Thank you so much for joining us today for the Surviving Sepsis 2071. We have an extraordinary complement of speakers for you, faculty members from various areas throughout the United States. Our first speaker is going to be Hector Wong, and he's going to talk to us about biomarkers beyond procalcitonin. And he's going to be followed by Dr. Morawitz, with the title is Gut Check, a Microbiome and Disease. And Dr. Judith Hellman will follow with immunomodulation. And wrapping things up, as far as content initially, it's going to be Dr. Shamin Namati with Predictive Analytics, talking to us about artificial intelligence. And then I'm going to finish up kind of summarizing the faculty talks with the title, Now That We Know It, How Do We Implement It? And we thank you so much for joining us today, and I'll pass it off to Dr. Wong as our first speaker. Hello everyone. Welcome to this session on sepsis. I'm sorry we can't be together, but hopefully this format will still be informative for you. So I've been asked to speak about biomarkers beyond procalcitonin, what can we expect in the future. I'm happy to do that. These are my disclosures here, funding from the NIH. My institution and I hold a number of sepsis biomarker related patents that are either issued or pending, and I sit on a scientific advisory board of these two companies, which are both actually quite interested in sepsis-related biomarkers. So biomarkers and sepsis. My trainees ask me very often about, you know, what do I think about biomarkers and sepsis? And so usually my question is, well, what do you mean by that? In other words, biomarkers of what? And so most of us, when we think about biomarkers and sepsis, we think of diagnostic biomarkers. In other words, biomarkers that allow us to differentiate who has sepsis and who does not have sepsis. And those are very, very important, but it's not what I'm going to focus on today. What I'm going to focus on today are what we call prognostic biomarkers of sepsis. In other words, among those who have sepsis, who is more likely to have a bad outcome? And also what are called predictive biomarkers, that again, among those who have sepsis, who is more likely to respond to a given therapy based on underlying biology? And those concepts get back to this concept of enrichment, which I think I would submit to you is a key concept for enabling precision medicine. So generally speaking, enrichment refers to the use of any patient characteristic to select a population in which your intervention is more likely to have an effect compared to an unselected population. Prognostic enrichment is a bit more specific in that it entails selection of a patient population more likely to have a disease-related event, such as mortality, whereas predictive enrichment is the selection of a patient population in which your intervention is more likely to have an effect, again, based on underlying biology or physiology. And so in an ideal world, we can take a heterogeneous sepsis cohort here depicted as circles for individual patients, which are very heterogeneous, and we can start to separate those groups into more homogeneous groups through perhaps prognostic enrichment, and you have your low-risk patients that are probably going to do fine with standard of care, and then you have your higher-risk patients that you could perhaps separate them out further based on predictive enrichment and then target specific therapies based on biological mechanisms. Granted, this level of enrichment is a much bigger lift and a harder task, but something we're working on. And so this is something we've been working on for a while. You know, our approach has been using high-throughput transcriptomics and bioinformatics, and what I'm going to tell you about first is our identification of prognostic biomarkers or biomarkers for prognostic enrichment, and this has been distilled down to something that we call PERSEVERE, which is an acronym for the Pediatric Sepsis Biomarker Risk Model. This is actually a model based on serum protein biomarkers that we identified through transcriptomics. It's based on classification and regression tree methodology, and what it allows us to do is to provide a reliable baseline risk of mortality among kids with septic shock. Again, this is this concept of prognostic enrichment. And so this is what the decision tree looks like. It's really not as complicated as it might seem. Those are all the patients and what are called the root nodes, and then using this concept of binary recursive partitioning using biomarker cut points, patients are subdivided into different categories until you can no longer subdivide them according to this tree. And so the way that this tree yields information is based on the terminal nodes, as highlighted there. Patients that are classified to these four terminal nodes have a very low risk of mortality at 28 days, whereas patients who are allocated to those two nodes have an intermediate risk of mortality, if you will, as listed there. And then the patients allocated to these three terminal nodes have a much higher risk of 28-day mortality. So we took this decision tree and tested it prospectively in a new cohort without changing any of the cut points. In other words, we locked the rules of the tree and applied it prospectively and had this kind of performance with an AUROC of .86 and that sensitivity and specificity. I would submit to you that that's probably about as good as one is going to get. I don't think it's reasonable to expect higher specificity than this or much higher because that would imply that what we do in the ICU doesn't make a difference. And so I would personally as a clinician like to believe that those kids that were predicted to have a high mortality risk but actually survived, did so because of interventions in the ICU. And that's why you have a false positive in that case. But we can debate that, if you will. Another way to look at these data is that these are the survival curves for the three classes of risk, low, intermediate, and high risk. And you can see very, very distinct survival curves there for these children. And again, this was a prospective validation of this model. If you're interested in the details, it was published about a year ago and this paper is listed here. We have recently adapted this PERSEVERE biomarkers to look at sepsis-associated acute kidney injury. And as we all know, if you have sepsis and you develop AKI, that's a bad prognostic sign. It really contributes to sepsis morbidity and mortality. And sepsis-associated AKI is a very complex interplay of both ischemia and oxidant stress as well as inflammation. And so we asked the question, can PERSEVERE or the PERSEVERE biomarkers estimate the risk of severe AKI associated with sepsis? And here, we identified severe sepsis as CADIGO stage 2 or greater by day 3 of septic shock. And so this work was done by one of our fellows, Natalia Stansky, who graduated from our program recently and joined us as junior faculty. This is her paper in the blue journal that just came out in the last year. And this is the tree that Natalia developed here. Again, using the PERSEVERE biomarkers and the PERSEVERE mortality probability, she tried to estimate the risk of severe AKI on day 3. And so one interesting thing about this model is that if you come into the ICU at stage 3, your chances of having severe AKI by day 3 is pretty high. That may seem intuitively obvious, but as far as I know, it's not been published at least in kids before. The other patients go to the left side of the tree, and these are the kids that have either 0, 1, or stage 2 AKI when they come in the door, and the model is able to segregate them into those with very low risk. Another way to interpret that is kids who can recover from initial AKI, and then it can also separate those kids with very high risk. And according to this model that Natalia developed, this was the performance in the derivation core at a very high AEC of 0.95 with the other diagnostic test characteristics as listed there. She saw a tenfold cross-validation summary AEC of 0.88 and a validation in a very old historical cohort of 0.83. So we're excited about this. It's going to be a really nice opportunity for Natalia. She's in the process now of prospectively validating this model. And if this model holds up, I think it could be very effective for selective application of renal protective strategies and also a way of enriching patient populations for clinical trials because we can effectively estimate who is going to still have AKI on day three of septic shock. So now we're going to turn to predictive enrichment biomarkers, and again, those are biomarkers that allow one to select a population with common underlying biology, and we used the very same approach, transcriptomics followed by bioinformatics. And we did this to try to discover endotypes, and endotypes are subclasses of diseases or syndromes as defined by biological mechanisms. And so if one thinks of septic shock as a syndrome, it implies the existence of endotypes, and we reasoned that perhaps they could have distinct gene expression patterns and we could identify this through transcriptomics. And so the short answer to that question is yes, we think so. These are the two endotypes that we published about five years ago in the Blue Journal. These are what we call endotypes A and B. What those graphics there are what we call gene expression mosaics that are depicting 100 genes that allow us to identify the endotypes. These are reference mosaics. We classify patients using computer-assisted image analysis. The mosaics on the bottom are individual patients. What's interesting is that these 100 endotype-defining genes reflect adaptive immunity and glucocorticoid receptor signaling, and so because of these properties, the fact that they're repressed in endotype A, we think that they may be a valuable predictive enrichment tool. And so some may say, well, this just reflects lymphopenia, right, it's just a fancy CBC, but as it turns out, endotype A patients actually have higher absolute lymphocyte counts compared to endotype B, so we don't think that that pattern is a reflection of just simply lymphopenia. And so endotype A patients, they have greater organ failure and greater mortality compared to B, but more importantly, if you account for illness severity, age, and comorbidity burden, allocation to endotype A is independently associated with increased mortality, and even more importantly, if you're an endotype patient and you get corticosteroids, corticosteroid prescription is independently associated with increased mortality, so steroids seem to be bad for endotype A patients. These are some temporal considerations, right. Our patients are very dynamic, and so that's an example of a patient who starts out as a B and becomes an A, that's a patient who starts out as a B and remains so, that's a patient who starts out as endotype A and by day three is a B, and this is a patient who stays A throughout the entire course, and that pattern is very bad. If you persist as an endotype A, your odds of poor outcome are very high, and they're even worse if you get corticosteroids. So again, further evidence of endotyping potentially informing decision making. And so all of that data are in children, all of that data are also observational, meaning that the steroids weren't standard or randomized in any way. So this is the VANISH trial, which involved adults with septic shock who were actually randomized to hydrocortisone and placebo, and fortunately the investigators put out their transcriptomic data in a public database, so we downloaded that data, assigned the patients from the VANISH trial into either pediatric endotypes A and B, this is something of what it looks like, again the top are reference mosaics, and then the middle and lower panel are examples of VANISH cohort patients and what they look like with respect to these pediatric septic shock endotypes. And so we asked the question, you know, here's a setting where it's an independent dataset, we allocate them to our pediatric endotypes, is there an effect of corticosteroids? And so we tested that using logistic regression to look for an association between hydrocortisone endotype and 28-day mortality. The reference group here is endotype B that gets placebo, and as we show here, that interaction of endotype A and hydrocortisone, the odds of mortality are higher, with a p-value that's approaching significance there, or right on the edge of significance, reflecting a pretty small cohort. But nonetheless, we were really excited about this, because again, this is an external cohort, when we apply the same strategy in a setting where steroids are randomized, we see a very similar effect. And so it's becoming increasingly apparent that corticosteroids might be harmful among endotype A patients. If this is true, this could account for the heterogeneity that we see in responses in the various clinical trials. And so we've also asked the question, it's one thing to identify who probably should not get steroids, but what about, can we identify who would benefit from corticosteroids? And so here, we combined the two strategies that I told you about. We used Persevere for prognostic enrichment, and endotyping for predictive enrichment, and what we have reported is that if you're an endotype B patient, and at baseline you're at intermediate to high baseline mortality risk, you have a more than tenfold decreased risk for poor outcome when you receive corticosteroids. Again, so these are examples of combining predictive enrichment with prognostic enrichment. Again, these data are based on observational data, or these results are based on observational data. So now, we're trying to test this prospectively in this thing called the SHIPS trial, which is a multi-institutional trial in children with septic shock who are being randomized to placebo versus hydrocortisone, and we're collecting biological samples for prognostic and predictive enrichment. And we're unable to use enrichment to inform enrollment, but we have a pre-specified post-hoc analysis. So more to come on that as things progress, but we're excited for that opportunity. So the take-home messages here, the intrinsic heterogeneity of sepsis can potentially be managed using precision medicine concepts. Prognostic and predictive enrichment strategies based on biomarkers are fundamental tools for enabling precision medicine. Persevere and gene expression-based endotyping represent potential enrichment strategies for septic shock, and I would submit to you that future trials in sepsis have to incorporate these enrichment strategies based on biomarkers or other strategies. Thank you very much for listening. I hope that that was informative and worth your time. Hi, thanks for the opportunity to participate in this panel. The topic of my talk will be the microbiome in disease as it relates to sepsis in the year 2071, when I hopefully will be 99 years old. As a framework for this short talk, I'm a general and thoracic pediatric surgeon who's trained as well in adult general surgery and trauma. So I'm not an intensivist, but I've spent many hours in ICUs with septic and critically ill patients of all ages, and as well, I have a research interest in the role of the human microbiome in disease, but also in health. There's extraordinary interest at present in harnessing the power of the human microbiome to improve outcomes for conditions ranging from schizophrenia to cancer to critical illness. Specific options include the use of probiotics, fecal transplantation, and even the topical application of maternal vaginal secretions to babies born via cesarean section rather than vaginal delivery. However, the culture and the current state of affairs in ICUs and arguably in all of clinical medicine is still quite different. I like to say that the GI tract has a branding problem as illustrated by its reputation as the undrained abscess of multiple organ failure or the motor of critical illness, and it's this type of idea that has given way to clinical strategies such as digestive tract or oral decontamination with non-absorbable antibiotics. As I'll describe, this idea is certainly not without merit, but it's notable for the failure to ask the question, what will come next after decontamination? This figure is from a manuscript from my group a few years ago. Like other groups, we showed that the composition of microbes within fecal samples from ICU patients differed sharply from samples from healthy volunteers. This wasn't surprising, but it was novel in that it was one of the first studies to use high throughput sequencing to characterize the highly complex microbial communities in the gut rather than laboratory culture. So in 2021, there's a tension between the current scorched earth strategy of modern medicine when it comes to the gut microbiome and the new appreciation for the contribution of commensal organisms to human health during homeostasis and the potential to harness microbes to improve clinical outcomes. So how do we get here and what can we do about it before 2071? Well, first, virtually everything taught in medical school in the 20th century, all of the tremendous advances and all of the ongoing challenges were by necessity discussed without consideration of the microbiome. This was nobody's fault, but as some experts have proclaimed, it's now time to support a new microbial literacy, not just among physicians and scientists, but amongst everybody, even beginning in childhood. So for example, once the pandemic is over, it may be time to scale back on hand sanitizer. Some may disagree with me on this subject, but the enormous emphasis on molecular biology has probably come at the cost of studying and thinking about big ideas. There are many big ideas out there that are certainly to infiltrate clinical care in the next 50 years, and some of these are listed below, but my focus, what motivates me and my lab is to work until clinical decisions are made with consideration of patient microbiome, just like we do currently for laboratory values and vital signs. It actually wasn't that long ago, at the end of the 19th century, following Leeuwenhoek's discovery of microbes under the microscope, that the germ theory of disease took hold and remains prevalent today. The germ theory of disease, also popularized by legends such as Pasteur and his contemporaries, states that specific diseases result from infection by specific microbes. Since antibiotics became discovered and widely available, it made perfect sense to treat infected individuals with antibiotics. However, it's now clear that this paradigm overlooked the fact that causative pathogens coexist with commensal organisms, and current antibiotics are not capable of selecting for pathogens without disrupting the commensal microbiome, which is important because the administration of antibiotics is a risk factor for subsequent bacterial or fungal infection. What might, and I stress might, be replacing the germ theory of disease is what can be called the microbial theory of health, and this theory acknowledges that microbes, along with the sun and circadian rhythms, have profound effects on human life, and these effects are mostly but not entirely positive. When these processes are disrupted, illness can result, and disrupting them further with medical treatment can complicate a patient's clinical condition further. There are many emerging frameworks for thinking about these ideas, but a particularly compelling framework is Balmer's germ organ theory of non-communicable or non-infectious disease. A key consideration of this framework is that the normal gut epithelium promotes the growth of strict anaerobes that generate metabolites that are beneficial locally in the gut, but also systemically as far away as the brain, short-chain fatty acids being the best studied at present. Patients that are sick can develop intestinal inflammation, the net result being that the strict anaerobes die and promiscuous pathogens, with which you're all familiar, can thrive. The important takeaway in the adjacent figure is that disease of any kind can lead to a vicious cycle in which dysbiosis can worsen disease, which in turn worsens dysbiosis. So a very good target for future developments is to identify opportunities to disrupt this vicious cycle. Understanding that the gut microbiome has systemic impacts and that unlike the human genome it can be dynamic is important for understanding the physiology of our critically ill patients. This applies really to all organ systems, but one vivid example for the purpose of today's discussion involves cytokine production by circulating leukocytes. This has now been shown clearly to vary across individuals, in part reflecting variations in the microbiome. So all of this, of course, is somewhat overwhelming and changing rapidly, which leads me in my own clinical practice to try to incorporate the Hippocratic Oath, but also President Obama's foreign policy doctrine, which stated, don't do stupid stuff. But this is easier said than done, since each of us can disagree on what might constitute stupid stuff. And it's difficult when we don't know what we don't know, and an illustration of this was provided by a landmark article in Nature a few years ago shown here, documenting the unrecognized effects of common medications, non-antibiotic medications, on the commensal human microbiome, notably antipsychotics and antihypertensives. And this was a stark reminder that even seemingly benign medical interventions can lead to undesirable side effects. Ultimately, in my view, the field of antibiotic stewardship over the next 50 years should give way to a microbiome stewardship. There are numerous ways to steward the human microbiome among septic patients, but here in the last few minutes, I'll touch upon three specific topics, diet, antimicrobials, and the cleaning of the physical ICU environment. Sixty-five years ago, an article was published in JAMA in which the authors declared that natural foods surpassed synthetic pre-prepared formulas in what they referred to as a prolonged trial. Nevertheless, due to a variety of reasons, partly the interest in reductionist science, this practice of using real food in the ICU disappeared, and today, we use chemically defined commercially available formulas that often are notable for high sugar and low fiber content in addition to the presence of emulsifiers and preservatives, which have been shown to be pro-inflammatory. Thirty years ago, one of my own mentors, Dr. John Alberti, reported that bacterial translocation from the gut in rodents was far higher with synthetic formulas than with natural rodent chow or formulas supplemented with vegetable fiber. Almost 30 years after that, in 2018, Dr. Alberti opined, correctly in my opinion, that a major flaw regarding nutritional support of the critically ill is the traditional view that the intestinal tract is a mere conduit for the absorption of nutrients. To the contrary, the diet feeds not only the human patient, but also the microbiome, and perhaps more importantly, it's not enough to know the concentration of protein, fat, and sugar without knowing the source of these macronutrients. Over the last few years, my own lab has sought to build upon this research. We've found that pureed natural food sources reproducibly yield superior outcomes when compared to commercially available, commonly used synthetic formula. We've largely studied this to date in murine models of intestinal inflammation and are now extending this work to models of critical illness. Thus far, the only scenario in which a plant-based natural diet is not effective is in germ-free animals or in animals with a weakened innate immune system, but not adaptive immune system. We are looking currently to clarify the clinical relevance of these findings. The status quo of one-size-fits-all antimicrobials is probably not sustainable, in part due to obvious resistance patterns, but also because it may be harmful to patients in other ways. There are numerous controversies regarding antibiotics in the ICU, and these are beyond the scope of this talk, other than to mention that the ongoing debates about the length of treatment required for bacteremic patients, whether or not antibiotics should be started for all ICU admissions, including viral pneumonias, and whether the popular quality metric of antimicrobial de-escalation is actually counterproductive, will likely serve as topics for clinical trials of the future. One of the current problems with antibiotic decision-making is that we choose antimicrobials based upon culture and sensitivity results, and then administer antibiotics, hopefully with a good clinical response, for a somewhat arbitrary length of time. Other point-of-care diagnostics regarding the microbiome could help in this regard. Shown here is a representative figure from a study from our group. Different colored bars in this graph represent different organisms present in the gut, on the skin, or in the oral cavity of a child that presented with enterococcal sepsis. Enterococcus can be seen as bright pink, which was obviously abundant in the gut and on the skin. Although not easily visible, it was also present at low abundance in the oral cavity at the time of admission, which is not normal. She improved clinically with antibiotics. And over time, the abundance of enterococcus decreased and eventually, in fact, disappeared. If such information were available in real time and at a reasonable cost, which it currently is not, then we may be able to, in the future, design rational, safe, and cost-effective treatment plans. In a related experiment, we monitored the expected collapse of microbial diversity shown on the Y-axis in mice that were exposed to broad-spectrum antibiotics while being fed normal rodent chow. After one week of antibiotics, the animals were randomized either to the natural plant-based puree or commercially available synthetic formula labeled as CEN. As you can see, diversity decreased in fecal samples from animals exposed to antibiotics. But those receiving plant-based nutrition, or PBEN, quickly recovered their diversity, whereas those on conventional formulas did not. Remarkably, as shown in the red curves, among animals that did not receive antibiotics but were randomized to CEN, diversity also collapsed, whereas this did not occur with animals that didn't receive antibiotics and were randomized to the plant-based nutrition. We believe this is clinically relevant because of the risk of secondary infection after antibiotic treatment. In another experiment, we gavaged antibiotic-treated animals with vancomycin-resistant Enterococcus while they were fed with either plant-based or conventional nutrition. At 24 hours later, both groups were colonized by high levels of VRE. However, at seven days following gavage, the abundance of VRE in fecal samples from conventionally fed animals was orders of magnitude higher than samples from mice that were fed the plant-based diet. A final subject worth mentioning in my view involving the human microbiome and septic patients is a favorite topic of mine, and this is the microbial overlap between the ICU room environment and the deranged human microbiome of ICU patients. Our group has looked at this in NICUs and PICUs, and without fail, a good 25 to 30% of patients demonstrate the identical organism verified by genomics within fecal samples, endotracheal tube samples for skin swabs as the organisms that can be seen in the room sinks and heavily touched surfaces within the ICU. This makes sense when considering that antibiotic treatments, whether enteral or systemic, essentially obliterate the gut microbiome, creating a blank canvas that can be occupied by opportunistic pathogens in the nearby environment. And this is why future strategies might pair antimicrobial therapy with plans to rehabilitate the microbiome after treatment. But in parallel, there's increasing recognition that the heavy-handed methods for cleaning hospital rooms unwittingly selects for unusually hardy, resistant organisms that are obviously problematic once they colonize patients. This recognition has given way to a new understanding of targeted hygiene or bidirectional hygiene labeled bi-gene in this journal article, which refers to the concept that maintaining environments that are reasonably free of pathogens, but not necessarily free of all microbes may lead to a healthier human microbiome than the current practices. To conclude, the widely published graph of pro and anti-inflammatory responses in septic patients may not be useful, in my opinion, without consideration of the microbiome, which is probably altered in most ICU patients at admission and becomes even more altered due to medical interventions. To optimize sepsis outcomes in 2071, I've argued here that a new framework for thinking about the microbiome is necessary, which in turn will require better and faster diagnostics, as well as novel approaches to nutrition, antibiotic administration, and cleaning of the physical ICU environment. With these developments, therapeutic options that will include fecal microbial transplantation, pre and probiotics, as well as others that I've not been able to cover in this presentation will undoubtedly be introduced in parallel and hopefully will further improve clinical outcomes. I'd once again like to thank the organizers of this meeting and this panel for the opportunity to present today. And I'd like to acknowledge my lab members and numerous collaborators at the University of Pittsburgh and elsewhere without which none of this work would be possible, especially my valued colleagues in the intensive care units at Children's Hospital of Pittsburgh. Thank you again for your time. Hi, I'm Judith Hellman. I'm Vice Chair for Research in the Department of Anesthesia and Perioperative Care at the University of California in San Francisco. I wanna thank the Society of Critical Care Medicine Congress Planning Committee for this opportunity to present at the 2021 SCCM Congress. I have no financial conflicts of interest with the contents of this presentation. I have several goals for this talk. For the first part, I'll review the definitions of immunomodulation, describe immunomodulation as it pertains to sepsis pathophysiology, and touch on immunomodulatory therapies that have been or are currently being investigated in humans with sepsis or that are in the preclinical pipeline for sepsis. In the second part of this talk, I'll delve into endogenous immunomodulatory systems focusing on non-leukocyte-mediated regulation of immune function. I'm going to then segue into a discussion of research that my lab has been doing on the immune modulating effects of the endocannabinoid system and the pain receptor TRPV1, which is actually part of the endocannabinoid system. Immunomodulation can be broadly defined as a change in the immune system that's caused by activation or suppression of immune function by intrinsic or extrinsic factors. Immunomodulatory therapies induce, amplify, attenuate or otherwise reorient the host's immune response. And this is for the purpose of treating specific diseases or to promote wellness. There's also endogenous immunomodulation whereby the host fine-tunes immune responses to beneficial rather than harmful levels and under optimal conditions, restores immune homeostasis so that the host is ready to effectively deal with future infections and tissue injuries. There's also another form of extrinsic immunomodulation that I've decided to call bystander immunomodulation. This results from off-target effects of medications or environmental factors on immune cells. Catecholamines and opiates are examples of medications that are regularly used in ICU patients that have off-target immune effects. Before I go through this slide in detail, I wanna show you this final product mainly to highlight the complexity of the host's immune response to sepsis and to point out how many potential points there could be to intervene in sepsis pathophysiology. During infection, microorganisms and their products interact with host cells, including white blood cells in the circulation and tissues, but they also interact with non-leukocyte populations such as endothelial cells, epithelial cells, and neurons. This leads to inflammation characterized by the upregulation of cytokines and chemokines, the development of oxidative stress, dysregulation of coagulation, and so forth. If things go well, the immune response is optimally balanced and the host has a healthy response with clearance of infection, inflammatory resolution, and recovery. Conversely, pathologic responses can also develop. The production of injurious mediators leads to the development of multi-organ failure. Immunosuppression leads to secondary infections or the inability to effectively clear the primary infection. How this process of inflammation and immune suppression unfolds over time has proven to be extraordinarily complex and it's generally believed that immunosuppression and pro-inflammation can occur simultaneously. Impaired inflammatory resolution further amplifies the pathology by potentiating the production of injurious mediators and by further suppressing immune function. All of this becomes a vicious cycle with more microbes and also host mediators now interacting with host cells and amplifying the immune dysfunction and organ injury and immune suppression. The goal of immunomodulatory therapies is to drive the system towards healthy responses that lead to clearance of infection, recovery, and immune homeostasis. Numerous clinical trials over several decades have focused on modulating acute inflammation to treat sepsis. Various agents have been tested that fall into several broad categories and these include agents that block or reduce early pro-inflammatory responses such as antibodies to LPS or TLR antagonists, TLR4 antagonists. Agents that neutralize or block activation by pro-inflammatory mediators such as antibodies to TNF or IL-6 or receptor antagonists or soluble receptors. And finally, corticosteroids which have broad effects on immune function. These strategies were not successful in multiple clinical trials largely but retrospective analyses do suggest that some of them may have been beneficial in patient subgroups. There are several strategies that are being explored for reversing immunosuppression in sepsis. Based on strong preclinical evidence, clinical trials have focused on using immunoadjuvants to boost T cell functions. Two immunoadjuvants, IL-7 and the immune checkpoint inhibitor anti-PDL1 antibody have been tested in small clinical trials. Both seem to be safe and effective at restoring lymphocyte numbers and function. And Richard Hotchkiss has been sort of the pioneer of using T cell, trying to boost T cell function using adjuvants including IL-7 and anti-PDL1 antibodies. Another approach to reversing immunosuppression is stimulating immune cell proliferation with factors such as GCSF and GMCSF. The use of agents to restore innate immune function continues to be another promising potential means of reversing immunosuppression. An interesting concept that is still in the preclinical stages of development is using trained immunity and metabolic reprogramming to improve innate immune responses. Now I'm going to discuss endogenous immunomodulation in sepsis. That's the process whereby endogenous factors and pathways fine tune innate immune responses and inflammation, which leads to beneficial responses including the clearance of infection, wound repair, and the restoration of immune homeostasis. Endogenous immunomodulation is mediated by conventional immune cells, including those shown on the slide, which are classically thought of as regulating immunity. So far, human sepsis studies have primarily focused on circulating leukocytes and blood biomarkers. And this makes sense in that leukocytes are key players in immunity and also because circulating leukocytes and biomarkers are readily accessible through blood draws. In contrast, the behaviors of other cell populations that are not easily accessible aren't well-defined. And this includes tissue leukocytes and fixed cells such as endothelial cells and neurons. Although not classically viewed as being important in immune function, there's increasing evidence that non-leukocyte populations play active roles in immunity. Endothelial cells, neurons, cardiomyocytes, and epithelial cells have all been reported to express innate immune receptors, including the toll-like receptors, and to produce cytokines and chemokines. Furthermore, multicellular immunomodulatory networks have been described in sepsis that involve interplay between neurons and leukocytes An example of this is the cholinergic anti-inflammatory pathway that was first described by Kevin Tracy's group, which regulates systemic inflammation via activation of the efferent limb of the vagus nerve. I'm now going to shift gears into a research project in my lab on the immunomodulatory effects of the endocannabinoid system. By way of background, I'm an intensivist doing basic translational research on sepsis and critical illness. My research is focused on the innate immune system in general, and on endothelial immune signaling in sepsis and acute inflammation. Over the last decade, my lab's also been studying the effects of cannabinoids on acute inflammation. I will now briefly describe some of our findings from this project, which I hope will illustrate the concept that cells other than circulating leukocytes may play an integral role in the endogenous regulation of immune function and immune homeostasis in acute inflammation. I first became interested in the immune effects of cannabinoids after reading a lay article in Science News describing the use of marijuana for chronic inflammatory conditions and for multiple sclerosis and cancer. This article prompted me to wonder how cannabis might affect sepsis and acute inflammation, and some searching to understand more about cannabis. And this led me to the discovery of the endocannabinoid system, which was first really discovered in the 1990s. The endocannabinoid system plays roles in neurological and psychological functions, as well as in regulating vascular tone and blood pressure. Interestingly, despite the use of cannabis products to manage symptoms associated with inflammation, the immune effects of cannabis and the intrinsic role of the cannabinoid system in regulating immune functions are not well understood. Receptors of the endocannabinoid system include the G-protein coupled cannabinoid receptors, in particular CB1 and CB2R, and a family of cation channels known as the transient receptor potential or TRIP channels. These receptors are activated by endocannabinoids, which are cannabinoids that are actually produced in the host, plant-derived or phytocannabinoids, and synthetic cannabinoids. There are multiple endocannabinoids that have been identified, including anandamide and 2-AG, which have been the most extensively studied. Among the endocannabinoids are some acyl-dopamines, including anarachidonil dopamine, or NADA, and oleoil dopamine, or OLDA, and these are produced in the central and peripheral nervous systems. We initially tested the effects of anandamide, 2-AG, virodamine, nolidine ether, and NADA on endothelial cell activation by LPS and other inflammatory enzymes. And we found that this one, NADA, had potent anti-inflammatory effects. In subsequent studies, we found that another one, OLDA, also has potent anti-inflammatory effects. And I'm not going to discuss OLDA more in this talk, and I'm gonna focus more on our studies on NADA, which have been published. Here's some of the early data on this project demonstrating the anti-inflammatory effects of NADA on endothelial cells and on leukocytes. In these experiments, we treated cells in a dish with LPS in the presence or absence of NADA, and we quantified IL-6 in culture supernatants at six hours. In this panel, I show results from studies done with human lung microvascular endothelial cells that were activated with LPS and treated with increasing concentrations of NADA, and we can see that there's a concentration-dependent reduction in IL-6 production by those endothelial cells. We also found a similar result for IL-8 production. In analogous experiments using leukocytes, PBMC, and bone marrow cells from wild-type mice, we found that NADA also down-regulated IL-6 production induced by LPS in these cell populations. In subsequent experiments that were published in this manuscript in the Journal of Immunology in 2017, we found that NADA had potent anti-inflammatory effects in endotoxemic mice. In particular, it caused a profound up-regulation of IL-10 in endotoxemic mice. We went on to look at the effects of NADA in bacterial sepsis, and here I show data that we gathered in a model of sacral ligation and puncture, which is analogous to bowel perforation or the component of ischemia in humans. And mice underwent sacral ligation or sham surgery, and were treated after five minutes and two hours with intravenous NADA. And at six hours, we analyzed their plasmas for pro- and anti-inflammatory cytokines. And I show here the data for IL-6 and IL-10 to show that there's a down-regulation of IL-6 levels in the blood of CLP mice treated with NADA. And a substantial up-regulation of the anti-inflammatory cytokine IL-10 in CLP mice treated with NADA. We also found that there was a reduction in other pro-inflammatory cytokines, including TNF-alpha, IL-1-beta, and CCL-2. We went on to show that NADA, that TRPV1 mediates the anti-inflammatory effects of NADA in endotoxemic mice by showing a lack of up-regulation of IL-10 or down-regulation of pro-inflammatory cytokines in endotoxemic mice that lack TRPV1. So here we see that there's an up-regulation in NADA-treated endotoxemic wild-type mice that goes away in TRPV1 knockout mice. TRPV1 plays a role in pain and thermosensation. It's activated by capsaicin, the active ingredient of chili peppers, by heat protons, and by endogenous lipids, and actually NADA is an endogenous lipid, and it's also putatively the most potent endogenous TRPV1 agonist. TRPV1 has mostly been studied in neurons, but expression of TRPV1 has been reported in non-neuronal cells. We hypothesized that TRPV1 outside of the hematopoietic system might be responsible for the anti-inflammatory effects of NADA based on its high levels of expression in non-leukocyte populations, in particular neurons. Therefore, we have performed experiments using mice that either lacked TRPV1 in cells produced by the bone marrow, or in all of the non-leukocyte cellular populations. And so this is a complicated experiment. I'm actually gonna go through it in some detail because it's illustrative of the point of non-leukocyte mediation of immune function. And here we treated, we did bone marrow transplant where we transplanted either wild type or TRPV1 knockout bone marrow into wild type mice or into TRPV1 knockout mice. So we generated four different types of chimeric mice. I wanna bring your attention to these two parts of the graph that are squared in red and blue to show you that mice that have TRPV1 in all of their cells, except bone marrow cells. So these mice lack TRPV1 in their bone marrow cells. They continue to respond to NADA. You can see this up-regulation of IL-10 by NADA in endotoxemic mice. In contrast, in mice that have no TRPV1 in their non-circulating cells, so ones that have been transplanted with normal bone marrow, but lack TRPV1 in all their other cells, we lose the anti-inflammatory effect. And I show this for IL-10, but we also see this for the pro-inflammatory cytokines that mice that have not got TRPV1 in their cells other than their bone marrows are not responsive to NADA. So what cells could be responsible for this non-hematopoietic TRPV1-dependent anti-inflammatory response? We think that it's probably neurons, but there are other possibilities, including endothelial cells and microglia and astrocytes. And we're sort of actively exploring some of these possibilities, in particular the neuronal possibility at the moment. And I'm only going to show one slide on the THC project. There's other data that we published recently in the Journal of Immunology in 2020. And here, I just want to show a study in which we found that THC also upregulates IL-10 production in endotoxemic mice. So here we treated mice with LPS alone or LPS in the presence of THC. They received intravenous THC at five milligrams per kilogram. And we found that there was a market upregulation in IL-10 that was accompanied by downregulation of the pro-inflammatory cytokines. To our surprise, we found that TRPV1, in fact, does not mediate THC's anti-inflammatory effects, but rather cannabinoid receptor one does. And we showed that by using a cannabinoid receptor one antagonist, Ramonibin, which gets rid of the upregulation of IL-10 in THC-treated endotoxemic mice. To summarize our experimental findings, we have found that the endocannabinoid NADA has anti-inflammatory effects in endotoxemia and sepsis, and in particular, causes substantial early upregulation of IL-10. NADA exerts its anti-inflammatory effects via TRPV1 expressed outside of the bone marrow. In data that I did not show here, we found that in endotoxemic mice, NADA improves the functional status at 24 hours and decreases early mortality. We found that similar to NADA, THC from the plant cannabis has potent anti-inflammatory effects in endotoxemic mice. THC strongly upregulates IL-10 and improves the functional status and reduces acute lung injury in endotoxemic mice. I didn't show the functional data, but this was also published in the Journal of Immunology in 2020. In contrast to the results that we found for NADA, THC mediates its anti-inflammatory effects via CB1 receptor and not TRPV1. We're in the process of defining the role of neuronal TRPV1 and cannabinoid receptor run expressed by the central nervous system and peripheral nervous system in the anti-inflammatory effects of NADA and THC respectively. We're trying also to understand the relationship between TRPV1 and cannabinoid receptor one in driving the anti-inflammatory effects of cannabinoids. We're further exploring the functional effects of the endocannabinoid system in infection models of sepsis. And finally, we're thinking hard about how to study this system in human beings. We really want to move from these very, very early preclinical studies across this sort of translational arrow into the human arena. And that's really, I think, the next challenge that we face with this project. Ours and others' data suggest that cannabinoids have substantial immune effects and may play a role in modulating acute inflammation in sepsis and injury. This matters for a few reasons. Cannabis-derived cannabinoids are now widely being used for medicinal and recreational purposes. We need to better understand their effects and their mechanisms of action in order to use them wisely. Furthermore, our data suggests that the endocannabinoid system may have important intrinsic immunomodulatory products, properties rather. Our data using bone marrow chimeras of TRPV1 knockout and wild-type mice and emerging data from our lab point to a neuroimmune connection that's mediated through the endocannabinoid system. There are potential implications, not only for sepsis, but also for other processes, such as trauma, ischemia, reperfusion injury, and pain. And finally, the endocannabinoid system may represent a novel target for immunomodulatory therapies. I'd like to briefly acknowledge the folks that have contributed to the cannabinoid project, including current members of the lab, in particular, Jeremy Joffrey and Erica Wong, who are currently leading different aspects of the project, Kevin Wilhelmsen and Samira Lawton, who did much of the early work on NADA, and my collaborator, Mark Schumacher, who is chief of the UCSF pain service and has a long-term research focus on TRPV1 in pain. Thank you. Hello, my name is Shamim Namati. I'm an assistant professor of biomedical informatics at UC San Diego. Thank you for attending this session. All right, today we are talking about predictive analytics. I have no financial conflicts of interest to disclose. All right. Okay, so the objectives of these talks is to provide an introduction to application of predictive analytics to early prediction of sepsis. I will briefly talk about machine learning and artificial intelligence and some of the advanced methods that are being used for development of predictive analytic models. And then I will describe how this all fits into this notion of clinical decision support systems. Then, I will introduce some of the topics related to interpretability and the issue of trust. You know, this type of machine learning based models that have been criticized for being black box. So, in order to improve acceptance of these type of models, interpretability and trust is super important. The next is the issue of reducing false alarms. We all know that cognitive burden on clinicians at the bedside is quite high and, you know, acceptability of these type of models also, you know, relies on reducing false alarms and providing accurate information to the bedside clinicians. Next, I will introduce the concept of generalizability and external validity. Closely related to this notion is the issue of fairness and inclusiveness. Finally, I will wrap up by talking about regulatory oversight of these type of algorithms and what are some of the considerations there. Alright, so you are all familiar with sepsis. Worldwide, anywhere between 30 to 40 million patients every year, they catch sepsis. Mortality rate of sepsis is very high. The estimates for United States is somewhere between 1 to 1.5 million patients against sepsis and it's a very expensive condition. One of the things that we know about sepsis is that early treatment helps. We know that for each hour treatment of sepsis is delayed, sepsis-related mortality increases by 4 to 8 percent, depending on which setting you look at. So, an analogy I really like is this notion of, you know, you have a house, you can think about it as a healthy house, and then at some point you might see some smoke coming out of this house and, you know, often it's not clear if you look at early signs of, let's say, a smoke, whether this is due to a cooking incident or this is really a, you know, full-blown fire. By the time that you see the smoke inside the house is starting to burn and then by the time you see the fire and you try to do something, the end result is, you know, basically the house goes down in flames. So, similarly with sepsis, what we are trying to do is to catch early signs of sepsis and intervene as soon as possible before we proceed to septic shock and death. So, predictive analytics deals with methods of extracting information from data and using it to predict future trends and behavior patterns. Predictive analytics is not a new concept. Even the simplest logistic regression methods in statistics can be used to develop predictive analytic models, but more recently with the sort of implementation of electronic health records and availability of massive volumes of data and also availability of computational resources, now we can develop and deploy much more complex predictive models. So, a question that we asked a few years back was, can we predict onset of sepsis in hospitalized patients using electronic health record data as well as bedside monitoring data? So, let's look into that and, you know, as we think about development of predictive analytic models, as I mentioned, one of the important issues is this issue of false alarm versus true alarms. That's really one of the big trade-offs that we will get into that. All right, so, you know, you all have heard about big data in healthcare. It comes in many different flavors. You have patient-level information, demographics, you have contextual information about the patient, diagnosis about patients, interventions, cellular, molecular, imaging, natural language, processing, and finally the physiological data. Much of this physiological data comes from bedside monitors or they are charted by clinicians, and more recently we also see some of the physiological data collected through wearable technologies. The focus of the type of algorithms I'm going to talk about is on physiological type of data, although, you know, they can use demographic information, other types of biomarkers, etc. So, just to give you a sense, you know, here what you see is, you know, one of the predictive analytic frameworks that talks to EPIC electronic health record and grabs certain patient information from laboratories, from pharmacy, from other resources, as well as bedside monitoring data, and all of this complex data has to be integrated together so that you could utilize it for development of these type of models. Okay, so, you know, going forward, in 2018 we had a paper in Critical Healthcare Medicine. It was titled an Interpretable Machine Learning Model for Accurate Prediction of Sepsis in the ICU. This particular model was developed using 30,000 ICU encounters from Emory University Hospital, and we externally validated this model on around 50,000 patients from the publicly available MIMIC database. This particular algorithm utilizes around 65 different variables commonly measured in electronic health records, as well as some of the features from bedside monitoring data. What you see in this figure is examples of, you know, one patient. Each of the dots on this figure is essentially one hour worth of data. In the middle panel, you see patients' heart rate, mean arterial blood pressure, respiratory rate, temperature, O2 sat, and the Glasgow coma index, and sometime around the end of this record, by the way this is about 36 hours or so, somewhere around the end, a clinician recognized that something is going on with the patient and put the patient on antibiotics and ordered some blood cultures. The predictive model that you see here in magenta color, it picks up, crosses the threshold several hours in advance. Essentially, by analyzing patient data, it was able to figure out something is wrong with this patient, and basically sounded the alarm as many hours in advance. We termed this algorithm an interpretable machine learning algorithm because in addition to the risk score, it was also able to show the, you know, top contributing factors. So again, you know, the inputs are, you know, vitals, labs, etc. The outputs are the risk factors as well as the top contributing factors. That's the notion of explainability or interpretability of these models. All right, so implementation of this algorithm, as I mentioned, requires data integration. Sometimes the data is already available in EHR and it can be applied. Sometimes you will require going to the bedside monitors, other type of devices, and integrate. So today I'm not going to talk about that, but that's the entire field itself. It has to do with health IT and data integration is, you know, is a challenging task. Once you have the data available and you build your model, you would need to create a user interface. In this case, the user interface that we developed includes, you know, simple panels here. The patients are sorted according to the risk score. It's small, it's hard to see, but you have the risk score over here. You know, the user interface shows by how much the risk score is changing. And also, if you click on one of these panels on the back side of that, it says here are the top reasons why this algorithm basically thinks the patient is becoming septic. All right, so this gets me to this issue of the five rights of clinical decision support. You know, eventually what we want to do is to get these sort of riskier scores and incorporate them into a workflow and build a decision support system around it. So what we want to do here is to get the right information to the right person using the right CDS intervention method and to the right channel and right time. So these are called the five rights of CDS. So in the context of our sepsis prediction algorithm, the right information was reliable predictions, you know, the top reasons, context. The right person for us wasn't the bedside clinicians who are super busy. It was the tele-ICU patients. So again, it's super important to think about what are the, who are the end users of this type of tools. And then the right CDS intervention was for us a user interface that can be displayed on the screen. In other cases, it could be text, apps, etc. And the right channel for us again was electronic health record or right time. So this is what are called the five rights of CDS, but I want to, you know, talk about some of the challenges that exist. In particular, this notion of interpretability, reduction of false alarms, generalizability, external validity, fairness, and regulatory oversight. Okay, interpretability and trust is super important for adoption of this type of models. One of the factors that contributes to trust is this notion of explainability, you know, telling the clinician that the algorithm thinks the patient is becoming septic because of the not following factors. So as I mentioned earlier on, some of these can be incorporated into user interfaces. As you see in this case, you know, temperature, heart rate, and platelet counts were abnormal. And once you click on one of these panels, you see those in the top of the, in the back of the panel. And then the clinicians can look at those factors and independently evaluate the patient for themselves. There is this really nice study by Patel et al that looks at this notion of hierarchy of concepts. In particular, you know, the type of inputs that goes into a machinery algorithm, we call them baseline concepts. Once these concepts are aggregated together, we have something like clinical findings. A clinical finding could be something like tachycardia or, you know, hyper hypothermia. And then once we have the clinical findings, we can combine them together to create these intermediate constructs, things like Sears, or things like SOFA, etc. And then once we have those, we can combine them together and come up with diagnostic hypotheses. One of the issues that currently we are encountering is most of these interpretability models in machine learning, they are at the level of these basic observations. In order to come up with predictive models that are more actionable, we would need to get a little bit higher up to maybe diagnostic hypotheses to say that, I think this patient is going to become septic for the following reasons. That makes these algorithms more actionable. But also this notion of trust requires, I believe, that early partnership between clinicians and machine learning experts. Because if you co-develop an algorithm system, I think clinicians are a lot just more likely to adopt and accept these type of algorithms. The next topic is that of reduction of false alarms. I'm not going to get into all the details, but there are quite a number of strategies that one can use to reduce false alarms. Snoozing of alarms is one strategy. There are other strategies that allows you to basically figure out in what scenarios these type of algorithms are more likely to generate false alarms. And whenever you encounter one of those scenarios, you basically say, you know, I'm not sure, and I'm not going to produce a risk score. I'm just going to say, hey, I don't know. There is another common reason for false alarms is this notion of multiple causes. Again, I can tell you that this patient is showing signs and symptoms of sepsis, but it might be hard to say why. And a potential strategy for reducing false alarms could be to look at differential prediction in the sense that you predict multiple outcomes, and then you look at the one with the highest potential probability. And finally, there are ways to reduce false alarms through interactions between clinicians and AI systems, where the AI system says, here is my confidence on my prediction, but if you order this extra lab, I might be able to improve my confidence level. All right, so this brings me to the next topic, which is generalizability and external validity of these type of algorithms. Under generalizability, there are two main topics. One of them is the notion of external generalizability. This is when an algorithm developed in one hospital system is taken to another hospital system and is being tested. There are some notions such as data distribution shifts and drifts that limits generalizability of these algorithms. And building proper inductive biases into the algorithms so that they are generalizable, that's super important. But even within a given center, patient characteristics can change, and different type of measurement devices can get introduced, and collectively this can cause data shifts and drifts. Another issue is the issue of fairness and inclusiveness. Whenever you introduce predictive models and influence decision-making of clinicians, you have the potential to create echo chambers or systems that can potentially reinforce biases in the data. So I'm going to skip through that, but that's the important topic. Finally, one of the important factors are the FDA oversight. There are two reasons that these type of systems can cause potential harm to patients. One of them is false alarms that can increase cognitive burden on the clinicians, and if the clinicians are overwhelmed, that can potentially cause patient harm. The second one is the issue of misdetection, which together with automation bias, if the clinicians become too dependent on the tool, that can cause patient harms. So there are guidelines from the FDA that basically lays out under what conditions clinical decision support systems can be excluded from the definition of software as a medical device. You can read about that. But essentially, when you go through that, you see that most of these predictive models, they do require oversight. Even though currently we are not seeing some of that, but I expect to see more oversight in the coming years. All right. With that, the conclusion that I like to make is that predictive modeling and machine learning can help with improving situational awareness through improvement of data integration, in particular with respect to multi-modal data, multiplicative risk factors, and temporal patterns in the data. There are some really exciting areas for research, for improvement, to improve adaptability of adaptation of this type of models. And finally, I want to emphasize the issue of marriage of quality improvement and AI predictive modeling. Within current hospital systems, there's an entire group of patients work on QI, and then you have the machine learning people work on predictive modeling. I think there are a lot of opportunities at the intersection of these two groups towards building a continuously learning healthcare system, which my prediction is maybe by 2050, we will have this type of systems. With that, I'd like to acknowledge my funding sources, as well as my collaborators across UC San Diego, Emory University, and other hospitals. Thank you for your attention. I'm happy to answer any questions that you may have. Good afternoon, and thank you so much for attending our session, Surviving Sepsis 2071. My talk is the closing talk, and the title that I was given was, Now That We Know It, How Do We Implement It? And as I was preparing the talk, I wasn't really sure that I knew exactly how to implement, so I kind of modified the title just a pinch. So I entitled it, Now That We Know It, How, When, and Can We Implement It? So it's my pleasure to be the moderator for this session, as well as to present this particular talk. I do not have any financial disclosures. However, I am a member of the Surviving Sepsis Campaign Steering Committee, and I serve as a group head for long-term outcomes and goals of care for that particular section in the 2020 Surviving Sepsis Campaign Guidelines. So as most of you are aware, in 2017, the World Health Organization recognized sepsis as a global health priority, reporting that in 2017, there were almost 49 million people with sepsis, and by the end of 2017, there were about 11 million patients who died, unfortunately. And that information is from a RUG publication from 2020. And in the resolution, the World Health Organization resolution, they had actions for reducing the global burden of sepsis. And the resolution actually included 17 recommendations and suggested actions for reducing the global burden of sepsis, with one recommendation to promote research to develop innovative methods to address sepsis to develop innovative methods to prevent, diagnose, and treat sepsis. However, it's easier said than done. And oftentimes, when we think about sepsis research, you know, we recognize that clinical trials are really difficult to implement, and just caring for sepsis patients is quite complex. So I like this term here, simplexity, as we care for patients and continue to do research in this area. The objectives for this particular talk today, I'm going to go through three different sections. We're going to talk about the past. In order for us to move on to the future, we certainly have to understand a little bit about the history. I'm going to talk about research, you know, for the past 30 years or so, and then talk about the present, based on what our faculty has shared with you this afternoon, and then just touch on the future of sepsis. So the final section will give us a little bit of a feel for what I think we may see in the future. So what were you doing 30 years ago? I was working in a non-medical field, and I had zero knowledge about sepsis. So certainly things have evolved over time. In the 1990s, I guess when you stop and think about what you were doing, some of you may not have even been physicians or nurses at that particular time, just like me. However, we need to look at the past to inform the future. So for the past 30 years, we've been unsuccessful in finding effective, innovative therapies for sepsis. This is likely due to inappropriate animal models to test therapies, which seem promising, and then they go on to be ineffective in humans. Also contributing to ineffective therapies is inappropriate patient selection, which is more complex than most of us could have imagined 30 years ago. Additionally, there may be other reasons why clinical trials failed, one being how sepsis was defined in clinical trials. In this paper published in 2014, the author reviewed the process of how sepsis definitions were derived over time, and indicated that with several consensus conferences that felt that the process was flawed, and that that particular definition was based on a guess. It was a heavily flawed guess in his opinion, and also indicated that research using this definition will consistently render variable and unreliable results. Yet that was all we had at the time, and as we know, the sepsis definitions were again redefined in 2016, which we're currently using as a sepsis 3 definition. The concern here was that using a simple definition may actually inflate or give a false sense of improved outcomes, and using particular diagnoses without any diagnostic tools, they felt that this was really flawed. And I think we probably could do a better job, but we certainly were just using the tools that we had at the time. And again, it's hard to imagine how many hours, how much money, and how much blood, sweat, and tears that scientists, sponsors, and investigators, research coordinators, and patients contributed to sepsis trials. Yet we have so much work to do before we can actually even get to implementation. So first, I think we need a strategy. And before we can actually implement, you know, there's a cut point here where we think about strategy and implementation, but they really need to be thought of separately. And before you have successful implementation, you have to have a strategy that delivers a successful product. So this strategy always comes before implementation. So now we're going to talk a little bit about, a little bit about and review some past information to inform our strategy. Strategy always comes before implementation. So now we're going to talk about and review some of the past information to inform our strategy. As we review the strategy, a vision and mission should evolve. So why have we experienced lack of success? These failures are the steps in the learning process that we have been climbing for years and hopeful that success is in our future. But some of the failures that have occurred include lack of precision in sepsis definitions, a heterogeneous nature of clinical diagnoses, lack of accurate diagnostic methodologies, lack of predictive biomarkers to actually direct clinical trial design. Over recent years, we've implemented an adaptive trial design, which I think is helpful, but again, we still need predictive biomarkers to help us in this process. And last but not least, diagnostic criteria for sepsis are vague and can be confused with other diagnoses. And again, when we think about how we've worked and improved and looking for success and finding the magic bullet, unfortunately we have episodes where we do fail. We continue to learn. We may fail again. And again, it's a learning process every time we fail, but our hope is that we are building a process here where we will have success in managing and treating this diagnosis for improved outcomes for patients. So how do the talks today represent promise? In Dr. Wang's talk, he shared information related to biomarkers and how they may be most beneficial in dealing with heterogeneity. So using predictive enrichment uses patient characteristics to study the population and determine which drug or intervention is most effective in a particular population. Prognostic enrichment strategies to select patients where they're more likely to respond to an intervention or drug based on biologic or physiologic mechanism. And again, these enrichment strategies are highly effective potentially in heterogeneous syndromes such as septic shock. And he shared his data based on pediatric septic shock population. And here we can see that measuring biomarkers will help predict and better match a particular therapy. You heard from Dr. Morowitz on how leveraging the microbiome may be effective to fight sepsis. Again, showing that there's a place for microbiome and stewardship, highlighting diet, antimicrobials, and ICU hygiene. And Dr. Hellman, she spoke to us about immunomodulation. It's now recognized that following the pro-inflammatory response in many patients, there's an emergence and persistence of an anti-inflammatory response that may be predominant and clinically meaningful, and that immune enhancement therapy could be beneficial in those patients. Although many people think this is only coming into prominence in the past few years, this particular reference is actually from 1996 from Dr. Roger Bone, who was really a pioneer in sepsis research and diagnosis. So again, we are hopeful that these new therapies will potentially be successful as far as immune enhancement therapy. There's also been a recent resurgence related to immunotherapy due to the success of immune enhancement therapies in cancer. Here we see that using immunotherapy in cancer, as a little diagram here, cancer is not attacked directly, rather the immune system army is released to battle with all its weaponry. And perhaps this type of therapy may be helpful in some patients with sepsis as well. And finally, Dr. Namadi shared how real-time implementation of predictive analytics using artificial intelligence models need to be fair, inclusive, trustworthy. So the team has to be able to use this tool and trust that it's appropriate and that we can interpret the information and as well, that it's important that it's valid and generalizable. So one of the key points that our faculty pointed out is that the secret to success is patient selection. There's an increased emphasis on personalized medicine, the right patient, the right time, the right dose, and one size does not fit all. A procedure of future individualized medicine is presented here. It involves a more complicated approach to include stratified medicine, precision medicine, and theranostics to get the right patient, the right drug, at the right dose, at the right time. So where is the hope of success for 2071? We're looking for faster, more reliable diagnoses, perhaps using machine learning or real-time AI. Biomarkers for diagnostic, predictive, and prognostic outcomes. Improved clinical design to stratify or identify patients that would benefit from treatment using prognostic or predictive enrichment. And new targeted therapies, such as immune therapies and microbiome. Again, we're looking for the hope of success at this point. I'm not sure how many of us are going to be around in 2071, but we need to have hope that over the next 20, 30 years that we can actually improve care for sepsis and outcomes. So surviving sepsis in 2071 may include several ideas here that I have. It may be that we revisit failed trials. There may be some hidden gems. Perhaps there were trials that we did 30 years ago that potentially would have been successful if we had biomarkers to identify the appropriate patient. Again, some drugs, some devices may work in certain patient populations. I'm sure you've all seen in our clinical experience, our research experience, it appears that maybe a drug is working, but yet the drug fails to show statistical significance with mortality. And again, when we think about how things occurred years ago, you know, with improvement and biomarkers and stratifying patients, we may actually find that some drugs potentially could be effective. And then the use of biomarkers and artificial intelligence may be helpful in diagnoses, treatment, and outcomes. We may actually be looking at outcomes very differently. For many years, we focused our attention on 28-day mortality. We've expanded that in some clinical trials for sepsis in 90 days, and even in some trials, we're looking at mortality at a year. However, there may be other important outcomes that we should be interested in as far as quality of life. So thinking about how we look at clinical trials in 2071, I'm hopeful that it's not only the quantity, how many years a patient lives, but what their quality of life is after surviving sepsis. And last but not least is a thought that, you know, maybe we're actually all going to wind up with some sort of implantable microchip as a health tool. They do have microchips that people are using implanted as a fob to get in and out of buildings. But it may be quite interesting in 2071 if we have this microchip, and the microchip is able to identify that we've been invaded by an organism, and it tells us on our smartphone or whatever type of phone we have at that particular time that we need to go to the nearest emergency room because we've been invaded by an organism. So it sounds pretty far-fetched, but you just never know. I'd imagine in 2071, technology is going to be incredible compared to where we are today. So this is certainly a possibility. So with that, I thank you for your time, and I appreciate you continuing to stay along here, and we will move on to the Q&A session. Thank you so much.

sepsis

biomarkers

PERSEVERE model

microbiome

immunomodulation

immunomodulatory therapies

endocannabinoid system

TRPV1 receptor

inflammation

endocannabinoids

anti-inflammatory effects

animal models

cannabis-derived cannabinoids