Thank you. I, too, did not receive my assignment and the title of my talk until after I agreed to do it. So somehow I got stuck with the hardest step, which is kind of like when you're absent from a meeting and you get assigned the worst task, but I agreed anyway. And fortunately, sepsis diagnosis is like my favorite thing to talk about, so I wasn't too mad about it. I do receive some grant funding to study sepsis care delivery, but I don't have any conflicts of interest related to this talk. So I'll ask you to start with sort of this imaginary clinical scenario that you've probably been in hundreds of times before, and that's imagine you're called to see a patient who is complaining of shortness of breath and cough. He has a temperature of 99.9, he's a little tachycardic, O2 sat's 88%, his blood pressure's a little squishy, and this is his chest X-ray. So you're thinking, hmm, he might have bacterial pneumonia, and there might be some signs that there's some organ dysfunction coming on, so he might have sepsis, but he might have heart failure, he might have influenza, or he might have some other condition that also presents really similarly to those other conditions, but you don't have time to really figure out definitively what this patient has. You're asked to treat him quickly in order to improve his survival. So you give this guy antibiotics now, do you give him some diuretics and see if that helps matters, or do you work him up for a viral infection and provide supportive care? You can do all of these things at one time, or you can do none of these things, or a combination of these things, but really the idea is just to think about how often we're in this position where we really don't have any idea what's going on, but we're asked to make a treatment decision. And this is really the enduring, vexing challenge of sepsis diagnosis, that it is a fickle syndrome with these signs and symptoms that overlap so dramatically with all these other conditions that can also be serious and have different treatments, sometimes conflicting treatments. And we're asked to make a decision really fast on how we're going to treat the patient while there is still quite a bit of diagnostic uncertainty. So indeed, this is the hardest step. And so how can we better identify sepsis? Hospitals and health systems have really thrown a lot of effort into improving their identification of sepsis, and lots of different strategies have been used to improve the detection of sepsis. For example, automated surveillance of electronic health record data to detect clinical signs of sepsis early, maybe before an astute clinician might pick up on this. Lots of work on artificial intelligence and unsupervised learning techniques that can identify patterns of clinical data that represent unrecognized constructs that can predict sepsis, even before that's observable to a clinician. Lots of hospitals have implemented, most hospitals have probably implemented standardized sepsis detection and diagnostic pathways. And then in the lab, we're still working really hard. I'm not working. Some people are working really hard to develop these rapid assays that can detect biologic markers of host or pathogen response so that we have more of a definitive way to diagnose sepsis. Unfortunately, though, none of these things have really shown effective translation into clinical practice. Raise your hand if your hospital uses at least one of these things to detect sepsis. Raise your hand if you use more than one of these things. Yeah. So they're widely distributed, widely used, and really in the absence of a lot of rigorous testing and the absence of good trial benefit of these strategies. And in fact, for randomized controlled trial and sepsis detection efforts, really randomized evidence to support that is almost nonexistent. These clinical alerts, in particular, have been widely distributed really in the absence of prospective evaluation. Most of you are probably familiar with this widely publicized article by Andrew Wong, who's a superstar resident at the University of Michigan. And their team did an external validation of this proprietary electronic sepsis prediction model. And they found that it really wasn't very good in real world data. So the EPIC sepsis model only picked up 7% of sepsis cases that had not already been identified by clinicians. So treatment had not already started. It missed 67% of true sepsis cases, despite the fact that it fired on almost 20% of all hospitalizations. And so this is not surprising to any of us that interface with these type of clinician alerts. They fire all the time. You get all sorts of alert fatigue. But they're really not providing a ton of marginal benefit over what you already know and already do at the bedside. Another thing I want to point out with these prediction models is there is concern and some evidence that suggests that if not carefully attended to, these models can propagate biases in care delivery. And so that really needs to be something that a lot of attention is paid to when we develop and implement these prediction models. You know, it'll be great when we have a biologic diagnostic test to definitively rule in or rule out sepsis, kind of like troponin for MI or imaging and stroke. But currently, the state of the science for these type of approaches really continues to be limited by financial, regulatory, logistic barriers. So this is not quite ready for implementation in clinical practice. And so I'm going to cop out on my assignment a little bit. And I'm told to answer, how can we better identify sepsis? I'm going to argue that that's not really the relevant question for sepsis care delivery based on the science that we have right now. We can't get better at identifying sepsis because we don't have the right tools right now. So really, the appropriate question for sepsis care delivery right now is, how do we risk stratify patients in a way that helps us know when patients will benefit from aggressive sepsis treatment and when that benefit will outweigh the harm of aggressive sepsis treatment? So it's really a risk stratification problem. And so if we kind of accept that we're not going to definitively know whether patients have or don't have sepsis and we're going to have to act under some uncertainty, this really becomes kind of a decision analysis problem. And we can make the best choice if we have information about how likely the patient is to have infection. And we have good estimates of what the benefit of aggressive early antibiotics are and good evidence about what the harms of aggressive early antibiotics are. Let's skip this. So what do we know about the probability of survival with and without antibiotics? In other words, what the benefits of aggressive antibiotic timelines are. We're pretty definitively sure that patients who are really sick, septic shock patients, the faster you get the antibiotics in, the better. We're pretty confident in that. We're less confident in how important it is to really rapidly deliver antibiotics for patients who are less sick, so patients without shock. And we're even less sure about what to do with the broad group of patients who may or may not have sepsis, whether rapid antibiotic delivery has a population benefit to that group. Most of the estimates that, or all of the estimates that we have to answer that question come from observational studies. Some have better methodology than others, but they all have some methodological limitations related to observational nature. Fortunately, we should have some information in the next few years that's actually randomized trial evidence for the benefit of aggressive antibiotics at three hours versus one hour. So the AIMS trial, and I think I saw Mitch Levy in the audience, is led by Mitch Levy. So this is an NIH-funded study that is evaluating a three-hour antibiotic target versus a one-hour antibiotic target. This trial's going to enroll, I think, 10,000 patients across 20 hospitals. It's cluster-randomized, and so we should have some answers about whether a three-hour antibiotic target or a one-hour antibiotic target is better for 28-day mortality. So really excited to get the results of that a few years down the road. If you don't know, SCCM serves as a data and clinical coordination center for that trial, which is really cool. Okay, so we're hoping to learn more precisely about what the benefits of antibiotics are, but we also need to kind of get a better sense of what are the harms of these aggressive antibiotic targets that we have for sepsis. I think this is something that's been understudied until recently. It is a polarizing topic, and the discourse tends to be, in my opinion, either everyone should get antibiotics. It doesn't matter. There are no harms, nothing to see here. So we're not even going to look for harms. Or I forget what you said about the meme that you said, but basically, antibiotics are going to be the end of civilization, and we shouldn't give them to anyone. It will wipe out your children and your children's children. No one should get antibiotics. And so we've had this kind of polarizing rhetoric, but no one has actually dived in to see what are the actual harms, and if we can quantify those, then we can really make a better choice about the tradeoffs. There's some evidence that we can be a little reassured about aggressive antibiotic targets actually don't lead to indiscriminate, willy-nilly antibiotic use. This comes from work done by Hallie Prescott and Vinnie Liu. They combined a huge data set of VA patients and Kaiser Permanente patients and evaluated temporal trends in time to antibiotics. So over the time period, hospitals, on average, got faster in their delivery of antibiotics, which is great, and there was no evidence that these hospitals that got faster actually treated more patients with antibiotics. So this was kind of reassuring. It gave us the idea that hospitals can increase the delivery of antibiotics faster without indiscriminately treating everybody with antibiotics, or at least in these two healthcare systems that can be done. On the other hand, there's data that suggests that we need to be careful about using antibiotics, that even one dose can have an effect on the microbiome that can be really significant. And if you talk to Bob Dixon for more than five minutes, you may never use Zosyn ever again. This results from what I think is a really elegant study in the European Respiratory Journal. They evaluated a sample of about 3,000 mechanically ventilated patients, compared groups that had received an anti-anaerobic antibiotic, so Zosyn, or a not anti-anaerobic antibiotic. And what they found was really interesting. So patients who had gotten anti-anaerobic antibiotics had their microbiome obliterated, just one dose, obliterated the microbiome, and there were clinically significant effects, including fewer ventilator-associated pneumonia-free days, fewer infection-free days, and lower 30-day overall survival, all depending on whether you got an anti-anaerobic antibiotic or not. So there's some evidence that we have to be really careful that there are effects that we don't even know about or we're just learning about when we think about giving patients broad-spectrum antibiotics. And then also thinking about harms from kind of a 30,000-foot view, not just individual patient adverse effects, but what are the opportunity costs of these aggressive antibiotic targets? So anyone who's been involved in trying to meet a one-hour sepsis bundle target knows that it is a lot of work, a lot of moving parts, a lot of things have to happen just right to get that to happen. And we think about what is the opportunity cost, in other words, what could benefits of using those resources to do something else, and how do those tradeoffs shake out? So within sepsis, Dr. Simpson mentioned that there's a whole bundle of things we have to think about in the first hour. We think antibiotics are probably the most important, but if getting those antibiotics in really, really fast delays source control or delays restoration of hemodynamics, maybe that's an opportunity cost we don't necessarily want to do. I don't know the answer to that. We need to study that more, but something to think about. And then another thing that we have to think about is all of this effort that we put in for aggressive sepsis metrics, what other conditions are other patients are missing out? And this really came up in some of our qualitative work where we had an emergency department nurse say this, when a code sepsis is called, you drop your other patient, your other family, your other conversation, and you run to do the code sepsis. It's extremely disruptive. And so in cases where the patient has septic shock, that disruption is probably really important and the benefit is worth the risk, but in cases where the patient may be less likely to have sepsis or ultimately does not have sepsis, we've really sacrificed the attention that we could have given to the cardiac patient in bed four or the trauma patient in bed six. And so we really need to understand more about the whole picture of what we're doing. And it sounds like I'm really sounding the alarm here on harms. I don't mean to do that. I just think that we have a lot more research to do and a lot more focus to understand where the balance lies. And so while all the great scientists in the room are trying to sort out those questions that I pose that we don't know the answer to, what is a rational decision maker supposed to do when you're presented with someone with suspected sepsis? And I really love this conceptualization. I use it in all my talks. Hallie Prescott and Jack Awashina put this together in Annals of ATS. And the idea is that your antibiotic timing depends on both the severity of illness and the likelihood of infection. So for patients who are really, really sick or really, really likely to be infected, those patients should be prioritized for rapid antibiotics, ideally within an hour. Patients who are not that sick or you're not very sure whether or not they have infection can wait a little bit longer while you can gather more information and decide whether or not they truly need antibiotics or maybe tailor the antibiotic therapy to a more narrow spectrum. And this was really well reflected in the updated treatment guidelines from 2021 that Dr. Simpson put up. So a surviving sepsis campaign recommends this sort of risk stratified approach to antibiotic timing, where if you have shock, those patients should be prioritized for antibiotics within one hour. It doesn't matter whether you think sepsis is kind of likely, sort of likely, totally likely, whatever. These patients don't have time to waste. Antibiotics should be prioritized within an hour. But if shock is absent, and probably if they even don't have shock but not that sick, we need to make that decision based on how likely we think infection is. So if sepsis is definite or probable, you should get antibiotics on board quickly. If sepsis is only possible, then you have more time to kind of consider and tailor your therapy. So I was really excited to see this. I felt like this was a big move towards kind of rational prescribing of antibiotics. One question that I had when I saw this is, what does probable mean? What does it mean for sepsis to be possible or probable? I'm kind of a numbers person. I want to know, what is that threshold where we would decide to treat? What threshold would tip you over to treat in a patient that wasn't that sick? And so we did some decision analysis work where we had clinicians respond to vignettes so that we could understand what their threshold was for treating sepsis, what infection probability would they start antibiotics. So we presented these vignettes that presented physiologic and laboratory data over time, and we had clinicians indicate at what hour they would start antibiotics. And we used some preexisting infection probability models to assign each hour a certain infection probability. So we could ultimately end up with an infection probability that was the threshold where clinicians would start antibiotics. And on average, that threshold ended up being about 69%. So across eight vignettes, 153 clinicians and like 130,000 hours of decisions, on average, clinicians would start antibiotics when infection was 69% likely. But this varied, as you can imagine, significantly with severity of illness so that in patients with high severity of illness, the threshold for starting antibiotics was 55%. For patients with a low severity of illness, the threshold for starting antibiotics was 84%. And that fits with kind of how we practice, and it fits with the other conceptualizations that we had talked about. Thinking about trade-offs, I find this a little bit easier to convert these percentages to odds. So for a patient with a high severity of illness, clinicians in our study said that they were willing to treat one patient inappropriately for every one patient that they treated appropriately with antibiotics. On the other hand, for patients with low severity of illness, clinicians indicated that they were only willing to treat one patient inappropriately for every four patients that they treated appropriately. So these numbers really help us kind of think about the trade-offs associated with sepsis treatment, can help populate decision tools, help us determine where to set thresholds, and inform kind of the clinical usefulness of our decision tools. I love to put this up because it always riles up the crowd a little bit, that we also looked at differences in thresholds across specialty. And these varied in exactly the way you'd think they would. So emergency medicine physicians had the lowest threshold for starting antibiotics. Emergency medicine clinicians started antibiotics at 50% likelihood of infection. Infectious disease clinicians had the highest threshold for starting antibiotics. They waited until, like, pus is oozing out of someone's body. They waited until a probability threshold of 88%. There's probably a lot of good reasons, rational reasons why there are differences here, but it's always fun to point out. There are also differences in years of experience. So less experienced clinicians had a lower threshold for starting antibiotics, whereas more experienced clinicians were willing to wait a little bit for a higher threshold of infection before starting antibiotics. And so I've spent most of the time talking about individual-level diagnostics or individual-level decision-making with regard to sepsis diagnostic, but really the truth is sepsis diagnostic excellence is a organizational problem. So there's individual clinician diagnostic accuracy, but there's so much driven by organizational climate, culture, processes that determine how diagnostic excellence ends up playing out. You can imagine an organization, and yours might be one, that really promotes sepsis over-treatment because of its policies and cultures and climate. You can imagine an organization that really promotes under-treatment because of all of those factors. And so understanding how those organizational features work and promote one or the other and finding the optimal balance that maximizes sepsis diagnostic accuracy is really the cornerstone of our team's current work. So I hope to get to share more of that with you next year. Some of our early work, just kind of measuring, like Dr. Simpson said, you don't measure it, you can't fix it. So we measured kind of trends in false positive sepsis rates over time alongside these sepsis initiatives and found that the rates of false positive code sepsis increased along with the increased detection that we were having. Not particularly surprising, may not be bad, but important to measure so you know what's happening and maybe surprising that a third to a half of our code sepsis activations actually ended up being false positive. This is really similar to the results from a smaller study in San Diego that had a sepsis alert system rollout and about a third of those activations were false positive. So again, thinking about the trade-offs, this may be okay, but certainly something that you want to measure and think about when you're doing your sepsis quality improvement. And so this is what I would recommend is that when we think about sepsis diagnostic excellence, we continue to monitor sepsis under treatment and I put that in ostentatiously large font because I think this is the most important, this is the work that we're already doing, but I think that organizations also need to measure sepsis over treatment and find the right balance. This will align sepsis programs with antimicrobial stewardship programs and make sure that the work we're doing is balanced so that we have overall good population impact. And I'll end here with this quote because these authors can say this way more eloquently than I ever could, but this is really the take-home message of this talk is that achieving diagnostic excellence for sepsis is less about determining which patients definitively have sepsis and more about improving how to accurately identify who among those with the precursor condition requires intervention to prevent evolution towards fully established sepsis. So it's not so much an identification problem where the science is right now, it's a risk stratification problem and organizations need to get better at doing that so that we can properly treat our patients. Thank you.

diagnosing sepsis

improve sepsis detection

early antibiotic treatment

septic shock

organizational factors

intervention