Good afternoon, so we're going to switch gears and talk about traumatic brain injury, TBI. I'd just like to acknowledge the SCCM for inviting me, as well as the co-moderators of this session, co-panelists who are colleagues and good friends. So these are my disclosures, non-relevant to this talk. So when we talk about the objectives of this talk, one of them is to distinguish the causes and consequences of heterogeneity, this is going to be a recurring theme in this session. We want to learn that solving heterogeneity could be a core objective for achieving precision medicine in general, but also for patients with TBI, and also make the specific inference that subphenotyping is a very powerful tool to enable biologically anchored reclassification. So I work at Johns Hopkins, which is located in Baltimore, Maryland. We have several different campuses, the primary one is on Homewood, which is both for arts and sciences, and then we have also a medical campus shown as the dome on the left, and in the middle is the Peabody Conservatory, which is for music, which is very dear to my heart since several of my children learned their instruments in this place. So when we talk about TBI, this is a big global problem. It's estimated that between 30 and 69 million cases of TBI occur annually across the world. It is the leading cause of neurologic death, as well as disability. In the US alone, it's estimated that 176 people die of a TBI-related death every single day. So this is a considerable problem. And the issue is that we don't have any treatments, right? There are no TBI-specific therapies. We're not very good, actually, at detecting TBI, especially in the milder cases and concussions. We certainly don't have a biologically anchored classification scheme, and we're not good at predicting both either short-term or long-term outcomes, and also not very good at delivering cost-effective care. So when we look at prognostication, the very best predictive models can discriminate between outcomes with an AUC of 0.8 to 0.85. So that means 15% to 20% of the time, these models fail to accurately discriminate between outcome categories. And this is maybe not so surprising, because the outcomes after TBI are complicated, right? They're the result of biological factors, psychological factors, social factors, ecological factors that contribute to the ultimate outcome. If we look at the right of this slide, you can see that there are a number of variables that have been implicated in the natural history of TBI, including genetic sources of variants in the host response, the secondary injury that occurs in TBI. And then after TBI, there's this very interesting paradigm where people are both subject to plasticity, where the brain is trying to repair itself, but also to neurodegeneration. And so these two are competing with one another over the course of the weeks and months and years after TBI. So if we focus even just on the secondary injury, this is an extraordinarily complicated affair, right, with multiple different signaling mechanisms. But the primary sort of signature injury of TBI is this stretching or shearing of axons that occurs with rapid deceleration or rapid rotation, right? So these axons in the central nervous system don't like to be stretched in this way. And as a result, they express a number of different markers and signaling molecules that end up in either neuronal dysfunction or neuronal death. And this is only looking at neurons. A very, very sort of analogous process is going on also with glial cells, with astrocytes and oligodendrocytes. So how do we measure treatment effect in TBI? I like this very simple metric, which is the number needed to treat. You recall from Statistics 101 that this is the reciprocal of the absolute risk reduction. And so if we consider some of the therapies that have been examined in large multisite randomized control trials of TBI, one of the ones that's really attracted a lot of interest is decompressive craniectomy. So if we take severe TBI patients, either who are prone to elevations in ICP or who already have established ICP, and we remove the skull, are we having a measurable impact on their outcomes? Well, unfortunately, the results of existing trials are very inconclusive. So the number needed to treat to prevent one death goes from 16 to infinity, infinity meaning that there are some trials that were inconclusive. If we look at hypothermia, a huge amount of interest in the possibility of temperature management as a therapeutic tool to treat patients with moderate and severe TBI. It turns out that it doesn't work. The number needed to treat is between 22 and, again, infinity, and I use that infinity carefully, but it's true that there are a number, in fact, the most recent very large randomized trials of hypothermia for TBI work inconclusive. Another drug that attracted a huge amount of interest, two very large studies published in the New England Journal of Medicine a few years ago looking at progesterone as a neuroprotective agent, and what we know is that the NNT is even higher. It's anywhere between 50 and infinity, right? So 50, in the best case scenario, that means you have to treat 50 individuals with progesterone to prevent a single death, all the way to infinity, which is probably the reality, which is that we know that it doesn't work in undifferentiated populations, and so this brings me to this often quoted citation, which is that traumatic brain injury might be the most complicated disease of the most complex organ in the body. It's quite a challenge. The question is, is it time for a new paradigm? I think many of us agree that heterogeneity is probably the biggest issue facing both TBI researchers, TBI patients, people taking care of TBI. Heterogeneity of treatment effects is really only the manifestation of underlying biological heterogeneity, and it's imperative that we figure out ways to solve this very complicated issue, right? So if we consider where we are right now, which is the one-size-fits-all medicine, where we focus on clinical syndromes, where our classification schemes are non-mechanistic, one step forward might be a stratified approach, which is what this session is about, where we might target therapies to specific subgroups of patients who may be more homogeneous than these undifferentiated samples. But the ultimate goal, of course, is to achieve precision medicine, where we can tailor therapies to the individual characteristics of each patient. And of course, there's a lot of thinking around how we can achieve precision medicine, but it looks sort of like this, where we combine domain knowledge with high-dimensional data, with statistical modeling, and we achieve more accurate prediction and classification. We make better decisions, and we achieve better outcomes. So in the realm of TBI, there are two very, very large-scale prospective observational studies that have been conducted just in the past decade, looking with the aim, essentially, of achieving precision medicine for patients with TBI. These are the Center TBI, funded by the EU, which concluded enrollments in 2017 with more than 5,000 patients, and TRACK-TBI, which is actually run out of UCSF here in San Francisco. It's a multi-center American study with a little bit more than 3,000 patients. Very, very detailed, deep phenotyping of patients, all of whom underwent imaging, extensive imaging, many of whom had genomic analysis, biomarkers, you name it, as well as longitudinal trajectory analysis. So when we think about TBI classification or reclassification, there's really two approaches. One is empiric subgroups, so using a priori knowledge of TBI, and the other is to employ data-driven methods, sort of agnostically, which is what Dr. Sinha was mentioning for ARDS. So empiric TBI subgroups are the ones that many of you are familiar with. It includes looking at the mechanism of injury. Is it blunt, penetrating? Is it due to deceleration? Is it blast injury? The severity, it's very commonly used. We talk about mild, moderate, and severe TBI based on the Glasgow Coma Scale. The clinical pathway, which was the stratification used in the Center TBI study, where they looked at patients in the ICU or patients in the hospital, patients seen in the emergency department and sent back home. The pattern of injury on imaging, for example, contusions versus traumatic subarachnoid hemorrhage, et cetera. This is an example of the heterogeneity of CT scans on admission in patients with TBI, and you can see across the spectrum of severity, there are many, many different permutations. If we look at the stratification by the Center TBI criteria here, you can see the outcomes as measured using the extended version of the Glasgow Outcome Scale, and you can see that it tracks more or less. Patients admitted to ICU, as expected, had worse outcomes. The problem with empiric classification is that it doesn't often capture the biological complexity of these patients, and empiric subgroups are often poorly predictive of outcome or of treatment response, and so we do need other methods, and this is where we come in with data-driven subphenotypes. Here the methods are model-free, they're agnostic. We leverage probabilistic unsupervised machine learning techniques, such as clustering latent variable models. The limitations here is that you need a lot of data, and you need high-quality data to train the models, and the other problem is that the outputs are correlative. You can never establish causality using these unsupervised machine learning techniques. The sources of data are diverse. A lot of the research has been done using simple electronic health records data or imaging or biomarkers, but I think the future is very bright because we increasingly will have access to the multi-omics data, the immune signaling data, the high-frequency physiological time series, as well as quantitative phenotyping, which is something that we're doing in my lab. The sort of endgame, in a sense, or the modeling framework is to leverage all these different types of data using, again, this unsupervised learning, which consists essentially in exploring data sets without any specific labels, to identify these specific more homogeneous subgroups, and then to generate disease insights and more effective therapies. So just to sort of run through a couple of studies, this is some data from Johns Hopkins. We've taken data from Center TBI, from our own Johns Hopkins precision medicine analytics platform, as well as data sets that are publicly available, such as MIMIC and EICU, and we've derived clinical variables, physiological time series, the variables, imaging, biomarkers, and we're running these through hierarchical clustering analysis algorithms. And so here are some preliminary results. These data are not yet published, but we did identify four distinct clusters characterized by differential survival rates and differential neurological outcome rates as well, and this was found in EICU, and it was confirmed in Center TBI. And then when we looked at the specific features that were driving the subphenotyping, we saw that they actually were very, very well separated. So here we're comparing subphenotype A, which had the best outcome profile, and subphenotype D, which had the worst outcome profile, and you can see across these different features a number of differences that are quite marked between the different groups, and this is a so-called standardized mean difference plot. In here, there's a heat map showing the similar information you can see, so the individual rows represent different input features, and the columns represent individuals in the study, and you can see that we've separated subphenotype A, B, C, and D, and you can see that the magnitude of expression of these different predictive features was very, very different across the phenotypes. This is data from Center TBI. In this case, the input features were imaging, demographics, clinical severity. The group identified four subphenotypes, each associated with specific outcome probabilities. The most important drivers of the subphenotypes were the cause of the injury, whether there was major extracranial injury, and the level of consciousness measured using the Glasgow Couma scale. This is a more recent study, also coming out of the Center TBI group. The inputs were demographic, clinical, physiological, laboratory, and imaging. This group identified six subphenotypes and analyzed them in detail, and they were able to identify specific groupings in terms of level of consciousness on admission, as well as whether the degree of metabolic stress that was occurring within these different groups. So this was a very, very interesting recent study conducted by a group from Sweden. And here, you see the patterns of physiological abnormalities across these six different TBI subphenotypes. So the question then becomes, what do we do with this? So we identify subphenotypes, where do we go from here? As Dr. Sinha was suggesting, there's a lot of work to do. Almost everything that I showed you today was retrospective work. In fact, a lot of the work I think that Dr. Sinha mentioned, as well as some of the work that Dr. Seymour is going to be showing, is retrospective. So we need to prospectively validate these subphenotypes. Prospective validation will allow us to demonstrate biological veracity. It will allow us to determine whether we can use these subphenotypes for treatment selection, so this is predictive enrichment, and whether we can use them also for prognostication, which is prognostic enrichment. So in summary, I think we've seen that TBI carries a very high global burden of death and disability, and there's no known therapies right now. Non-conclusive trials reflect, of course, underlying heterogeneity. However, the good news is that there are distinct biologically anchored TBI subphenotypes, which can be isolated using unsupervised machine learning, and research should focus on demonstrating the value of these subphenotypes for treatment selection and trajectory prediction. So with that, I'll stop, and I'm happy to take maybe a pressing question, or I'll stop at the end.

traumatic brain injury

precision medicine

subphenotypes

treatment selection

outcome prediction