 Well, I think we can get started. OK, thank you all for coming. So welcome back to this Integrative Research Seminar. So the name of the seminar series looks fancy. Integrative, the idea is to achieve a better level of integration in the department. Once we achieve that, we'll change the name of the seminar. Actually, talking about names, OK, so we have a talk today. What is our nonlinear time series analysis group on linear? Here are puzzles, OK, about the titles. So you have to wait no longer, OK? So we have a talk by Raf. So just a few words before he gets started. So Raf got his PhD in 2001 in Germany in the physics department working in the area of neurophysics. He then did the post-doxy in Germany, also in this area. And since I think 2005, OK, he came to Barcelona He was the beginning of a post-doxy in Gustavo's group on computational neuroscience. And then he was Ramon Kehal, fellow. And I think 2011 he became a subject professor in the department, OK, so the rest of the story so Raf will talk about it, OK? OK, thank you very much, Hector. Can you hear me? Is that maybe too loud, or? OK, so I will have a feeling that might be too loud, but they will probably take care of that. So I will come back to that title in a minute. But before that, I wanted to thank Hector, Aurelio, Judith, and Vanessa for organizing the seminar series. And for me it's a pleasure to start our second round of the seminar. And I believe that these seminars are a great initiative because we have all different backgrounds and the research done at this department is highly interdisciplinary. And then we have undergraduate students in different engineering degrees who will go to many different directions after their studies with us. And last but not least, we also receive postgraduate students and postdocs from all over the world and again with a plentitude of different backgrounds. And to integrate all this is, of course, quite a challenge. And I believe that these integrative research seminars, so I do like the title of these seminars, are first a very important step in this direction. And when we look at the format of the session, so the text written here, we find that we should present our research in a way that is accessible to colleagues, students, and researchers from other areas. And that truly took that as a guiding principle when preparing this presentation. And certainly there are different ways in which you can address a general audience, like we already have seen the first round of these seminars before the summer. And the particular approach that I will use is the following. So I will spend quite some time explaining one straightforward example. And I selected this particular example because I believe it allows me to explain some of the basic principles of our work. And then only in the last 15 minutes or so I will give you a broader overview about our work and a little bit of context. So the point which I'm also trying to make is please be not disappointed when I exclude very advanced and very specialized topics. OK, so when we are online and click on this link, we find the different research groups that we have in our department working in these different areas of research. And among them, you also find the nonlinear signal analysis group down here. And in fact, the correct name of our group is nonlinear time series analysis group. But that's OK because the terms time series and signals are equivalent. And now that brings me back to my title because the question is, why is our nonlinear time series analysis group not linear? So what does this term nonlinear refer to? Does it refer to the signals that we analyze? Well, in the strict sense, any signal that is not a line is not linear. So that is not what this term is referring to. So does it refer to the analysis tools? Are the equations that we use for our analysis not linear? Well, that does not mean either. In fact, many of the equations we use are linear in the sense that we just take averages over some local quantities. So that is also not meant. So let me show you the solution. Nonlinear time series analysis is a short form of saying analysis developed for the characterization of signals measured from nonlinear dynamics. And since this long expression didn't fit up here, we stayed with the short version. Now let's put this expression up here. And now you might, of course, say, well, then you have to tell me what is the nonlinear dynamics. But that is easy because for linear dynamics, the strength of an effect is proportional to the strength of the course. Whereas for nonlinear dynamics, strength, course and effect are related in some nonlinear way, also typically leading to a more complicated temporal evolution of these dynamics. And now let me ask you the dynamics you're looking at in this department. Are these linear dynamics or are these nonlinear dynamics? So maybe people working with the heart, would you say the heart is the linear dynamics or would you say that the heart is nonlinear? Nonlinear? People from music technology, where would you put your systems? Also on the nonlinear side. Very good. When you asked me where to put the brain, the brain is certainly also not a linear dynamics. So when you think about how neurons work, you have to push them above a certain threshold before they fire. And then they would saturate upon additional input. So there you have certainly a nonlinear relation between input and output and on all different scales of neuronal organization. The brain has many nonlinearities. That is why a priori, it seems like a promising approach to apply nonlinear signal analysis to neuronal dynamics or other dynamics that we study in this department. And the specific example, which I will use as the following, so don't be now shocked, I will use a surrogate corrected nonlinear prediction error to localize the epileptic focus from seizure-free electroencephalographic recordings in epilepsy patients. At first, that looks terribly complicated, but if we take some time and dedicate some time to each of these items, this will turn out to be quite straightforward, hopefully. And to avoid any overlap with the research seminar I gave some time ago, I at first considered replacing this example with a more elaborated example, which would be even more closely related to our most recent research. But in the end I decided to return to this straightforward example, because that is the one example which allows me best to explain the principles of our work. And thereby, it would also be the example which allows me best to address a general audience, which should be the format of this session. Okay, so now I will go step by step through these different items. And let me begin with by asking how can we distinguish between deterministic and stochastic dynamics? Let's begin with deterministic dynamics. If we know the present state of a deterministic dynamics, we can perfectly predict the future evolution. For example, if we plan an excursion in France, we can look up the exact time of sunrise and sunset on any given day, for example, in Toulouse. And that is because the solar system is deterministic. We can perfectly predict it. However, we cannot predict with certainty if we will actually see the sun, or whether it would be covered by clouds, or whether it will rain. And that is because on the time scale of just a couple of days, the weather has a strong stochastic component. And therefore, even if we know the present state of a stochastic dynamics, we cannot predict the future evolution and stochastic dynamics shown erratic and random temporal evolution, whereas deterministic dynamics show regular non-random evolution. Now, we do signal analysis. So how can we conclude from signals whether the underlying dynamics is deterministic or stochastic? So let's look at these examples, and let me ask you, for which of these examples do you believe I took a measurement from a regular deterministic dynamics? From this one, from the red one? Yes, I do agree, but what feature do you use to make that distinction? What would we, what do we look at when we say, okay, that is regular? There's a pattern, and this pattern would repeat, right? And if you, that's the very feature, because if you think about it, we see these patterns and these patterns repeat. And that means that if we know the present state, we can predict the future evolution. If we have similar present states, they will have similar future evolutions. That's the very notion of a deterministic dynamics. And so now I can tell you that T indeed I used so-called Lorentz dynamics. So I used a deterministic differential equation to generate this signal. Now, how about the other three signals? Would you see any structure in the remaining signals? That's, it's much less clear. Maybe one could try to find some structure, but let's agree that we are not certain about these signals. So let's try some linear signal analysis and let's look at the power spectra. So these are still my same signals and these are now the power spectra or in more technical terms, the pyridograms which show you the energy of these signals in the different frequencies. And I suppose you only see the power spectra, so you don't see the signals. What can we conclude? Well, if we look closely, we see in fact that the power spectra of this signal, which we already identified as deterministic, is identical to that one. And also these are very, very similar. And so using the power spectrum, we wouldn't make any conclusive distinction. But don't get me wrong. I'm not saying that the power spectrum is not useful in general. What I'm saying is that here for the problem at hand which is distinguishing between deterministic and stochastic dynamics, the power spectrum does not help. But now we have these nonlinear techniques. So let's look at one of these nonlinear techniques. And the first thing we do in most nonlinear signal analysis techniques is to use delay coordinates to get a reconstruction of the underlying dynamics. So we embed the signal in some higher dimensional space. And for this purpose, we put the signal to the first axis, the signal shifted by a time delay to the second axis, shifted by twice this time delay to the third axis and so on. So here for this plot, I can only plot three dimensions, but for the actual computation, I use much higher dimensional spaces. And now when we look at our deterministic signal, we see that it has a very nice structure. We see that locally, these trajectories are aligned. They all point to the same direction. And if you think about it, that again means that these patterns repeat. That means that if we have similar present states, they will have a similar future evolution. So we have local alignment. For the other three signals, we are not so certain. So maybe here in the black signal, you do see some degree of local alignment. Now, local alignment, that's the bridge to the next slide because that is what is quantified by this nonlinear prediction error, which I show here for the four signals identified by their colors, independence on the delay time that I use for the reconstruction. And so it quantifies the degree to which these trajectories are aligned. So it's a prediction error. So if you have a high predictability, you have a low value and vice versa. And we see as to be expected that our deterministic signal clearly stands out. It has the highest predictability. And that is already more than what we got before by looking at the power spectra. So this you have seen before, just to remind you that here we couldn't really distinguish this signal from the black one. So but when it comes to the other three signals, we don't really get much more information. So because here the results overlap, so we cannot see if there's any structure in one of these remaining signals. So we got one step further, but we haven't reached our final goal yet. But fortunately, there's a solution to this problem, namely the concept of surrogates. Surrogates are very powerful and useful tool in nonlinear signal analysis. They allow us to test different null hypothesis about the dynamics underlying some experimental signal. And here we'll introduce this concept by just in a talk way by using one particular example for these surrogates. So the problem is we need a baseline for this nonlinear prediction error. The question is what would we expect by chance? What would we expect if these signals were just random noise? And the answer is it depends because these signals have different degrees of autocorrelations. They have different degrees of smoothness. And the high autocorrelation leads to a high short-term predictability. However, and the high short-term predictability would lead to a low value of the nonlinear prediction error. Now the problem is that a high autocorrelation does not mean deterministic. Stochastic systems can have whatever degree of autocorrelation. So what we need to do is that we need to estimate the value which we would expect for our signals supposing that these are random signals given their autocorrelation. And this can be done using surrogates. And what you do here is that you start with the original signals and then you use a constrained randomization to generate surrogates from your signals. And you constrain your randomization in such a way that these surrogates have the same autocorrelation as their corresponding original signal. Now apart from this constraint, the surrogates are completely random. So if there was any structure in the original dynamics, this will be destroyed by the randomization. And that corresponds to the null hypothesis that my signals are just a linear stochastic stationary correlated process. And very importantly, I do not assume that this is just white noise. Quite in the contrary, like I just said, I explicitly include the linear autocorrelation into my null hypothesis. And in particular, each system has the autocorrelation of its original signal. So each surrogate can provide the baseline for its own original signal. And so what we now still need to do is to run the nonlinear prediction error not only for the originals, but also for the surrogates and results are shown here. So this plot, these four plots are the same like the big plot which I had before. Now simply I have one for each original signal. And so here you have the nonlinear prediction error independent on this delay time. And you have values for the original signals and these gray bars correspond to the range obtained for a set of surrogate signals. Now if you see that the surrogates match the originals, then apparently there was nothing beyond the linear autocorrelation leading to this predictability. Then putting the autocorrelation into these surrogate signals was sufficient to reproduce these results. Then you cannot reject the null hypothesis that this is just noise. If however you find the mismatch between the original and the surrogates, then yes, the autocorrelation was not sufficient to explain the predictability of your signals. And now looking at our concrete examples, we see here for the signal which we already identified as deterministic, we get a very, very clear mismatch between our original and the surrogates. However, here for the black and the blue signal, we find that the surrogates perfectly match our original. And now I can tell you that here indeed what I used was noise. So the blue signal is just an autoregressive signal and the black signal is a stochastic signal which I constructed to have the same power spectrum like the Lorentz dynamics. But they're completely stochastic. And now comes the most important point because here for the magenta signal, we again see mismatch, although not very strong, between the original and the surrogates. And now I can tell you indeed this signal has deterministic components. What I use was a Lorentz signal and I put a lot of noise onto it. So this means that once I combine my nonlinear prediction error with the surrogates, so when I focus on the nonlinear properties of my dynamics, when I correct for the linear properties of my dynamics, then I can make the distinction between deterministic dynamics, possibly superimposed with some noise on the one hand and stochastic dynamics on the other hand. And please note again that using only the nonlinear prediction error, this would not be possible because if you look closely, here the blue signal has actually lower values than the magenta signal. So here you are below 0.8 and here you are above. So under the scope of the nonlinear prediction error, the blue signal would win. But the surrogates allow us to reveal, okay, that was simply due to the autocollation of the blue signal. And just one last reminder, again, so when we were trying to do this with the power spectra, this was not possible. But it becomes possible once we combine the nonlinear prediction error with surrogates. Okay, so time for a turn. Now we know what is a nonlinear prediction error and we just motivated the importance of the surrogate correction. And now I should give you some information about epilepsy and why we measure the electron cephalogram in epilepsy patients. And to begin with, oftentimes, we would define a disease by the main symptom. So let's look at the definition of an epileptic seizure, which is a transient occurrence of signs and or symptoms due to abnormal excessive or synchronous neuronal activity in the brain. And I will come back to this definition in just a couple of minutes. And these seizures are more common than one might believe because some 5% of the population will have a seizure at some point in their life. Importantly, having a single seizure does not lead to the diagnosis of having epilepsy. Epilepsy is defined by a chronic condition of the brain leading to the chronic occurrence of epileptic seizures. And in the most general sense, one can say that the cause of an epileptic seizure is an irritation of the brain. And such an irritation can, for example, be an acute head injury leading to a single seizure or it could be some permanent disturbance, some permanent brain damage leading to the chronic occurrence of seizures. And our brains are very diverse. So, for example, we have cognitive processes, so we can use our brain to solve very, very complex problems. But also our brain is controlling the body, so it allows us to make the most complex movements. And finally, also our brain allows us to communicate and to coordinate with others leading to yet another level of complexity. And as you all know, our brains are organized such that different functions are organized by different brain regions. Therefore, the exact symptoms of an epileptic seizure will depend on where in the brain the seizures start and which brain area later gets involved into the seizure. And the main symptoms are listed here, but it's not always such that the patient would fall to the ground and would show some very extreme symptoms like violent movements and severe dysfunctions of the autonomic nervous system, but rather these symptoms can be very subtle and they can go along with very subtle symptoms that might go even unnoticed by patients, by persons next to the patient. And now, in order to give you an idea, I would like to show you a video recording of an epileptic seizure, but don't be afraid. So this will be an absolutely non-shocking case. I only have to indicate to our technician not to record this because I do have permission to show you this video in a scientific context, but I don't have permission to put it into the public domain. So that's why we shouldn't include it in the recording of this session. So let's first look only at the video. And now, this patient will have a seizure in just a short while. If you believe the seizure started, then raise your hand. If you believe, okay, the seizure is ongoing, then raise your hand and say, I believe the seizure already started. Okay, so, actually, I see some hands. Okay, very good because the seizure actually already started. And what will happen now is that the medical doctor who was talking to this lady in the off, he will now approach the patient and will talk to the patient and will test the patient. He will ask her to raise her hand. He will ask her what day of the week we have, et cetera. And that is because the symptoms that they can now observe are very, very valuable information for the diagnostics of the patient because the type of symptoms allow you to draw conclusions about what brain areas affected, what type of seizures this was. So now he will make an effort to get as much information as possible from this seizure event. And so, they were talking in German and he will ask her for the day of the week and that will be the only word which she will understand. So please listen carefully. Give me your right hand. Give me your other hand. The right hand. Can you give me the right hand? The right hand? Which day is the week? Which day is the week? Can you give me my hand again? One, two, three. Good. Give me your right hand. Left hand? Which day is the week? Can you give me your right hand? Friday. Okay, great. Did you see anything? I don't know. Did you notice anything? I don't know. Okay, so she said there was something strange and there was another lady sitting here. She was taking her breakfast and if the doctor wouldn't have been there probably the other patient would not even have noticed and her colleague just had a seizure and they were talking in German but did you understand when she was saying what day of the week it was? She was saying it in French and in Switzerland it's also bilingual or a country where many languages are spoken and so she replied in French, and only later she completely recovered and said Freitag, which is our word for Friday. Okay, so now let me close this one and let me go... Now you can again record everything. Okay, very good. This was only one example and epileptic seizures are very, very diverse phenomenon and let's only distinguish two main types, namely focal seizures, which have a localized origin, the so-called epileptic focus that would always trigger the epileptic seizure and that is located in different parts of the brain for different patients and the counterpart would be generalized seizures where you do not have a localized origin but rather a simultaneous onset in both hemispheres and the data we are using is mostly from patients with these focal seizures and to underline a little more the relevance of this disease let me give you some numbers since we have approximately or we have at least 50 million patients worldwide and in a typical European population between 0.5 and 1% have active epilepsy meaning that in Spain we have about 400,000 patients and the total cost of epilepsy amounts to some 5% of the total health budget so that's quite a big part of the health budget and the good news is that for some 60, 70% of the patients the seizures can be stopped completely by anti-seizure drugs so they are fine, they take drugs and they don't suffer any seizures but also means that for the remaining patients they still have seizures despite an optimal medication and for these patients with so-called medically refractory seizures epilepsy surgery is an option and what you do in epilepsy surgery you basically aim to cut out the epileptic focus from the brain so that this epileptic focus can no longer trigger these seizures and it seems difficult to estimate the percentage of candidates among these patients because estimates range from 10% to 50% and when we look at the actual number of patients in which surgery is performed this study from the United Kingdom this number is actually still below this 10% and the ideal outcome of epilepsy surgery is that the patient is completely seizure-free, would have no severe deficits and also could actually stop taking these anti-seizure drugs could stop taking the medication now in order to plan the surgery for an individual patient of course we have to localize the epileptic focus as precise as possible and apart from new imaging techniques the most important diagnostic tool in this context is still the recording of the electrical activity of the brain by means of an electroencephalogram and what you always do in the beginning is that you measure the electroencephalograms the EEG from the surface of the scalp with a non-invasive surface EEG however for some patients this non-invasive technique does not allow you to localize the epileptic focus with the necessary precision and certainty and in these patients the implantation of such intracranial electrodes is medically indicated and so here we see a schematic drawing where we have a grid electrode with 8 by 8 contacts so you would then have 64 channels and these electrodes are used when you suspect the focus to be located somewhere in the covered brain region but you don't know where exactly and here's a picture of these grid electrodes where that illustrates that they're not stiff they are flexible so that they can smoothly cover the curved surface of the brain and you also see that they come with different numbers of contacts depending on what you need for the individual patient and if the focus is suspected in some more interior brain region you have these needle-like depth electrodes and often you would also use a combination of these electrodes and always the implantation scheme is tailored to the individual patient so the medical doctors make a lot of pre-diagnostics and then they already have quite a good idea where the focus is most probably located and that's where they put the electrodes to find the exact spot where to tell the neurosurgeon okay, this brain region must be cut out okay, now during a period of like two weeks these electrodes remain implanted and are used for permanent recording of the EEG and of course during this period the patient has to stay in the hospital and that is what was done with the ladies that we saw in the video and this is not yet what I wanted to show you okay, this is... because now we will return to the same patient and we will again watch this video but now in combination with the EEG now in combination with the information that the medical doctors have so here you see in the small screen up there you see the video the recording of the EEG hello that's the very start of the seizure so here you see the resolution is not very good but if you see it on a screen used by medical doctors you can very nicely see that here you have a very rhythmic oscillatory activity at each individual channel and in the definition of an epileptic seizure they use the two words of activity that is what they mean by these terms in the definition of epileptic seizures and now you see that the seizure starts and will propagate to further reach this is the very point when the medical doctors said okay that is the end of the seizure so now the seizure has no clear sign any longer in her brain but it takes some time for her to recover completely and this post-ictor period can last many tens of minutes in this case it will last only a couple of seconds she will be back okay so what was done in this patient they could localize the epileptic focus she underwent epilepsy surgery and she is completely seizure free ever since so she was successfully cured with this scheme which I just described and so what we just saw is the most important information for the medical doctors a recording of an epileptic seizure because if they find that these seizures always start in the same region of the brain then this region can be identified as a target region for the neurosurgeon and this region can then be resected provided that it is in a brain area where you can actually operate seizure freedom and that was achieved for this patient and many more patients so the rate of success in this surgery is in fact quite high so now I can again give a signal to our technician that we can again record everything okay now we at first learned about the nonlinear prediction error and the importance of surrogate correction and now we learned about epilepsy why we would record the electroencephalogram in epilepsy patients and now remember the specific question which I wanted to ask is whether we can localize this epileptic focus only looking at the seizure free recording so for the medical doctors the seizure is the most important information and therefore we try to see can we localize it using only seizure free recordings so we deliberately do not look at such recordings of seizure activity but we only look at windows where we have the seizure free interval of the patient and now what we do if we have such a window we analyze each channel individually so that we have one result for each individual channel and then we move our window forward in time because the recording is much longer than these individual windows and then we have such a result profile where we here have time and here we have the channel so each column corresponds to the result of one time window and each row corresponds to the results of one individual channel and for this particular patient the focus was contained in the right hemisphere and here we look at a measure from linear signal analysis the so called delta energy which is simply the energies in the slow frequencies up to four hertz and we do see some variability across time and also with regard to the different recording channels and maybe also some asymmetry with regard to the hemispheres now let's look at our nonlinear prediction error which is shown here for the same EEG recording and again we see some variability with regard to time and with regard to the different recording locations and now once we apply the surrogate correction once we focus on the nonlinear properties of our dynamics by correcting for the linear articulation we get a very high contrast between the channels here in the focal hemisphere as compared to the channels in the non-focal hemisphere and let me remind you of what is the difference between these two last approaches so let's go back to the previous result and let's put this icon on top of it to remind you of the first block so here we look at the raw values of the nonlinear prediction error and please remember that when we look at the raw values of the nonlinear prediction error this did not allow us to make a clear distinction between linear deterministic dynamics possibly superimposed with some noise versus stochastic dynamics what we do here when we use the surrogate correction is that we do not look at the raw values but we only look at the difference between the original results and the surrogates which would be high here and which would be zero here and for our models this was the key step to arrive at the clear distinction between nonlinear deterministic and stochastic dynamics and returning to the EG this also turns out to be the most important step in arriving at this clear contrast between the focal and non-focal hemisphere and now to see if this is only one example to see if this is some general finding we studied a group of 29 patients which has the advantage of being a very well-defined patient group because all patients were implanted with such a pair of so-called intra-hippocampal depth electrodes in all cases leading to the diagnosis of a unilateral medial temporal lobe epilepsy and in all patients one of the two hippocampal formations was resected leading to complete seizure freedom in all cases and in total we analyzed 66 hours of the seizure free interval of these patients and results are summarized here where we have the mean values taking over time separately for the focal and for the non-focal hemisphere for our 29 patients and we look at the same measures like we had before the linear delta energy the non-linear prediction error and the non-linear prediction error with the surrogate correction and you see that here we get the clearest contrast between results from the focal and the non-focal hemisphere the contrast is much higher than here because here in 27 out of 29 cases we get higher values for the focal hemisphere as compared to the non-focal hemisphere and that finally allows me to answer the question which I asked in the beginning why is our non-linear time series group not linear that's why that's why because we really have to focus on the non-linear properties of our signals by correcting for the linear articulation to arrive at the advanced characterization of this dynamics that is the key in the model systems and that also holds the key when we look at dynamics such as recordings from the brain of epilepsy patients and so that is the bridge to this title because we by focusing on non-linearity we really can arrive at an approved characterization of the epileptic brain and this study line of trying to localize the epileptic focus using non-linear signal analysis that's a continued study line of mine which I pursue since a couple of years or since many years and results are published here from a more clinical point of view and here from a more physics point of view and these results really can have a clinical application and we are not alone in this research line because recently we contributed our results to this multi-center study which will be published next week in November. Again here we have a team of people working in different quantitative EEG analysis approaches the first four authors and then we have a medical team so medical doctors working different epilepsy centers in Switzerland, in Italy and in France and our common goal is to arrive from the current retrospective analysis of our data to some prospective application in a clinical setting. So now this was the example which I announced in the beginning so now we saw that how we can use these techniques to successfully localize the epileptic focus from seizure free recordings in epilepsy patients and as you saw I took quite some time to go through this example and therefore now in the end only briefly I would like to give you a short overview and a little bit of context about our work and the first thing I would like to clarify is where do I get oops not so fast where do I get my the pointers right next to the button with which you forward where do I get my data from so this type of data is only recorded in specialized centers like this epilepsy clinic here in Bonn and very importantly this is only done for an optimal diagnostics for each individual patient so this is completely independent from our retrospective of the data of the data and I'm never involved in the measurement of the data so that's all done by medical doctors for an optimal diagnostics for each individual patient and to see how I got connected to this data we have to go all the way back to my PhD which was at that time based on the cooperation between this department of epileptology and the physics institute at the University of Bonn in Germany so that's how I got connected to this type of recording in the first place and I kept close contact with this department ever since and in recent years we also established very active cooperation with Kasper Schindler and Christian Rummel from this epilepsy center in Bern in Switzerland and also locally we want to further strengthen the cooperation that we have with the newly established epilepsy unit here in Barcelona in the Hospital del Mar and actually Adrià I saw that he came in now I cannot see him there are some people who already have a quite successful cooperation with Rodrigo who is the head of this epilepsy center here in Barcelona and now so from my personal experience having data available was never a limitation quite in the contrary there are many clinicians such as Kasper Schindler, Christian Delga, Rodrigo Rokamora and the medical doctors from this multi-center study I showed you before that are happy to share their data because they share they also believe in the potential of this analysis and we share the common goals the common vision that these results can at some point be exploited in the clinical setting so that these results can be used for the benefit of the diagnostics of the patient and or for the treatment of the patient okay so before we looked at the localization of the epileptic focus and another aspect was covered in this paper which is currently in press where we looked if we can predict retrospectively the surgical success for individual patients based on our analysis and again this team is composed of people with a physics background like Christian and myself and medical doctors working in neuroimaging neurophysiology, neurosurgery and neurology in in general now we also looked at the predictability of epileptic seizures and as you can guess from these titles here the results were much less promising because if you ask me if seizure prediction works any better than chance the answer is no it does not work and seizure prediction certainly was a very very hot topic in epilepsy research and that is because if one could predict epileptic seizures this would have a huge impact that would have a huge impact on the quality of life of millions of patients and this would be a billion dollar market no doubt the problem is it does not work seizures seem to be as unpredictable as earthquakes and if you want to we can discuss this later on in more detail now these clinical or these recordings in epilepsy patients are very valuable from a clinical point of view and from a research point of view and having direct access to these recordings is a very good or very high privilege but I believe in order to advance our field as a whole just like many initiatives from the music technology group we share the data and the source code we use in our studies with the communities and this is the so-called Bern-Bass-Lohner EEG database which we published in 2012 and already by now there are some papers by others which studied this database and approaching the end so when you have these recordings in epilepsy it allows you to study epilepsy but it also provides you with a unique window to the human brain and thereby it allows you to study questions from basic neuroscience and from cognitive neuroscience however these cognitive aspects were certainly never a main target of my own work. What was the main target and will always be is to develop new signal analysis techniques because recordings from the brain often come in modalities where we have no technique available for example you sometimes measure the response of the brain to some stimulus and then what you do is that you measure the brain activity time trigger to repeat it presentations of the stimulus and here we came up with a special technique to detect couplings in this event related settings now returning to the time continuous case you sometimes have hybrid measurements where you measure an amplitude variable such as EEG and at the same time you measure the spiking activity of individual neurons and we came up with the technique again with regard to the couplings to address this mixed scenario and also work done together with UPF students is often going in the direction of developing and improving techniques from non-linear signal analysis so for undergraduate students if you're interested in this type of work please feel free to contact us you will be very welcome to join our team now when it comes to the application recently we focused very much on this point process and that is because I believe that they will become particularly relevant because nowadays these intracranial electrodes are very often extended by little micro wires which would allow you to measure the spiking activity of individual neurons and that is of course yet a completely different dimension in these recordings and the department of epileptology is one of the leading centers in this in this technology and we already developed some new techniques to study the data and analyze some small data sets and currently we seek to analyze larger data sets and large data sets is the bridge to the next thing because our analysis can be extremely time-consuming so we have these non-linear measures which are very time-consuming we have many channels from very long recordings from many patients and many parameters to test and so this analysis can literally require like hundreds of days of computation time and the only way to solve this is using distributed computing and as you know our department offers a very good infrastructure in this context namely our high performance computing cluster but even on this scale the workload of our analysis can be very high since if you look at the workload of our cluster in 2014 you see these two peaks and I declare myself together with my previous student Daniel Guilty of these two peaks but they also led to two very nice studies so my point is here that this is really a very useful infrastructure that is provided by our department here I started running some simulations for our new European project in which I worked together with Ernest Montbriot and this European project has a very modest name it's called COSMOS and COSMOS stands for complex oscillatory systems modeling and analysis so we do signal analysis but in order to validate new signal analysis techniques we use models so in order to see if our techniques work fine we at first test them on models but our models are very simple and people that seriously work in dynamical systems such as Ernest Montbriot they study models of much more complex systems and one of the main aims of this European project is to look for synergies between the analysis and the modeling of such complex systems because these systems can be very very challenging and very important test for my techniques and the results from my analysis can help to characterize these models and can help to to develop new models and now really coming to the end so this will be my last slide so you might say well I have all your papers almost all of them are published with people from abroad so how can your work contribute to the integration within our department because that's what is on our agenda right and I do believe that we do that we can contribute here and as an example I would like to show you two papers done together with our colleagues Joan and Xavier from the music technology group which also involved a research study of Joan at the Max Planck Institute in Germany at the group of Holger Kahn and these are just two examples with which I would like to say it doesn't always have to be the EG of epilepsy patients I believe that nonlinear signal analysis can be quite useful for the study of many different types of experimental signals hopefully including signals that you're looking at and so the message would be that we are very open to look at other type of data that we might have in our department and with that I believe I took already too much time so I would like to thank you for your attention and these institutions for funding thank you very much thanks Raph so now we have time for questions, comments and so you can probably it's very interesting the result that with this oh sorry hello it's very interesting the result that with this analysis you can tell apart between between the seizure when that happens I mean you cannot predict it but you can visualize it the question that I have is why do you think that the brain behaves more nonlinearly during a seizure as compared as to when it's behaving normally because the brain is so complex system that presumably I will have pursuing the other way around that when it's normally using all the machinery to perform computations you have the seizure maybe it becomes more linear so my prediction will have been the other way around that's true so we have two aspects we have on one hand the seizure I have a feeling that I speak very loud now we have on one hand the seizure versus the seizure free period and we also have the seizure generating area versus the healthy brain area during the seizure free period let's say that both of these are related to the epileptic process and what would happen in the epileptic process is that in the definition of the seizure we had the word hypersynchronous so there's the hypersynchronous activity of neurons which you would typically not have during in a normal brain area when there's no seizure and this hypersynchronous activity can somehow lead to reduction of the complexity of your overall dynamics so that it somehow the deterministic structure that is already there is de-masked and you can better see it so for example you mentioned that we go to Camp Nou on Saturday and you put a microphone just in the middle of Camp Nou and everyone would lead to deterministic communication so they would talk to each other and it makes sense what they say but if you put a microphone you would just hear noise but then when they start to sing for example when a goal is scored or when they would support their team then they sing in synchrony and then all of a sudden despite that you have 90,000 spectators you would see there is some non-randomness so by introducing synchrony you can reduce the overall complexity and then you can see there is something non-random you would hear more than noise and in a similar way when we look at at at recordings I have to go all the way back of the seizure free interval you would also sometimes see outside of the seizure already activity that seems like coming from some synchronous activity of neurons and this would also be related to this this type of dynamics so here the seizure didn't start yet and that could perfectly be in the middle of a seizure free period so and no seizure would take place afterwards but already here you see some synchronous activity which you would not suspect to find in our brains yes please couple questions so first so you said it's important to find the site of the seizure yes so what do you do once you find it so where is it important so if the medical doctors so first of all once I said that we first and foremost look at data from patients with focal seizures so that means that these are patients where indeed there's one well defined brain region where the seizure starts that is good because if you have epilepsy and the seizures always start instantaneously in the entire brain then this is more difficult to treat if you find this one brain region where the seizures would always start and if you can then remove this brain region then this patient can become seizure free and you cannot always remove this brain region because it might overlap with some functional area so if it's right in the middle of your speech area you probably don't want to operate on this area and also they were in the history some big mistakes when they for example resected both in the first patients they resected both hippocampal formations the patient had no seizures ever since but also the patient could not form any new memory so you really have to make sure that you can operate in this very spot but if you arrive to resect the area which triggers the seizures the patient will have no seizures and that is the aim of this type of surgery the second unrelated question so the area of non-linear temporal prediction so I guess these days you hear a lot about machine learning techniques, deep learning these and that and so on and of course I tend to think that many people have applied artificial neural networks okay for doing predictions of these sorts of non-linear processes I mean it's that technique that is used in the area yes it's used and it's also used people try to use this on different levels for the problem of seizure prediction so they apply it to the signals themselves and then you can also apply it in a sense so if we have let's suppose that here a seizure would start right so you have your raw signal but then you have the feature which you extract from your signal and then people would use tools from machine learning to extract the feature which would allow me to make that prediction so not analyzing the raw signals but the features extracted from these signals and they would try this using machine learning techniques to see if that would allow to predict the epileptic seizure but to the best of my knowledge also this was not successful yes I can hear you well okay expert analysis are based on the assumption of the stationarity of the signals so of course these methods don't work well and easy signal especially with epilepsy so I don't know if you have experience about that we apply time frequency analysis because we can also detect dynamic in this signal especially with the epileptic signal I don't know if this kind of information can be used like patterns and apply for some prediction model and also probably we can obtain good results I personally never tried time frequency analysis myself people apply it to EEG recordings a lot also with regard to the problems that I mentioned and you said that it's based on the assumption of stationarity that's very correct and the same holds true for any of the methods that we use so always we assume that the window would be stationary over time something which is not always fulfilled and one way of dealing with that is time frequency analysis that's very true but I personally never tried this technique yes I just want to ask the localizing the starting point of this issue as you mentioned is very important especially if you are going to operate on it and why not but probably the cost is an issue but fMRI has much more spatial resolution less temporal resolution has there been work on trying to apply similar techniques but with signals or series of activations detected by fMRI? yes so in general new imaging techniques fMI and further new imaging techniques are becoming more and more important and in more and more cases the implantation of such electrodes is not needed any longer which is very good because this is of course somewhat risky undertaking to put these electrodes into a brain and also you have to take into account you will only cut out one of these two brain structures the other brain structure should stay there but you put a needle in it so if you can avoid that you would avoid it and new imaging techniques are increasingly important because the quality of these images becomes better and better so for example patients with this type of epilepsy often they have some sclerotic change so that neurons would die and this would change the tissue in the 90s you couldn't see it on an MI scan now you can and so really like I said apart from new imaging EG is very important but the field of new imaging is getting bigger and bigger and which is very good for the patient population as a whole more doubts ok so I think that we are done so thanks for coming and bon profit