SCCM Resource Library-Improving the Precision of Early Sepsis Detection

Video Transcription

Hi, my name is Tiffany Osborne, and I'm here to discuss pre-hospital sepsis detection and treatment.  I want to thank the Society of Critical Care Medicine and Dr. Elliott Kruser for inviting me to participate in this conversation today. Thank you.  So pre-hospital sepsis detection and treatment, the first question might be, why are we even talking about this?  What does pre-hospital care have to do with patients that we take care of in the intensive care unit?  Well, a lot of our patients present to the emergency department and to the hospital by ambulance. In fact, in 2017, over 20 million ED visits arrived by ambulance.  Additionally, when looking at the sepsis population, one study that evaluated 850,000 ED sepsis  visits found that over half of those patients arrived by ambulance, and over half of those patients were also 65 years of age or older.  The age is important, because in a sample that represented around 29 million ED visits, they found that as age increases, there's an increased likelihood that patients will  utilize pre-hospital services. So then, how do we find sepsis in undifferentiated patients when we're in a pre-hospital environment?  This has been challenging, as in a systematic review in 2016, they found that there was a lot of variability and unclear ability to differentiate between sepsis and other  etiologies. In 2020, the authors re-evaluated and updated their data, and there were significant improvements.  However, there is still no single scoring strategy that has high sensitivity and specificity.  Additionally, there is significant variability, not only between individual scoring systems, but actually among individual scoring systems.  And that variability is really based on the fact that their ability to correctly identify and differentiate sepsis patients had to do with the number of the individual criteria  within each score that were met. And this could be translated to mean that potentially sicker patients were easier to be identified in a pre-hospital environment.  That stands to reason because the same is true within a hospital environment. So in general, we can say that looking for sepsis patients within an undifferentiated  patient population is difficult within a relatively stable environment of the emergency department.  When you take that environment, now you're changing it where there are increased distractions,  danger, time, and even more pressing time-sensitive element, it is not inconceivable that it would be very difficult to identify these patients.  So how might we improve our clinical risk scores in the pre-hospital environment? So the data that we're going to talk about is still developmental.  It's more hypothesis generating, but really important because it helps us to determine or to inform future more rigorous studies.  So one evaluation looking at around 300 sepsis alerts decided to utilize not only their clinical sepsis screen, but to add end-tidal CO2 to that.  And the end-tidal CO2 was specifically less than or equal to 25. And they found that patients where the protocol were compliant, where EMS providers were compliant  with utilizing the protocol, these patients were more likely to be correctly diagnosed with sepsis or severe sepsis than those who were not compliant with the protocol.  End-tidal CO2 also was significant in helping to identify sepsis, you can see the AUC here,  as well as severe sepsis, and was an important contributor to the prediction of 30-day mortality.  So a Bayesian clinical model of a pre-hospital sepsis score looked at utilizing this pre-hospital  sepsis score that included clinical elements, including shock index, and then to that added lactate.  And they found when they added lactate that they improved their likelihood ratio for being able to accurately diagnose sepsis, as well as accurately diagnose those who did not have  sepsis. And in fact, their absolute gain increased from 6% to 12.5%. In a letter to the editor, Bryant et al. looked at 452 pre-hospital patients, and they'd calculated  their pre-hospital clinical risk score, which was a validated clinical risk score. And they measured eight biomarkers in the hopes of determining how that would impact  their primary outcome of identifying community sepsis. And community sepsis was present in around 13% of 57 patients, and what they found was  that by adding the biomarker to their clinical risk score, they were able to more accurately identify and differentiate sepsis.  As you can see, this is their best-performing panel and the individual components along with their AEC.  The same authors looked at whether or not pre-hospital treatment might be able to impact specific inflammatory markers.  So in a prospective cohort evaluation, these authors looked at 345 patients and drew IL-6, IL-10, and tissue necrosis factor both pre-hospital and upon emergency department arrival.  Of those patients, around 88 or 26% had received a fluid bolus, and that impacted IL-6 by reducing IL-6. However, it had no impact on IL-10 or TNF.  So how does all of this potentially impact patient care? Well, one is by demonstrating that what happens in the pre-hospital environment also contributes  to severity of illness and helps to potentially predict outcome. So in this evaluation of over 1,200 patients, the National Early Warning System looked at  how the ability to predict mortality would change over time, but then how that changed when they also added lactate.  So here you can see that they looked at the new score pre-hospital, ED, and ward, and they demonstrated that when it remained consistent, that the mortality increased, and that that  was specific when the news was greater than versus less than 5. When they added lactate to the mix, they were able to demonstrate an increase in mortality over the new score alone.  Additionally, another evaluation looked at how we could impact CMS sepsis core measures in a pre-hospital environment.  And so what they looked at was the pre-hospital sepsis assessment tool. And they trained 120 paramedics who were called on over 1,100 sepsis alerts and engaged with  over 56,000 patients. And those paramedics were doing things such as obtaining IV access, drawing lactate, blood cultures, and administering antibiotics.  And the admission diagnosis for this population, half of them were sepsis, 15% were severe sepsis, and 9% were septic shock.  Now interestingly here, less than 5% of the blood cultures that were drawn in the pre-hospital environment were contaminated, less than 5%.  Almost 20% of those uncontaminated blood cultures had bacterial growth that was treated. And EMS administered antibiotics.  Those antibiotics matched the cultures, meaning that they were effective antibiotic choices in 72% of patients.  Now we also have seen a randomized control trial of over 2,600 patients in their intention to treat group where antibiotics administered in the pre-hospital environment did decrease  their time to antibiotic delivery by a median of 26 minutes. It did increase awareness of the septic patient's arrival across the acute care chain, but there  was no difference in mortality. So in summary, a lot of this data is preliminary and can be used to inform more rigorous studies  in the future, but it is very important in our initial assessments. It helps with identification, and identification may be improved with clinical risk scores added to biomarkers.  Early detection appears to be associated with early treatment, which could impact our outcomes. And pre-hospital treatment appears to be timely, capable, and effective.  Given that this is our 50-year anniversary with SCCM, one of our founders said something that was relevant.  Dr. Peter Safar many times had discussed the fact that critical care is really a continuum, that we're not hidden behind the golden walls of the ICU, and that continuum begins with  out-of-hospital care, it continues with ED intervention, and it culminates in ICU management.  So I believe that the idea that he was putting across is that whether we're looking at pre-hospital  providers, the emergency department, the intensive care unit, or the ward, we all have something  important to bring to the table, and our collaboration is very important in the outcome of our patients.  So in general, our patients need us to be more than individual service lines. They need us to work collaboratively to provide a cross-continuum service, and never has that  been more clearly demonstrated by critical care providers in this past year. And I just wanted to say thank you to everyone for all the work that you've been doing.  I know that it's been challenging. Each and every one of you have met that challenge, and I just want to say thank you for making this a career to be proud of. Thank you. ♪♪♪  Thank you, Dr. Osborne, for your excellent presentation. I want to ask everybody to  wait till the end of the presentations for the Q&A session, and I'd like to move ahead to the  second talk by Dr. Rishi Kamalavaran, who is an assistant professor of biomedical informatics at  the Emory School of Medicine and also has an appointment in the Department of Biomedical Engineering at the Georgia Institute of Technology, and we're going to learn something  about machine learning for the early identification of sepsis, and I think it's a very interesting angle on that, even using autonomic kind of signals and physio feedback.  So with that, I'll have you, Rishi, please take it from there.  Thank you. I'll be sharing my screen.  Okay, thank you for the nice introductions, Dr. Kirshner. So this is a presentation on big data biomarkers. So this will be slightly different  from what you may have been listening to in the last few presentations.  Disclosure, I am a consultant expert consultant for BlueSpark Technologies, which is not related to the content that I'll be presenting today.  The objectives of this presentation is one to understand the need for an artificial intelligence led intelligent clinical surveillance system  in the ICU to review the challenges of developing and the results of an AI driven sepsis screening tool that's based on physio markers  to introduce methods for capturing and the analysis of what we call temporal big data  and review the methods of applying artificial intelligence to predict the onset of pediatric and adult sepsis and finally introduce an integrated physio marker  and biomarker approach for real-time intelligent learning systems and critical care.  So as a background, many of you would have certainly heard of the word big data artificial intelligence machine learning. It's now become more than just a buzzword.  It's actively starting to be used in many of our systems in our hospital systems,  in the electronic medical record systems and so forth.  So as a background, the derivation of biomarkers from sensors or continuous data sources  such as your electronic medical records has become easier than ever before.  Almost all the medical hospitals  in the United States have now developed  a robust clinical data warehousing tools that allow for users to access, derive and extract meaningful data from their medical systems  in order to use it in the form of quality improvement or other research tools.  There's also been interoperable middleware that has been developed to support the integration of heterogeneous silos, largely these systems that have never really been integrated  from a forward speaking sense, such as your radiology packs, your medication pharmaceutical databases and so forth.  The integration of these disparate sources of data has resulted in novel biomarkers that serve as indicators of disease.  So this presentation, I'm going to focus on one of the aspects of those various sources of data  and I'm going to be focusing very specifically on continuous data that's largely prevalent within the critical care environment and that is data that's collected  from patient monitoring tools, ventilators, infusion pumps and so forth. The derivation of biomarkers from the sensor or continuous data sources has become an area  of major interest in recent times.  Asynchronous biofeedback from various components  of the autonomic system has been shown to indicate the onset of disease. One of this, as the images that shows on the right,  you can see that there's often an intricate communication between the various components of our autonomic nervous system that indicates sympathetic parasympathetic activity  and modulations of these pathways have been shown to occur in the early onset conditions, primarily in sepsis.  About 20 years ago now, there was a seminal work that came out of the University of Virginia led by Randall Mormon and his team that showed that altered heart rate variability  prior to the onset of neonatal sepsis was actually very predictive of longitudinal trends such as mortality and complex disease trajectories.  So this really allowed for a number of work across both the pediatric and the adult spaces,  which indicated that the high fidelity data could themselves afford a sense of situational awareness that just is not available in slower, lower changing modes of data.  This here is an image of an infant in the NICU that has connected to a number of data sources, which includes the electrocardiogram, the IV pumps, the oxygen saturation monitor,  the ventilators and so forth. If you aggregate all of this data, which is data that's being produced at hundreds of samples every second,  you actually can get several millions of data points that's aggregated per patient per day. So in this figure, conservatively speaking,  we're talking about 40 plus million data points that are recorded per patient per day. What we often find in the ICU is that that volume of data gets summarized  as vitals that are entered on an hourly basis into the electronic medical record. So there's a significant amount of data that's lost on a routine basis.  What is big data? So typically when we think of big data in the clinical sense, we think of genomics and so forth. But just as a comical reference of sorts,  if you were to just look at the haploid archival genotype of patients, you can actually fit that into a CD-ROM.  If you actually look at a sensor data, for example, you can see that just this four second snippet of data can actually accumulate to several hundreds of megabytes  that's acquired per patient per day. That's in just simplistic terms, that's one whole human genome that's produced every four days.  So what are these events and how do they appear? So typically from a computing perspective, we refer to these events in a number of ways.  Some of these are indicated as primitive events, multidimensional events and physical events. And these can be either present within a single sensor, such as a single lead ECG,  or it can be present across multiple leads of sensors, such as the pulse oximeter, the infusion pump, the incubator, so on and so forth.  So there's a number of sources of sensors that exist and each of them may provide a different information about the health status of a patient. For example, the ECG may indicate  whether there's atrial fibrillation or whether there's tachycardia or bradycardia. The PPG sensor may tell us whether there's low oxygen saturation and so forth.  So again, these are all different modalities, which if we integrate together within a hierarchical system, we can actually draw quite unique inferences.  In a typical sense of the analysis world, people have often analyzed data as though the data was fixed at a certain point in time.  Instructions were sent to the database, which then would churn through the instructions and provide a result. But the assumption underlying is that the data does not change  because if that data changed, that result would no longer represent the underlying totality of information. The recent paradigm over the last several years  has been to see data as data in motion. And what we mean by data in motion is that data is represented only within its temporal characterization.  And when instructions are sent to the data, they are applied within an epoch and then results are provided back, which informs the person that's submitting the query  of the actual changes that occurred within that time period.  We have now started to become more robust in how we can capture this data. Not only can we capture this from a variety of different sources, both static and streaming,  we can also apply machine learning models within this infrastructure so that we analyze the data in real time, we extract insight from that and then feed that to the end users  so that they can view this information and perform some sort of response back to clinical care.  Bringing the analytics to the bedside is the ultimate goal, our overarching intent. And for that purpose, what we're hoping or what we're wanting to do is capture the data,  extract these meaning, these physio markers from the underlying sensors, such as your heart rate variability and so forth,  and extract them on a continuous basis so that we can identify when there are abnormal signs and do all of this before the onset of sepsis and the theoretical onset of sepsis  and the clinical recognition that's associated to it and the subsequent outcomes such as death  and take therapy, which is currently lagging clinical recognition and potentially allow for us to take that prior  to where that knowledge currently exists. And so early initiation of therapy is our ultimate goal.  We have developed some of these infrastructures within the real-time space to abstract the continuous data  from the bedside monitor on a concurrent basis. So as it's generated at the bedside, we would capture it, we would analyze it, we would extract meaning from that,  and then we would provide some classification. So in this case, we're referring to proximal atrial fibrillation, and that would be an alert that gets sent out.  But underlying that is a system that captures this data continuously and models and abstracts information from it.  We've also seen how this data, especially the streaming data can be incredibly useful, even when you take it at the most fundamental element of this.  So without including any other data, such as clinical or labs, just by looking at the high-fidelity physiologic data, we're able to see some very interesting insights.  Here's some results where we show that you could take a model that you trained using just routines, the clinical data that's sampled and entered into the electronic medical record  once every 60 minutes as routinely performed. And if you just double that and sample that data automatically from the bedside monitor once every 30 minutes,  you start to see an increase in model performance. And when you do that subsequently to data that's sampled automatically from the monitor at once every minute,  you see a significant increase in the overall performance of our machine learning models. So the take-home message from this is that higher granularity actually allows us  to see more important information about the physiologic stability of that patient.  In some of our prior work, we've shown that when we apply this machine learning pipeline within children  who had severe sepsis, we can see a very robust performance. We've shown this in the form of results that we acquired from logistic regression and random forest.  We see the sensitivity specificities as fairly robust both across both models with logistic regression performing a little bit better  on the sensitivity and somewhat less on the specificity.  But ultimately the model seeks to use what we had defined as an event that was clinically defined as an event and using data from two hours prior to that event  to eight hours before that event, we trained this machine learning model to recognize those characteristics. And it was then able to use that subset of data  to generate our machine learning alerts.  When we, in the adult space, when we apply the sepsis three definition, we then took the sepsis three definition on about 6,000 adult patients that were in the ICU.  We had about 10% sepsis in terms of the cohort.  And when we apply this same pipeline using just the minimal set of physiologic data, heart rate, blood pressures, respiratory rate and temperature,  sorry, not including temperature, but just those subsets, we see that we are able to identify the onset of sepsis as defined by the sepsis three definition  up to 17 hours before the onset of suspicion.  The AUC is fairly low here, which is about 0.78.  But considering that this is just data from the vital science and not including any labs or any clinical assessments, we think that this is kind of the baseline.  And then once we include more rich contextual information, we can improve on this work.  Interestingly, when we actually look at the data, you see a very robust physiologic model that starts to show up. And if you see, for instance,  we show the model, the importance of a certain variable from about one hour prior to the onset of sepsis all the way back to 21 hours before the onset of sepsis.  And what we see here is that heart rate has a very early signature and that signature continues to dominate until about the seventh to fifth hour before the onset of sepsis,  during which time respiratory rate starts to take over. And respiratory rate is one of the key indicators of the dominant features predictive of sepsis  in the hours immediately prior to the onset of sepsis. You see that there is not a significant fidelity  in the blood pressures, but they do show up near the immediate onset. These blood pressures were derived from cuff measurements.  And so we think that if we had continuous measurements of blood pressure, we actually could perform better and that we would start to see some more interesting interactions here.  So that brings us to the problem that we have in continuous measurements of blood pressure. This is a major challenge because less than a third  of patients admitted to the ICU are on an ART line. And if we're looking to predict sepsis early  and if blood pressure is indeed important, then having access to a non-invasive source or an estimation of blood pressure becomes very important.  So there is significant literature now over the past decade or so, where different authors have presented or proposed physio markers that predict,  that estimate blood pressure from the integration of the ECG sensor and the pulse oximeter. So one of those key biomarkers here is this pulse arrival time,  which is sampled from the R peak of the ECG up till the peak of the PLEC signal. And that tells you how long it took for the volume of blood  to leave the heart and enter the peripheral capillary at the time that this pulse was detected.  So this again is very interesting information and it allows us to then from a modeling perspective,  combine these two information and propose a non-invasive blood pressure estimator. And this is something that we've done using methods called domain adaptation  that are fairly recent. And you may have seen this in some examples where you submit a picture of a zebra and then you get that zebra to morph into a horse.  It's the same type of domains, relatively speaking, but the invariant features of those domains get manipulated and you now suddenly can change between those two domains  fairly robustly. That's a theory that we have in our estimation is that you can actually take a pulsatile signal like arterial blood pressure and a surrogate pulsatile signal,  which is the pulse oximeter. And by combining those two pulsatile signals, although there is some phase difference because one is further upstream the other,  we can actually derive some of those common information and use that to estimate blood pressure.  We've done some of this work that's ongoing work that we're doing. And if you look on the left side, you start to see that a naive model of estimating blood pressure is presented  and you can see that the true arterial blood pressure is in blue and the estimated blood pressure is in red. And you're seeing that this is underestimating systolic  and overestimating diastolic in many cases. And you see that more prominently here that the diastolic components are being significantly overestimated.  But if we apply some of these modeling adaptations through multiple sensors and changes in algorithms, we start to see somewhat of a closer approximation,  but still this is a very challenging problem. And I think we still have a bit of ways to go before we can get to that medical grade solution.  Tying that all into the most heterogeneous space of disease that we know, the space of sepsis,  I think it's a very challenging problem because there are just so many potential pathways and states that exist.  So our goal is to identify these biomarkers that predict the outcomes, but we're often limited by when we sampled those biomarkers, the timing of that sampling becomes really key.  And because we oftentimes don't know the underlying biological processes  that are going on in real time, it's hard to know precisely where the patient is on any of these trajectories. So one of our ongoing goals from a biomarker perspective  is to look at longitudinal multiple measurements of biomarkers and to see how those change over time. And you can see here in this figure, this is a certain biomarker  that exists in the pediatric sepsis world, the CCL3,  one of the components of Hector Wang's Persevere models. And here, what we see is that when you sample this biomarker on day one, close to the onset of sepsis,  you actually get a very robust separation. But if you just miss that by a day, that separation narrows and further narrows until it's completely gone. So looking at these dynamics  and looking at how this is changing over time may potentially allow us to present when biomarkers should be sampled, what's the appropriate time  that which the sampling should actually occur to maximize our predictive and prognostic ability. So this is, again, another area of major challenge,  but something that we're looking to contribute in. In summary, the overarching goal is to tie the physiomarker and the biomarker space together  to allow for that earlier targeted therapy  and do all of this even before the clinicians suspect the onset of sepsis. That's the overall arching aim. And I think that inch by inch, maybe we're getting closer to that.  Thank you so much for your attention and I'll turn it back for any comments or questions.  Thank you very much. That was excellent. I actually followed that. That was really good. Very typical concepts sometimes to follow. I think most of us aren't very familiar  with machine learning, but we've all heard about it. I thought you did a really nice job of making sense of some of that.  I wanted to start off the discussion by going back a step.  When you talk about machine learning,  sometimes it's a black box that is being generated and you don't know what is the secret sauce in there. And as such, does it really change?  Do you have to validate it for each patient population? For instance, if you're in the emergency department and you train it, your algorithm, now you're moving to the post-surgical,  do you have to retrain it? Is this thing learning constantly or do you set it at some point?  That's an excellent question.  So when we say machine learning, actually the correct term should be data-driven algorithms.  And so by algorithm development driven by data, you start to appreciate that the algorithm is only as good as the data that it was trained on.  And if the algorithm was trained only on ED patients, then the chances that it would generalize to beyond the ED is extremely slim.  Unless you do what we somehow have called transfer learning.  You take all the learning that you've put into one model and then you extract that and then you retrain it based off of the prior knowledge.  You do something called transfer learning and you hope that that transfer of information  would enrich it. In some cases it does, in some cases it doesn't. That is an area of ongoing research within the artificial intelligence community.  We also are getting to a point where explainable machine learning is becoming incredibly important. I think we spent the last decade  just convincing people, the skeptics, that machine learning actually can be useful. And I think the next decade we're going to be spending in actually showing that we can make sense  of this high dimensional abstraction. Many times, some of these neural networks and so on  are seen as black boxes only because the number of parameters or variables that they use to make a decision can actually go into the hundreds of thousands or millions.  And so it becomes very, very confusing for humans  to say, okay, well, you're taking in a hundred input and then you're using a million permutations to come up with an understanding and then you're just giving me one output,  which is a probability of this or that.  So how does that make any sense? And I think that's the challenge within these deep neural networks is where you have this incredible number of parameters. So to alleviate that,  what many of my colleagues in the machine learning communities are doing are building these lightweight models, architectures that can perform just as well  using a minimal amount of parameters as some of these deep, vast algorithms do with millions of parameters. It's just impossible to try to estimate what that is,  even from a human perspective.  So I think that work is actively ongoing and I think we'll see more work come in the years and decades ahead.  So that was gonna be one of my other questions. It seems very complicated. It sounds like you need a lot of horsepower to generate and decipher all the data,  but you've just told me you can simplify it so that a smaller computer located at my institution might be able to drive it as opposed to some supercomputer at Emory  or something like that. Exactly, yeah. I think that's a goal and there's been a lot of interest in what we call edge computing. So instead of sending all of your data  to the supercomputer somewhere to perform, to analyze and to churn the data, what you're doing is actually you're doing all of the analysis at your site.  And where this becomes very interesting is in low resource environments. So I'm Canadian, I'm from Canada. We have the Northern territories where access to resources,  especially neonatal care is extremely slim. So identifying which patients should be transferred is important. But if we are hoping to apply some of these algorithms,  especially some of these video analysis algorithms, then we cannot go with the assumption that the data will transfer through the fiber lines and reach some major city urban center  and then get back to that low resource environment. So we do need other methods to drive that. So I think that's where the ongoing work  is really starting to pick up to simplify the models so that they can actually be embedded into a smartphone. So you can just take your phone and plug in some wire  and then that becomes your computer.  It seems to me machine learning is what all of us do over time in our careers. We're basically accumulating pattern recognition software  or we learn from our input and we start to recognize trends. How important is it also that the providers, it's 17 hours before sepsis onset. This patient's looking pretty good,  but your algorithm tells me, I think your patient's gonna be septic later.  Maybe I even treat that patient and they never get septic.  It's like, I don't believe your algorithm anymore.  So we're gonna be, I think confronting this issue of we're so far ahead of the curve. Maybe we don't see sepsis anymore and that'd be awesome.  But I think people's behavior is shaped by the events  that they're exposed to. So I think it's gonna be a tough sell if you're actually really effectively identify people so early and you can actually mitigate sepsis.  But let's hope we have this problem, right? Yeah, I totally agree. I think we're far away from that. And I think some of the data I showed is again,  the AUC is 0.78, it's doing decent, but it's not there.  And this is analogous to a conversation I heard on the radio  about machine learning applied to agriculture. So there's a lot of robots right now that are being built to pick strawberries, but the humans that actually pick the strawberries  pick the strawberries far better, far more efficiently and a faster rate than the robots do. And why is that? It's because the humans recognize color  as it's adjusted based off of the different sunlight conditions and different lighting conditions and so forth. They also have very tactile locomotion to go on  and grab and pull at the right pressure so you don't jerk or squish the strawberries. And also they understand when something is ripe and not ripe and they'll move on to others  to offset some of the wastage. This is a challenging problem. I don't think we're anywhere close to solving it. I think that this is still going to be a work in progress  in the years to come until we get to a point where we can match the tactile and the contextual,  the computer up over there, I think is, you know. Pretty good. Yeah, it's pretty good. But I think this gets into another, something you didn't discuss, but it's a practical issue.  Yeah, we're pretty good if we're watching the patient constantly, right? I've got 20 patients on my service or in the ICU or wherever I am.  I'm paying attention to them intermittently. I'm not constantly monitoring them. I think if I was, I'd like to think I could catch them earlier in the sepsis progression.  But the problem is I'm spending time with them in the morning, I come back later. In the middle, I haven't been looking at the data.  So I think one advantage of this type of an approach is a continuous input is being provided or analyzed. And a timely alert could be provided to me  who is otherwise distracted by other things. And then I use my judgment, I hope, to come to the conclusion that, yeah, I agree.  I think this person is septic and I should treat them. So I think it's underrated. It's like Forrest Gump.  Just being there is half the battle, right? If you're not there paying attention, you will never make the diagnosis.  So it kind of gets into the pre, you know, when I was gonna ask, Dr. Osborne was unable to join us for the QA session, but I had a question lined up for her about what is the role,  because there's a lot of data reported from the pre-hospital screening for sepsis in the, by EMS, et cetera.  They talk about various parameters that could be used to detect like end tidal CO2, this kind of thing. But what is the role of the history, right?  The pretest probability of a person having sepsis, if they had cancer, they're on immune suppression drugs, you know, they have lung disease, et cetera.  I have to think that the squad that is trained to detect sepsis, their brains are trained. They're gonna perform a lot better than the squad. It's not thinking that way.  It's just trying to get that patient right to the hospital. So I actually think that maybe a machine learning algorithm isn't as effective in a setting  where you have continuous contact with the patient, but it's really effective in a place where, remote place or a place where you don't see patients very as often.  So- I think also the use case where it may be really interesting is to rule out certain things.  You can, you know, when an alert goes off, you can maybe just glance over and like, oh, you know, it's the EKG lead had fallen off. So something of that sort.  I think it becomes a useful tool for that. Yeah, so that's another thing I was gonna bring up. Alarm fatigue is a big issue, a big distraction. There's a lot of false alarms.  So I think that one of the mandates would be  for us not to be, maybe be a little more specific and also train them on this kind of sepsis that needs to be intervened upon very quickly.  There are some brands of sepsis that you could kind of, it doesn't matter if you give the antibiotics now or next, you know, an hour from now,  that person, we need to be smart about how we train  our algorithms to tell us who needs intervened upon very, very quickly to their benefit. So, go ahead. So, Dr. Kucera, I think that, you know,  one of the spaces where I think the opportunity for some organization like the SCCM or others is actually in identifying who is actually  a real sepsis patient or who was or who was not. Because, you know, when you have these quantitative methods like, you know, there should be an antibiotic 30 hours  before there's a blood culture, and then there has to be a two point change in SOFA within this, all of that is fine, but it's only as good as the electronic medical record  that's capturing all this information. And it's only as good as the interpretation of the person that's applying the quantitative methods.  I mean, even within SOFA, you have so many variables within, you know, the system, you have so many variables. So, the ambiguity against the time of onset is clear  on the quantitative side. I think that that's a challenge. And even more the ambiguity against who is actually really septic and who is not becomes even more problematic.  And so then when you give this kind of ambiguous end goal to a machine learning system, which thinks the world is either black or white and, or maybe somewhere in between,  you know, it's a major challenge.  That's right. I mean, some of the criterion we use for sepsis are clinical observations. They aren't necessarily linked specifically to the pathophysiology of the disease.  So, I think better biomarkers will be fed into your machine learning and yield better, more accurate diagnoses. I was gonna ask another kind of similar kind of question  in that the nuances of sepsis were kind of at the point where sepsis is not all one thing. There are people with sepsis, you know, different locations in the body  with portending a different outcome, maybe a different bacteria based on where the location is. I wonder if there would be a way of training  machine learning such that we start to break down sepsis into sub phenotypes. Sort of like Hector Wong has done with genomics, but using other data that is more readily available.  What do you think of that? I think that's exactly where we should go. And I think if we do that, we will improve the performance of these algorithms.  I think where the algorithms really get confused is where, you know, there's a, you know, that line is a little bit blurry.  It's unsure as to, is this pneumonia or is this sepsis or whatnot?  And I think that once we, if we can sub phenotype  these patients, then, you know, perhaps you could identify this is, you know, the probability of sepsis within this patient is high. It's clear that this is a class one or class two  or whatever it may be. And you could do something. And then where that alert fatigue, you know, comes in potentially is where the probability is a little bit in the borderline.  And you're not sure whether it's this way or that way. And you, you know, you can enriching that phenotype  may actually allow that machine learning algorithm to say, okay, clearly this is in a different basket of patients. And so I can now more easily segregate them  once I know that I don't have to compare these patients to, you know, patients that are clearly very different from in their physiology.  Right. And yeah, so that, is there a way to, cause some of these physiological variables are modified at least to some degree by drugs that are taking this kind of thing.  Does that help? For instance, your heart rate may not go up if you're on a beta blocker. Exactly. These kinds of things. Is that something that you can tune into the algorithm  so that the heart rate becomes less important? Can you have missing data and still have a working algorithm?  You're hitting the nails like on the right places. I think this is a major challenge. Missing data is a key, key challenge.  What we often do is we actually include binary indicators to represent whether a patient is on a presser, whether they're on an inotrope.  And we assume that the model will be able to in that space of, you know,  the manifold, it will be able to differentiate between who is inotropic, who's not. And then using that information further isolating patients who are septic and non-septic.  So I think this is something that we do. There's certainly things we could do better because, you know, again, we're at the mercy of our order entry systems and what they tell us.  And they may not be as accurate as the physicians at the bedside, you know, in terms of knowing exactly what went in and what didn't.  So that, again, is an opportunity for learning systems and learning algorithms to get better at.  So 10 years from now, where do you envision this continuum?  You know, we're talking about people at home. Could they be at home? And we'd be testing them at, like,  we know they're a neutropenic, they have cancer, they're a neutropenic, we send them home. Or we even send them home from the ED. We weren't sure if they were septic or not,  but we're gonna put some kind of monitor on them and we'll remotely monitor them. Is that something you see could happen in the not too distant future?  One of the things that I envision is, you know, everyone has these Apple Watches now  that sample ECG, SpO2, you know, a lot of these activity.  So wouldn't it be great if it were to wake you up one day in the middle of the night and tell you, hey, you know, your temperature has been going up,  your heart rate's been going up, your saturation's going down. I've sampled these biomarkers using your sweat.  I think that you are headed down this pathway, already alerted your family physician. They are, you know, ordering these medications.  A drone is gonna drop it off, take them at this time.  Here we are. You know, I think it's appropriate. It's the 50th anniversary of the SCCM. We should be looking forward to these kind of things. I think that's, and I would buy stock  in that company, by the way.  I think that would be amazing. Yeah, so, you know, I think we get caught up  in the here and now and the things that, they're not quite ready for prime time yet, but you gotta, I think we gotta look ahead to the opportunities that are there  and kind of slog through the brush till we get to a clearing where it starts to really pan out for us. I really think this is gonna be the future.  And I think you're gonna be one of the first to have presented it in a coherent form at SCCM. So I really appreciate you doing that. Thank you.  And I'd like to thank Dr. Osborne also for having made time to put her presentation together.  Apparently she's very busy right now. She's a surgeon and she's an emergency medicine doctor. And I think they have her preoccupied at the moment and she couldn't make it.  So I'm gonna give her a pass for not being able to show up today, but really appreciate her putting together a really nice talk about pre-hospital care  and detection, early detection of sepsis.  I don't know if there are any other questions from the audience. I don't see any here.  But I would give people an opportunity to go ahead and add your questions here.  I don't know if you have any other comments you'd like to make. We have a few minutes left to talk about. Yeah, thank you for the opportunity. I think this is fantastic.  And I look forward to the rest of the conference and hearing other great talks. I guess I have one other question. You are a PhD, am I correct? Yes.  And what is the role of an MD in your,  how do you collaborate with the MDs and how should we leverage each other to get things done? I think that partnership is extremely important. I think that, you know, for me,  I don't have eyes on the ground. So they are often my eyes and my hands in terms of telling me, you know, this is what I need to be looking for.  And sometimes the algorithms spit out the most bizarre things. And, you know, someone who has the knowledge, the clinical knowledge to say, no, this doesn't seem right,  or maybe you should look into this a little bit more is extremely valuable. So I think that partnership is almost, you know, things would be just not possible without that.  I'm going to ask you one more thing I meant to ask earlier. One of the most useful tests we use is a test, a D-dimer for instance, for pulmonary embolism.  If it's negative, actually, it's more useful than if it's positive. Do you feel like the machine learning could reliably  eliminate the likelihood of sepsis such that we move on to other diagnoses? I think that might be something we could use. It would be very useful for instance,  in many cases in the hospital. I think this is, again, where it's driven again by what data we have and what data is available.  But it would be a very interesting modeling exercise to do and to see what, you know, what that shows. Off the top, I don't think I can say, but certainly I think this is something  that we can look into.  Well, thank you very much. I think our time is actually officially up. I appreciate your answering all my questions. You had to take them all. Thank you so much. It's a pleasure.  I really enjoyed them. Right, thank you. Thank you so much, Dr. Fisher.  Thank you.

Video Summary

In this video, Tiffany Osborne discusses the challenges of pre-hospital sepsis detection and treatment. She begins by highlighting the significant number of patients that arrive at the emergency department by ambulance, with many of them being elderly and at risk for sepsis. She explains that detecting sepsis in pre-hospital environments is difficult due to the variability and lack of clear criteria for differentiation. However, recent studies have shown improvements in pre-hospital sepsis detection strategies by incorporating additional physiological markers such as end-tidal CO2 and lactate. These markers have been found to improve the accuracy of identifying sepsis patients and predicting outcomes. Osborne also emphasizes the importance of collaboration between pre-hospital providers, emergency departments, and intensive care units in order to provide comprehensive care for sepsis patients. She concludes by stating that early detection of sepsis is associated with early treatment and better outcomes, and that integrating physiological and biomarker data can enhance the development of real-time intelligent learning systems for sepsis management. Overall, Osborne emphasizes the need for continued research and improvement in pre-hospital sepsis detection and treatment in order to provide the best care for patients.

Asset Subtitle

Sepsis, 2021

Asset Caption

The Society of Critical Care Medicine's Critical Care Congress features internationally renowned faculty and content sessions highlighting the most up-to-date, evidence-based developments in critical care medicine. This is a presentation from the 2021 Critical Care Congress held virtually from January 31-February 12, 2021.

Learning Objectives

Critique prehospital screening techniques that perform best for early sepsis detection
Explore new biomarkers that improve the accuracy of sepsis detection during the initial emergency department encounter and guide delivery of sepsis care bundles
Identify widely available parameters during the initial clinical encounter that could lead to earlier detection and treatment
Apply computer-based machine learning algorithms to electronic medical record data to detect vital sign and laboratory patterns that are highly predictive of sepsis

Earlier Sepsis Detection and Treatment Prior to ED Admission
Tiffany M. Osborn

Big Data Biomarkers
Rishikesan Kamaleswaran

Meta Tag

Content Type Presentation

Knowledge Area Sepsis

Knowledge Level Foundational

Knowledge Level Intermediate

Knowledge Level Advanced

Membership Level Select

Membership Level Nonmember

Membership Level Associate

Membership Level Professional

Tag Scoring Systems

Tag Shock

Year 2021

Keywords

Tiffany Osborne

pre-hospital sepsis detection

pre-hospital sepsis treatment

physiological markers

end-tidal CO2

lactate

sepsis patient identification

outcome prediction

collaboration

early detection of sepsis

	Society of Critical Care Medicine 500 Midway Drive Mount Prospect, IL 60056 USA Phone: +1 847 827-6888 Fax: +1 847 439-7226 Email: support@sccm.org	Contact Us About SCCM Newsroom Advertising & Sponsorship DONATE	MySCCM LearnICU Patients & Families Surviving Sepsis Campaign Critical Care Societies Collaborative	GET OUR NEWSLETTER
© Society of Critical Care Medicine. All rights reserved. \| Privacy Statement \| Terms & Conditions The Society of Critical Care Medicine, SCCM, and Critical Care Congress are registered trademarks of the Society of Critical Care Medicine.