Among his work, he's probably best known for his attempts to illustrate some of the masterpieces of historical literature. We see here is rendering from, from Milton's Paradise Lost called the fall of man. The depicts Milton's description of humanity falling into hell. Milton called darkness visible. I love that term or the phrase, it's doing. What good art should do. It's trying to communicate something that we don't quite have the right language for, that we don't have the right words for another person who like that turn of phrase was a guy by the name of William Steven. He's an author, he's a novelist. If you know him, you probably know him from his novel called Sophie's choice. But you may not know is that after writing Sophie's Choice many years later, he fell into a deep depressive episode. Hospitalized, nearly took his own life, emerged out of it. And several years later wrote a book called darkness visible, where he described in some of the most engaging and lucid detail the actual experience of being depressed and coming out of that depressive episode. I'm sorry to insult you by reading passages on the screen here, but I don't know how it's showing up online. So I well, in depression and depression, this faith and deliverance is absence, the pain is unrelenting and what makes the condition intolerable is the fore knowledge that no remedy will come, not a day, an hour, a month or a minute. It's hopelessness, even more than pain that crushes the, so. I'm just really move the needle on the national conversation about mental health. When it came out. Why don't you to hear in their own words from some patients that are experiencing a major depressive episode. Oh wow. Very, very dark. There's no light. And it takes doing much worse. It gets heavier. I can't keep going. I know the audio is probably quiet here in the room, but hopefully you can hear much of that. There are few things that I hope you can take away from that. One is the way they use analogy to describe what it is that they're feeling. They don't quite have the right words for it, but they keep using it's like this, it's like that. Also the physicalness of the analogies may talk about in physical terms. Can't move bottom of a well, things like that. Also. You can see it on their face. Here's one of those patients, a clip from that video and a clip from that person a year later when they're out of that depressive episode. You can see their face during the depressive episode almost looks Parkinsonian, right? It's a flat affect, no emotion. You can see this in sirens writing to, instead of pleasure, I was feeling my mind a sensation close to but indescribably different from actual pain. The pain rarely describable. Most of the suffers from this ancient affliction would have been able to confidently depict for their friends, loved ones positions from the actual dimensions of their torment. Perhaps elicit a comprehension that's been lacking. For myself. The paint is most closely connected to a grounding or suffocation. Even these images are off the mark. I fell onto the bed and like gazing at the ceiling, nearly immobilized, and then a trance of supreme discomfort. So you see the same things. You see him saying, I don't have the right language for this, but I'm going to try through my analogies. You see, I'm describing it very physically. Inability to move, brown, suffocation, it's a disease of the brain and the body. And finally, unfortunately for me, you see that he's actually telling us some 30 years ago that dimensionality reduction is going to be the answer. Here, is asking us to find some of the actual dimensions of his format. So fortunately, he's giving us some algorithmic guidance to What I'm going to describe to you today is an effort that in the context of depression, but I think applies more broadly, an effort to try and peek inside the black boxes of these complex systems. So if we take a step back from depression for a second, I would I would offer to you that I think a lot, maybe not all, but a lot of neuroscience experiments have a similar sort of set of challenges to them. We start with some external factor. If you're a sensory neuroscientists, this is the stimulus. If you're a motor neuroscientists, this is some task. If you're an effect of neuroscientists, this is some environmental factor that's causing a change in you, but it's measurable, it's external, it's your stimulus parameters. You have. This goes inside a black box of biological intelligence that has multiple components to it. As neuroscientists, we often think of the second of those, the neural activity and how that's a black box that we want to peek into. And that's a technological challenge of being able to record from neural populations that we've made great progress on as a community. I'm going to focus more on some of the algorithmic challenges. Often what we do is we try to correlate or correspond our neural activity to those external factors that we have nice and quantified. More and more we're doing that with advanced algorithms that we'll call kinda broadly artificial intelligence. Right? These are very, very powerful algorithms. Alright? But there are little inscrutable. In many cases, we don't know why they're behaving, how they're behaving. Which for commercial application has all sorts of issues of trust and fairness and equity. But in a scientific domain, it limits our scientific understanding that can come from these algorithms. Sometimes we'd really like to be able to explain what's going on in those algorithms. And in fact, that's a big focus of the next generation of artificial intelligence tools, called the so-called third wave of artificial intelligence tools. Where we look toward things like explainability, right? We try to add onto these tools or develop them from scratch. You try and give some information, some explainability about them. I'm actually going to offer that there's another black box that we hope to peek inside in this diagram that we don't often think about much. And that's the internal thoughts. I'll say very broadly. Again, if you're a sensory neuroscientists, that's the internal perception. Your motor neuroscientists, that might be the, the, the task motivation, the task intention, I should say maybe. We often don't try to correlate our neural activity this, because frankly, we often don't know it. We can run behavioral experiments sometimes, but those are very limited, right? And they're limited for a few reasons. One is that we don't necessarily know what the natural feature space is. For these internal thoughts. Though, when we're working with animals who don't have any language, it's kinda very difficult to know exactly. So we use very impoverished behavioral measures that often correlate to our measurement variables of the external factors, you as best as we can. Even with humans, they have language that we just saw in cases like this where we're talking about interoceptive feelings, fact, the feelings. They don't always have clear language for it. So we're just going to touch on that briefly today we're going to spend most of our time talking about explainable artificial intelligence. But the end, I'm just going to tell you a little bit about some new tools that I think are actually going to help us peek inside this other black box a little bit as well. We'll spend most of our time here today. Find a peek inside this black box and use some tools from this next generation of artificial intelligence tools, explainable AI tools, in order to help us better understand recovery from depression under deep brain stimulation. Well, I'll make sure to cover the most important slide first, it is a joy to work with all the collaborators that I get to work with. If many of them listed here want to call out a few specifically, one is Shankar Gottman. He's a research engineer that I work with here also works with Rob you, Tara. And many of the things that I'll show you today are due to his skilled hands. Push the patients in the clinical team that cares for them. Some key collaborators from Mount Sinai, Helen Mae Berg, and two people on her team, key sung Choi and Steve hazing, which did some of the key analysis it'll show you as well. And Patricia Rive a post say who is the lead of this team along with Helen and I'm on the project that I'm going to tell you about. Want to make sure to tell you about any potential conflicts of interest that could color what I'm going to tell you today. I won't belabor this because I think everyone probably has a good appreciation for the cost of depression, but just to put some numbers to it, for non-communicable diseases, neural, psychiatric disorders are almost 30 percent of the overall disease burden. About 10 percent of that comes from major depressive disorder. Clinical depression, also called unipolar affective disorder. It's an enormous disease burden. Today, there's probably 300 million people globally suffer from major depression. In the US, it's estimated that the cost just to employers is over a 100 billion dollars per year. The leading cause of disability, certainly in the developed world. And there are 20 to 30 percent of these patients that are treatment resistant, meaning that at this point there is no approved therapy and it gives them relief on their symptoms. We're in the middle of the presentation right now, so we'll handle another time facts. I think no one needs convincing that this is an issue to work on. And like my family, many of you have probably been touched by depression in your own families. Don't want to belabor this. But again, just to tell you that for me personally there are two communities of people that I care about deeply affected by this. One is that there is a huge influence of income level on diagnosis of depression. Huge influence. Again, not to put too fine a point on it, but I grew up here. I grew up in poverty and much of my family remains in power. This is my grandmother, is 96 years old. That's a picture from last summer. I point her out because she is the reason that I grew up in a family of Native American heritage. He's an active member of a native sovereign nation a few miles from where I grew up. Native American mental health has a 2.5 times the suicide rate for their young adults. Their children and teens have the highest major depressive disorder rates of any ethnic or racial group in the country. 22% of Native American females have attempted suicide. Almost one in four people. Astounding number. So again, I don't think that I need to probably convince you that it's an important problem to work on. Let's talk about what it is and how we diagnose and how we treat. These will be important things as we try to examine. And this is the DSM-5 diagnostic criteria. This is how depression is diagnosed. There are two core symptoms, depressed mood and anhedonia. Lack of an ability to experience pleasure, joy. You must have one of those two core symptoms over an extended period of time. He must also have some of the additional symptoms. Weight-loss, decrease in appetite, insomnia. Psychomotor retardation is the common one. Psychomotor slowness, fatigue, loss of energy, guilt, worthlessness, diminished cognitive capacity, ability to concentrate, thoughts of death, suicidal ideation. So clinically diagnosed by matching these criteria through clinical interviews, there are ways that we try to measure the severity of depression. I'm going to give you an example of one scale here. It's called the Hamilton Depression Rating Scale. It's probably the most common of these. These are subjective survey instruments. So a clinician would fill out a survey after meeting with you and answer, in this case, 17 questions about how you're doing. I want to be very clear. This is not a diagnostic for depression. It was never intended to be developed in the 1950s. It was developed to measure how well inpatient depression patients were doing in response to the first generation of antidepressant drugs. So is very focused as you see here on the physical symptoms, right? If you look at the core depression symptoms, depressed mood, and work and activities, which is essentially are you doing things that you find pleasurable? There are two of the 17. There's a lot of stuff on here about different types of anxiety in some insomnia has three categories on here. So what you can see is that this is a measure that's heavily weighted by a lot of different non-specific symptoms, many of which occur in the co-morbid co-morbidities, the abs, and come up with depression. So this is a numerical rating score. It is often used in drug trials, but it is not meant to be a diagnostic and it's not meant to be applied all the time. Originally the guidance was every two weeks you do this. Now the, the the way to use this is at the beginning and end of say, a drug trial to measure whether you've made an improvement overall. But this can be a deeply confusing measure with a lot of subjectivity. Major depressive disorder is a network disorder. We aren't necessarily going to go through this diagram in a lot of detail, but just wanted to show a few references that state kind of very clearly. Some of the major symptoms of depression have entire networks associated with them. So anhedonia, the loss of pleasure, deeply impacted by the reward network. Though areas like nucleus accumbens, things that modulate dopamine production in the brain. Cognitive control has both limbic and default mode network elements impacted dysphoria. So feelings of unsettled NIS like something doesn't feel right in my body as has a lot of impact from the Olympic networks as well. Rumination, default mode network comes up a lot internally focused thought. You see here on the right that these networks, these major kind of networks are modulated and depression default mode network seems to increase its activity some more inward focus thought. Things like the salience network that helps you assess sensory-motor information and adjust cognitive control seems to be downregulated. Dopamine is down-regulated. The serotonin and dopamine networks are deeply connected to all of these areas that seem to be mediating the symptoms that we're showing you here on the left, right? These are often found through correlational studies. I think the picture that I want you to take away is it's a network disorder and it's complicated. It's difficult to know how these pieces all fit together to get a picture of what's going on. While there are many depression treatments, I'm not going to go through the history of that. We're going to focus today on deep brain stimulation in the sub colossal singular. This is an experimental therapy. It is meant for treatment resistant patients. It's obviously very invasive. This has pioneered by Helen Mae Berg, first paper reporting this was in 2005 with a cohort of six patients up in Toronto. Again, there's a really, really interesting long history to this that I won't be able to go through today. But I'll just say that they identified a target called the subcostal cingulate or the subgenual cingulate. They did it in a data-driven way from imaging data, PET imaging mostly at the time. And they implanted deep brain stimulators and have done so in several cohorts of patients as they refine the therapy up to the present day. The current target has been defined also in a data-driven way. It's done now through individualized DTI imaging on the patients, but was found to a common map of responders in large cohorts. The key thing to take away are there for white matter bundles. And what we're actually trying to hit is the intersection of those four bundles. There's in yellow there, the cingulum bundle in red, the forceps minor in blue, the uncinate fasciculus, and in dark blue there it's hard to see in this, in this view, but frontal striatal fibers that hit subcortical structures. Just to help a little bit, we drew a quick diagram of this. So you can see in the frontal striatal you're hitting subcortical things. Thalamus nucleus, nucleus accumbens, very important in arousal and reward regulation. I've put up two gray autonomic function, wrap I nuclei, critical nucleus and the serotonin network seems to be involved in homeostasis. A serotonin in the limbic network from the uncinate you have things like insula, hippocampus amygdala. In the cingulum bundle you have anterior, middle, and posterior. There seems to be some disagreement about how many divisions there are in the cingulum bundle, but certainly affected in things like cognitive control, cane movement to some degree. And then the forceps minor headed out to prefrontal cortex, especially ventral medial. So headed to the frontal pole, you see key areas in both the salience network and the default mode network. So it's not clear that subclones the sink. Scc is actually the site of dysfunction. But it is a network way station essentially where a lot of really important nodes are crossing through. So it's unclear exactly where the deficit is, but this is the target was identified and people are getting better from what we see. And this is, this is a longitudinal study over a cohort of 28 patients from Emory, where the majority of patients get better and stay better. You can see in this cohort they have over 75 percent long-term, our clinical responders, the response in this case is defined by a 50 percent reduction in their symptoms according to the scale like the Hamilton. And they have about 50 percent of people that are in complete remission at the eight year mark. This stands in contrast to the fact that actually this was the subject of a clinical trial earlier that was halted midway through. And I'll be sure to use the right word there. It was halted with an indeterminate outcome. Breathless headlines that the trial failed. They actually never got through the collecting their data. The company that was running it decided partway through they ran if utility analysis and they had halted it. There are probably many reasons for that. Again, if you want to talk offline, I can talk about them. There are a lot of sources of variability and some things have been done since then to try and reduce that variability. A key thing is that they needed a more objective way to assess what was happening. If you look at the longitudinal data, they had a number of patients, I believe is almost 75 patients that kept the implant in. They turned on the controls, that they turned on the sham group. So everyone who's receiving stimulation two years into stimulation, they actually had a 50 percent response rate. This is after the trial was halted. So response seemed to take longer. May have been due to in precise targeting. That's a story for another day. The key thing is they didn't know that they were on their way to a good result. If you're using the Hamilton isn't indeterminate measure. There are a lot of things going on. They needed an objective way to know that these patients were headed in the right direction. Or if they weren't to be able to do something about it, like adjust the dosage of their device. What we're searching for today, a potential biomarker here. Another way to measure how a patient is doing through this therapy that would enable closed-loop modulation. And if somebody work on closed-loop modulation, this might be particularly. Closed-loop modulation, like adjusting dosage on week timescales can also tell you when to administer adjunctive therapies, like new medications, psychotherapy, both of which are important part of getting people better and deep brain stimulation. What we need here, we need two critical things out of a putative biomarker. We need to correspond to a relevant behavior that we're trying to measure and we needed to move. Those are kind of hard to baseline thing. So if we're going to have a biomarker, it's gotta measure something about what we want to capture. And it's gotta move when we apply a therapy that is a factor that will keep those in mind. The primary data I'm going to tell you about is a cohort. We implanted 10 patients or treatment resistant depression patients are implanted at Emory, implanted with Medtronic PC plus S device, which is an experimental device that has never commercially available, but was the first device that really let us record from the stimulation say, oh, I'm sorry. Well, so over six months of therapy with patients, we ended up with six patients of usable data. So everything I'm going to report to you from here on out today, isn't it? So a reasonably sized cohort for an invasive invasive procedure like this. They were all done with the kind of standard procedures, the state of the art and how this team knows how to do it. So individualized DTI targeting. You can see here on the upper right. This is a rendering of one of the patients. We can measure what the volume of tissue activated is some stimulation. We can measure the volume of tissue recorded. It's a differential recording on two electrodes around the stimulation contact. We can do computational modeling to understand what fiber bundles are being engaged. Especially in the stimulation. He's patients came in weekly for their clinical interviews. During the time that they were n, we collected data from them, seven minute long recordings. We did stimulation on and stimulation off. Everything I'm going to show you today is with the stimulation off, so there is no possibility of a stimulation artifact. And what I'm going to show you broke that up into 10 second blocks, calculated the standard suite of local field potential features. So band-limited power, coherent, all the usual sort of things. The other key thing is they do a clinical interview while they're there. And that is videotapes that the key piece of data. You can see here, the computational modeling on the bottom from, from just the sagittal view in this patient you can see we're getting good engagement of cingulum, uncinate, four steps minor and the frontal striatal fibers. So we're seeing kind of all the things that we want to see, to know we're hitting in the right place. At 24 weeks. Five of the six people are what we're going to call typical responders. Meaning they got better for those six are in remission, mean they're Hamilton is lower than eight. The four that we don't have good LFP data for, we're also typical responders. So overall in this course, cohort nine of 10 for typical responders, we have one relapse responder, which is a very interesting case. She started after surgery, she started in remission when most of the way through the trial and then had a combination of things happen, some personal things, and fell into a deep depressive episode right around the five month mark. Do you eventually was was treated and at the seven month mark was also a responder, but at our six month end point, she was not. So what's really interesting here is we have a cohort were almost everyone got better. So we can study the differences and how they got better and the time course of that trajectory. And then we have one person that basically drove that road and reverse started better and then got sick. So we have this really interesting data set. If you look at the individual traces, you can see there's tremendous variability. Okay? Some people get better quickly. Some people take much longer. We see in red there are relapse responder to see one interesting feature which is after an initial decrease, the Hamilton score, which is what I'm showing you here, often has a bump up. A few weeks later. We'll talk about that. So idiosyncratic trajectories. What do we do with this? Shouldn't we be trying to find a biomarker for the Hamilton? Hamilton has a number of known issues, bias issues, the non-specific symptoms that we talked about. Essentially, it's capturing things like their local mood day-to-day. In addition to their disease state. The Normally we look for a biomarker, we want the biophysical signal on one axis and there are ground truth on the other axis. But it brings this question, but does our ground truth? If you look at the clinical note, here's one patient. They seem to have an increase in their Hamilton score. Those red dots are where the clinician decided to increase the dosage. You see highlighted here is an area where they're Hamilton went up and the clinician didn't increase their dose. You look into the notes, they say, yeah, I'll paraphrase here. The clinician essentially says, yeah, their score is going up, but they're really just anxious. They're feeling the depression go away and they're not sure who they are without it. And they're worried that they're going to feel alone. This person. They're not getting worse than their depression. They had something else going on that's going to take therapy to recover. This person has the same sort of thing. Depression metric spiked up the Hamilton spiked up. If you look at the clinical notes, there's a very stressful personal experience for this patient that they describe in detail. And then the clinician says Essentially, yeah, this is anxiety. Like this person had a really bad week. They had some family drama going on. Sick, they just had a bad week and it gets reflected in these measures. So what did we use for our ground truth measure? Well, we know that in the first month, all of these patients are sick. If you look at their aggregate Hamilton scores, there is no clinical disagreement that they are all sick in the first month. There is also no clinical disagreement that they're all well in the last month, in month six. And I'm talking about the typical responders here. We'll take our relapse responder out for a minute. What we're gonna do, we're not going to try to regressed against the Hamilton. We're going to build a sick well system. That's what we want our biomarker to, to key. Is this person sick? Are they, well, where are they at in between those two states? So we're going to use that data. It's not clear whether we're going to see a change in the sub close to a singlet because it's a network disorder. But we're going to build a neural network classifier to try and see if we can detect it. In particular, we're going to build one neural network classifier for all six patients. And what you see is we can actually get quite good discrimination performance. Though there is something going on in that data that is a marker of whether that patient has thicker as well. Okay. Area under the curve here of 0.87, quake the perform. And again a single classifier. So whatever is happening, there are features in common across all five of these patients. It doesn't tell us what those features are. We could try a probe. There are ways to probe and say What's your most important feature, but that doesn't really give us a full picture of what's happening. So we need now to turn to trying to explain this classifier that we open our machine-learning textbook. Alright? And we say, well, how do we explain things? And the standard answer you're going to get is black box. We don't know. All right, so we're going to move to a notion of what does explainability me that we need to sit down and think about this. We're going to abstract our problem here. Do a classifier that's going to classify these fruit shapes for a minute and then we'll get back to some brain data. But this will be easier to illustrate. If we're putting a bunch of images of fruit into a blackbox classifier that we don't know what that classifier is actually using to classify. We just get the class labels on the output. You might ask yourself, Well, I don't know what it is. Is it color, is its shape, is it some other feature that I don't have a word for? I don't know. What does it even mean to explain this classifier? We're going to use the language of causal statistics. We would like to identify aspects of the data that have a large causal effect on the classifier. Meaning if we intervene, if we were to change that aspect of the data, then we would change the classifier's output. Okay, so graphically, what we mean here is we take our data, we would like to organize it in a way that different aspects in this case, classifier relevant aspects and non-class fire relevant aspects are on different axes that you see here, divisions by color and shape. And if we take one of these, say take the orange and say keep the shape. There'll be an apple, the roughly the same shape, but moves the color from orange to red. Then look at what happens when we probe that classifier C. Oh, you're labeled changed. He went from one class to another. Whereas if we would've changed from carrots are orange, we wouldn't have seen the same change. This is notionally what we want to do here. Okay? So how do we do that? Well, the tool that we've developed is called the generative causal explainer. Essentially we set up a causal statistical model where we have two sets of latent variables, alpha and beta. Have in mind your favorite dimensionality reduction approach. Think PCA, if that's easy for you, we're going to go into the principal component space. We're going to divide the principal components into two sub-parts, Alpha and Beta. Those are going to combine together to form a piece of data, right? So we have two different aspects that are going to get combined together to form a data point. And then that data point is going to get put through a classifier. X is going to get put through to find a class label y. Yeah, we're not looking for an individual feature. We're looking for something like an axis in a dimensionality reduction algorithm, something that could be complicated. So this is actually how we set it up. We train another machine learning system to work cooperatively with the black box. So the classifier is trained, that is fixed. We're not changing that now. What we're going to train another machine learning system whose job is not to classify. Its job is to generate data. And it's just going to keep generating data and pumping it through the classifier. But we want to train that system. Okay, So again, think about calculating PCA. And we want to train that system so that alpha, the dimension that we selected, either classify a relevant one actually has a causal impact on the output of that classifier. The other one, beta, is used just to help capture our variance in our data. So in this case, we're not going to use PCA, we're going to use something more non-linear. It's called a variational autoencoder, but very kind of non-linear neural network-based method. But essentially you can conceptual, conceptualize it like that. Dimensionality reduction, where some of our features. I've been trained in order to have a causal impact on the output of the classifier. And the others are left alone so they can just capture data variance. I'm not going to belabor this because we started a little bit late, but I'll say that we evaluated this and the kind of standard way that we evaluate machine learning methods like this. So we put data in, we trained it up on a classifier for digits. If we move the relevant variables, we get things changing their character, like from an eight to a three. If we move the non-relevant variables, we get line thickness changing and slant changing. And we can do the same thing with different datasets. And we can do ablation studies where we take those variables out and make sure that it's actually what the classifiers depending on. So for now, we're just, we're just going to say that we evaluated. This is a whole separate paper. So let's go back to our data from these patients. We've already shown you the train classifier. Now we're going to do is we're going to apply this to the classifier that we learned on the local field potential data from these patients. Then we're going to look at what those discriminative components do during the intermediate weeks. So the weeks when we don't know whether a patient is sick or well. So SDC is the acronym here, Spectral discriminative component. So that is the dimension in this generator, right? That is the explainable variable. And we call it spectral because the features that we're using are largely spectral features here. Okay, what we see, I'm showing you all six patients here on the right. In green, is there SDC? So this is the LFP based marker and in red. All right, sorry, orange, is there Hamilton score? So you can see first of all, roughly they're moving in the same direction. They seem to have good qualitative correspondence. There are some really interesting individuals to point out. Here's our relapse responder. You can see the Hamilton crashes around Week 15 or so. So notice that we could have identified that Crash coming approximately a month to maybe six weeks before they actually showed symptoms that were, that were picked up clinically. Something was happening with this patient in their brain. And this marker, even though, let me remind you, this marker was not trained on this patient at all. All of the training for the data pipeline that I've shown you was done on the other five patients. So this is a complete generalization from the training that we did to a patient that are our algorithms have never seen. Although remember our patient that was having anxiety because it was their depression felt like it was lifting and they were uncertain how they went forward. We can actually dig into the Hamilton sub scores and we can see that that's what's happening. Notice that their brain is still moving in the right direction. There's still headed down, which in this case is going from sick to well, just like the Hamilton. And we can see before this point, they're depressed mood was a much bigger factor in their overall Hamilton score than their anxiety. After this point, anxiety is on par with their depression. So something is happening with them, but it's not a worsening of their depressed mood. And the brain reflects that correctly. If there's something else going on. If we look in the aggregate, we can see across patients, we have a signal that has lower variability than the Hamilton. We see. An interesting point around week six or so where there's an early transition where they dip below 0.5, they kind of dip below threshold and look a little bit more well, and then they come back to being sick for a while until they eventually decrease into stable wellness. Note that if we try to correlate this on a person by person basis with their Hamilton scores. It does not correlate. We're not tracking the individual fluctuations in the Hamilton. If we look instead at a notion of stability. So take our stability criteria. Remember in the Hamilton, if we dropped by 50 percent, it's considered a response. So we're going to define stable wellness as where do they get below 50 percent and stay there for at least three weeks. So they've gotten into the response range and then they stay there. We're going to do the same thing on our LFP marker. And so you see here on the right the individual sick transition to well, for all of our typical responders. Okay. You can see there are a few patients that are off by a little bit. I can talk about them offline, but overall accuracy is almost 80 percent. And identifying week to week whether they're sick or whether they're unstable wellness. So here we have our first clue that we asked for in our biomarker, right? We're corresponding to something that is behaviourally relevant. We're corresponding to the transition. Sickness to stable wellness. Let's take a look at actually what's going on. Because I promise you we're going to explain something, right? So we can actually look at these features. We can put a signal in, get into our latent space, and then we can move, we can go to one principal component effectively, and then we can move along that axis and say what happens. What are the features that seem to be the classifier seem to be Kenya. So here you see a plot of all those features. And the ones that are statistically significant are in the left hemisphere, Alpha and low and high Beta. And on the right hemisphere, high beta and gamma. Those are the features that over the six months of therapy seemed to be what the classifiers kingdom. We look just on the left SAR instance. We can see that the five typical responders all see substantial increases in their left low and high Beta. Patient. One there in the yellow is our relapsed responder. So she is supposed to be going the wrong way. Right. Because that's actually what happened with her clinically. I want to point out, first of all, we weren't sure we would see anything here. But second of all, this is certainly not what was expected from interoperative data. So when you're acutely recording, you implant the electrode, you provide therapeutic levels of stimulation in the operating room. And people were able to measure and see actually what you see is a decrease in beta. So the acute effect of stimulation when you first apply it is a decrease. If we look at our one month data, we see that as well when they get turned on chronically. And in fact, that decrease in beta corresponds to how much they're Hamilton has decreased one week later. That's what this graphic infinity et al, is showing you here. If you didn't know anything other than the acute response in the operating room, you would have expected that we should have been chasing a decreased and Beta. If you're building a closed-loop system, you would've been going the wrong way. You would have been chasing the wrong direction. The signal that we need to get people stable well, is the opposite direction from what happens with effective acute stimulation. Long been suspected that the recovery of depression patients DBF is due to some type of additive additivity, some type of plasticity. And this is the first concrete evidence that we have of it, that something different is happening over six months then happens over the minute timescales of acute stimulation. We can actually look and see what happens when there's a therapeutic intervention. We have time points where the clinician decided to increase the dose on the patients because they weren't doing well clinically. What I'm showing you is all of those points of stimulus change aligned to the week 0 time point. So Week 1 is one week after stimulation change. Xk minus one is one week before stimulation was eventually changed. What you see is they're Hamilton score does not causally move in the next week or two after stimulation has changed. So according to that gold standard metric, they actually have no causal effect in the week or so after stimulation change. Whereas there is a statistically significant improvement in the LFP in one week after stimulation change that we see is the second thing we needed, which is movement with a therapeutic intervention. So we have at least are two basic ingredients to try and have a biomarker that it's actually useful to tell us objective what's going on. I want to make just a couple of connection points. I told you we also have the face datasets. We have videos essentially of these patients during their clinical interviews. And we all saw it. You can see when they're sick or when they're well, we don't need fancy tests for it. We did essentially the same pipeline, but using machine learning classifiers on their face. And we can identify features. This is a patient in the beginning and in the end, first and last month, you can see his face looks different. But we can also identify which action units, which muscle groups seem to be changing the most as he gets better. We actually have quite good classification just based on the face. Based on the face, we can tell whether an individual is ticker as well. I do want to make one note here. There's much more idiosyncratic character to the face. This is an individual classifier built for each person. Not the same as the brain signal where we were able to build one classifier that worked for everyone. But notice we have a lot of things happening here. We've colored green things that are in the upper face, purple things that are in the lower face. Though notice it is not that they're just smiling more, but it's a complicated change in their face. We see here in the aggregate, I won't belabor this, but if we look at that face classifier next to our brain signal, they seem to be moving in concordance. Even including our relapsed responder. If we look at the time point where their face becomes unstable and their brain becomes agent. With the concordance, we get a Kendall's Tau of almost 0.9. So rank order, the people that look better first also have a brain signal that gets better. Burst. We can also look at their pre-surgical DTI imaging. We can look at the health of their white matter tracks. So measure that with fractional anisotropy FA, some measure of white matter. How you can see here if we look just within our networks, you can see again forceps minor. I'll try and point out. The blue areas are areas where there is significant correlation with the time to achieve stable wellness in the patient. The big piece of damage here in the left, forceps minor almost out to the frontal pole. We have a big piece of damage here in the posterior cingulate cortex. Another one here down by hippocampal formations. So essentially what we can see our areas, and that's what's shown on the plot in the bottom. Where the more damage you have in those areas, the longer it takes you to get better. We don't know exactly that it's a remyelination process. That's certainly one possibility on the table. But the more damage there is, the longer it takes you to get better under stimulation. So what we see is this is the first longitudinal study of chronic stem in an SCC DBS patient. We didn't know what we would find. Our explainable AI techniques have shown a putative biomarker, which is a really, it's a critical step towards scaling this therapy and moving it toward potentially another clinical trial. There's evidence that we have multiple things going on. We see acute relief then increase in some types of symptoms, but they seem to be core anxiety symptoms and other things rather than core depression and everyone. And then a longer-term recovery. There's a lot to learn about this. There are potentially multiple mechanisms going on. This would actually make a lot of sense given what we know about the pharmacologic agents, we have fast acting agents like ketamine that last two to three weeks then have to be re-administer. We have slow acting agents like SSRIs that take four to six weeks to reach their full effect. They have different types of responses for the patient. They also act on different mechanisms. There are also some interesting correlations with the specific frequency bands that we're seeing, the Beta range and the Gamma range. Certainly it's been hypothesized that there are connections between those ranges of the things that happened in depression. So broadly, if you're familiar with that literature, you might look at this and say that this is evidence supporting that an initial response is due to a disinhibition globally. And then a slow a slow restructuring of inhibitory control over the, over the time period of recovery. Again, this is really just kind of hints. We don't know anything about this, but if you're an expert in those areas, just giving you some hints in case you want to talk about it later. We seen here a really interesting connection between the electrophysiology, the behavior, and the anatomy, right? All these pieces are coming together. There's more to explore. But I just want to remind you for a second that we saw in the face that what was mediating the change in their face to look more well, we're mostly upper facial features. And then the imaging we saw, among other things, deficit in the post a posterior cingulate cortex, especially more so the anterior singular cortex is just again, just a hint at something that we haven't explored fully. There are different control networks for the upper and lower face. This has been shown in humans, it's been shown in non-human primate. The upper, called the primary emotional system. This is kinda the real emotion that you're feeling. Lower is the social emotional system. This is, you're forcing yourself to smile. It's the fake emotion. This is why people tell you smile with your eyes, right? And you're getting a picture taken. Different networks that mediate it. I will just point out that the networks that mediate the primary emotional system, and here on the right as shown in non-human primate, but it's very, very similar, are higher motor areas. Okay? So supplementary motor area, M3 and so on in human M2, M3, M4. Okay. These are adjacent to or part of cingulum bundle as you head toward the posterior single and try. The social emotional network is controlled more by premotor cortex, an M1, both can have some limbic involvement, but these are further away from our stimulation network, right? So certainly weren't expecting this, but in hindsight, it certainly makes some kind of notional sense that we might see more recovery and upper facial features than lower facial features. So to kinda step back for a second, we've talked a lot most of our time here together about peeking inside the artificial intelligence black box. But I do want to say there is still this problem of how do we actually measure behavior from these people. We have these kind of imperfect scales, these Hamilton scales. And again, they're trying to describe how they're feeling and they don't have the language for it. So how do we deal with situations? Where there's a limited language to describe the internal thought, the internal percept, whether it's a human or an animal. When we have to deal with things like this in the machine learning literature and buy things like this. I mean, we want to characterize data, but we don't have a good feature space for it, and we may not even be able to learn a good feature space for it. What we do is we look at techniques like similarity embedding. We say listen, I don't know what the right features are, but what I can do is maybe I can lay this data out such that things that are similar are close together and things that are dissimilar are far apart. And similar and dissimilar are completely subjective judgments. So that's, that's being given to you by a human, right? For instance, if a, if a machine where to lay out all these images, they put things together, maybe like this, you see a lot of color coordination. But if you asked a person, you to organize things by gustatory similarity, what tastes good with Sabi and guacamole very far apart. And the computer would have put it very close together. We have no way to get at that relationship there then querying the person. The problem is it's very difficult to query people for things like this. This actually is very similar to a notion that's used in neuroscience. I'm really a lot in the last 10 years, which has representational similarity analysis. This is used when you have electrophysiology data and a computational model like a deep neural network. You might say, Listen, the feature spaces are probably constructed very differently. What we want to see if there's a similar organizing principle. We can't just put them side-by-side, but let's see if two things that are similar to stimuli that are similar and the electrophysiology are also similar in the model. And that's the way that we're going to try and correspond these representations together. It's very similar to that. Only it doesn't get used with behavior very much. Again, because it's difficult and time-consuming to collect this data. We actually sort of know what to do, right? This is a psychometric sort of task. The classic task would be what's called a triplet query. You pull three data points out of your dataset. You make one the reference. All right? And you say, here are two possibilities. Which one is closer? It's more similar to the reference, right? And you have them pick between the two, so they have a binary choice. This works if you collect enough of this data, then you have essentially a dissimilarity matrix and you do dimensionality reduction and you can get this type of similarity embedding. You can make a mental map of the space of the, of the person's subjective percent. The problem is, it takes, for ever, for any reasonably sized data set, it's entirely impractical. Impractical. I'm not going to dive into the details of this, but I'll just say we have new state-of-the-art methods that we've developed for doing this better. One key thing is that they're active learning methods. We do is we look at the answers that the subject has given us before, and we pick new queries using information theoretic criteria to say what's the most informative query we could ask? Well, right now, so lets us ask many fewer queries. We also extended the query type to be these rank-ordered lists. So instead of just a binary choice, we can give them a list and say, we rank order these with reference to similarity to this reference to this anchor object. And we shall people can do that out to lists of 789 objects and doesn't seem to be a problem for them at all. And we get much better data. That way. It makes the people be more self-consistent when they're answering. And so it's not just that you're more efficient and collecting the data, it's actually higher quality data. We've done this. We collected a dataset of almost 700 thousand queries, but that's like 200 thousand queries from 700 subjects. This is just to show, like our methods are much better than the randomized query selection that had been done before. So it opens the door to doing this practically. We have to really notice the AI system doesn't touch the data at all. We don't need the AI system to know anything about the images, about the taste of the images. The only data that we work with is the response from the individual. So this could be anything. This could be your favorite sensory stimulus. This could be a motor situation that you want people to compare. Hey, was that movement similar to that movement? Even though your joint angles, we're all very different. It could be An interoceptive sensation. That feeling you're having right now. Is that more like yesterday or more like the day before? Now, really the algorithm doesn't care what sort of data it is. It's just going to figure out which points in that dataset to ask to most efficiently make a mental map of the space for the subject. Easiest with humans, not impossible, that you might be able to do this in some animal perhaps, but doesn't necessarily have to look exactly like this. We have some modifications to it that we've talked about. Actually possibly doing with animals. He wouldn't have them rank order list, but maybe still picking most similar sensory stimuli could be something that they could do. This just wanted to show this to show how absolute pilot set that we just collected. We barely analog. This is asking people to take day-to-day situations and choose how similar they are in terms of their perceived effort. Remember, core symptoms of anhedonia, lack of interest in doing things. Psychomotor slowness, this huge perceived effort. We're trying to get at whether we could characterize depressive symptoms with methods like this. Just the basic clustering applied to it. I just really wanted to show you that we can go to very, very abstract queries like this. Some people can do it. We don't know yet what it means. There are a few kind of idiosyncrasies in the clustering that we don't know if they're real results are not that happy to talk about that offline. I think what I'm trying to show you here is a way to try and use AI techniques to give voice to people who don't have the language to describe their symptoms. Remember I said that that's also what good art can do. So as we try to put these pieces together, I think we need to remember that not all of our solutions are technical, that aren't still has a role. So just a brief advertisement here that I've been working for the last year and a half on a newly commissioned piece of modern dance with GT arts and the local modern Ballet Theater called terminus. And a choreographer, I'm sorry Schumacher from New York City Ballet. Making a piece of modern dance that's focused around neuroscience, neuro technology, and neuroethics. Trying to give voice to those things that we don't yet have language for. Some of you have met with them. It's basically an artist in residence sort of process that they'd been going through. We'll have the premiere here in the fall and be sure to advertise that broadly. It'll be here in Georgia Tech's campus. And I'm really excited about just want you to hear from some patients what it sounds like when they come out. Incremental changes, things got a little bit easier. And things got a little bit easier. Brush a little bit easier to get out a pen, and a little bit easier to have. Just started a cascade of positive instead of the cascade. And that's the difference. Was happy. Ebs didn't solve all of my problems with how I was thinking or how are misbehaving or hi lack of boundaries and relationships and solve those things that needed to be worked on was make it possible for me to not have this huge burden preventing me from being able to solve all. I walk more than I did before just because it places, I mean, I don't even think twice about walking to every village anymore. Here. It's kinda like but I'm not worried. Going to lead back into some kind of awful which life? Hopefully you see it in their faces. You hear them describing movement. Here, especially the second patient describing stability. I'm not worried. More. Stevens words, it can close this out here. For those who dwelt in depressions, dark wood and known it's inexplicable agony. The return from the abyss is not unlike the ascent of the poet, trudging upward and upward out of hells Black Death. The poet here of course is Dante alligator. He's describing the Inferno. He quotes from the Inferno. And so we pick came forth once again be held the stars in this doorways illustration of that moment there. So as we search for our virtual, I hope we can use some of these technical advances as well as our connection to the arts and humanities to try and help these patients find a way out of it. Thank you for your time and attention. I know we had some technical trouble. I appreciate you sticking with. Thank you. Alright. So if you have questions, we can beg them. You can just unmute and ask or you can put them on the chat. I can't see the chat it all himself as quite yet, nobody from mom and one monitor the chat here. I can ask a question on online. Burst. Hey, Kris, maybe crazy people that are here. You can repeat the question so that yeah, her up there. I was wondering I got here sometimes in the sum, at least these trials with depressed patients who have TVM, sorry t and that accumulation that there's a significant placebo effect in the placebo group, something like 25 percent. And I was curious whether or not, you know, this well, not pregnant, that you trained on the actual patients receive stimulation, you might be able to be applied. Patient get better and CBO trial. Maybe we can, certainly lucky if you're talking about DBS, placebo group because we couldn't get our data for this TMS trial because I need the basic recording. Yeah. Curious or whether it might reveal. A problem with the Hamilton scale, even if you maybe, yeah, maybe certainly there are a lot of things we could do there. I should say we are looking at EEG correlates that wouldn't necessarily need the invasive recording, but we're not there yet, but we're looking at those sort of things. Maybe it'd be very interesting to kind of look at all the sub scales in the Hamilton. See if we're seeing factors in the brain signal that correspond to different, different sub-factors. And the Hamilton score. We've looked at that a little bit. I can't say anything groundbreaking has come from it. But yeah, I mean, placebo effects are real. I mean, these patients, they desperately want to be better. They, if they want this to work, this is a treatment of last resort. They're working closely with the clinical team and they want, they're willing subjects, they want to please their clinical team. So there are all sorts of reasons that you can get placebo effects here to be very interesting to know. Right now we have no placebo patients, we have no Shams in this group. Be very interesting to know if we had that data. Whether when we see potentially placebo induced improvements in the sham group, if their brain is also moving in the same direction that the treatment group is moving. But we don't have that data. It would take another clinical trial that get them done and go for it. Just because I think that's a repeat the question. Yeah, I'll do better at that. I'm sorry. I people and yet actually we could hear the question pretty clearly. I'm wondering. Do you want me to ask Hey, Simon? Is that yeah, go ahead and start. I can hear you now. Okay, Great. Hey Chris. Thanks. That was it. That was really fun. I really enjoyed the talk, although it has some some sobering messaging as well. But I was really, really interesting. I was thinking a little broadly when you were talking about this stuff that depression is one of these diseases that you were talking about that is diagnostic. It meaning that you could have a lot of different physiological origins and signatures and causes and comorbidities and such. And as you know, a lot of, a lot of diseases are like that. Migraines, things like that. And so for these kinds of diseases and problems, do you think that you could discover with some of these methods structure? Obviously, not what the small group that you did this intervention that you were talking about what, but if we got this to much larger populations, can we use some of these techniques to discover patterns and features latent, latent diseases within those sort of diagnostic, broader diagnostic categories. And it's anyone thinking about doing that? Yeah. All right. Space. And it's a really, really great question. Time. And so, you know, I mean, at one thing, even if you just narrow your question back to depression for a second. Yeah, there's a core question there which is, is every patient different? Or are there other phenotypes that we can group people into that would respond to similar treatment and they have similar symptom clustering in some ways. The last work that I showed is one attempt to sort of say, maybe we can talk to them behaviourally and find some symptom clustering, rather than just using one dimensional numerical scales and understand what the phenotypes are. I think the data, maybe we could look at that, we certainly haven't. And as you point out, it's probably too small of a group to do that. I will say the next generation of devices from Medtronic has more steer ability on the current. And so it will open up an adjacent question which is, you know, we don't know that you need actually all four of those bundles. And it's possible that different phenotypes of depression or different co-morbidities with the depression would correlate to different parts of the network. And so in a larger trial with more steer ability, It's something that you maybe could dissect out a little bit. And we're certainly not there yet. But I think that is a super, super interesting question, definitely on my radar for for next steps. Thanks. One last question here. Yes. Yep. Yeah. My my question is in relation to you mentioned earlier, beyond the statistic were based on socioeconomic status. Depression overwhelmingly affects those in poverty. I was curious to know kind of what the demographics are for this particular population. And yeah, I guess if that is homogenous, what sort of efforts can be made to make sure we include those of low pop up in poverty. Again, these sorts of trials to try and understand differences that we observe in a large-scale clinical trial? I'm certain that I believe I am unmuted, yes. In a large-scale clinical trial, I would hope an imagined that would be part of what they try to achieve and getting a representative sample in this population of six. I actually don't know what the socio-economic variability is. I know of at least a couple of other patients in the next cohort that I didn't report on that are in very challenging economic situations. Of course. By the time they show up at our door, many of them have some affluence because they'd been seeing doctors and doctors upon doctors. They've gotten to the point that their treatment resistant, they haven't abandoned the notion of therapy. Thank you. For most patients, they have some level of resource in order to navigate through that. I can give you a chart, the demographics, otherwise here there's a pretty broad spectrum of age range. I believe the six that I showed you were for female to male, which is broadly reflects the gender the gender bias that you see in depression diagnoses anyway. But I don't have socioeconomic data on them. I would guess that it is more homogeneous than what one might like in a large-scale clinical trial. But good question, certainly important.