So, my name is Usma Samadhani, and I'm really grateful for this opportunity to talk about state-of-the-art for neurotrauma. My topic today is a welcome to robots, and it's about how big data and algorithms are going to transform brain injury care in the very near future, primarily from a diagnostic perspective, but ultimately also from a treatment perspective. So, I'd like to thank the AANS for giving me the opportunity to share this talk with you. I'm from University of Minnesota. I am in private practice, and I also work at the VA. These are my disclosures. The ones that are relevant to this are the intellectual properties related to assessment of concussion and brain injury and assessment of chronic effects of neurotrauma and dementia, as well as to treatment of intracranial hemorrhage. So, the current classification scheme for brain injury, if you Google brain injury on Wikipedia, basically it's mild, moderate, and severe, and it's based on Glasgow Coma Scale and duration of amnesia, as well as duration of loss of consciousness. And the problem with this scale is that it's flawed. Basically, loss of consciousness can occur for many reasons. There can be intoxication or polytrauma, and lack of loss of consciousness does not equate with milder injury. So, we know, for example, that some patients who come in with an impaired GCS or an impaired level of consciousness, they don't have a brain injury at all. They have something else wrong with them. And then other patients come in, and they say, well, I'm not even really sure I hit my head, but I definitely didn't lose consciousness. And then they turn out to have a brain injury with sometimes potentially even fatal consequence. And this is an example of that. This is a 37-year-old woman who fell at home. She had absolutely no symptoms at the time of her fall. She subsequently, over the next two weeks, developed a mild headache and then found she had word-finding difficulty. She remained GCS-15 the entire time, so her GCS was normal. She had no amnesia. She had no loss of consciousness. And ultimately, she developed a chronic subdural, and she essentially had what could potentially be a fatal brain injury if it weren't treated. And so, obviously, this is tricky, because if someone has a brain injury that could potentially be fatal, and we're not even classifying it as mild, moderate, or severe, then we're missing out in our classification scheme. So, what makes brain injuries so hard to diagnose and define? Well, the first part of it is that no two brain injuries are the same. You can have patients presenting with very similar symptoms, but they have very different underlying pathophysiologies. So, we know, for example, that when someone gets hit in the head, they often have a comorbid neck injury, and the neck injury can cause similar symptoms to brain injury. They can both cause headache. They can both cause dizziness. There are many other cerebellar symptoms that are classically associated with neck injury. In this particular case, you can see that the patient actually has a spinal cord injury. About 17% of brain injuries are comorbid with spinal cord injury. There can be inner ear dysfunction. There can be vestibular dysfunction coming from the inner ear. There can be endocrine dysfunction, so you can have a patient who flexes and extends rapidly and they shear their pituitary away from their hypothalamus, and that can present in a delayed fashion. So, this is a patient who comes into the emergency department with a normal head CT and then is found to have a bigger problem. Cortical spreading depression, this is classically one of the problems that gets worse with time. So, the patient may present to the emergency room not having a lot of complaints and then subsequently develops abnormal, basically electrophysiology, where they have an area that is having sodium channels or potassium channels that are not functioning normally. And then through what's called kindling, this actually spreads, and it's called cortical spreading depression or spreading depolarizations. And that is one of the most common causes of post-brain injury headache and seizures. And then there's what I call the neurosurgical spectrum, which is the things that we get called for, scalp injury, skull injury, bleeding on the surface of the brain, above or below the dura, within the brain at subarachnoid hemorrhage. Traumatic subarachnoid hemorrhage is one of the most common findings on a head CT after brain injury, intraventricular hemorrhage, diffuse axonal injury, anoxic brain injury. The problem with anoxic brain injury is it can be comorbid with other injuries and it may not be detectable. And obviously, the prognosis is very poor. The reality is, is brain injury is like a salad and these are all mixed. They may be present to varying extents all in the same patient. So obviously, when you have a problem this complex, you're going to need complex algorithms to solve it. What else makes it hard is the fact that no two recoveries are the same. This is Kevin Pierce, the famous snowboarder who would subsequently go on to have a brain injury during training for the Sochi Olympics. They made a movie about his recovery. And what's very clear from that movie is that Kevin was not a normal person. He had superhuman resilience. He was much, much more energetic than the average person pre-injury and post-injury. And that suggests that there are both genetic and environmental contributors to recovery. We know there are at least 12 genes, catecholamethyltransferase, alleles, BDNF levels, APOE receptors, retinoic acid receptors that impact recovery. Genetic risk factors, these range from pre-injury, so the genes that make you have an accident in the first place, the ones that predispose to risk-taking behaviors impact your likelihood of sustaining a brain injury. Some call them Kennedy genes or risk-taking behavior genes. Then there's the immediate impact genes, the ones that affect cerebral edema in the immediate time after the injury. There's the delayed impact genes. These are the ones that cause secondary edema and problems in the three to seven days after injury. And then finally, there's the long-term impact genes, and these are the ones that predispose to poor recovery and sometimes chronic effects of neurotrauma. The other problem is that not every patient who has a brain injury was actually hit in the head. We know that blast is one mechanism by which patients can have a brain injury without having a direct blow to the head, but there are others. A blow to the body can cause a change in venous drainage, so it increases abdominal or thoracic pressure or both, and then it decreases jugular venous drainage so that you have a transient elevation in ICP, and we've seen this. And then there are also things like flexion extension injuries that can cause acceleration and deceleration of the head and lead to what's called an inertial brain injury. So it's not always clear from the history that someone has a brain injury. Ultimately, neither imaging nor loss of consciousness tell a whole story. So a physiologic brain injury is the elephant in the room, and a portion of it is visible on imaging. That's the part in yellow, and some of it is associated with loss of consciousness, but not necessarily all of it. And then obviously we don't know all of the pathophysiologic components of the elephant in the room. So there's inflammation, edema, diffuse axonal injury, endocrine dysfunction, spreading depolarizations, iron toxicity, hypoxic or anoxic injury, and astroglial scarring that can all be present to varying extents, and some of these actually get worse with time. So things like astroglial scarring and spreading depolarizations can get worse over time. Steam deposition resulting in iron toxicity can lead to cellular apoptosis, which obviously gets worse over time. So what do you do when you have a super complicated problem? Well, historically in science, we've always used conventional approaches. We make a hypothesis, and then we test our hypothesis, and then we evaluate whether or not that hypothesis is true with conventional statistics. And this was great for things like the discovery of insulin, because obviously it was sort of a relatively simple problem compared to the complexity associated with brain injury. When problems are super complex, the better way to solve them is probably using more modern techniques such as machine learning. Ultimately, the machine comes up with a complex explanation of the data that does not require human understanding, and there may be factors in there that we would not necessarily have predicted or understood. And we have to test this, obviously, on an independent data set. But the classic example of this is the car that drives itself. It learns from its mistakes, and it learns from the experience of other cars. So the capability of a machine to mimic human intelligence, ultimately artificial intelligence, is what's going to be helpful for these super complex problems. And I would say three of the most complex problems in the neuro field are headache, chronic effects of neurotrauma, and dementia. And my talk is basically focusing on brain injury and chronic effects of neurotrauma. But you can imagine that some of what I say could be applicable to those other problems. Where has this been useful in the past? Well, something like face recognition software. So when you go to the grocery store, you'd think that the reason they don't suspect you for stealing anything is that you don't look suspicious. Well, it's actually much more complex than that. Not only do they use face recognition software to know exactly who you are, but they know how much you spend every time you go to the grocery store and if you're an unknown person. So this enables consumers to spend less money on security at the grocery store. But imagine if we had a technology like this where someone could come into the emergency department and we could characterize their brain injury using this type of knowledge. So I would argue that robots will ultimately augment and then replace doctors. And this will begin with diagnostics, physical examination, radiology and pathology, and ultimately include therapeutics such as surgery. Diagnostics will be first. So something like a CT scan, right? A CT scan is currently read by a radiologist. In Europe, the BLAST CT algorithm was recently published. And basically what this is, is a deep learning image segmentation tool for traumatic brain injury in 3D CT scans. And it was published based on the center TBI data. And basically what it does is it enables us to analyze four different aspects that are classically associated with brain injury in an algorithmic way on CT scans. So things like edema and hemorrhage in various spaces are all measured using this technology. Other things that may potentially contribute to classification of the nature of injury are serum markers. So here you can see that this is from the TRACK TBI study and it's a review put out by that group. And basically what it shows is that different serum markers are associated with different types of cellular injury. So neuronal cell body injury is associated, for example, with UCHL1, whereas an axonal injury is associated with other proteins being released. An astroblial injury is associated with GFAP. And then once an injury occurs and the proteins cross the blood-brain barrier, you may get antibodies to them that you can detect later. So one might contemplate characterizing the nature of injury on the basis of these serum markers. These are present in various phases from acute to subacute to chronic. Basically you have patients who have an initial injury, it's exacerbated very acutely, and you see changes in serum levels of particular proteins very acutely. And then over time, these gradually resolve and other factors are released in response to the injury. So how is this helpful for us? Well, we can look at different types of injury and characterize the nature of the injury on the basis of serum markers. So here we take approximately 100 patients. Some of them have sustained a trauma. Those are in blue. Some have sustained a spontaneous hemorrhage. Those are the orange triangles. Some have had a cardiac or respiratory arrest. Those are the asterisks. Some have had a negative head CT after a high-velocity trauma. Those are in purple. And then others have not had any trauma at all, and those are in green. And so you can see that they cluster differently based on their mechanism of injury. And so you can classify the patients. So if you take combinations of serum biomarkers and CT analysis, you can actually determine what type of injury someone had, whether it's spontaneous hemorrhage, cardiac or respiratory arrest, or CT, negative high-velocity trauma, which is consistent with sort of an inertial injury or an accelerational injury. And you can characterize the nature of the injury. And one might imagine that this would allow you to create a classification system that might potentially help you differentiate which patients would benefit from particular therapeutics. This is another study that can be used to predict severity. So this is looking at patients who ultimately progressed to brain death. So brain dead versus CT positive. So controls are CTL. Non-traumas are NT. Those are patients who presented to the ER with a headache. CT negative are traumas that are CT negative. CT positive are traumas that are CT positive. And what you can see is that when you compare all of those other groups versus the ones who progressed to brain death, that you can actually predict it on the basis of either serum markers alone or serum markers plus head CT with a relatively high sensitivity and specificity. So it's possible to prognosticate an injury on the basis of serum markers and CT scan without ever having a physician assess the patient. And if the patient progresses to brain death due to cardiac or respiratory arrest, you can similarly characterize the nature of the problem solely on the basis of imaging and algorithmic imaging analysis and serum markers. So one could see that potentially these types of biomarkers will help classify the nature of injury and predict who is going to do poorly. So there are differences. This is the same data presented in a different format. But there are differences in UCHL1 and GFAP between brain dead, brain injured, and controls. So the brain dead patients are the blue circles. The brain injured patients are the purple triangles. And the controls are the silver diamonds. But you can see that the serum marker levels of these patients do differ. Finally, there are differences in serum markers among brain dead patients due to cardiac or respiratory arrest, high velocity trauma, and being found down. So the cardiac and respiratory arrest patients are the little blue stars. The patients who are in a high velocity trauma are the triangles. And then the patients who were found down are the squares. So one can see that you can actually classify the nature of injury on the basis of UCHL1 and GFAP. So after you use algorithmic measures to classify the nature of injury or make your diagnosis, the next step is using robots to perform your physical examination. How might one do this? Well, there's eye tracking. So with eye tracking, a person can watch a music video. The music video plays inside an aperture. The aperture moves around a remote monitor or a video monitor. And what happens is that a camera records the eye movements of a person while they're watching the video move around the viewing monitor. And it's actually quite powerful. You record the movements of the pupils over time. So you get the X, Y Cartesian coordinates of the pupils over time. And then you can characterize whether or not someone has difficulty moving their pupil up or down. So the simplest case, obviously, would be a third nerve palsy. And what happens when someone has a third nerve palsy is they are unable to move their pupil up or down because both the inferior and superior rectus are innervated by the third nerve. So that impacts box height. And then another simple case would be a sixth nerve palsy. So they cannot move their pupil out to the side, and that impacts box width. What that means is that if someone has a blow to the head, for example, and they have elevated intracranial pressure, you could possibly detect it with eye tracking. So here we show in this paper that, basically, when someone has an elevated ICP, their eye tracking becomes abnormal. So here's a patient with an ICP of 5, and you can see that their boxes are square. Then the ICP goes up. One of the boxes becomes narrow. And then the patient recovers, the ICP is no longer being recorded, and they go back to normal. Here's another example. The ICP is 3. Then it gets worse. The box gets narrow, and then the ICP is not recorded. And the interesting thing about this is that it's actually the change in ICP that's relevant much more so than just the actual number, which is consistent with data that was seen out of Europe. So eye tracking can then be used to screen or diagnose for concussion. What you can do is you can plot the Cartesian coordinates of the pupils over time. So the x-axis is time and the y-axis is the Cartesian coordinates. So when they're watching the video, the left and right eyes should be directly superimposed and in both the horizontal and vertical planes. So here's the horizontal plane on top and the vertical plane is on the bottom. And one can see that when someone is concussed, the pupils are not superimposed. They're actually moving without a complete level of coordination. And so as they recover, you can actually watch that go back to normal. So it's actually very powerful to assess when someone might have a potential concussion. And then what happens is, is that if you look at the normal distribution of healthy normal patients, eye tracking, the vast majority of them have eyes that move together. Whereas patients who are injured, they start out fairly bad. Some of them get a little bit worse initially, and then most of them get substantively better. So the distribution of patients improves over time. And this is seen not only in the emergency department setting, but also in the sports medicine setting. This is work that was done by Children's Hospital of Philadelphia and Boston Children's Hospital collaborators. This can also be useful for something like BLAST. So this is the Minnehaha Academy School. There was a natural gas explosion, and there was a blast, and there were multiple people who were injured. There were two fatalities. Seven patients were admitted to the hospital, and there were 36 exposed survivors evaluated in the lab with eye tracking in SCAT 3. The mean age was 35.6. The range was 13 to 70, and 23 of them were females. And what you can see from this analysis is that there were five eye tracking metrics that were significantly different between survivors and age-gender matched controls. And the eye tracking was able to distinguish between individuals exposed to BLAST versus those who had not been exposed to BLAST. This is the heat map of the school. So the blast occurred here. These were the two fatalities. The darker the circle, the more abnormal the eye tracking. So the people who were closest to the BLAST site had the most abnormal eye tracking. Some of the girls who were in the volleyball gym also had very abnormal eye tracking because BLAST waves travel down corridors, and they can actually be amplified by narrowing of a corridor. The kids who were outside on the soccer field were least affected because BLAST waves sort of dissipate when they're not confined. So one can imagine that those people were least likely to have major consequence. So to summarize eye tracking, basically one can use eye tracking as a cranial nerve monitor. It measures a number of metrics, including pupil size and dilation, so it functions essentially like pupillometry, assessing cranial nerves 2 and 3. And then it also assesses motility, which is 3, 4, and 6. And then finally, it assesses blink, which is 5 and 7. So what else? Now what can you do when you have all this information, right? So you have genetics that we know impacts recovery from brain injury. You have eye tracking, which is a proxy for physical exam. You've got MRI, you've got CT, and you've got serum biomarkers. You can put them together, and you can build an algorithm that will help classify the nature of injury. And the goal is to make this less biased or inequitable than current measures. Currently, we rely on physicians to assess who might have a brain injury. And the problem with that is that we know that physicians have bias. We tend to look at certain people, and we make assumptions about what their etiology is. You may have a person who was in a car accident, and they come in, and there was a beer bottle sitting next to them on the car seat, and they smell like alcohol. And you don't even scan them, for example, in the emergency department for a couple of hours because you assume that they were intoxicated as opposed to brain injured. And then sometimes it'll happen that that was a patient who actually had a primary brain injury. They were not necessarily intoxicated, and that brain injury was missed because they were not prioritized given the fact that there was an assumption that they were intoxicated. What throws this off is the vast majority of patients who come in as traumas at certain hours in certain centers are intoxicated with one substance or another. Sometimes it's as many as 50%. And I've heard from some hospitals that between 11 p.m. and 2 a.m., it's almost 100%. So trying to distinguish between brain injury and intoxicants may seem like a trivial thing, but actually it may not necessarily be. And our biases about who these patients are may cloud our judgment. Ultimately, surgery will be algorithmic and automated. So the era of doing things sort of blind or freely will change. So this is a chronic subdural. We've all seen them. We think of these as easy cases, but the reality is they're in hospital longer than a brain tumor. In the VA system, it's 9.3 days versus seven days. They have an 11 to 20% recurrence rate. The one-year mortality can be as high as 32%. And the mean survival is 4.4 years versus six years from actuarial tables or what would be expected. The incidence of chronic subdural is rising as the population ages. So as the percentage of the population at age 65 or over increases, and this is actual and this is projected, you can see that the number of chronic subdural cases is also expected to increase. And chronic subdurals will be the number one indication for a neurosurgical intervention by the year 2030. Can a machine do a smarter surgery than a human? Well, quite possibly. A machine can identify the centroid of the subdural, and then a robot can drill it. Machine learning can optimize the drill site location, and it turns out that the centroid is not necessarily the optimal drill site. This is data that's presented where residual hematoma is a percent of pre-procedural hematoma and optimization improves the drainage. So this is with the HoloLens, and you can see that the suite site for drainage here is actually not the centroid of the subdural. It's a little further forward and lower down. Ultimately, the hope is that we'll have better outcomes through automation of image analysis, and optimization of drill site and visualization. So we'll have the surgery robot, and the reality is they actually don't need to wait till morning. This will enable us to have 24-7 care because robots are unlike humans and they don't need to rest. So I think ultimately the algorithms are coming. It's not an if, it's a when. It's a matter of patient perception, safety, and cost as a function of volume and time. Complex problems demand it, and it's going to happen. So that's where we are with this. Why is this a good thing? Well, it's good because we don't have to assume that the problem or solution will be simple. So for example, there's long been an assumption that the amyloid hypothesis is what contributes to dementia, and multiple medications were trialed, and none of them were found to be particularly helpful. The takeaway from that was that neurologic decline is complex with multiple causes and not just brain injury, and the pathophysiologies and treatment of a single pathology, especially one with dubious connections to symptoms, is unlikely to be successful. At the end of the day, when you can't accurately measure and classify a problem, you can't count on being able to treat it. So that's the mistake that was made in the dementia field, and we don't want to make that mistake with chronic effects of neurotrauma. We don't want to assume the problem is simple. This was a failed trial for verrubecostat for prodromal Alzheimer's disease, and these are the sites of action of many of the failed therapeutic approaches to Alzheimer's disease. So what you can see here is that they're targeting the formation of amyloid plaques, and different medications have targeted different steps along this pathway. Atacanumab is one of the ones that has been approved. However, there's a lot of controversy associated with its approval. And what this suggests is that this pathway is not solely responsible for the development of dementia. And this is how much money gets put into these. Basically, this is from Biogen, and you can see the number of drugs that fails is incredibly high. But, you know, obviously Atacanumab was approved, and that was a Biogen med. So you know, they may have the occasional success from an FDA perspective, but it doesn't necessarily translate into a commercial success. Ultimately, I believe big data will be needed to solve these big problems. If you look at the risk factors for dementia, the largest risk factors are education. The midlife one is hearing loss. The later life ones are smoking, depression, physical inactivity. Social isolation plays a role. Diabetes plays a role. But, you know, ultimately, these 11 risk factors for dementia, many of them are modifiable. So other than the ApoE allele, the rest of them are modifiable risk factors. So we need to be targeting those from a public health perspective to solve these problems. How is big data helpful? It can help us figure out which medications are going to be problematic. This article in CNN was based on this study, 41,000 patients, 284,000 controls, and it revealed that those who took anticholinergics and other medications had a 30% increased risk for developing dementia. So are we getting closer to having, you know, sort of face recognition software for how we treat, you know, brain injury in both its acute and chronic phase? I would say yes. We should be able to classify the nature of brain injury and then treat it appropriately. Who's going to do this? Well, I would argue that it's the people who have a very vested interest in keeping their employees healthy. And that's the, you know, the large employers, Amazon, Microsoft, Google, the people who employ literally, you know, hundreds of thousands of people in the United States will have a very vested interest in developing algorithms that promote the health of their employees. As there are more and more mergers between insurance companies and pharmaceuticals, obviously there's going to be some integration. And, you know, patients are not looking to go to the doctor anymore. They want to get all of their information on Google and they want to get treated at home. And that's just the reality of what we're dealing with. What is that going to mean? It's going to mean that if patients are direct, if patients are getting treated as a direct to consumer, they don't need the doctor, they don't need the hospital, they don't need the administrators and they don't need the insurance companies. So that that's probably a good thing. When you look at the elephant in the room, it's actually the administrators. I mean, the cost of health care is is astronomical and it's not sustainable. So a direct to consumer approach is the future. And whether we accept that or not, it's essentially inevitable because this is unsustainable. How else will that be helpful? Well, it has the potential to reduce inequities, brain injuries unique among maladies. There are people who still don't recognize it as a real disease because there are so few objective diagnostics. There's guilt or blame associated with its infliction. When someone has a brain injury, they don't often the family doesn't go out and have a fundraiser. They sort of blame the victim. Well, what was he doing, driving intoxicated? These things happen. What was the kid doing riding their bike without a helmet? I specifically told them not to. Well, I thought the other parent was watching them. And so there's there's a lot of psychosocial secondary victims of brain injury that prevent this disease from being treated as effectively as it as other conditions are. And so how do we even start to fix this? You know, there's there's serious problems with who even goes to the hospital. We know that white people are most likely to be entered into the trauma databank, which suggests that African-Americans and Hispanics and Latinos are less likely to go to the hospital in the first place. Who gets treated aggressively? Well, obviously, there's huge psychosocial impacts on that. You know, if you have a patient who's sitting in their room all by themselves and sometimes they don't even have a name, they're just called John Doe and there's no family present. You know, the amount of time that a clinician spends in that room is less than if you have a patient who when you go into the room, there are multiple family members present. Sometimes they're educated. Sometimes they're on Google. Sometimes they have other friends who are neurosurgeons in other parts of the country. Often they have a religious person that they can call. That patient is more likely to get attention. There's a tendency to judge our patients. So when you have a patient that does something that results in a brain injury because of their own reckless behavior, you know, our treatment of that patient with derision or scorn or judging them impacts not only our own level of commitment to that patient, but also the level of commitment of the entire health care system, because if we don't respect our patients, why would anybody else? You know, we have to we have to take these patients as seriously as we take patients with brain tumors. Ultimately, at the end of the day, the neurosurgeons who treat these patients, the majority of us are white and male. Most of us come from highly educated backgrounds. So we don't even have the perspective of having family members who are similar in nature to our patients that we're treating. Most doctors come from a high socioeconomic status background and most doctors have cultural homogeneity. So we don't see the world the way our patients do. We think of certain behaviors as risk taking. They may think of it as completely normal. And then ultimately, there's the bias against trauma. Even academic neurosurgical centers may have no trauma specialist and the culture of medicine in the hospital, the hierarchies preclude correction of that. Who goes to rehab versus a nursing home? Well, I hate to say this, but it's independent of insurance. It's basically a function of how white your skin is. If your skin is more white, you're more likely to go to to rehab. Race is an independent predictor. So minority status is the single greatest predictor of whether or not someone will go to rehab. Insurance obviously is a predictor, but even independently. Who gets follow up care? Well, you know, we were proposing doing a study of children with brain injury. And if you want to study low socioeconomic status children, it's very, very difficult because who's going to bring them in for follow up visits? The parents cannot get off work. You know, and this is a serious problem because, you know, they they may be able to they may have vacation days and they can't use them. And, you know, if they if they stay home from work enough days, they'll lose their job. So so that becomes a very serious problem. You know, there's there's huge socioeconomic discriminators in who gets follow up care and who gets studied or research. Ultimately, many IRBs exclude patients specifically due to some factors that can lead to biases and outcome results. And then there's the the patients themselves may have distrust of research due to past wrongs, their education and cultural factors that that preclude their participation. And then obviously time travel. And not every patient has phone or Internet service. So ultimately, we in order to create solutions for this, we need diagnostics and therapeutics that are objective and unbiased and can potentially be brought right into the patient's home. We've made progress. The first step is realizing we have a problem. The solutions are socioeconomic. We need insurance companies to contribute to raising the standards of care. Thank you. And if anyone has questions, they can always reach out to me on email or Twitter. Thanks.

neurotrauma

big data

algorithms

brain injury care

diagnosis

machine learning

inequities