.151 --> 00:51:48.711
its pragmatic added value is not there anymore.

00:51:49.091 --> 00:51:53.051
And I would not support this approach to go any further.

00:51:53.731 --> 00:51:58.371
But one thing you did to give more structure to your mean field model interpretation

00:51:58.371 --> 00:52:03.491
was to move towards very distinct models of different kind of oscillators.

00:52:04.429 --> 00:52:08.969
And then to use these to sort of, if you want, constrain the function interpretation

00:52:08.969 --> 00:52:14.629
and then use that again to go back to your epilepsy physiology to say,

00:52:14.749 --> 00:52:17.869
I don't need to model this as some mean field network.

00:52:18.169 --> 00:52:24.109
I can actually re-describe the mean field model as an underlying oscillator

00:52:24.109 --> 00:52:27.229
with different characteristics, even state variables.

00:52:27.229 --> 00:52:30.489
And the question is now becomes, well, this whole family of oscillators,

00:52:30.629 --> 00:52:38.349
which one best describes the specific bit of epileptic state that I'm trying to describe, right?

00:52:38.609 --> 00:52:42.309
So why did you move in that direction? Why did that become then the next step?

00:52:42.429 --> 00:52:47.429
Because in some sense, you then gave up the literal whole brain mean field model,

00:52:47.529 --> 00:52:52.029
which is no, no, I'd use that to calibrate my oscillator model.

00:52:52.169 --> 00:52:54.609
And the oscillator model becomes now my reference interpreted data,

00:52:54.729 --> 00:52:59.929
right? to actually have collapsed again the mean field model of cortex into

00:52:59.929 --> 00:53:00.789
these oscillatory models.

00:53:01.029 --> 00:53:06.969
So has that given you leverage or is it sort of more like an experiment that's

00:53:06.969 --> 00:53:09.589
underway right now and it might not really work out?

00:53:11.709 --> 00:53:16.589
It's interesting that you ask this question. It's actually the way how science

00:53:16.589 --> 00:53:21.929
sometimes goes, it can be very exciting when you look backwards.

00:53:24.900 --> 00:53:31.780
We started off with mean field models, again, driven by a particular question

00:53:31.780 --> 00:53:37.760
in the context of epilepsy that is a propagation through the network.

00:53:37.860 --> 00:53:49.200
I recognized that an important fundamental characteristic is missing in these mean field models.

00:53:49.380 --> 00:53:56.620
It's simply not in there. Through the averaging process, through the methodology

00:53:56.620 --> 00:54:00.640
that is being applied, you essentially lose it.

00:54:01.740 --> 00:54:05.960
It is an additional dimension,

00:54:06.300 --> 00:54:14.160
an additional degree of freedom that we refer to as the slow variable that acts

00:54:14.160 --> 00:54:21.400
upon another timescale that is not part of the classic mean field averaging,

00:54:21.400 --> 00:54:27.840
But that has certain characteristics that guides the, on a slow timescale,

00:54:28.060 --> 00:54:33.040
the mean field dynamics, if you wish, through a sequence of behaviors.

00:54:33.160 --> 00:54:39.160
And then also the system from seizure onset through the evolution during the

00:54:39.160 --> 00:54:42.700
ictal state all the way to seizure offset.

00:54:42.700 --> 00:54:46.200
Said and that was not

00:54:46.200 --> 00:54:49.660
something we could overcome we could

00:54:49.660 --> 00:54:53.140
look into physiology we knew there are slow processes

00:54:53.140 --> 00:55:01.440
classics are extracellular potassium the work of uwe heinemann etc there is

00:55:01.440 --> 00:55:07.340
knowledge in there but that would require an additional mechanistic bottom-up

00:55:07.340 --> 00:55:11.080
approach developing this And it's multifactorial, as we know.

00:55:12.220 --> 00:55:17.060
And we ask the question, can we,

00:55:17.080 --> 00:55:21.640
since we are interested in the network question, can we find another...

00:55:22.830 --> 00:55:29.330
Another perspective still capturing this additional feature in there justified

00:55:29.330 --> 00:55:33.650
by a scientific perspective, but maybe a different perspective.

00:55:33.910 --> 00:55:37.290
And there we turned to mathematics.

00:55:37.710 --> 00:55:45.270
And since we were looking for a slow variable, we tapped into the theorems of

00:55:45.270 --> 00:55:47.950
nonlinear dynamic systems, fast-slow systems.

00:55:47.950 --> 00:55:51.030
And we made use

00:55:51.030 --> 00:55:59.630
of that to abstract entirely away from the physiological interpretation of these

00:55:59.630 --> 00:56:08.250
variables and looked at the dynamic structure of the representative of this

00:56:08.250 --> 00:56:09.390
mean field that we needed.

00:56:09.390 --> 00:56:14.250
And we did this very, very systematically.

00:56:15.190 --> 00:56:18.790
And we learned a lot out of this.

00:56:18.890 --> 00:56:23.390
But it gave us also a new perspective about how to look at seizures,

00:56:23.690 --> 00:56:27.770
of what features to look at seizures, not physiologically motivated,

00:56:28.030 --> 00:56:32.410
not mechanistically motivated anymore, but purely dynamic features.

00:56:32.410 --> 00:56:36.970
And that started a trend in the community in terms of slow variable,

00:56:37.250 --> 00:56:39.970
onset bifurcation, offset bifurcation, etc.

00:56:40.250 --> 00:56:43.990
So that is becoming a language now. This is exciting.

00:56:43.990 --> 00:56:52.810
But then, looking back, I still remember how I went to my collaborator, Christoph Bernard,

00:56:53.050 --> 00:56:58.890
who did the physiological testing in the hippocampus, and we were discussing

00:56:58.890 --> 00:57:07.570
these dynamic features that the meta mean field or this expanded mean field

00:57:07.570 --> 00:57:11.610
representation expressed in terms of phenomenological variables should have.

00:57:11.610 --> 00:57:17.470
And I said, Christoph, I have a problem with that.

00:57:17.510 --> 00:57:26.570
There should be a baseline jump that you see in your data in the time series

00:57:26.570 --> 00:57:31.150
because the bifurcations, the dynamics that we find, there is a baseline jump.

00:57:31.910 --> 00:57:35.710
And it's not in the data. Can I somehow sweep it under the carpet,

00:57:35.850 --> 00:57:39.070
maybe hide it in the signal-to-noise ratio?

00:57:39.070 --> 00:57:41.890
Show or what can we do about this

00:57:41.890 --> 00:57:45.370
and he said victor what you're ah but look these

00:57:45.370 --> 00:57:48.470
are ac data that everyone

00:57:48.470 --> 00:57:51.690
is recording an ac but i can make the same recordings in

00:57:51.690 --> 00:57:58.310
dc yeah so he went back to the lab yeah did the same recording in terms of dc

00:57:58.310 --> 00:58:02.450
and because i told him about the slow variable at the same time he measured

00:58:02.450 --> 00:58:10.090
also oxygenation and atp consumption and came back to me and showed me the same recording AC.

00:58:11.793 --> 00:58:16.613
There you didn't have a baseline jump. In DC, you had exactly the baseline jump

00:58:16.613 --> 00:58:18.553
at seizure onset and offset.

00:58:19.533 --> 00:58:24.113
Hippocampus, in TUTO, in the red, exactly as I expected, very beautiful.

00:58:24.233 --> 00:58:30.193
And the oxygenation and the ATP consumption traced out very beautifully the

00:58:30.193 --> 00:58:34.053
time course of the slow variable that we expected.

00:58:34.453 --> 00:58:37.833
There you could say this is not too surprising because it's,

00:58:37.833 --> 00:58:41.553
of course, linked to energy consumption and it has to happen.

00:58:41.793 --> 00:58:46.313
When the tissue fires very fast.

00:58:46.593 --> 00:58:50.753
Well, but it's, and it's, I'm sure it's not the slow variable,

00:58:50.953 --> 00:58:54.073
but these are representations or expressions thereof.

00:58:54.213 --> 00:58:57.293
But it became consistent and things came together.

00:58:57.513 --> 00:59:02.213
And then we went into human tissue. We contacted other institutions,

00:59:02.593 --> 00:59:06.393
Ikeda in Japan, Milan Brush Deal in Brno.

00:59:06.473 --> 00:59:11.473
They had also DC recordings from the human tissue. And then for the seizure

00:59:11.473 --> 00:59:14.453
types that we were looking at, we found the baseline jumps.

00:59:14.593 --> 00:59:17.733
And this was coming from the dynamic structure of the system.

00:59:17.913 --> 00:59:25.713
And this was really unexpected, very beautiful, and gave us confidence.

00:59:25.933 --> 00:59:33.353
Different elements started coming together. And now people, my colleagues,

00:59:33.513 --> 00:59:39.653
are looking into possible realizations of what we call the slow variable.

00:59:39.833 --> 00:59:45.993
I mentioned extracellular potassium, definitely glial activity are good candidates

00:59:45.993 --> 00:59:52.073
that are linked to the neuroelectric discharges, etc. Right.

00:59:52.193 --> 00:59:56.333
But do you believe that it's the slow variable that is also of most relevance

00:59:56.333 --> 00:59:58.813
with respect to the propagation through the network?

01:00:02.504 --> 01:00:05.584
I have to hypothesize. You ask believe.

01:00:05.844 --> 01:00:12.844
Yes, I do believe that the slow variable plays a particular role in this.

01:00:12.944 --> 01:00:17.164
We know that the glial network is tightly connected to the neural network.

01:00:17.604 --> 01:00:22.764
It's evolving on a slow timescale, much slower than the neuroelectric discharges.

01:00:24.624 --> 01:00:28.504
We do not know enough about that.

01:00:30.704 --> 01:00:36.264
However, I'd like to point out that when seizures propagate and spread,

01:00:36.704 --> 01:00:44.764
sometimes everything discharges, the internet network discharges at the same time.

01:00:45.064 --> 01:00:50.544
Sometimes it starts discharging at a particular location, one,

01:00:50.644 --> 01:00:55.844
two seconds later it recruits another area, then it disengages and recruits another area.

01:00:55.964 --> 01:00:59.064
So you have a spatial temporal organization on the timescale of seconds.

01:00:59.064 --> 01:01:02.584
What are the mechanisms that are linked directly to this?

01:01:02.764 --> 01:01:08.004
You need to have some timescale hierarchies there,

01:01:08.084 --> 01:01:15.844
not necessarily separate and expressible in physically separable quantities

01:01:15.844 --> 01:01:19.864
such as region 1 and region 2, but more informational.

01:01:20.064 --> 01:01:26.024
But there must be a timescale hierarchy in order to describe this multiscale behavior.

01:01:26.024 --> 01:01:30.964
But the consequence of that would then be that you also look at a new generation

01:01:30.964 --> 01:01:37.644
of network or whole brain or whole cortex models where each node in a network

01:01:37.644 --> 01:01:41.344
becomes a distinct oscillator, right?

01:01:41.984 --> 01:01:45.084
Coupled through whatever the connectomics has told you.

01:01:45.084 --> 01:01:54.644
Plus a slow variable that is locally connected to this oscillator through the

01:01:54.644 --> 01:01:58.144
connectomics and then this oscillator is coupled through the connectomics as you point out.

01:01:58.144 --> 01:02:04.904
So we look at a co-evolving, a new form of mean field models,

01:02:05.124 --> 01:02:13.004
co-evolving on two timescales, fast and slow, localized in tissue, in space,

01:02:13.344 --> 01:02:21.104
communicating at least through the connectome, maybe in addition to this,

01:02:21.204 --> 01:02:23.144
also at least locally through a glial network.

01:02:23.784 --> 01:02:27.204
That would be a new generation of mean field models. Exactly,

01:02:27.264 --> 01:02:27.624
yeah. Yeah, absolutely.

01:02:28.364 --> 01:02:34.004
And would every node now cycle in a state-dependent fashion through different

01:02:34.004 --> 01:02:36.204
models, through some different oscillatory models?

01:02:37.441 --> 01:02:40.981
Or would it be just one oscillatory model that is just pushed around in its

01:02:40.981 --> 01:02:43.761
state space because of the external perturbations?

01:02:44.161 --> 01:02:49.441
I do not know. I cannot tell you. Because the oscillator you take for the node

01:02:49.441 --> 01:02:55.161
where you have the seizure origin is in some sense a pathological node, right?

01:02:55.221 --> 01:02:58.201
Where you see the DC shift and the low variable and so on.

01:02:58.661 --> 01:03:02.961
This might not necessarily be the right oscillatory model for all the other nodes in the network.

01:03:03.621 --> 01:03:08.081
Yes. Okay. So there is now, that might be interesting. So it might look like

01:03:08.081 --> 01:03:13.621
every node, the dynamics is potentially represented by a certain set of oscillators,

01:03:13.741 --> 01:03:18.301
but now depending on the state of that specific node, I'm either in one or the other.

01:03:19.721 --> 01:03:26.521
The way we have looked at that so far is we looked at topological equivalence

01:03:26.521 --> 01:03:27.461
of the parameter spaces.

01:03:27.461 --> 01:03:34.341
I don't want to make it too abstract, but it's like you can identify certain

01:03:34.341 --> 01:03:38.361
features in the properties of the oscillators.

01:03:38.441 --> 01:03:44.161
And then you know that in the neighborhood of this feature, in this case,

01:03:44.161 --> 01:03:48.861
it's a type of bifurcation, a co-dimension 3 bifurcation, Tarkin's Bogdanov.

01:03:49.601 --> 01:03:54.381
You know, in this neighborhood, all the behavior and all the classes are more or less the same.

01:03:54.381 --> 01:03:58.381
Yeah so it's not multiple oscillators but

01:03:58.381 --> 01:04:01.281
the qualitative behavior of the same

01:04:01.281 --> 01:04:04.421
oscillator changes as you move

01:04:04.421 --> 01:04:08.001
around in this type of neighborhood so it's

01:04:08.001 --> 01:04:10.881
a local statement and if you wish to

01:04:10.881 --> 01:04:13.681
call it a pathological oscillator i'm comfortable with that

01:04:13.681 --> 01:04:17.761
uh uh however i'm

01:04:17.761 --> 01:04:24.401
not comfortable with making statements that are not somehow in the local neighborhoods

01:04:24.401 --> 01:04:31.701
but that jump into completely different behavioral uh repertoire somewhere else

01:04:31.701 --> 01:04:36.021
i simply cannot make any statement about that yeah i do not know,

01:04:37.243 --> 01:04:41.103
But you're free to speculate. I would be. I'm free to speculate.

01:04:41.223 --> 01:04:42.543
You want me to speculate? Of course.

01:04:44.863 --> 01:04:56.903
There is, the way we looked at it is we talk about oscillators and we built

01:04:56.903 --> 01:05:02.363
a generic canonical model around a very characteristic bifurcation point,

01:05:02.443 --> 01:05:05.223
which is a kind of an anchoring point in a parameter space.

01:05:05.223 --> 01:05:10.643
This oscillator has no idea that it's supposed to be pathological or epileptic

01:05:10.643 --> 01:05:16.543
at the end of the day it's a mathematical object that has certain properties

01:05:16.543 --> 01:05:24.863
that we are manipulating and that are based on certain proximity principles.

01:05:26.703 --> 01:05:32.363
There are consequences for its behavior and actually there is just a finite

01:05:32.363 --> 01:05:35.763
number of ways how bifurcation lines can collide.

01:05:35.983 --> 01:05:41.163
And hence there are not so many types of different behaviors that can occur.

01:05:41.383 --> 01:05:46.823
It's almost evident that there is a limited repertoire of behaviors in there.

01:05:47.103 --> 01:05:54.923
So having said this, and they are interdependent, so having said this,

01:05:55.883 --> 01:06:02.183
why, I'm speculating, this line of reasoning does not apply only to epilepsy.

01:06:02.183 --> 01:06:09.683
What's about physiological oscillations? What's about up and down states?

01:06:09.943 --> 01:06:14.763
What's about spindles that occur?

01:06:14.903 --> 01:06:26.583
Can we apply this to physiological objects or physiological oscillations then

01:06:26.583 --> 01:06:29.643
characterize them in a similar way?

01:06:31.041 --> 01:06:34.601
Maybe, maybe, but it's a little less controlled.

01:06:34.641 --> 01:06:39.521
They appear sometimes, they do not. We know they are linked to sleep and wake

01:06:39.521 --> 01:06:44.661
changes, the different sleep stages, there are different behaviors.

01:06:47.441 --> 01:06:54.881
A similar way of thinking should apply to that also and may even,

01:06:54.961 --> 01:06:59.741
and I'm speculating even further, require the need of a slow variable.

01:06:59.741 --> 01:07:05.521
So I would even be comfortable to speculate that we need to generalize this

01:07:05.521 --> 01:07:10.801
way of thinking of a co-evolving,

01:07:10.801 --> 01:07:20.381
at least two-tiered temporal scale mean field model to physiological applications

01:07:20.381 --> 01:07:23.401
and say it should apply actually also to that.

01:07:24.361 --> 01:07:28.641
That's very beautiful. But now look, so we made quite a tour here,

01:07:28.801 --> 01:07:30.361
trying to understand epilepsy.

01:07:31.501 --> 01:07:35.001
The important, the critical step was to move to a network perspective,

01:07:35.261 --> 01:07:38.421
right? To a network science perspective, for medicine perspective on epilepsy.

01:07:38.901 --> 01:07:42.321
And now, of course, then we're thinking about what, at the heart of that stance,

01:07:42.401 --> 01:07:46.261
called computational model grounded approach to get what are these models.

01:07:46.861 --> 01:07:52.641
But now it was up to you. I mean, really think about the full-fledged deployment

01:07:52.641 --> 01:07:54.681
of this way of thinking to the clinic.

01:07:55.321 --> 01:08:00.461
So we have infinite time and infinite resource, now it's there, in the hospital.

01:08:01.941 --> 01:08:08.821
What do we see? How will this translate to a technology really at use in the clinic or at home?

01:08:09.161 --> 01:08:13.441
How do you see this realized in the real world down the line?

01:08:14.361 --> 01:08:17.941
Beyond epilepsy? Well, let's start with epilepsy. Mm-hmm.

01:08:19.500 --> 01:08:32.680
The way I would like to see it is as a decision-making system that in the real world,

01:08:32.700 --> 01:08:36.000
we're talking about a real-world application now that it's a clinician,

01:08:36.260 --> 01:08:43.900
enters in the decision-making process during the patient management conferences.

01:08:43.900 --> 01:08:51.420
They develop confidence upon this software.

01:08:51.640 --> 01:08:58.500
It will probably express itself as a software and feedback their competences

01:08:58.500 --> 01:08:59.940
in order to make it better.

01:08:59.940 --> 01:09:07.720
But at the moment, it's extremely primitive where I would like to see it would

01:09:07.720 --> 01:09:14.080
be as an in silico platform where we can perform testing,

01:09:14.200 --> 01:09:17.000
maybe optimization of procedures.

01:09:17.100 --> 01:09:22.800
But that requires validation that requires confidence that the predictive value

01:09:22.800 --> 01:09:28.580
is sufficiently strong enough there. there we cannot bypass the animal.

01:09:28.880 --> 01:09:33.600
However, we can see it as a means at some point, once we have generated the

01:09:33.600 --> 01:09:40.640
confidence, to bypass animal experiments, because they will be substituted by in silico experiments.

01:09:41.520 --> 01:09:47.980
And then we will have in silico brains of individual human beings where we can

01:09:47.980 --> 01:09:55.080
optimize procedures, maybe rewire, maybe discover new therapeutic interventions, but.

01:09:58.260 --> 01:10:08.700
But we have to lean out of the window, beyond what we are doing right now.

01:10:08.860 --> 01:10:14.820
Like with a lesion, as we talked about earlier, lesioning in an area that is

01:10:14.820 --> 01:10:20.300
not impacted by the impairment.

01:10:21.760 --> 01:10:28.520
So we need to go beyond that. And this, that we need to find good strategies how to do this.

01:10:28.620 --> 01:10:34.740
Otherwise, we will never have trust and faith in an in silico platform for a particular patient.

01:10:35.800 --> 01:10:41.460
So this requires a strategy, confidence, trust building. But especially with the clinicians.

01:10:42.080 --> 01:10:47.240
So you're not going in a direction of having devices that will work directly

01:10:47.240 --> 01:10:51.120
with the patient to help them to control their epilepsy.

01:10:51.120 --> 01:10:56.100
You see it's really more going into the clinic to decision support for the clinicians

01:10:56.100 --> 01:11:01.760
to decide on interventions, right, that might mean non-local interventions in

01:11:01.760 --> 01:11:03.160
a network affected by epilepsy.

01:11:03.500 --> 01:11:07.660
Just as an example, the non-local part, but there may be other ways of dealing with it.

01:11:07.740 --> 01:11:12.480
That would be if I generalize on that, which I have no difficulties in doing.

01:11:14.387 --> 01:11:19.087
Maybe that also means that at some point we do need to carry a model of our

01:11:19.087 --> 01:11:22.467
brain as part of a medical record because we might have a lesion somewhere.

01:11:22.827 --> 01:11:25.987
And then to figure out what could go wrong, I need the reference.

01:11:26.067 --> 01:11:30.767
And the reference is the model of a cat is a cat, preferably the same cat,

01:11:30.867 --> 01:11:31.987
would certainly hold for us.

01:11:32.727 --> 01:11:37.487
So this might be then also an automatic consequence of the approach that you're following now.

01:11:37.707 --> 01:11:40.407
Which would be wonderful, wouldn't it be?

01:11:40.407 --> 01:11:47.647
Then we have a reference brain for you that may have been fingerprinted,

01:11:47.667 --> 01:11:53.567
data fit, where the parameters through a battery of cognitive tasks,

01:11:53.887 --> 01:12:00.807
maybe stimulation paradigms, your brain has been, your virtual brain of you

01:12:00.807 --> 01:12:06.287
yourself has been calibrated and the range of parameters has been identified.

01:12:06.287 --> 01:12:11.927
Then the model parameters themselves become a biomarker for your health.

01:12:12.787 --> 01:12:18.807
Because maybe we could imagine an ongoing online calibration of the parameters

01:12:18.807 --> 01:12:23.687
of your brain model on your favorite app.

01:12:24.607 --> 01:12:33.207
And then as you get tired, depressed, I don't know what, the parameters get out of range, etc.

01:12:33.427 --> 01:12:35.967
You could imagine, now we are dreaming, now we are speculating.

01:12:36.287 --> 01:12:42.207
But yes, a virtual brain model entering in your medical records,

01:12:42.547 --> 01:12:45.887
I think this is not necessarily a bad idea.

01:12:46.227 --> 01:12:52.627
But the cool thing is that automatically and for free, you achieve the post-humanist

01:12:52.627 --> 01:12:55.667
dream of downloading your mind into a computer.

01:12:56.447 --> 01:12:58.247
Who has talked about the mind?

01:12:59.607 --> 01:13:04.927
Well, you know, we think it's isomorphic. But then that mean field model you

01:13:04.927 --> 01:13:06.767
have in my medical record, better be correct.

01:13:06.987 --> 01:13:12.347
The mean field model is a model of a regional activity expression,

01:13:12.627 --> 01:13:14.007
and you know that very well.

01:13:14.647 --> 01:13:18.387
I'm just provoking you. Yeah. But look, so Victor,

01:13:18.567 --> 01:13:26.647
this is really very exciting territory, and you made huge tries in really redefining

01:13:26.647 --> 01:13:30.107
the challenges and making real concrete progress on answering those challenges

01:13:30.107 --> 01:13:32.747
and understanding the brain and also having clinical impact.

01:13:33.427 --> 01:13:39.707
So if we would like to follow in that direction, what would be Victor's law

01:13:39.707 --> 01:13:41.107
that we have to adhere to?

01:13:44.207 --> 01:13:48.487
It would be more specific. If we would follow into this direction,

01:13:48.647 --> 01:13:50.967
which direction? Virtual brain. The science.

01:13:51.927 --> 01:13:55.667
Virtual brain, yes. What's Victor's law that I should follow?

01:13:56.927 --> 01:13:59.287
In order to do? To make progress.

01:14:01.707 --> 01:14:07.767
To achieve the dream you just declared, a virtual brain and every medical record. Mm-hmm.

01:14:12.578 --> 01:14:24.918
I don't know about Victor's law, but the go beyond states, think processes.

01:14:26.158 --> 01:14:31.918
I'm trying to formulate it as a law. But go beyond states, think processes,

01:14:32.178 --> 01:14:34.138
let them be spatial temporal or not.

01:14:34.138 --> 01:14:38.198
But we need a dynamic way of thinking

01:14:38.198 --> 01:14:41.338
or thinking of the dynamic

01:14:41.338 --> 01:14:45.718
brain and i would looking

01:14:45.718 --> 01:14:53.078
back at the history of science this state's approach has hampered much of the

01:14:53.078 --> 01:14:59.458
translational capacity that neuroscience could have had so let's think in terms

01:14:59.458 --> 01:15:03.718
of dynamics and not just jargon but make put it to use,

01:15:03.878 --> 01:15:12.618
and explore the powers that we have in terms of dynamics to ask different types of questions.

01:15:13.358 --> 01:15:18.518
Right. So the last question for me is, so you run a big center there in Marseille.

01:15:18.918 --> 01:15:23.618
You have a lot of machinery and a lot of support, so you can really advance

01:15:23.618 --> 01:15:26.298
the science at quite a pace.

01:15:26.298 --> 01:15:31.358
So four years from now I'm going to go visit you in Marseille to see whether

01:15:31.358 --> 01:15:33.418
you falsified or verified,

01:15:34.258 --> 01:15:39.978
a key prediction that you're going to share with me now so in your program what's

01:15:39.978 --> 01:15:44.278
the most central hypothesis you want to see tested in that four year time frame.

01:15:50.938 --> 01:15:55.058
I would be yeah Yeah.

01:15:55.558 --> 01:16:04.318
The most central hypothesis in a four-year time frame, in the way how we think

01:16:04.318 --> 01:16:07.518
in Marseille about epilepsy,

01:16:07.718 --> 01:16:12.938
and here I'm focusing it on epilepsy, is that we can...

01:16:16.530 --> 01:16:20.510
Use the network description to a

01:16:20.510 --> 01:16:24.030
patient-specific connectomes within the networks that

01:16:24.030 --> 01:16:29.770
it has a real explanatory

01:16:29.770 --> 01:16:33.750
power and we

01:16:33.750 --> 01:16:37.230
can make use of this explanatory power to

01:16:37.230 --> 01:16:43.550
help the patient and improve the outcome

01:16:43.550 --> 01:16:52.210
of the therapy which means if we use patient-specific and not generic connectomes

01:16:52.210 --> 01:16:58.970
to build virtual brains we can make better predictions about what should be

01:16:58.970 --> 01:17:02.210
done in order to help the patient.

01:17:02.470 --> 01:17:08.870
When you come in four years into my lab in Marseille,

01:17:09.070 --> 01:17:17.350
I would like to present you with three to four hundred patient cases where I

01:17:17.350 --> 01:17:21.350
have convincing metrics.

01:17:22.450 --> 01:17:30.450
Convincing statistical significance that demonstrates unambiguously that using

01:17:30.450 --> 01:17:34.730
this patient-specific connectome linked to dynamics,

01:17:35.230 --> 01:17:42.790
dynamic mean fields, has actually improved surgery success or interventional success.

01:17:43.530 --> 01:17:49.250
But wait, let's be specific now. Today's success is 50%, you're saying, right?

01:17:49.930 --> 01:17:54.350
Average across epilepsies, etc. So where's the number going to go, the success rate?

01:17:55.430 --> 01:17:57.610
Do you want to pin me down on a number?

01:17:58.730 --> 01:18:04.330
60%. Improvement by 10%. I'm really shaking this one out of my sleeve.

01:18:04.430 --> 01:18:07.610
You know that very well. Of course, but you know, we need the bet.

01:18:08.770 --> 01:18:14.130
Okay, 10% improvement, 400 patients in four years. Fantastic.

01:18:14.590 --> 01:18:17.910
Victor Jesra, thank you very much for this conversation. Thank you, Paul.

01:18:18.970 --> 01:18:24.870
The CSN podcast was produced by the Convergent Science Network of Biometrics

01:18:24.870 --> 01:18:31.270
and Biohybrid Systems, a project funded by the European Sevens Research Framework Program.

01:18:32.830 --> 01:18:38.150
For more interviews, recorded lectures, or upcoming conferences in the field

01:18:38.150 --> 01:18:44.390
of biometrics and biohybrid systems, go to csnnetwork.eu.

01:18:44.670 --> 01:18:46.570
And thank you for listening.

01:18:47.910 --> 01:18:51.530
Thank you for watching!