Thanks so much, Dr. Heise, and thanks to the organizers for inviting me today. For what it's worth, we are not de-adopting proning. I was just on service back in December, and we were still aggressively proning our patient with terrible pneumococcal pneumonia in ARDS. So my opinion, I agree. I think it's here to stay. All right, so I'm going to talk about subphenotypes in ARDS. Is it time for personalized medicine, including for COVID-19 patients? Do I click on? Left click, okay. Thank you so much. These are my disclosures, which are not especially relevant to the talk today. So what does it mean to phenotype or subphenotype ARDS? When we talk about this, and Dr. Heise very nicely laid the groundwork for this topic earlier in the session today, what do we mean? Well, it's just, I think, a fancier way for saying lumping versus splitting, which is something we've been talking about in medicine for decades, really. We can call it precision medicine or personalized medicine, but it's really the same idea of taking a heterogeneous group of patients and breaking it down into less heterogeneous or more homogenous subgroups. And just like there are many different ways to slice an onion, we can talk about many different ways to phenotype ARDS, and people have been talking about this, honestly, since ARDS was first described as a syndrome in 1967, and continuing to debate it in the over 50 years since that time. And most of the discussion has really focused on clinically obvious phenotypes, and this is just a few of the different ways we think about it, right? So ARDS risk factor, trauma versus sepsis, direct versus indirect lung injury, we've all heard that discussion, right? Severity, mild, moderate, and severe, diffuse versus focal on the chest radiograph, and then most recently, of course, COVID-19 versus typical ARDS, and this is not an all-inclusive list. But I'd say that the utility of these clinically obvious phenotypes is variable, and in some cases, controversial, and why do we care about phenotypes? Why should we be trying to do this? I think we care because we're trying to find groups of patients that respond differently to therapies, right? I mean, we can, it may be interesting academically to think about different ways to subdivide patients, but we really care because we want to find new ways to improve outcomes for our patients. So which of these phenotyping strategies have some evidence that they may be helpful in that regard? Well, I'd say we really don't have much, actually, for traditional ARDS risk factors, like trauma versus sepsis, or direct versus indirect lung injury. I think we have some pretty compelling evidence that treating patients on the basis of severity may be useful. Diffuse versus focal I'll talk about in a moment, and then, of course, COVID-19 versus typical ARDS we'll spend some time on today. So let's dive into that a little bit more. Like I said, ARDS risk factor, direct and indirect lung injury, in practice are often overlapping, right? And it can be really challenging to tease these apart in clinical practice, as patients often have numerous simultaneous risk factors. Now severity, of course, as Dr. Heise alluded to early on, is embedded in our current definition of ARDS, and clinically useful. Proning, we just heard from our prior speaker, we think is beneficial when the PF is less than 150. And of course, ECMO, we only really apply in our most severe patients. So there's an example of where distinguishing different subsets may be useful clinically. Now, what about COVID-19 versus quote, unquote, typical ARDS? I'm actually not going to talk about ventilation because Dr. Sahedia was going to talk about that, but just to touch on the issue of medications, I think, certainly IL-6 receptor blockade and JAK inhibition, we really only use in the setting of COVID ARDS. And corticosteroids, a constant controversy in ARDS, but I think we could at least say that the data is more conclusive of their benefit in COVID-19 ARDS versus typical. So there's some evidence where, you know, it may be useful to think about splitting our patients in COVID versus typical, though I have to say overall, I agree with Dr. Heise that COVID is another type of ARDS. Now, what about this diffuse versus focal ARDS? And I bring this up, not because this is something that I use clinically, but because I think the study around this I'm going to tell you about is a really nice way to try to answer some of these questions about when it is and isn't useful to phenotype patients with ARDS. I'm going to tell you about the live trial that was conducted by Jean-Michel Constantine and colleagues in France. And I think you can see the reference down there published in Lancet Respiratory in 2019. Randomized 420 ARDS patients to what they described as a personalized strategy in which there were two different ventilation algorithms based on whether the chest X-ray or chest CT showed a diffuse or focal radiographic pattern versus standard care, which was essentially a lung protective low tidal volume ventilation strategy. And it showed no difference in overall outcome. So you may say, well, this is not very exciting. But then when you actually excluded the 20% of patients that were misclassified from the personalized group, that strategy appeared to be superior to the usual care arm. And you can see that from these survival curves on the bottom. So when you did the per protocol analysis, which excluded those misclassified patients, you saw a survival benefit. And I think this brings up a couple of different really relevant points here. Number one, there is potential perhaps, if you believe this analysis, for a more personalized strategy actually being useful. But it also brings up the challenge of misclassification. These were folks who were doing a clinical trial on this with the whole idea that they were going to identify the phenotype based on their chest radiograph and CT. And still 20% of the time when it was compared to central adjudication, the consensus was they didn't get it right. So it's not easy to do this in practice. Okay, now those are clinically obvious phenotypes. Might there be other phenotypes that are less apparent that might actually be relevant in terms of how we should be treating our patients? And I'm gonna tell you a little bit about work our group and others have done on what I'm gonna call latent phenotypes or hidden phenotypes. And that's sort of subsets or patterns in the data that have been observed using a technique called latent class analysis that asks, if we see this data distribution, is it more consistent with the idea that this is one homogenous group as shown by the solid line, or actually three underlying groups as shown by the dotted lines? And with that approach, our group has identified a hyperinflammatory phenotype characterized by higher levels of plasma inflammatory biomarkers that we've seen now in about 10 different cohorts of patients including a recent cohort of pediatric patients that has consistently much worse clinical outcomes. In secondary analyses of randomized controlled trials, these latent phenotypes appear to respond differently to mechanical ventilation with low versus high PEEP, to fluids conservative versus liberal fluid management, and to pharmacotherapy with simvastatin. Now that's all secondary analysis though, right? So suggest they might be interesting in this way that I previewed at the beginning of responding differently to treatment, but we really need prospective confirmation to determine this. So is there a way to identify these hidden phenotypes at the bedside? And that's the downside of these types of phenotypes compared to those clinically obvious ones I mentioned earlier, right? That we can all see with our eyes clinically is that these are hidden and harder to identify. So we've been working over the past few years to try to develop approaches to identifying these phenotypes so that we can actually test a personalized medicine strategy in our ICUs. And this is work that's been led by Dr. Pratik Sinha who's a junior faculty member at Washington University in St. Louis. And what Dr. Sinha found is that we can actually use just a handful of variables, three or four variables, as you can see here, to identify these phenotypes compared to the gold standard of phenotype by latent class analysis, which is that complex modeling system I described earlier. You can see that when these three or four variables are used compared to the gold standard, it might be kind of hard to see. Sorry, guys, that's kind of small, but the area under the receiver-operator curves here on this figure on the left is 0.94 to 0.95, which suggests that these perform with pretty good accuracy, actually, for classifying patients. It turns out that multiple different models actually performed quite well. So this is a table that shows the top six variables which were useful in classifying these hidden or latent phenotypes and the area under the curve in a validation dataset. And you can see that different combinations of three or four of these variables actually performed pretty well for identifying these phenotypes. All right, so let's get back to COVID, right? Because that's kind of the rationale here is how do these phenotypes perhaps apply to COVID? Are they relevant at all? And so I'm gonna show you some work that was again led by Dr. Sinha in collaboration with Danny McCauley from Queens University, Belfast, and Tamas Sakmani from the University of Cardiff in the UK. And this was a very small study of 39 patients with COVID-19 ARDS that were all enrolled during the first wave of COVID in the UK. So cast your mind back to spring 2020. And what we did in this study was to prospectively classify these patients into one of those latent phenotypes that I described previously with a novel point of care assay that could measure interleukin-6 and TNF receptor one in real time. And then combine that data with the patient's clinical bicarbonate data and assign the patients in real time to one of these latent phenotypes or the other. And the clinicians in this case were kept blinded to the patient's phenotype assignments so that no clinical care would be affected. And what we found was using that model that Dr. Sinha had developed and cut points from prior studies that between 10 and 20% of the patients fell into that hyperinflammatory phenotype. That that group, that hyperinflammatory phenotype had markedly higher mortality than the hypoinflammatory phenotype though. This was the first wave of COVID so mortality was not great in either phenotype. And this is a very small data set so I don't want to make too much of it because it's really only 39 patients. But I think it demonstrates that real time phenotype classification may be feasible in the very near future. And Dr. McCauley is leading an ongoing study called the FIND study to test the feasibility of this approach on a larger scale and has now enrolled about 400 patients doing that with results to come probably in the next year or two. Wouldn't it be nice though if we could identify hidden phenotypes or classify our patients without the use of biomarkers, right, because that's certainly a barrier to the models I described to you previously. So I've been working with Dr. Sinha and with Matt Chirpek from University of Wisconsin here to develop a model that uses only clinically available data in a machine learning approach to classify patients into the one or the other of these two phenotypes. The algorithm is called XGBoost and it's kind of like a fancy random forest for those of you who are familiar with that type of analysis. And it basically takes all the clinical variables that were used to identify these latent phenotypes and combines them in a sort of black box classification algorithm. In clinical trial datasets, this approach has an area under the curve of 0.94 to 0.95. In observational cohorts, a bit lower, 0.88 to 0.92, but nevertheless suggests pretty good classification accuracy for identifying these groups. Recently working with Dr. Sinha with the Lung Safe Group and with Manoj Madali, who's a fantastic pulmonary fellow nearby here at Stanford, we were lucky enough to get to apply this model in the Lung Safe Cohort, which as you guys probably know is probably the largest epidemiologic recent study of ARDS. Now this was all pre-COVID, published by Giacomo Bellani and JAMA in 2016. And what we found when we applied that clinical classifier model to the Lung Safe dataset was that patients who were classified in that hyperinflammatory phenotype had dramatically worse survival than those in the hypoinflammatory group. And interestingly that the pattern of response to PEEP in this observational dataset, now keep in mind it's not randomized, seemed to be different in the two phenotypes in the same way that we had observed in the alveoli trial. So remember I said before, we were interested in these phenotypes because they appeared to respond differently to PEEP. And what we saw here was that if you were in that hyperinflammatory phenotype, mortality was lower in the highest PEEP tertile and the opposite pattern in the hypoinflammatory group. And this was a statistically significant difference though important to emphasize it was not randomized. Okay, now how does this potentially again relate to COVID? We took this clinical classifier model that I described to you and working again with Dr. Sinha, with Dave Ferfaro, who's now at the BI Deaconess, Max O'Donnell and Dan Brody, applied this to a cohort of COVID ARDS patients enrolled at Columbia in spring of 2020. Now this was, I really just want to highlight the work done by my Columbia colleagues here, incredible work, they enrolled almost 500 patients with ARDS in a two month span at one medical center. And again, I mean, I think we all lived through that pandemic and you can only imagine the kind of exceptional strain they were under and how challenging it must have been to enroll these patients. We applied this clinical classifier model to this data set to identify phenotypes and found that the classifier model identified two groups with different prognosis. So this is 90 day mortality, significantly higher in patients classified to the hyperinflammatory than the hypoinflammatory phenotype though mortality really quite high in both groups. I think we were really sort of surprised and interested by this finding. So this cohort, as I mentioned, was all enrolled in March and April 2020 before use of corticosteroids was routine. And what we found very interestingly was what appeared to be a difference in response to corticosteroids. I say appears to be because we've got a lot of caveats here. Steroids were not randomly allocated, right? So there clearly could have been confounding by indication but they were equally administered across the phenotypes. With that caveat in mind, we saw that that hyperinflammatory group had a lower mortality when treated with steroids and the opposite pattern was observed in the hypoinflammatory group and this difference was statistically significant. Now, how relevant this is to the COVID patients that we are all treating today, I really don't know because I think we all know clinically how different these patients are than the patients we were taking care of early on in the pandemic. We have some data that we're preparing for publication now based on patients enrolled later in the COVID pandemic that actually indicates these patterns are quite different and almost all of the patients that we're enrolling now are falling into that hypoinflammatory group. Now, whether that's because of changes in the variants, changes in the way we're treating the patients is really hard to say, but I think this may be really quite different in our patients that we're treating now. But I think the point I want to make with this is that I do think there's considerable biologic heterogeneity, even within these patients that we clinically see as very consistent with the clinical definition of ARDS. So what are the next steps? I was asked to say, to address in this talk, is it time for personalized medicine? How do we get there? Well, I think there's a few important steps. We're working on prospective validation of those parsimonious biomarker models using real-time assays in the FIND study led by Dr. McCauley that I mentioned, as well as in the ICE by COVID network here in the U.S. We're working on developing an EHR interface to implement a clinical classifier model prospectively, and really importantly, to evaluate how stable that model is over time for testing and clinical use. And then I think really critically, before we think about personalized medicine, I want you to think back to that live study that I showed you of diffuse versus focal ARDS. All of this is really theoretical and retrospective right now, right? I think we really need to test the utility of splitting our patients into subgroups prospectively before we can even consider changing our treatment. How should we do that? Well, we could certainly randomize patients to a usual care arm versus a personalized strategy, as in the live trial, or I think perhaps in a more innovative approach, a stratified randomization with a plan to adapt depending on interim analyses. And this type of approach might allow you to learn as you go regarding the optimal treatment for each phenotype. If you're interested in this type of adapted design, I'd refer you to this nice publication from the New England Journal in 2016. But basically the idea here is you could enroll a group that you know to be non-homogenous with a plan a priori to analyze patients in two different subgroups, do an interim analysis in which you evaluate the treatment in each subgroup, and then potentially stop in both groups, continue in both groups, or only continue in one group depending on your findings. All right, so take-home points. I do think, I agree with Dr. Heisey that ARDS is a very heterogeneous syndrome. We are already treating patients with different ARDS phenotypes differently. ARDS severity and certainly COVID-19 versus typical ARDS in terms of our medication regimen. I do think changes on the horizon, new methods for identifying latent molecular phenotypes, including ones other than the ones I've talked about today just in the interest of time, have been developed. But clinical testing is really required prospectively to test whether or not this approach will actually work before clinical implementation. Thanks so much.

subphenotypes

acute respiratory distress syndrome

personalized medicine

COVID-19 patients

phenotyping