 Hello everybody, thanks for the invitation to speak here, the background's been set up perfectly for me by Yeejang to talk about epilepsy and by my time to talk about connectomics. So I'll skip most of this except just to say that hopefully you're convinced by now that epilepsy is a real societal problem. There's a lot of patience with epilepsy and it's a major economic burden amongst other reasons why it's a bad thing. So focal seizures, a particular type of seizures, focal epilepsy is a particular type of epilepsy and patients with focal epilepsy have seizures which are essentially mainly constrained to just a part of the brain, just an epileptic focus and the longstanding thought was that this is caused by focal abnormalities to the brain tissue and in many cases this is the truth so it can be caused by a brain tumor for example just a focal part of the brain, abnormal tissue and this can lead to focal seizures that are spatially constrained in their onset. However recent literature has suggested that it's actually a bit more complicated than this and that epilepsy is an emergent phenomenon of abnormal networks which are more widespread than just the observed seizure focus which is spatially constrained and so there's been lots of literature lately saying that these abnormal networks can be what's leading to the seizures. So in this talk I'm going to talk about structural networks and in particular structural networks inferred by diffusion tensor imaging. So quick summary of what I mean by this. So what we can do is we can infer network nodes using a passillation scheme here and that can be obtained from passillation of the structural T1 image and we can also infer connections between those nodes using diffusion tractography for a diffusion image. These two can be registered to the same space and then connections between one node and another can be inferred and we can generate a connection matrix something like this where we have nodes and connections between them and these can be represented in a matrix. One thing you'll notice is that not all nodes are the same so here's the superior frontal gyrus so this is a very large node and at the front here you've got a temporal pole which is one of the smallest nodes and so to make fair comparisons it's not quite so easy. So we have different nodal properties and we have also different edge properties so connections so there's different ways in which we can define the edge in the network or the connection network. This can be done by a connection weight which is kind of normalising between the density of streamlines which takes into account the node size and the number of streamlines connecting or we can just use the number of streamlines which is done in a lot of studies or the length of the streamlines. There's any number of ways that you can use to define your edge strength if you like and what a lot of studies do is they take this weighted connection matrix they set a threshold and then binarise it but all of these lie around the assumption that edge weight or strength or a weighted measure is representative of what's underneath it what's going on and as you saw in the previous slide there's many studies have shown that there's alterations in these edge strengths or these connectivities between patients and controls and my question here that I'm trying to investigate is what's driving this is it alterations in the gray matter node size or is it alterations in this white matter connectivity or is it a combination of both so just as I've said that the all nodes are not equal somewhere bigger than others all nodes that are the the same node in different subjects can also be different so the same node can be bigger in one subject than another and this can also influence the connectivity so I'm interested in finding out what's driving these network changes that have been found in the literature so we begin by just looking at the the node size so in this case we've looked at the surface area of each of the 82 nodes in this case picked a widely used passillation scheme that desiccant kill any free surfer a passillation scheme and we've looked at 22 patients with temporal lobe epilepsy or focal epilepsy taught to be constrained to the left temporal lobe and 39 age and gender matched controls and we essentially just plot the t statistics or the afar apart these two distributions are for each of the different regions and we see widespread atrophy in pretty much the whole brain and this is because patients have smaller brains than controls in terms of the total surface area this has been reported before in literature this is not particularly new this is we've known this since probably the late 80s when early MRI studies were done in this but it's just important to to highlight this the fact that there are differences in nodal properties in patients that may be driving the connectivity changes that seen in previous studies so what about if we just look at the connectivity values so we can look at the number of streamlines as a measure as he's done in many studies and here we look at the so this is a t statistic of each of the edges in the network and you can see it follows a pretty normal distribution there's no huge outliers here on either side if you were to plot the this is equivalent distribution for the surface areas it would have been very skewed to the left because they were shrinked a lot smaller in patients and when we plot the top 10% of these just an arbitrary number we can see that there is no obvious spatial profile to the left hemisphere as we may expect and these are not significant after correction for false discovery rates so multiple comparisons and so using this measure of just the number of streamlines the connectivity is actually not different really I would say I've been seeing this what we can do is look at a different measure that's widely used in the literature and this is the connectivity weight and many of you will be familiar with this study I would expect yes no it's one of the key studies that kicked out of the whole structural connectomics revolution that we're seeing in the recent years and it essentially takes a connectivity weight which is a function of the density of the number of streamlines for the surface area normalized by the length of the streamlines that interconnect those two areas and Hagman used this in his study in healthy subjects which I think is fine there's nothing wrong with that and I tried this in the patients to see if this would result in significant differences and it does even after FDR and so we find that in a subset of the connections there are when using this measure this definition of connectivity weight as many studies do in the literature we find these significant differences but I would argue that these differences are actually really being driven by the alterations in the nodes in the network rather than necessarily the connections between them and so I would say if you're if you're doing any connectomic studies looking at differences between groups so whether it's male or female it's well known that females have smaller brains than men generally or younger and old all the people have thinner cortices than males or disease cohorts where there's no differences I would say one should really take this into account and at least be aware of this in any structural connectomic studies so that's all well and good but that doesn't really tell us anything useful for treating epilepsy clinically with with consultant neurologist in the hospital and I'd just like to draw your attention to this number that Yi Jiang touched upon in her talk so around 30 to 50 percent of patients who have focal epilepsy will go into surgery they have their outcome is not good so they will still have seizures even after having this invasive surgery to remove what is thought to be the epileptic causing part of the brain and the reasons for this are not fully known but it's been suggested in recent years that epilepsy is a dynamic disease there's seizures come and go seizures evolve over time they may start off with fast oscillations or high amplitude spikes they may come and go at different times of day so some patients have more seizures during nighttime or different times of month so it's I would argue that there's strong evidence to suggest that epilepsy is a dynamic disease in addition to aspects of structural properties of the brain such as the network that was described in the earlier studies and so if we could if we take both of these aspects into account the dynamics and the structure the goal really is to ultimately see if we can predict surgical outcomes for patients before we do the operation before we remove remove the part of the brain in this invasive clinical work that's done so here we've used the same data so 22 patients and we've included the connectivity so this is the connectivity matrix M into the model with the realistic time delays and we have a essentially the model is a it exists in a bi stable state so there's a stable fixed point and an oscillatory limit cycle around it and the likelihood of transitioning from this state to the epileptic form state is dependent upon these parameters which are inferred from the patient data so M which is the connectivity matrix the delays and also the degree of atrophy in the patients and this is all node specific and derived from the patient specific data and so the first thing we can do is simulate this model and see if it gives us something reasonable looking for a seizure so here you've got a time series of a subset of the network nodes so there's eight nodes here now overlaid on the brain and when one goes into a seizure state you'll see it pop up on the brain so there's the first one that's left temporal I think and then we've got right thalamus popping up and others as the seizure spreads quickly throughout the brain so that's great we can simulate the seizure spreading in the patients can we predict the surgical success rates or can we explain that so as I've said a few times it ranges between 30 to 50 percent the success rate the reason for this variance is one of the reasons is because it depends on where the resection is so if you have a temporal lobe seizure and you have surgery for that it's probably going to be like 70 percent if you have frontal lobe seizures it's probably going to be more like 50 percent so here we're only looking at temporal lobe patients so we would expect the success rate to be about 70 percent and here's the literature that explains all that and we essentially ask the model what happens if we remove those nodes from the network and resimulate the model do we get an improvement in the seizure likelihood following that and that simulated resection and if so in how many patients do we see that significant improvement and actually in 72.7 percent of the simulated patients we see that improvement so we're matching the clinical data in two regards first of all the seizure spreading and second of all the success distribution if you like and the reason for that is this initiation at nodes with high atrophy and at network hubs and the spreading through what we think are possibly healthy connections or normal connections so the model can reproduce this essentially likelihood this probability of having a successful outcome what about individually what about if we test this on individual patients can we predict their outcome well this is preliminary work that's currently in review on two patients that I'm showing here and patient A we've plotted the simulated seizure likelihood so which nodes go into that seizure state sooner than other nodes and we've plotted this spatially and the red nodes are the ones that go into the seizure state very quickly using the patient data again and on both panels and the black nodes are those that were actually clinically resected and you can see that in this patient they overlap quite a lot and in this patient they don't so the predictions from the model would be that this patient would have a very good outcome because you're essentially removing those nodes from the network that are causing the seizures and in this patient you would have a very poor outcome and we went back to the data we did this blind to the to the actual outcomes and we went back to the data and we found actually yes that patient does have a good outcome and actually that patient did have a poor outcome they had seizures even after chopping out that part of the brain as brutal as it sounds and then we did this for another 16 patients so we've got 18 patients in total and we get 83 accuracy which when you think that the clinicians only get around 70 accuracy I think we're doing pretty well there importantly the probably one of the best figures from this is 100 sensitivity and what that means is that in all the patients where we predict there will be a bad outcome they actually had a bad outcome and that's crucial because that means that we can go on and say to the the surgeon here's what we think you should be doing instead we think you should be operating on this part and maybe the surgeon will take that into account and investigate that a little bit further so to conclude there's kind of two bits of work here that I presented but I think it's and I think the first bit is is very important to bear in mind for many or most connectomic studies is that if especially if we know that there's differences in volumetric measurements or differences in surface area between subject groups if you if you're doing a connective study comparing groups I think that this is crucial that these aspects should be considered especially in these groups for example where we know there's differences and dynamical simulations may be prove may be useful for providing mechanistic interpretation so why are the seizures spreading in the model why are they spreading in the patients and why might seizures come back and accurate predictions are possible and we can investigate alternative surgery strategies with these techniques so with that I'd like to thank Nishan and Francis who were the students which did a lot of work on this they were master students with me in Singapore and in Newcastle and just do a shameless plug there for the conference that we're doing next year thank you for your attention thank you rather than DTI based structural connectivity what what do you see is that different or does it slavishly follow structure so I skimmed over that very briefly so this is actually using the functional connectivity between ECOG channels no I meant fmri functional connectivity I'd love to try it and so there's there's a bit studies showing that there are differences between patients and controls in functional connectivity as you might expect so sorry but if you've got a large well I don't know if you've got a data set of patients with resting state functional connectivity and you've got the last 200 patients who failed or you accumulate them from data sets across the world you could ask the question where the local connectivity it'll be something a lot closer than you get from ECOG but if you could get something it would be very very clinically useful yeah I agree could be useful typically fmri and resting state fmri is not clinically acquired so getting that data would be a big challenge and a better option might be to do MEG resting state MEG and source localize that so there was a paper out in brain last year or the year before I think it was Engelot which looked at that and he found similar things to was actually these kind of network changes that correlate with what are seizure focus seizure focus eye in good outcome patients good idea it's a I think that's a a way to take this absolutely thank you