Thank you, that introduction you heard about me was not true, other than that I work at Wash U. So I'm gonna talk to you about subphenotypes in ARDS. I have some disclosures, but none relevant to this talk. Whoops, how do I go back? You know what, that talk, that slide wasn't very important, it's just disclosures. Okay, so here's the question. I'm gonna start by asking the question, what is ARDS? And really, this is a philosophical question. It means many things to many people. And I'm sure if I took a vote, all of you would say different things about it. But here's what we understand of it as a clinical entity. And these constellation of symptoms are what describes ARDS. This is a Berlin definition. And actually, it's fairly frequently encountered in critically ill patients. And if we look at its component parts, the PF ratio is quite clunky. Chest X-ray interpretation is very, very junky. And unsurprisingly, we get a lot of heterogeneity subsumed in our population when we classify patients based on this. And importantly, I think it's important to acknowledge that there is no link to etiology, direct link to etiology or biology. And so the other question to ask is, well, why is heterogeneity a problem? Well, partly because we have had a bunch of negative trials. This is a list, and by no means is it exhaustive, of those interventions, biological interventions we've trialed in RCTs. And all of these are negative. And so what is phenotyping? I was broken down very basically. It's trying to get a population that's heterogeneous and finding subgroups within it that are more homogeneous, and then you study them a little bit better with the end goal of trying to treat the right patient at the right time. And really, it's a pathway to precision medicine. And I'm gonna lean on one of my old mentors here. This is Hippocrates. He's all our mentors. And actually, you'll find that precision medicine, it's a very old phenomenon. And really, what Hippocrates described was to take illnesses and break them into the component parts. And if you see here, medicine is not absolute. Thus, its direction cannot be generalized to everybody. Each human body and organism is different and responds differently to therapy. And the physicians should choose the appropriate treatment depending on the patient's individual characteristics. And I believe we all probably do this individually. However, how our syndromes manifest themselves is really quite generic and doesn't quite capture what we really do at the bedside. And phenotyping isn't really old in ARDS either. There are multiple pockets of phenotypes that have been described in literature. This includes histology, etiology, radiology, clinical phenotypes, and in physiological phenotypes, some of them have led to success in terms of randomized controlled trials simply using the PF ratio. People have broken down phenotypes of ARDS and shown efficacy of therapy, prone positioning. The PERCEIVA trial is a good example of that. But there aren't that many specific biological phenotypes of ARDS. And I have got one more disclosure that because of time constraints, I'm mostly going to focus on the work of our group, which is looking at multivariate approaches to a systems biology problem. And here's another very old mentor of mine, Dr. Carolyn Kalfi. Very young, yeah. And really the idea here was to take a systems biology approach and to take multivariate, multivariate solutions to this complex problem of ARDS. And the idea was to use unbiased approaches to let the data do the talking. And here we use latent class analysis, probably a little bit too complex for a Tuesday afternoon. So, is it a Monday afternoon? Monday afternoon, you see, it's that complicated. But really what we did was we used a composite of biomarkers and clinical data to try and uncover what the underlying data structures were in randomized controlled trials. Importantly, a hallmark of a good phenotype should be that it should be useful. And that use may be biological or clinical. It should provide some novel information or at least some treatment solutions. And as these phenotypes get discovered, one of the things that you have to show is that they're robust, reproducible, and generalizable. And finally, another important step, which is really where the field is heading, is to try and identify how we can implement these phenotypes at the bedside. And so we, using latent class analysis, have identified two phenotypes of ARDS. They're called the hyper and the hypo-inflammatory phenotype. What you see on the red line here represents the hyper-inflammatory phenotype. On the blue line, the hypo-inflammatory phenotype. On the x-axis are all the variables that we used in the model to try and identify these phenotypes. I will emphasize again, we don't use outcomes like mortality or the interventions as part of the modeling. And on the y-axis, you have the standardized values for each of these variables. And to cut a very long story short, we call the hyper-inflammatory phenotype because it is associated with elevated levels of inflammatory biomarkers such as IL-6, IL-8, soluble TNFR receptor 1. Conversely, it was also associated with lower levels of bicarbonate, protein C, and platelets. Across sort of multiple clinical trials now, we have identified about 30% of the population is hyper-inflammatory. And it'll perhaps come as no surprise, given the profile, the biological profile of these patients, that mortality is consistently and significantly higher in the hyper-inflammatory phenotype. And subsequently, we have gone on to show that these phenotypes also exist in observational cohorts of ARDS. And more recently, in a pediatric cohort of ARDS. It seems to be conserved in adults as well as children. So what's so exciting about another prognostic schema? I think part of the reason why people have taken an interest in this work is that we've observed differential treatment responses to these ARDS phenotypes in the alveolar trial, which tested high-PEEP versus low-PEEP. We observed differential treatment responses. I will touch upon that a little bit later. And in the FACT trial, which looked at conservative and liberal fluid management in ARDS, we also observed heterogeneity of treatment effect with more fluids being beneficial in the hyper-inflammatory and a conservative fluid being more beneficial in the hypo-inflammatory phenotype. And in the HARP2 study, which tested the efficacy of simvastatin versus placebo in ARDS, we noted that with simvastatin therapy, there was a survival benefit in the hyper-inflammatory phenotype compared to placebo, which wasn't observed in the hypo-inflammatory phenotype. And all of these trials were negative and showed no benefit. And it's also important to emphasize that when we split this population up by markers of severity, other markers of disease severity, such as PF ratio or the Apache scores, we didn't see the same treatment effects. And very quickly, I'm going to touch upon other people who have also used similar types of biological schema to try and identify phenotypes. This is work from Luebos and colleagues, and they used a slightly different approach to what I've shown you previously. Their approach was to use biomarkers only. They used k-means clustering as opposed to LCA, which is a sort of more of a data science approach rather than a modeling approach. They also identified two phenotypes to best fit their population. It's called the reactive and uninflamed groups. Mortality in the reactive group was higher because it was, again, associated with higher protein biomarkers associated with inflammation. And they have shown this schema exists both in ARDS and patients who are mechanically ventilated but haven't been diagnosed, haven't really met the diagnostic criteria for ARDS. These are kind of similar to the hypo and hyperinflammatory phenotype, reactive being hyper, uninflamed being hypo, but they're not quite the same. So the next logical step after we've shown, I hope you agree that this is robust and generalizable, that how do we implement this at the bedside? And this is slightly more complex than it sounds, in part because these LCA models are quite complicated. They comprise of 20 to 30 variables, in some cases a few more. And the other issue is that these variables are standardized to the population mean of that cohort. So when you're trying to apply it prospectively to a single patient at the bedside, that becomes a little bit complicated. So to counter that, we've developed some more simple models and I'm going to show you exactly how we've done that. The other problem is that we don't really have an easy way to quantify protein biomarkers at the bedside. So to counter that problem, we have built classify models that only use clinical data. So in terms of the parsimonious models, what we did was we used an array of machine learning approaches to try and identify the features that best classify the patients. And here I've shown you a three and four variable model, including IL-8, bicarbonate and protein C in its simplest and best model. And this model had an AUC of 0.95. And if you add vasopressor to it, the AUC was 0.95. So they're reasonably good performance metrics in terms of classifying the LCA phenotypes. We have also described a multitude of other potential permutations because at the time when we were developing these models, we weren't really sure which of these protein biomarkers would be available as a bedside test. So we've offered the field a few choices when it comes to developing such a device. And importantly, when we used one of these parsimonious models in the HARP2 study, which I'd shown you previously, we were able to observe the treatment benefit in the hyperinflammatory phenotype with simvastatin, and this was statistically significant. And in terms of feasibility over here, this is a very small study that we did with Danny McCauley and Thomas Zakimeni. And this was a small study in two UK centers of patients with COVID-19. Unfortunately, because of handling fluids, the study got kiboshed very quickly. But what we did was we used one of these parsimonious models based on a novel point of care device which quantified these protein biomarkers, IL-6 and soluble TNFR receptor one, in about an hour to classify patients into the hypo and hyperinflammatory phenotype in real time. Now, this needs proper validation, and the FIND study, which is being led by Danny McCauley, is in the process of doing that, where they're collecting these samples and using these devices in hundreds of patients in the UK. And so, moving on to the next approach, and here I'm very grateful to Dr. Matt Chirpek, who pointed us towards XGBoost, which is a terrific modeling algorithm which we applied to only clinical data. You can see these data up here. And these are all routinely collected data in most ICUs in critically ill patients. And using these variables as predictors, we were, again, able to classify these patients in a holdout validation set with reasonable accuracy. AUCs were, again, 0.93 and 0.95. And when we applied it where alveoli was the holdout randomized control trial to test this model, to test the trained model in, we found that mortality in the hyperinflammatory phenotype was 45% compared to 20% in the hypoinflammatory phenotype. Whereas in high PEEP, the mortality was higher in the hypoinflammatory phenotype, and it was lower in the hyperinflammatory phenotype, and this was statistically, this treatment interaction was statistically significant. And really, this mirrors what we found with the latent class analysis. So to reiterate, in this clinical classifier model, we only use clinical data, and it compared quite well with the latent class analysis findings. So subsequently, we have used this clinical classifier model further. This work has been led by Manoj Maddali, who is a pulmonary and critical care fellow at Stanford. And we wanted to evaluate how these models performed in observational cohorts of ARDS. And then we took it and pointed it at LungSafe, which is one of the largest observational cohorts of ARDS that's been ever collected in our field. And important point to make here is that in valid and early, we had the gold standard so we could look at the model performance, whereas in LungSafe, we were just evaluating how it performs as a clinical tool. And here's what we found. The AUCs were, again, pretty impressive. Importantly, the biological information that we wanted to capture was captured by these phenotypes using the clinical classifier model. In LungSafe now, we found that 26% of the patients were hyperinflammatory, so similar to what we've observed before, with mortality approaching 60% compared to 33%, and fewer ventilated free days. And I'm gonna show you some data where we stratified these patients, depending on how much PEEP they received between days one and three, so the mean PEEP that they received over days one and three. I will caveat this, that this was an observational cohort. However, we got the same signal that we did in the alveoli trial, where high PEEP was associated with a survival benefit in the hyperinflammatory, but is associated with harm in the hyperinflammatory phenotype. And this treatment interaction, adjusted treatment interaction was significant. So as I wrap up, what are the next big questions? Are these phenotypes of ARDS? I think logic would dictate that given that our measurement is mostly made in the blood, this is probably a systemic signal. So the question is, can we translate this to other non-ARDS syndromes? I think the work that I showed you from Louay Boss and Georgios Kitsios at Pittsburgh has also shown that it's applicable to at least people who are mechanically ventilated with acute hypoxic respiratory failure, and we are asking the same questions in sepsis. Are these molecular phenotypes stable over time? And that's another important unanswered question, particularly as we phase this into a potential clinical tool. And a question that I keep asking myself and other people, and I hope somebody will answer is, when can we do a phenotype-specific clinical trial? So to summarize, I hope I've shown you that there are two distinct, consistent phenotypes of ARDS based on circulating protein biomarkers. We have gone some ways to describe clinically implementable model. It provides a route to prognostic and potentially predictive enrichment in future clinical trials. But all of this is hocus pocus and smokes and mirrors until we have shown that we can prospectively identify these patients and show that there is a measurable treatment benefit. And until then, just this remains a very fancy research idea. I want to acknowledge many, many people, but these are just some of the people. I want to acknowledge Dr. Kalfi, of course, for all her guidance through these studies. And I really thank you for your attention.

subphenotypes

ARDS

acute respiratory distress syndrome

phenotyping

precision medicine