 So thank you very much for this opportunity. So I'm presenting the work of my PhD student Manisha who won't be here. So this is actually a hobby project to be very honest and I'm going to present the excitement as a end user of the tools that are developed rather than as a tool developer. So I'm going to talk a little bit of cognitive neuroscience as well as our excitement from a data science perspective. So this is all about meta-analysis in brain imaging and the whole idea here is that there is so much of data that it's very difficult to make sense out of it. There are tens of thousands of studies coming out every year and everybody claims everything. So chordate is involved in language, motor cortex may be involved in something else. So everybody, the ultimate idea of what do you do with all this data is functional imaging helping us a little bit more. All right, yeah, that's better. So we've had discussions about the traditional software-based approaches where people use statistical parametric mapping or an equivalent for analyzing brain images data. Or more recently from the last 5-10 years, people have been also using machine learning methods for reverse inferencing. So what people are able to do is rather than finding activation in the brain side given a cognitive state, you are now trying to decode the brain state based on the activations. Now I would like to bring together these two points here and try to address the issue of variability across participants, across experiments, and across different scanners or different softwares. Now I'm going to actually use these different experiments as a case study on decision-making my area of research and give you a couple of examples how I've used meta-analysis for the benefit of how do we address different paradigms, different people, different software issue. So meta-analysis, again as an end user, we are trying to find consistency, consistency of activation across multiple studies and specificity. Now this kind of a claim is only going to try to be expanded only when we are able to use machine learning approaches. Why I mean that is when I have a coordinate-based reporting of brain activations, I can definitely find consistency by pooling across studies, right? And there are very good developed algorithms, very well wetted maths behind it. And if we have access to the data, the more joy we can actually analyze the data directly. Now, when I talk about specificity, I'm also trying to see the functional imaging in the light of the neofrenology. Are we really talking about cognitive segregation? How much of integration is there? And if given a brain activity pattern, what is the most likely cognitive state the person is going to be? For this purpose if I can apply this Bayesian framework and Yarkoni and colleagues have done this remarkable software called Neurosynth where they are trying to apply a Bayesian inference. And if we want to apply Bayesian inference, you need to have n number of alternative hypothesis. Then you get the probability of one hypothesis being higher or lower, right? So what they have done really I would say a kind of a big data approach and I was very excited right from its early development. Of course I've been following the activation likelihood estimation of meta-analysis. Now the new approach that they have done is a large-scale text mining. They initially started with full text and I'm going to really talk about how we were trying to tweak it at that stage. And in the full text what they have done is through text mining taken up high-frequency words, at least one in every thousand and then they have done automatic coordinate extraction of the brain peak activation data and then try to come up with this very nice web interface in fact of analyzing the number of terms that they have in the database. Now this was a limitation from our perspective. So for example, I wanted to analyze reward-based studies and Neurosyn did have that as a term. So what I have done is I have taken the brain map database. Brain map database is a very well documented, hand curated, very well worked out by researchers themselves and entered into a database very carefully done. So they have these reward-based studies. So I have taken the brain map database, used the ginger ale tool, get automated likelihood estimation. So I have the consistent activation from the brain map. Then Torweger has this MKDA tool, multi-canal density analysis. So at a point of time when ALE was not very well developed so MKDA was supposed to be a better method. Now there is a work around for the limitations of ALE where you have more reporting bias is basically accounted for well in MKDA. But both are different methods of analysis pipelines. But what I was trying to see is given the same dataset, do I get the same activation patterns or not? So more or less with a different intensity, through different software, through different pipelines, the reward studies, the meta-analysis gives me the brain areas which I know of. So it gives me the mid-brain, striatum, medial prefrontal cortex and so on. Now what I do, I go to this Neurosyn software, do a few clicks, try to see, so this is a different subset of studies. And this is majorly based on text mining. So there are inaccuracies. The original paper by Yarkoni has done all the required validations. So I'm not going to worry about that part, about the validity of using it. But what I'm trying to do here is trying to see what the forward inferencing and what the reverse inferencing really mean. So forward inferencing would give me the consistency of activations, whereas the reverse inference could possibly point out to specificity of activations. And I do see differences here. This is expected. I might actually expect that the specificity brain areas will be lesser than what the consistent activations that are being reported. But surprise, now I have a side-by-side comparison of the two things that you have seen just now. So the left graph is for ALE and MKDA. So not all, so what I've done is for every brain area that I can get from an automatic labeling software, I have calculated, given the total volume of this brain area, how much percentage is active and what was common by binary conjunction. So I do find areas that only MKDA reported or only ALE reported and also a large overlap. So this may be just because of a different software. But in the forward and reverse inference, this is where I started getting more inquisitive about it, is, as I said, the reverse inference probably has to be a subset. But I do also find some areas which are popping up in reverse inference but didn't pop up when I was doing the forward inference. What it means is these areas are probably not reported that often. So like the superior orbital or some part of the middle orbital, the gyrosectors, olfactory, some parts are not very commonly reported. Okay, fine. A case study for its validity and comparison is good. So what I'm going to do is I'm really trying to use the power of neuroinformatics and try to see this wonderful area of decision-making. Why I say neuroinformatics can solve my problem is these are different worlds. We're all talking about decision-making. The paradigms used in perceptual value-based and social decision-making are vastly different. You cannot do them in a single study. It's not possible. We are all talking about common neural currency. We talk about, you know, reward processing by mid-brain neurons. We are talking about ventral stratum involved in decision-making. We are talking about a common neural currency in the brain. Where is it? The paradigms are different. We consistently report the same areas. So does that mean that there is a common neural currency? How do you actually bridge together? What we did that time is because I couldn't get all these from the Neurosynth database and I wanted to exploit the automatic coordinate extraction from this database. So what I've done is I have, at that point of time, created my own custom-made mini-neurosynth in my lab. So all I needed to do was download the abstracts from PubMed, take the abstracts instead of the full text. So now, more recently, Neurosynth, the entire thing is now migrated to abstracts only because that's the more relevant terms rather than full text. I'll give you some examples later on. And then the analysis I have done in Gingerelli, which is the tool that the brain map provides. So I'm just trying to make this into an automated analysis pipeline kind of a thing. Just as an illustration, so if I pick up these words from abstracts, I get 84 studies which have the term value. I don't know what it means, but that word exists there. 61 papers have social and 36 papers have perceptual. I had used my filtering criteria to have one of these terms, reward, decision, choice, value, social or percept. Just an expansion. This is the final data I'm actually presenting. So there were 8,061 studies in the database when I started with. There were about 2,500 studies which had at least one of these terms. If I included reward as an additional term, it's just 200 more studies. Of these, I filtered out and got 639 studies. Now just look at this small Venn diagram. There was just one study which had all the three terms perceptual, value and social. And when I looked at the title of that paper, it didn't make any sense to be a decision-making study. But nevertheless, as a data processing person, I decided let me fix my filtering criteria and proceed with this. Now what do I find? I do an ALE meta-analysis separately, three different analyses, right? And I find more or less again a consistent thing, but a different pattern of activation. So striatum, orbitofendocortex, the same areas we've been repeatedly talking about or probably repeatedly having published. So those are areas to pop up. Now after that what I do is I start to compare them pairwise. I do conjunction analysis. I do contrast analysis. And then try to see what is it that these meta-analysis results became. So common neural activations across all the three domains, not a big surprise, a little bit of motor areas and the common reward and decision-making areas, fine. So caudate vitamin pallidum and insula have just plotted data. So these are the three colors. So you can see where the conjunction of all of them is. So there are also some areas which are not common, right? So let me just look at this independent analysis, okay? So perceptual and social decision-making was found to activate anterior singulate, which was not there in all the three, okay? In all the three I have not claimed anterior singulate, right? So that's one area. Value and perceptual decision-making, I had left inferior parietal area. And for social and value-based decision-making, they were angular, gyrus, caudate, anterior singulate, medial or peripheral. So what's happening is it's not as if the common neural currency is seeing that there are no specific brain areas also. So a little bit more beyond that. This is just a side-by-side comparison of what is the level of activation, right? So 20% ish when it comes to all three domains, but when I take only the value and social-based decision-making, it's slightly more for the caudate. So this percentage is percentage of the original anatomical brain area, right? So I'm not comparing them yet. This is a qualitative analysis here. Just to see how much overlap I see. Now the domain-specific activations. This is also a simple math, okay? I'm really doing it like a binary Venn diagram sort of a thing. So some part of the amygdala was more active in social situations alone. And several other regions were for value-based alone. Probably monetary decision-making is more dominant paradigm. So we have a lot many more brain areas coming there. And parietal, precentral gyros, supplementary motor area, these were more often reported in perceptual decision-making tasks. So that probably says that there is some task-specific activity as well, right? So now coming to the comparisons, right? Now these are really contrasts, the statistical comparisons. So again, there is greater activity in frontal areas for perceptual versus value. And in anti-singulate and media-prefrontal for social versus perceptual. So two at a time, I'm just showing the comparisons. For perceptual versus value-based, value-based really activated a lot of these medial OFC, chordate, anti-singulate, right? And the perceptual was more of inferior frontal, inferior parietal, those areas, right? And social versus value, as I said, only amygdala was the more in social. But value-based decision-making has a lot more. So right insular is also more in the social decision-making, all right? So that's in the blue. I haven't repeated the legends. So what it means is if I have to use this, you know, data-based approaches or neuroinformatics approaches for such classical meta-analysis, I might have to supplement it with a lot many more, you know, evidences. So I need to say what happens when you have reward very specifically involved or not? What happens in a choice compared to decision-making studies? What about those studies? You know, there were like 393 studies which were not necessarily involving either choice or decision-making out of the 600 and odd. What about those studies? What kind of social stimuli they are talking about? What kind of value stimuli they might be talking about? So these are text-based terms and they make sense only when you, you know, do it. But meta-analysis, each meta-analysis even with the, you know, good tools in our hand will take some significant amount of time. And I don't have the luxury of doing meta-analysis of everything. So I think and I am very strongly excited that this Neurosyn type of method is the way to go forward. But we might just need to do a lot more to support our findings. That's what I think. So just moving forward, we were just scratching the surface. I'll just present a little bit more of the type of the work we've been doing. As I said, this is a hobby type of project. So we've been trying out a lot of things. So one of the things BrainMap database has done and in a very decent way is finding the intrinsic connectivity networks. So they have used independent component analysis along with the metadata descriptions of the studies they have done. And they came up with a very good, beautiful description of the different independent components in the brain, which they call as intrinsic connectivity networks. So like the resting state connectivity as well as task-based connectivity. So they talk about the intrinsic connectivity network. So what we have done is we have tried to follow the same pipeline but using the Neurosyn database. And so we take the activation sites, we construct images back, and then we do an ICA. Fine, well and good. And then do a hierarchical clustering on it. And apply this bioinformatics method of having a correlogram. So you have clustering on both the key terms. So there were 525 key terms in the version of the database that we have taken and 20 independent components of the brain activation. Right? And when you look at this, it's easy to interpret when you have 20 independent components. But when you have 525 text terms not curated by any cognitive term ways. So then it's very difficult to try to make sense of this. So we went ahead and tried to say, is there a good clustering algorithm that can use a text mining approach, simple text mining approach. So we are just trying to look at these frequently occurring words. What they mean? Is there something that comes together as more logically? So I'll present that piece of the work. So what we have done is we have taken the database, we have found the pairwise distances between all the terms. Right? Because what we have is the frequency of occurrence in the article. And that's our feature vector. And then we've constructed a graph, a connectivity graph of the words. This is not the brain, right? But these are the words, terms. And each of the terms became a node. And I have a connection between different nodes when the distance between them is the shortest. Because I find the distance to every other node. So I draw an edge between two nodes between whom the distance is the smallest. And that creates my graph of all the terms. Now in this entire graph, I'm trying to find subgraphs. So one of the interesting, so all to all sort of a connectivity when I try to see, there is one node which is connected to a large number of other nodes. And that term was written. And this is the sentence, written informed consent was obtained from all participants. Right? And this is invariably there in all papers. So there's no doubt that it's connected to most of the terms. Right? So things like this, what is my cognitively relevant terms and what are not, right? So we removed that term and redid it. So the histogram shows what is how many nodes per cluster. So I'm just trying to find different clusters, right? So a better distribution of clusters rather than having very highly connected one term. Right? So I had 248 nodes connected to one term. Rather than that, it's better to have a better distribution. So when I removed that term, I got still clusters of the size 75, 94. Right? One cluster, but still that's a larger cluster. So what I've gone ahead is, okay, so what those types of clusters were like? In fact, this method worked beautifully. It found the same stem words, right? Which you could also verify with a portostemming algorithm in natural language processing. So person's, person and self. So there are these terms that my cluster is picking up, right? This is simply text processing. There's no brain connectivity yet involved in here. So motivation, food and eating becomes a cluster. So just to give you an idea about how this method works, right? Then what we did, we went ahead and only took a high frequency. So we did just a thresholding on the frequency and only considered those terms, right? So in these, yeah. So in these terms, what happens is we use the jacquard distance which ignores if there are zero numbers rather than a Euclidean distance or the distance metric, okay? So what happens is my vectors become smaller and therefore the nodes that they are connected to are more robust. And again, I have good clusters identified. One example of a cluster of size 17 is what you see here, right? So this is a directed graph because I can be nearest to one neighbor but the nearest neighbor to that might be different. So just to give an illustration of the visualization of one of those clusters that we found. So what it means is there is a combination of methods that we need and we need to use this more often. And then we can probably come up with a good understanding of the brain connectivity networks. So that's my last slide. Thanks a lot. I think there would be a way to sort of, would there be something that journals, for example, could do that would make such a text search is easier if they have. Is it helpful if they have key words? They do, right? So L-Sphere has all the brain terms and the tabular coordinates as well as disease terms. So they do have, and even PubMed has started indexing them. So this is going in the right direction for sure. Thank you.